title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
sequencelengths 1
21
| type
stringclasses 3
values | abstract
stringlengths 413
2.52k
| keywords
stringlengths 4
397
| TL;DR
stringlengths 5
250
⌀ | submission_number
int64 2
14.3k
| arxiv_id
stringlengths 10
10
⌀ | embedding
sequencelengths 384
384
| github_url
stringlengths 0
126
| github_stars
int64 0
55k
| num_models
int64 0
82
| num_datasets
int64 0
7
| num_spaces
int64 0
100
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition | https://openreview.net/forum?id=41HlN8XYM5 | [
"Aliyah R. Hsu",
"Georgia Zhou",
"Yeshwanth Cherapanamjeri",
"Yaxuan Huang",
"Anobel Odisho",
"Peter R. Carroll",
"Bin Yu"
] | Poster | Automated mechanistic interpretation research has attracted great interest due to its potential to scale explanations of neural network internals to large models. Existing automated circuit discovery work relies on activation patching or its approximations to identify subgraphs in models for specific tasks (circuits). They often suffer from slow runtime, approximation errors, and specific requirements of metrics, such as non-zero gradients.
In this work, we introduce contextual decomposition for transformers (CD-T) to build interpretable circuits in large language models. CD-T can produce circuits at any level of abstraction and is the first to efficiently produce circuits as fine-grained as attention heads at specific sequence positions.
CD-T is compatible to all transformer types, and requires no training or manually-crafted examples.
CD-T consists of a set of mathematical equations to isolate contribution of model features. Through recursively computing contribution of all nodes in a computational graph of a model using CD-T followed by pruning, we are able to reduce circuit discovery runtime from hours to seconds compared to state-of-the-art baselines.
On three standard circuit evaluation datasets (indirect object identification, greater-than comparisons, and docstring completion),
we demonstrate that CD-T outperforms ACDC and EAP by better recovering the manual circuits with an average of 97% ROC AUC under low runtimes.
In addition, we provide evidence that faithfulness of CD-T circuits is not due to random chance by showing our circuits are 80% more faithful than random circuits of up to 60% of the original model size.
Finally, we show CD-T circuits are able to perfectly replicate original models' behavior(faithfulness = 1) using fewer nodes than the baselines for all tasks.
Our results underscore the great promise of CD-T for efficient automated mechanistic interpretability, paving the way for new insights into the workings of large language models. | Automated Circuit Discovery, Explainable AI, Interpretation, Machine Learning, Language Models, Transformers | null | 13,618 | 2407.00886 | [
-0.10275781154632568,
0.052587732672691345,
0.0531134307384491,
0.015017407014966011,
0.027642743661999702,
-0.03291475027799606,
-0.02761186473071575,
0.043499905616045,
-0.006415653973817825,
-0.03791051730513573,
-0.042171236127614975,
-0.10645272582769394,
0.03573743999004364,
0.05158495530486107,
-0.02577790804207325,
0.04040112346410751,
0.010060245171189308,
0.09465773403644562,
-0.05421849712729454,
-0.09797263145446777,
0.03786251321434975,
0.09325268864631653,
0.010894594714045525,
-0.03700212016701698,
0.085841603577137,
0.04805191978812218,
0.02169039100408554,
-0.06949852406978607,
-0.008892521262168884,
0.02656625211238861,
-0.01415231078863144,
-0.007477647624909878,
-0.01649867556989193,
0.02743537351489067,
-0.017102455720305443,
0.0858045294880867,
-0.06704144179821014,
0.03706485033035278,
0.059907231479883194,
-0.048723768442869186,
-0.00918485876172781,
0.039446618407964706,
0.05799957364797592,
0.056764356791973114,
0.08022727072238922,
-0.03366760537028313,
0.011223857291042805,
-0.040508922189474106,
-0.027051672339439392,
-0.08760623633861542,
-0.01003985945135355,
-0.009062808938324451,
0.029342668130993843,
0.12529338896274567,
-0.02919754572212696,
0.021142447367310524,
0.06716585904359818,
0.009587405249476433,
-0.003466396825388074,
0.04344711825251579,
-0.025765663012862206,
-0.08293192833662033,
-0.007146716117858887,
-0.04928939417004585,
-0.0741419643163681,
-0.00333287357352674,
0.0626039206981659,
0.004846277646720409,
-0.02098587155342102,
-0.002781882416456938,
-0.013970449566841125,
0.05600802227854729,
-0.05242175981402397,
0.03131747990846634,
0.06812980771064758,
0.046519145369529724,
0.11823111772537231,
-0.03530523180961609,
-0.020408187061548233,
-0.09137067943811417,
-0.024907223880290985,
-0.028005456551909447,
0.01891995035111904,
0.003794147167354822,
0.0974966436624527,
0.027420273050665855,
-0.016562601551413536,
0.004832446575164795,
-0.004800335504114628,
0.028677135705947876,
-0.07907991111278534,
-0.10781171917915344,
0.01877189427614212,
-0.05127957463264465,
-0.000913271214812994,
0.053903721272945404,
0.03917360678315163,
-0.04437434673309326,
-0.03873973339796066,
0.11457070708274841,
0.005120704881846905,
0.03616776317358017,
0.06079939380288124,
-0.0960782915353775,
-0.02404046058654785,
0.02434285171329975,
-0.005196770187467337,
0.04004061594605446,
0.03186510130763054,
-0.12704645097255707,
-0.015696872025728226,
0.06384297460317612,
-0.03471023216843605,
-0.016543665900826454,
0.06290880590677261,
-0.027171416208148003,
0.034501660615205765,
-0.027525242418050766,
0.016642674803733826,
0.0806499719619751,
-0.024197101593017578,
0.023856882005929947,
-0.0550132654607296,
0.08961057662963867,
0.03816458210349083,
0.018283817917108536,
-0.03695998713374138,
1.9236994438565506e-33,
0.008657645434141159,
0.030683908611536026,
0.03742891177535057,
0.04986434802412987,
0.10782274603843689,
0.0027739806100726128,
0.05940305069088936,
0.08423225581645966,
0.03284265846014023,
0.02317553199827671,
-0.055692099034786224,
0.02766576036810875,
-0.06644092500209808,
0.049748413264751434,
-0.01640893891453743,
-0.02188585326075554,
-0.06779538094997406,
-0.020490655675530434,
-0.006419532932341099,
-0.09624825417995453,
0.03476481884717941,
-0.06829898059368134,
0.028343329206109047,
-0.028605125844478607,
0.0036846434231847525,
-0.04146677628159523,
0.02592172659933567,
-0.03309061750769615,
-0.057232338935136795,
-0.017812132835388184,
-0.0964168906211853,
0.03971249610185623,
-0.0022671991027891636,
0.004475836642086506,
0.005170158110558987,
-0.04566410556435585,
0.04718216881155968,
-0.02557418681681156,
0.06932304799556732,
-0.008404809050261974,
-0.04767795652151108,
0.016111204400658607,
-0.06288384646177292,
-0.019552994519472122,
-0.05718567594885826,
-0.026519732549786568,
-0.00930252019315958,
0.017894769087433815,
0.048955030739307404,
-0.0013640180695801973,
0.03912108391523361,
-0.01761482283473015,
0.037706807255744934,
-0.05335795134305954,
0.07337108254432678,
0.030144600197672844,
0.031332142651081085,
0.05625168979167938,
0.1155879944562912,
0.09750822186470032,
0.0016013459535315633,
0.0698089674115181,
-0.0636182427406311,
0.08385626971721649,
-0.04916784539818764,
0.07600011676549911,
-0.04200759157538414,
0.04628991708159447,
0.03187618777155876,
-0.017249036580324173,
-0.07373511791229248,
0.028165165334939957,
-0.04790812358260155,
-0.07098309695720673,
0.11931043118238449,
-0.011764289811253548,
0.0638519749045372,
-0.014279081486165524,
-0.015336931683123112,
-0.006267326883971691,
-0.013356305658817291,
0.007972361519932747,
0.013316371478140354,
0.012869873084127903,
0.005685516633093357,
-0.023954132571816444,
0.014983966015279293,
-0.05566834658384323,
-0.03165333345532417,
-0.09900090843439102,
-0.05897070840001106,
-0.004078713245689869,
-0.04749172553420067,
0.00804782472550869,
0.03252696990966797,
-2.9630507940314855e-33,
-0.051070328801870346,
0.01631767302751541,
0.014188727363944054,
0.00427217734977603,
-0.037432678043842316,
-0.13262368738651276,
-0.06222466751933098,
0.011478912085294724,
-0.022312097251415253,
0.0013021813938394189,
0.0026275734417140484,
-0.021695250645279884,
-0.03250839188694954,
-0.051568854600191116,
0.021129582077264786,
-0.009611623361706734,
-0.04840463772416115,
0.00024006528838071972,
0.02012990042567253,
0.07758574932813644,
-0.007366950158029795,
0.111454077064991,
-0.1777188926935196,
-0.0345243364572525,
-0.10736148059368134,
0.014835817739367485,
-0.07757134735584259,
0.07086948305368423,
0.01577051356434822,
-0.04377662390470505,
-0.037553928792476654,
0.01605253294110298,
0.0011868082219734788,
0.016445886343717575,
-0.0582345649600029,
0.05631130933761597,
0.005499189253896475,
0.01904241368174553,
0.0020542119164019823,
0.010312739759683609,
0.046310313045978546,
-0.01858370751142502,
-0.09863071888685226,
0.07398351281881332,
-0.05429461598396301,
-0.009616587311029434,
-0.016771972179412842,
-0.00536664854735136,
-0.023121004924178123,
0.06075294315814972,
-0.048831261694431305,
0.022698206827044487,
-0.05118263140320778,
-0.06583265960216522,
-0.007166041526943445,
-0.0747794657945633,
-0.012145303189754486,
-0.015955764800310135,
0.09964263439178467,
-0.09548810869455338,
-0.051246803253889084,
-0.06446409970521927,
0.048756349831819534,
-0.03409762308001518,
-0.003634135704487562,
-0.045128483325242996,
-0.014716998673975468,
0.0221901535987854,
0.05317853391170502,
0.046836983412504196,
0.005208778660744429,
0.020022781565785408,
0.011440800502896309,
-0.07024387270212173,
-0.038930345326662064,
-0.07102019339799881,
-0.029226433485746384,
-0.058823246508836746,
-0.0016609926242381334,
-0.04407701641321182,
0.009410885162651539,
0.026080770418047905,
-0.028810907155275345,
-0.06816637516021729,
-0.014991581439971924,
0.06087891012430191,
0.022443624213337898,
0.06781816482543945,
0.08643966913223267,
0.03604857996106148,
-0.08997467905282974,
0.09252973645925522,
-0.00969440396875143,
0.09086170047521591,
-0.03241005167365074,
-4.486036075945776e-8,
-0.0040794843807816505,
0.10012178868055344,
-0.02596045285463333,
-0.03666377440094948,
0.09581045806407928,
-0.05828561261296272,
0.029414242133498192,
0.11536164581775665,
-0.04613412544131279,
0.014878151938319206,
0.10262381285429001,
-0.007154945284128189,
0.02370959147810936,
-0.059701867401599884,
-0.038031283766031265,
0.07679984718561172,
0.017020348459482193,
0.025661448016762733,
-0.019069325178861618,
0.01967758685350418,
-0.03449661284685135,
0.07134360074996948,
0.03802185505628586,
0.0437377505004406,
0.11611511558294296,
-0.09151824563741684,
-0.03381713107228279,
0.02510102279484272,
0.07225033640861511,
0.01537773571908474,
-0.03693239018321037,
0.03154470771551132,
0.0008452114416286349,
-0.025688108056783676,
0.007100222632288933,
0.10778962075710297,
0.029411310330033302,
-0.04358451068401337,
-0.01988149993121624,
-0.03091985173523426,
-0.03852468356490135,
-0.005844922736287117,
-0.06532406806945801,
0.03496615216135979,
-0.013099330477416515,
-0.03330783173441887,
0.0009050665539689362,
-0.04673723131418228,
0.03508392348885536,
-0.021816333755850792,
-0.00005657986912410706,
-0.030937325209379196,
-0.10292311012744904,
-0.007063608150929213,
-0.005741933826357126,
0.05730739235877991,
-0.04609015956521034,
-0.043660085648298264,
0.0037640875671058893,
0.04567980393767357,
-0.07097608596086502,
0.03944246098399162,
0.015526101924479008,
-0.027590589597821236
] | 0 | 0 | 0 | 0 |
|
How to Evaluate Reward Models for RLHF | https://openreview.net/forum?id=cbttLtO94Q | [
"Evan Frick",
"Tianle Li",
"Connor Chen",
"Wei-Lin Chiang",
"Anastasios Nikolas Angelopoulos",
"Jiantao Jiao",
"Banghua Zhu",
"Joseph E. Gonzalez",
"Ion Stoica"
] | Poster | We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback).
The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance.
However, this process is prohibitively expensive.
To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks.
These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains.
To investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowd-sourced human preference platform to view real reward model downstream performance as ground truth.
Ultimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we opensource for public use and further development at https://github.com/lmarena/PPE. | RLHF, RL, Reward Model, LLM, Benchmark, Dataset, Evaluation | null | 13,614 | 2410.14872 | [
-0.10671377182006836,
-0.12314127385616302,
-0.006068799644708633,
0.03931453824043274,
0.04506322741508484,
0.05315708741545677,
0.005695175379514694,
0.015503293834626675,
0.0029500345699489117,
0.008476882241666317,
-0.09340059012174606,
-0.095465749502182,
0.029009928926825523,
0.07484141737222672,
0.056024134159088135,
-0.025208937004208565,
0.11000309884548187,
0.003079055342823267,
-0.06618387997150421,
-0.07162190973758698,
-0.003912882413715124,
-0.016931356862187386,
0.047493137419223785,
-0.015125555917620659,
-0.02246008813381195,
-0.06546114385128021,
-0.03920583054423332,
-0.05001157149672508,
-0.004054593853652477,
0.0031336811371147633,
0.05054788663983345,
-0.016967585310339928,
0.013952394016087055,
-0.05509905889630318,
-0.11690443754196167,
0.07432077825069427,
-0.09405717998743057,
-0.08589890599250793,
-0.03212318941950798,
-0.053827472031116486,
-0.05817485600709915,
-0.051154814660549164,
0.01610502041876316,
0.024348611012101173,
0.055071599781513214,
0.02652318961918354,
0.03414566442370415,
-0.00838126614689827,
-0.028489375486969948,
0.010295286774635315,
-0.022788768634200096,
-0.0028441494796425104,
0.061823390424251556,
0.037085022777318954,
-0.013497140258550644,
0.04738461598753929,
0.03660508617758751,
-0.02192722074687481,
0.009234636090695858,
-0.05309312418103218,
-0.07315874844789505,
-0.14218194782733917,
-0.10646573454141617,
0.006360860541462898,
-0.057368356734514236,
-0.037030261009931564,
-0.029133247211575508,
0.06311333179473877,
0.0073066698387265205,
-0.005242908839136362,
-0.023843105882406235,
0.0366128534078598,
-0.023511363193392754,
-0.02728867717087269,
0.04560196027159691,
0.04737371578812599,
0.005775293335318565,
-0.019766557961702347,
0.07304796576499939,
-0.06161706894636154,
-0.004614907316863537,
-0.020028533414006233,
0.02936118096113205,
0.001845991238951683,
0.10471244901418686,
-0.007931968197226524,
0.06868626922369003,
0.017161408439278603,
-0.008469375781714916,
0.0015259444480761886,
-0.02140715904533863,
0.008442006073892117,
0.04652189090847969,
0.005017064977437258,
-0.03718024492263794,
0.0640423372387886,
-0.03470825031399727,
-0.1004379391670227,
-0.06703389436006546,
0.10574054718017578,
0.00988182332366705,
0.09161083400249481,
-0.018485277891159058,
-0.043647512793540955,
0.020923592150211334,
0.006465552374720573,
-0.08140838891267776,
0.09584329277276993,
0.04853826016187668,
-0.06624262779951096,
0.011225885711610317,
0.04183173179626465,
0.0044161174446344376,
0.027599865570664406,
0.029068417847156525,
0.010634114034473896,
-0.00219710567034781,
-0.054250430315732956,
-0.013318310491740704,
0.02440735511481762,
-0.0066014924086630344,
-0.017601683735847473,
0.06825097650289536,
-0.0034349991474300623,
0.05623003467917442,
-0.0035698923747986555,
-0.03743746504187584,
2.7316699757644804e-33,
0.06295303255319595,
0.05721362307667732,
0.0723043829202652,
-0.02464372292160988,
0.03607042878866196,
0.006136057898402214,
0.061942461878061295,
0.0670783519744873,
-0.07199680805206299,
-0.027517076581716537,
0.039938125759363174,
0.07048534601926804,
-0.06650994718074799,
0.10510042309761047,
0.01490162592381239,
0.000021535164705710486,
-0.05235627293586731,
0.008678161539137363,
-0.0014512925408780575,
0.004845329560339451,
0.1129176989197731,
0.022515511140227318,
0.00437880726531148,
-0.043234266340732574,
0.05022798851132393,
0.012751015834510326,
0.05128976330161095,
-0.07270245254039764,
-0.03178635612130165,
0.01319747045636177,
-0.012064374051988125,
-0.019394058734178543,
0.0728553757071495,
0.029349029064178467,
-0.04181760177016258,
-0.024314241483807564,
-0.06987451761960983,
-0.019531555473804474,
0.009212564677000046,
-0.04928602650761604,
0.028989113867282867,
0.005017956718802452,
0.0055166310630738735,
0.016271017491817474,
-0.1100904792547226,
-0.0871194377541542,
0.04337101802229881,
-0.0030215606093406677,
-0.015254261903464794,
-0.0011978972470387816,
0.08716858178377151,
0.04603130742907524,
0.016919424757361412,
0.020726632326841354,
-0.056723471730947495,
0.014081727713346481,
-0.012423819862306118,
0.07638630270957947,
0.011656166985630989,
-0.0011403788812458515,
0.026781577616930008,
0.04876284301280975,
0.02383749932050705,
0.03360382840037346,
-0.011728804558515549,
0.0753420814871788,
-0.04103056341409683,
-0.04457636922597885,
0.010706516914069653,
0.045869071036577225,
0.018144143745303154,
0.029221104457974434,
0.008637865073978901,
-0.037212491035461426,
0.05659693107008934,
-0.010693139396607876,
0.029578495770692825,
-0.07446388155221939,
0.1040320172905922,
-0.011603910475969315,
0.01494160108268261,
-0.01697119139134884,
0.009759308770298958,
-0.06446106731891632,
0.007781687192618847,
-0.061868079006671906,
-0.00483756884932518,
-0.13838402926921844,
0.005816214252263308,
-0.017467152327299118,
-0.0006719870143570006,
-0.039654094725847244,
0.0034803750459104776,
0.052782248705625534,
0.0028396379202604294,
-2.5380451176909256e-33,
-0.03148391842842102,
-0.03225274756550789,
-0.003448431147262454,
0.10597854107618332,
0.0012816655216738582,
-0.04704572260379791,
0.011333241127431393,
0.037843771278858185,
0.010237260721623898,
0.03548194095492363,
-0.037698350846767426,
0.016217220574617386,
0.06064731627702713,
0.04022917151451111,
0.042729701846838,
-0.10005932301282883,
0.019406570121645927,
0.02198173850774765,
0.0032252082601189613,
0.039503779262304306,
0.05039476975798607,
0.1169622614979744,
-0.08960628509521484,
0.10344238579273224,
-0.015639375895261765,
0.013506155461072922,
0.018982693552970886,
0.011212501674890518,
-0.04145369306206703,
-0.016544343903660774,
-0.035916995257139206,
-0.013413960114121437,
-0.09515274316072464,
-0.08441062271595001,
-0.049322422593832016,
-0.005114027298986912,
0.05447913333773613,
0.05857476219534874,
-0.10058313608169556,
0.13228391110897064,
0.03672442585229874,
-0.05315089598298073,
-0.09457988291978836,
-0.007233620621263981,
-0.020653072744607925,
0.0191019456833601,
-0.1505860835313797,
0.018725095316767693,
0.004271669313311577,
-0.0528918094933033,
-0.031151238828897476,
0.05541013553738594,
-0.04015657678246498,
0.04169796407222748,
-0.06011589989066124,
-0.051260802894830704,
-0.049471575766801834,
-0.045153699815273285,
0.07266628742218018,
0.010144725441932678,
-0.08235117048025131,
0.04887533187866211,
-0.017624208703637123,
0.030137093737721443,
0.007481226697564125,
-0.08441967517137527,
0.010797243565320969,
-0.0007883214857429266,
-0.013370899483561516,
-0.003723299829289317,
0.05247220769524574,
-0.008340680040419102,
0.027949664741754532,
-0.010802563279867172,
0.03440888598561287,
0.03371548652648926,
0.013841856271028519,
-0.016475815325975418,
0.03125004097819328,
0.006797686219215393,
0.07179123163223267,
0.027182726189494133,
0.10928154736757278,
-0.04187198355793953,
0.05786212906241417,
0.05187872797250748,
0.06217436119914055,
0.09738804399967194,
-0.03629785403609276,
0.007655102759599686,
-0.06444498151540756,
0.00017727349768392742,
-0.04369039088487625,
0.026423100382089615,
0.041220758110284805,
-4.201129399916681e-8,
-0.022488920018076897,
0.009340311400592327,
0.0140005461871624,
0.05471716448664665,
-0.059897299855947495,
-0.011026821099221706,
-0.004321688786149025,
0.03427368402481079,
-0.07027765363454819,
0.00288915797136724,
0.06934299319982529,
0.0854446068406105,
-0.02579144947230816,
-0.05527454987168312,
-0.04800768569111824,
0.09779059886932373,
0.010598212480545044,
0.043717656284570694,
-0.01220244262367487,
0.01587810553610325,
0.11528023332357407,
0.058369752019643784,
0.07054583728313446,
-0.09673160314559937,
0.023803988471627235,
-0.06969544291496277,
-0.010383746586740017,
0.08026816695928574,
-0.01666739583015442,
0.00036285899113863707,
0.003816284704953432,
-0.05623938888311386,
-0.017639966681599617,
-0.020457524806261063,
0.09523177891969681,
0.06906121969223022,
-0.025137675926089287,
-0.02966092899441719,
0.03168373182415962,
0.033256448805332184,
0.046934016048908234,
0.06905530393123627,
-0.04548173397779465,
0.007964180782437325,
-0.018412666395306587,
0.0008559278794564307,
-0.10345591604709625,
-0.0714733824133873,
0.0563451312482357,
-0.07374723255634308,
-0.001132800243794918,
-0.0664568543434143,
-0.10182319581508636,
-0.0005959435948170722,
0.05988696962594986,
-0.012662138789892197,
-0.04999059438705444,
-0.0790945291519165,
0.0036949256900697947,
0.05298693850636482,
0.09744717180728912,
-0.03371454030275345,
-0.04721513390541077,
-0.011342065408825874
] | https://github.com/lmarena/ppe | 45 | 0 | 0 | 0 |
An Efficient Framework for Crediting Data Contributors of Diffusion Models | https://openreview.net/forum?id=9EqQC2ct4H | [
"MingYu Lu",
"Chris Lin",
"Chanwoo Kim",
"Su-In Lee"
] | Poster | As diffusion models are deployed in real-world settings and their performance driven by training data, appraising the contribution of data contributors is crucial to creating incentives for sharing quality data and to implementing policies for data compensation. Depending on the use case, model performance corresponds to various global properties of the distribution learned by a diffusion model (e.g., overall aesthetic quality). Hence, here we address the problem of attributing global properties of diffusion models to data contributors. The Shapley value provides a principled approach to valuation by uniquely satisfying game-theoretic axioms of fairness. However, estimating Shapley values for diffusion models is computationally impractical because it requires retraining and rerunning inference on many subsets of data contributors. We introduce a method to efficiently retrain and rerun inference for Shapley value estimation, by leveraging model pruning and fine-tuning. We evaluate the utility of our method with three use cases: (i) image quality for a DDPM trained on a CIFAR dataset, (ii) demographic diversity for an LDM trained on CelebA-HQ, and (iii) aesthetic quality for a Stable Diffusion model LoRA-finetuned on Post-Impressionist artworks. Our results empirically demonstrate that our framework can identify important data contributors across global properties, outperforming existing attribution methods for diffusion models. | data attribution, diffusion models, Shapley values | By model pruning and fine-tuning, we efficiently estimate Shapley values and attribute global properties of diffusion models to data contributors | 13,612 | 2407.03153 | [
-0.004826139658689499,
-0.06977040320634842,
0.026173943653702736,
-0.03251180052757263,
0.047956112772226334,
0.014642397873103619,
0.0015411137137562037,
-0.07050187140703201,
0.039005015045404434,
0.015512299723923206,
-0.018805397674441338,
-0.06386777758598328,
0.07000463455915451,
0.026921752840280533,
-0.009374390356242657,
-0.018844950944185257,
0.13600504398345947,
0.10855139791965485,
-0.14950428903102875,
-0.069481261074543,
-0.02601160667836666,
-0.03510406240820885,
0.03651456907391548,
-0.038426268845796585,
0.01976834237575531,
-0.08161954581737518,
0.043709445744752884,
-0.028425905853509903,
0.014092431403696537,
-0.06769546121358871,
0.024596285074949265,
-0.009353662841022015,
-0.04469394311308861,
-0.021998541429638863,
-0.02505316026508808,
0.026721693575382233,
-0.06134969741106033,
-0.044946085661649704,
0.008541520684957504,
0.02084462158381939,
-0.007554302923381329,
0.007577164098620415,
-0.062479909509420395,
-0.004595690872520208,
0.00013770084478892386,
0.021514087915420532,
0.04586736857891083,
-0.04600168392062187,
0.02065056934952736,
0.01696251519024372,
0.026988305151462555,
0.004787902347743511,
-0.06856406480073929,
0.028405923396348953,
0.011800307780504227,
0.013238977640867233,
0.017143798992037773,
-0.010686108842492104,
0.018383631482720375,
-0.04669414833188057,
-0.11061036586761475,
-0.046566009521484375,
-0.0473141074180603,
0.018962129950523376,
0.05599242448806763,
-0.0349709615111351,
0.049540236592292786,
0.02457030862569809,
-0.08054835349321365,
-0.008634927682578564,
0.03699697181582451,
0.03633526340126991,
-0.10361646860837936,
0.007661239709705114,
-0.04111313074827194,
0.07107696682214737,
0.05577118322253227,
0.0020880992524325848,
0.018683962523937225,
-0.08930612355470657,
-0.003901956370100379,
-0.02617299184203148,
0.04444132000207901,
0.0126335509121418,
0.053702253848314285,
-0.07153136283159256,
0.0012104678899049759,
0.05268903821706772,
0.049693770706653595,
0.04775816574692726,
-0.017050310969352722,
0.11733024567365646,
-0.026319116353988647,
0.011835640296339989,
0.04646114259958267,
0.062303442507982254,
-0.018917381763458252,
-0.09620578587055206,
0.031074916943907738,
0.1339469701051712,
-0.06811926513910294,
0.05451621487736702,
0.006915656849741936,
-0.055199459195137024,
0.08805810660123825,
-0.02072606235742569,
0.04300606995820999,
0.047928620129823685,
-0.034199606627225876,
0.0075619095005095005,
0.017626168206334114,
0.03561224788427353,
-0.02060689404606819,
0.003325817873701453,
0.07937806844711304,
-0.034979723393917084,
-0.01891792193055153,
-0.03526555746793747,
-0.017169801518321037,
-0.023020701482892036,
0.03805841505527496,
-0.016299713402986526,
0.031842801719903946,
-0.0349840484559536,
-0.02980581298470497,
-0.07050581276416779,
-0.05825790762901306,
3.149758786205413e-33,
0.00975099764764309,
0.0076364148408174515,
0.09709683805704117,
-0.07108888775110245,
0.05777781829237938,
-0.013665005564689636,
-0.05733657628297806,
0.03472603112459183,
-0.060141995549201965,
-0.027696307748556137,
-0.06468523293733597,
0.05901990458369255,
-0.1264580339193344,
0.06082276627421379,
0.053443439304828644,
-0.005365083459764719,
-0.028176121413707733,
0.08530320972204208,
-0.02282012440264225,
0.06317941099405289,
0.05892485752701759,
-0.08760352432727814,
0.03650340065360069,
-0.002733970293775201,
-0.030590295791625977,
-0.019244858995079994,
0.04734431207180023,
0.008883964270353317,
0.02587302215397358,
0.029482021927833557,
-0.05823999643325806,
-0.029566222801804543,
0.024924838915467262,
-0.01784384809434414,
-0.010869608260691166,
0.014906407333910465,
-0.08876067399978638,
0.01720276102423668,
-0.007002490106970072,
0.042246993631124496,
0.02427225187420845,
0.03956630080938339,
0.006567676551640034,
-0.010308214463293552,
-0.13235962390899658,
-0.024790192022919655,
0.04620368406176567,
-0.08672800660133362,
-0.06144537031650543,
0.00454262038692832,
0.10472676903009415,
-0.0599588006734848,
0.020622914656996727,
-0.07595966756343842,
-0.09469739347696304,
-0.020907476544380188,
0.002863184316083789,
-0.08741495758295059,
-0.004463958088308573,
-0.04755092039704323,
0.023227494210004807,
0.04155813157558441,
0.004786165896803141,
0.034162405878305435,
-0.008199694566428661,
0.03869292140007019,
-0.0006524340133182704,
0.05250962823629379,
-0.029794396832585335,
-0.027012867853045464,
0.0024618911556899548,
0.08087943494319916,
-0.08001042902469635,
-0.06657171994447708,
0.03134908527135849,
-0.021692544221878052,
0.049503494054079056,
0.009098491631448269,
0.050802137702703476,
-0.0026852814480662346,
-0.045843228697776794,
-0.013210787437856197,
-0.05950631573796272,
-0.10444933921098709,
-0.0063234358094632626,
0.0014524937141686678,
0.03928244113922119,
-0.05922705680131912,
-0.04956984892487526,
-0.04278082400560379,
-0.051825933158397675,
0.016062524169683456,
-0.05284195393323898,
-0.019202446565032005,
0.009156279265880585,
-4.183632516376508e-33,
-0.06431887298822403,
0.05374717339873314,
-0.007771484088152647,
0.14820446074008942,
-0.03592914342880249,
-0.03775295615196228,
-0.01975122280418873,
0.04483698308467865,
0.054675646126270294,
-0.04882983863353729,
-0.037916794419288635,
0.000481378345284611,
0.031418994069099426,
-0.040056727826595306,
-0.027702784165740013,
-0.050220005214214325,
0.0008860136149451137,
-0.04169229045510292,
-0.050519634038209915,
0.023222265765070915,
0.021655911579728127,
0.06609705090522766,
-0.04979495704174042,
0.005860412959009409,
-0.0682879388332367,
0.09110041707754135,
-0.1072048470377922,
-0.009272528812289238,
0.06548189371824265,
0.04058025777339935,
-0.032109685242176056,
-0.014450288377702236,
-0.0655377060174942,
-0.08127649128437042,
-0.06519253551959991,
0.023455150425434113,
0.07478943467140198,
0.05921749025583267,
-0.06860978156328201,
0.1795579344034195,
0.048872921615839005,
-0.04339280724525452,
-0.11210408806800842,
0.07858999818563461,
-0.020601021125912666,
-0.06140177324414253,
-0.01750345155596733,
-0.03971317782998085,
0.09699276834726334,
-0.03984978049993515,
0.02874603308737278,
0.051644522696733475,
-0.016170989722013474,
0.16000297665596008,
-0.028363078832626343,
-0.020744869485497475,
-0.03196737915277481,
0.0030147272627800703,
0.055691253393888474,
0.03123287484049797,
-0.08795982599258423,
-0.013192868791520596,
-0.07202818989753723,
0.035800784826278687,
-0.05740942060947418,
-0.028076592832803726,
-0.02569577284157276,
-0.05377485603094101,
-0.019182369112968445,
0.014601158909499645,
-0.01238683145493269,
-0.012393228709697723,
0.007156579755246639,
-0.03687017783522606,
-0.01782999187707901,
-0.02299690991640091,
0.02343878149986267,
0.043195225298404694,
0.02533819153904915,
0.06606487929821014,
0.05654497817158699,
-0.03845676779747009,
0.023118916898965836,
-0.014768050983548164,
0.02632071264088154,
-0.06425093859434128,
0.040221329778432846,
-0.05381052941083908,
-0.03200915455818176,
0.033123139292001724,
-0.059563666582107544,
0.013498730957508087,
0.005318839568644762,
0.033187638968229294,
-0.09124074131250381,
-4.8250935691385166e-8,
-0.02462630905210972,
0.08225476741790771,
0.055636558681726456,
0.07381784170866013,
0.0067410822957754135,
-0.015075474046170712,
-0.026420947164297104,
0.05792267248034477,
0.010935426689684391,
-0.004838180728256702,
0.07939451932907104,
0.03552887961268425,
-0.0057460549287498,
-0.029969429597258568,
0.0026332982815802097,
0.15446457266807556,
0.009279834106564522,
0.03808079659938812,
0.013047793880105019,
0.07277143001556396,
-0.030948426574468613,
0.02320774830877781,
0.05588829517364502,
-0.05367845296859741,
0.06533020734786987,
0.009522175416350365,
-0.0252788495272398,
0.006538944318890572,
0.04786514863371849,
0.05134919285774231,
0.03420637175440788,
0.020807858556509018,
0.051811814308166504,
-0.09311357885599136,
0.050297290086746216,
0.03057721257209778,
0.017000462859869003,
-0.0328206866979599,
0.011752165853977203,
-0.017193777486681938,
-0.029242906719446182,
0.06536061316728592,
-0.05029381439089775,
-0.014661351218819618,
0.09137455374002457,
0.010046171024441719,
0.0859101265668869,
0.03872781991958618,
-0.06366968154907227,
0.042565423995256424,
-0.012200728990137577,
-0.033628422766923904,
0.009042507968842983,
0.022736284881830215,
0.028754087164998055,
-0.04739009588956833,
0.0493914969265461,
0.002597647253423929,
0.0439421571791172,
0.0545370951294899,
0.07186514884233475,
-0.08151700347661972,
-0.051677852869033813,
-0.017382649704813957
] | 0 | 0 | 0 | 0 |
|
Decentralized Optimization with Coupled Constraints | https://openreview.net/forum?id=AJM52ygi6Y | [
"Demyan Yarmoshik",
"Alexander Rogozin",
"Nikita Kiselev",
"Daniil Dorin",
"Alexander Gasnikov",
"Dmitry Kovalev"
] | Poster | We consider the decentralized minimization of a separable objective $\sum_{i=1}^{n} f_i(x_i)$, where the variables are coupled through an affine constraint $\sum_{i=1}^n\left(\mathbf{A}_i x_i - b_i\right) = 0$.
We assume that the functions $f_i$, matrices $\mathbf{A}_i$, and vectors $b_i$ are stored locally by the nodes of a computational network, and that the functions $f_i$ are smooth and strongly convex.
This problem has significant applications in resource allocation and systems control and can also arise in distributed machine learning.
We propose lower complexity bounds for decentralized optimization problems with coupled constraints and a first-order algorithm achieving the lower bounds. To the best of our knowledge, our method is also the first linearly convergent first-order decentralized algorithm for problems with general affine coupled constraints. | decentralized optimization, convex optimization, affine constraints | Optimal algorithm and lower bound for smooth strongly convex setup | 13,609 | 2407.02020 | [
-0.015341143123805523,
-0.03150751441717148,
-0.05584139749407768,
-0.019596664234995842,
0.00867367535829544,
0.01721452921628952,
0.021820848807692528,
-0.052143994718790054,
-0.038491588085889816,
0.02214420586824417,
-0.06524141877889633,
0.048165831714868546,
0.0264813844114542,
0.013045616447925568,
-0.04296873137354851,
0.13319063186645508,
0.09773985296487808,
0.002594443503767252,
-0.015703924000263214,
0.03394033759832382,
-0.03637731820344925,
0.047525521367788315,
-0.05942271649837494,
-0.010930467396974564,
0.03717949241399765,
-0.08156104385852814,
0.02309228666126728,
0.029553724452853203,
-0.029763633385300636,
0.011029211804270744,
0.05503030866384506,
-0.0720120444893837,
0.009349717758595943,
-0.025244638323783875,
-0.021361535415053368,
0.08308219909667969,
-0.023474181070923805,
-0.01323511078953743,
0.016063116490840912,
0.012412585318088531,
0.0461224690079689,
0.07696107029914856,
0.018551351502537727,
-0.042090900242328644,
0.007377587258815765,
0.010468952357769012,
0.01771404966711998,
0.04396326094865799,
0.03462032228708267,
-0.006672216579318047,
-0.04924850910902023,
-0.0391535721719265,
-0.057457249611616135,
0.017247142270207405,
-0.004094498232007027,
-0.08472643792629242,
-0.020589038729667664,
-0.032807357609272,
0.0014774507144466043,
0.03257263824343681,
0.002107206266373396,
-0.09095092862844467,
0.01097369845956564,
0.01955566368997097,
0.009453030303120613,
0.011506877839565277,
0.04624330624938011,
0.004326608031988144,
-0.027374058961868286,
0.028305212035775185,
-0.01094651035964489,
-0.007515754550695419,
-0.04743285849690437,
-0.018499476835131645,
0.0625460296869278,
0.000740660703741014,
0.09184680134057999,
-0.0012175198644399643,
-0.023948850110173225,
-0.022878369316458702,
0.05473753437399864,
0.10852599143981934,
-0.02239697426557541,
-0.021342096850275993,
0.031913042068481445,
-0.12150655686855316,
-0.02045498602092266,
-0.026149416342377663,
0.10017102211713791,
-0.036764368414878845,
-0.04804341867566109,
0.11732200533151627,
-0.08809609711170197,
-0.016594214364886284,
0.008356180973351002,
-0.06366795301437378,
0.0710395947098732,
0.01591363176703453,
-0.006926418282091618,
0.13885922729969025,
-0.06425128877162933,
-0.06434378772974014,
0.021417109295725822,
-0.002045966684818268,
0.005981862545013428,
-0.015422229655086994,
0.0014900172827765346,
0.03686213493347168,
-0.00829306710511446,
-0.09619903564453125,
0.0404079370200634,
0.051759347319602966,
-0.04090774804353714,
0.046937983483076096,
-0.04616597667336464,
-0.027429163455963135,
0.06572111696004868,
-0.014547942206263542,
0.07239818572998047,
0.027723167091608047,
-0.023954641073942184,
-0.05808335915207863,
0.08499859273433685,
0.05124323442578316,
0.009810877963900566,
-0.010201322846114635,
-0.13229085505008698,
1.5328314091608162e-33,
0.0061073279939591885,
0.00711075309664011,
0.02469349279999733,
-0.049412261694669724,
0.03440253436565399,
-0.01037711463868618,
0.05530696362257004,
-0.00822458229959011,
-0.06911032646894455,
0.05309149995446205,
-0.005983375012874603,
0.021046819165349007,
-0.03975539654493332,
-0.02140137180685997,
0.10083015263080597,
-0.049993209540843964,
0.04454032704234123,
0.016822710633277893,
0.06549258530139923,
-0.11211084574460983,
0.0835646390914917,
-0.12062102556228638,
-0.016858572140336037,
-0.02081591822206974,
-0.01867733523249626,
-0.07512889057397842,
0.09910903126001358,
-0.0014605156611651182,
0.04938479885458946,
-0.008085152134299278,
-0.028907516971230507,
-0.024957725778222084,
0.07582281529903412,
0.07022270560264587,
0.044023461639881134,
0.02996094897389412,
-0.04588289558887482,
0.0023397253826260567,
-0.0390743613243103,
0.025234876200556755,
0.01624935492873192,
0.062174953520298004,
0.026401808485388756,
-0.06659757345914841,
-0.04619450867176056,
-0.08201143145561218,
-0.06257619708776474,
0.05472354218363762,
-0.0019130688160657883,
-0.06384310871362686,
-0.052890997380018234,
0.028235526755452156,
0.008481813594698906,
0.010738499462604523,
0.004066447261720896,
-0.031021814793348312,
-0.004526013042777777,
0.0153883658349514,
0.006624932400882244,
0.07135449349880219,
-0.019543271511793137,
-0.08328942209482193,
0.027635494247078896,
-0.030759023502469063,
-0.0010403133928775787,
0.06314481794834137,
0.004476680886000395,
0.04750273749232292,
0.106624536216259,
-0.0820022001862526,
-0.01578744687139988,
0.04849318787455559,
0.08308430016040802,
-0.04767962172627449,
0.07996796071529388,
0.05369824171066284,
-0.0028698688838630915,
-0.0422666110098362,
-0.0640488862991333,
-0.032729372382164,
-0.0004816540458705276,
0.046250540763139725,
-0.028413740918040276,
0.05907418951392174,
-0.0692906454205513,
-0.032427169382572174,
-0.04253578558564186,
0.016764359548687935,
-0.015971321612596512,
-0.019521070644259453,
-0.09816073626279831,
0.0362941175699234,
-0.0015624483348801732,
-0.00966219138354063,
0.03375314176082611,
-6.873104908175857e-34,
-0.008930142037570477,
-0.05342019721865654,
-0.021907398477196693,
0.020185595378279686,
0.050689779222011566,
-0.047334689646959305,
-0.05623096600174904,
-0.057869333773851395,
-0.009744110517203808,
0.03709179162979126,
-0.034061968326568604,
-0.0549059733748436,
0.011750339530408382,
-0.02925032563507557,
0.026223396882414818,
-0.002447689650580287,
-0.03621422126889229,
0.04073513299226761,
0.030098415911197662,
-0.02712809294462204,
-0.050741784274578094,
0.08991624414920807,
0.04474613070487976,
0.003324126359075308,
0.06484895199537277,
0.04103764519095421,
-0.060994528234004974,
0.08229042589664459,
-0.01494628842920065,
0.06337188929319382,
-0.0714448019862175,
-0.02726501040160656,
-0.08676586300134659,
-0.007359959185123444,
-0.012222799472510815,
0.0565880686044693,
-0.09327432513237,
0.00591229647397995,
-0.0315479077398777,
0.11707834154367447,
0.049097053706645966,
0.033510759472846985,
-0.06783106178045273,
0.0010432405397295952,
0.05226986110210419,
0.014133143238723278,
0.0023850747384130955,
-0.016241053119301796,
0.05352098494768143,
-0.01011613104492426,
-0.034519895911216736,
0.015180934220552444,
-0.0715375691652298,
0.0021230815909802914,
-0.057917963713407516,
0.009114540182054043,
0.02027914486825466,
0.09259123355150223,
0.08119341731071472,
-0.08590252697467804,
-0.07321705669164658,
-0.07125899940729141,
-0.06425449997186661,
0.06870270520448685,
0.09311217814683914,
-0.007242011837661266,
0.060520969331264496,
-0.03946363925933838,
0.09194698184728622,
0.09413542598485947,
0.008602921850979328,
0.04845700040459633,
0.027541836723685265,
0.06918572634458542,
-0.001758999191224575,
0.016055045649409294,
0.1117299422621727,
-0.03418118879199028,
-0.039721596986055374,
-0.0030385234858840704,
0.06714218854904175,
0.007090921048074961,
0.07270994782447815,
-0.006291781086474657,
0.015322050079703331,
0.019271647557616234,
0.06591589748859406,
-0.017246006056666374,
-0.012835212051868439,
-0.021402375772595406,
-0.021169884130358696,
0.01788526587188244,
0.05170958861708641,
-0.054605256766080856,
-0.09277316927909851,
-4.009421417094927e-8,
-0.07434342801570892,
0.023623507469892502,
-0.09826353192329407,
-0.0103702237829566,
0.07232896983623505,
-0.01228456199169159,
0.05609073117375374,
-0.01619848795235157,
0.029590437188744545,
-0.03784115985035896,
-0.047202181071043015,
-0.006734700407832861,
0.02390105091035366,
-0.06605220586061478,
-0.1192520409822464,
0.01119706965982914,
-0.08211851865053177,
-0.028521476313471794,
0.01062380988150835,
-0.08848417550325394,
-0.06507541984319687,
0.006014667917042971,
0.011553806252777576,
0.029099293053150177,
0.00842366274446249,
-0.054543156176805496,
-0.006575625855475664,
0.028655141592025757,
0.1314825564622879,
0.11503186821937561,
-0.11452692002058029,
0.008534098044037819,
0.060035645961761475,
0.008838784880936146,
0.03831568732857704,
0.09363371133804321,
-0.06739505380392075,
0.01824379898607731,
-0.0585484653711319,
-0.036056943237781525,
0.04169902205467224,
0.07506848126649857,
-0.012326073832809925,
0.002886660397052765,
0.05932673066854477,
-0.022314393892884254,
0.026262467727065086,
0.016919037327170372,
0.056292444467544556,
-0.03674044460058212,
0.07122848182916641,
0.04914659261703491,
0.02210301160812378,
0.03916630893945694,
0.010057199746370316,
-0.036258764564991,
0.0030516304541379213,
-0.10848654061555862,
0.08777846395969391,
0.03624916076660156,
-0.026847299188375473,
0.0022186608985066414,
-0.07500612735748291,
-0.03974632918834686
] | 0 | 0 | 0 | 0 |
|
You Only Prune Once: Designing Calibration-Free Model Compression With Policy Learning | https://openreview.net/forum?id=5RZoYIT3u6 | [
"Ayan Sengupta",
"Siddhant Chaudhary",
"Tanmoy Chakraborty"
] | Poster | The ever-increasing size of large language models (LLMs) presents significant challenges for deployment due to their heavy computational and memory requirements. Current model pruning techniques attempt to alleviate these issues by relying heavily on external calibration datasets to determine which parameters to prune or compress, thus limiting their flexibility and scalability across different compression ratios. Moreover, these methods often cause severe performance degradation, particularly in downstream tasks, when subjected to higher compression rates. In this paper, we propose *PruneNet*, a novel model compression method that addresses these limitations by reformulating model pruning as a policy learning process. PruneNet decouples the pruning process from the model architecture, eliminating the need for calibration datasets. It learns a stochastic pruning policy to assess parameter importance solely based on intrinsic model properties while preserving the spectral structure to minimize information loss. PruneNet can compress the LLaMA-2-7B model in just 15 minutes, achieving over 80\% retention of its zero-shot performance with a 30\% compression ratio, outperforming existing methods that retain only 75\% performance. Furthermore, on complex multitask language understanding tasks, PruneNet demonstrates its robustness by preserving up to 80\% performance of the original model, proving itself a superior alternative to conventional structured compression techniques. | Model Compression, Large Language Models, Structured Pruning | null | 13,607 | 2501.15296 | [
-0.02769903466105461,
-0.00045192340621724725,
-0.053626228123903275,
0.02136971987783909,
0.059577617794275284,
0.027859490364789963,
-0.0446571446955204,
-0.03144628182053566,
0.02959693968296051,
-0.022220725193619728,
0.003870051819831133,
0.0029221170116215944,
0.04567640647292137,
0.033040206879377365,
0.011469729244709015,
-0.005681508686393499,
0.004085663706064224,
0.09517234563827515,
-0.11426102370023727,
-0.07356104254722595,
0.008076735772192478,
-0.013986199162900448,
0.015099809505045414,
0.042418964207172394,
0.09056321531534195,
-0.028547154739499092,
-0.022793805226683617,
-0.038393694907426834,
-0.0028763182926923037,
-0.029553167521953583,
0.05425179377198219,
-0.050224028527736664,
0.04292444884777069,
-0.006727713160216808,
0.0008355790050700307,
0.02095041051506996,
-0.039628129452466965,
-0.01901807263493538,
-0.04497500881552696,
-0.03309663385152817,
-0.023724950850009918,
0.046514060348272324,
-0.0072769043035805225,
0.03938184678554535,
0.04472743347287178,
0.0147716598585248,
0.0405549593269825,
-0.08264686167240143,
0.005889551248401403,
-0.003061255207285285,
-0.04634671285748482,
-0.018812227994203568,
-0.0042279064655303955,
0.10239748656749725,
0.006094945129007101,
-0.01274740882217884,
-0.0558217316865921,
0.012711782939732075,
-0.06836288422346115,
-0.008434001356363297,
-0.06501076370477676,
-0.11832360923290253,
-0.086145780980587,
0.016962425783276558,
-0.05043475702404976,
-0.01152102556079626,
0.09863121807575226,
0.03969714790582657,
0.005861734040081501,
0.04391846805810928,
-0.04162438213825226,
0.07908959686756134,
-0.00660055223852396,
0.12259415537118912,
0.03029344417154789,
0.02723255753517151,
0.07893291115760803,
0.01848546415567398,
0.05594877153635025,
-0.04857666417956352,
-0.04160671308636665,
-0.04830960929393768,
0.011439476162195206,
0.004431554116308689,
0.06742269545793533,
-0.04201095178723335,
0.0327095091342926,
0.026390016078948975,
0.04579273611307144,
-0.023956799879670143,
-0.04639755189418793,
0.006243569310754538,
0.06967382878065109,
0.0019300997955724597,
-0.002243185415863991,
0.04356444627046585,
-0.03575415536761284,
-0.05177869275212288,
-0.08647558838129044,
0.08227355033159256,
-0.011972909793257713,
0.04751027002930641,
0.02974328212440014,
-0.033671874552965164,
0.005601261276751757,
-0.07235368341207504,
0.00652418052777648,
0.0671389177441597,
-0.0037210953887552023,
-0.04030640050768852,
0.049074314534664154,
0.05393966659903526,
0.015832219272851944,
-0.012031308375298977,
0.07188016921281815,
0.014744829386472702,
-0.05357247218489647,
-0.07762027531862259,
-0.00207878858782351,
0.0672297552227974,
-0.08103645592927933,
0.02290063165128231,
-0.03902807831764221,
0.03497592359781265,
0.022296864539384842,
-0.008643590845167637,
-0.04505067691206932,
5.0972469296447446e-33,
0.09346892684698105,
0.10832122713327408,
0.013720027171075344,
-0.03657116740942001,
0.10022267699241638,
0.025386348366737366,
0.04927505925297737,
0.06619609147310257,
-0.0064728534780442715,
0.003752610180526972,
-0.07032204419374466,
-0.015044542029500008,
-0.059611305594444275,
0.04376504197716713,
0.06104881316423416,
-0.02456672489643097,
-0.02677251771092415,
0.11298231035470963,
-0.03421817347407341,
-0.021641328930854797,
0.04676888510584831,
0.0011932672932744026,
0.004303164314478636,
-0.09934938699007034,
0.02030583843588829,
0.03578213229775429,
0.09083101898431778,
-0.024829905480146408,
-0.1017540767788887,
0.029331771656870842,
-0.04470163583755493,
-0.0054458254016935825,
-0.04690271243453026,
-0.0067120082676410675,
0.06706167012453079,
-0.0664258599281311,
-0.03279901668429375,
0.01018970925360918,
0.03506861627101898,
-0.008337091654539108,
-0.01812889240682125,
0.038421813398599625,
-0.009079970419406891,
-0.0350048802793026,
-0.07426022738218307,
-0.10580908507108688,
0.038549549877643585,
0.0156706590205431,
-0.03780152648687363,
-0.02788688614964485,
0.10863659530878067,
0.018157927319407463,
-0.00819761399179697,
0.02372901327908039,
-0.10458442568778992,
-0.023297257721424103,
0.07163763046264648,
-0.02723814733326435,
0.03306394815444946,
0.03942527994513512,
-0.032098401337862015,
0.004678494296967983,
0.007028529886156321,
0.07312242686748505,
0.08156202733516693,
0.03831326588988304,
-0.04467872157692909,
0.09567283093929291,
0.02584720589220524,
-0.07552780210971832,
-0.05228893831372261,
-0.01802830584347248,
0.0021475874818861485,
-0.079854317009449,
0.0519217774271965,
-0.03107822686433792,
0.11880042403936386,
-0.09847929328680038,
-0.022784817963838577,
0.04048172011971474,
-0.058087434619665146,
-0.0007151782047003508,
0.02377459593117237,
-0.04957178607583046,
-0.02183409407734871,
-0.06366026401519775,
0.06401359289884567,
-0.07534407824277878,
0.017585568130016327,
-0.04996340349316597,
-0.03828521445393562,
-0.045901425182819366,
-0.04389342665672302,
-0.04486912116408348,
0.028935451060533524,
-4.046619833580539e-33,
-0.008645238354802132,
0.008081560023128986,
-0.03776192665100098,
0.07038820534944534,
-0.013012895360589027,
-0.09471660107374191,
-0.005697940941900015,
0.06361792236566544,
-0.0682963952422142,
-0.12680740654468536,
-0.042150527238845825,
-0.056463468819856644,
0.03558146208524704,
-0.013160279020667076,
0.009828565642237663,
0.024870431050658226,
-0.05457087233662605,
-0.035735610872507095,
0.014569912105798721,
0.004381238482892513,
-0.05421552062034607,
0.07992791384458542,
-0.07544390112161636,
0.02404746226966381,
-0.016471704468131065,
-0.017322881147265434,
-0.03400810435414314,
0.031217487528920174,
0.059199728071689606,
-0.020996447652578354,
-0.05992238223552704,
-0.018941298127174377,
-0.08341655880212784,
-0.07786200940608978,
-0.09212625026702881,
0.004607535433024168,
0.01084450725466013,
0.05986792594194412,
0.07376470416784286,
0.15610508620738983,
0.09707274287939072,
-0.02283596433699131,
-0.11292877793312073,
0.013687781989574432,
0.021062208339571953,
-0.00842005480080843,
-0.05607856437563896,
-0.06808927655220032,
-0.0034708138555288315,
0.05306587740778923,
-0.02519829198718071,
0.07866788655519485,
-0.04068322852253914,
0.03317805007100105,
0.01465979591012001,
-0.1100691631436348,
-0.1097848042845726,
-0.008693448267877102,
0.038255494087934494,
0.015125100500881672,
-0.04734420031309128,
-0.014233573339879513,
0.027356352657079697,
-0.049186449497938156,
0.01732281967997551,
-0.01483332458883524,
0.012131934054195881,
0.048319246619939804,
0.013249523937702179,
0.0010934383608400822,
0.031836383044719696,
0.011078044772148132,
-0.03345782309770584,
0.06895988434553146,
-0.040394432842731476,
-0.007870658300817013,
0.006922387983649969,
-0.0046669249422848225,
-0.015262485481798649,
0.02108241245150566,
0.04729722812771797,
0.022385139018297195,
0.012478606775403023,
-0.005137073807418346,
0.046102870255708694,
0.09655401855707169,
0.07255803793668747,
0.04885701462626457,
-0.002931908704340458,
0.021766886115074158,
-0.05858491733670235,
0.0579647496342659,
0.0820862203836441,
0.10959741473197937,
-0.03818989172577858,
-4.454490465377603e-8,
-0.0877687931060791,
0.04499047249555588,
-0.004751511383801699,
0.08538900315761566,
0.04735363647341728,
0.0011231203097850084,
-0.06702063977718353,
0.11127731204032898,
0.047819193452596664,
-0.05096403509378433,
0.10636676847934723,
-0.007062830030918121,
-0.07359475642442703,
-0.005433059297502041,
-0.03549189865589142,
0.07380698621273041,
0.039780810475349426,
0.0033955282997339964,
-0.014823149889707565,
-0.00504247797653079,
0.02280426397919655,
0.031883154064416885,
-0.01974588632583618,
0.014729620888829231,
0.03490130603313446,
-0.07851818948984146,
-0.009995595552027225,
0.13114969432353973,
0.0814894288778305,
0.02733575366437435,
-0.04256743565201759,
0.018749400973320007,
-0.013124224729835987,
-0.08103389292955399,
-0.0015049512730911374,
0.09939049184322357,
0.08315989375114441,
0.004151553846895695,
0.0038983484264463186,
-0.04006753861904144,
0.008562680333852768,
0.03238909691572189,
-0.05708646774291992,
0.021962983533740044,
0.003604589030146599,
-0.040559932589530945,
-0.007747844327241182,
-0.10973949730396271,
0.020900942385196686,
-0.006913372315466404,
0.045295391231775284,
0.036127883940935135,
-0.040105827152729034,
0.04865747690200806,
-0.02843291126191616,
0.04310952499508858,
-0.022254260256886482,
-0.006605029571801424,
-0.002850567689165473,
0.06350144743919373,
0.008436383679509163,
-0.04005274921655655,
-0.08084359019994736,
0.004745622165501118
] | 0 | 0 | 0 | 0 |
|
FreDF: Learning to Forecast in the Frequency Domain | https://openreview.net/forum?id=4A9IdSa1ul | [
"Hao Wang",
"Lichen Pan",
"Yuan Shen",
"Zhichao Chen",
"Degui Yang",
"Yifei Yang",
"Sen Zhang",
"Xinggao Liu",
"Haoxuan Li",
"Dacheng Tao"
] | Poster | Time series modeling presents unique challenges due to autocorrelation in both historical data and future sequences. While current research predominantly addresses autocorrelation within historical data, the correlations among future labels are often overlooked. Specifically, modern forecasting models primarily adhere to the Direct Forecast (DF) paradigm, generating multi-step forecasts independently and disregarding label correlations over time. In this work, we demonstrate that the learning objective of DF is biased in the presence of label correlation. To address this issue, we propose the Frequency-enhanced Direct Forecast (FreDF), which mitigates label correlation by learning to forecast in the frequency domain, thereby reducing estimation bias. Our experiments show that FreDF significantly outperforms existing state-of-the-art methods and is compatible with a variety of forecast models. Code is available at https://github.com/Master-PLC/FreDF. | Time series, Long-term Forecast | Learning to forecast in the frequency domain improves forecasting performance. | 13,602 | null | [
-0.10944011062383652,
-0.07572510093450546,
-0.03651871904730797,
0.09047984331846237,
0.0029323361814022064,
0.03673557564616203,
-0.05211567506194115,
-0.0010945270769298077,
0.03684474900364876,
-0.013218113221228123,
-0.11076126992702484,
0.016099127009510994,
-0.043070241808891296,
0.04545742645859718,
-0.023661406710743904,
-0.05333935096859932,
-0.0034517571330070496,
-0.01662171445786953,
0.01374045666307211,
-0.12056464701890945,
0.013184268958866596,
0.06903185695409775,
0.0034192565362900496,
0.03160387650132179,
0.06067251041531563,
0.0027914296369999647,
-0.029610011726617813,
0.06589818000793457,
-0.014492075890302658,
0.0380280464887619,
-0.010294047184288502,
0.036799702793359756,
-0.015511651523411274,
-0.06705797463655472,
-0.06010468676686287,
-0.03134286031126976,
-0.018906371667981148,
0.09206502884626389,
0.05255268141627312,
0.011463846080005169,
0.062526173889637,
0.028033893555402756,
0.03277033939957619,
-0.006097197998315096,
-0.0450185127556324,
0.06256495416164398,
0.07017165422439575,
-0.03850133344531059,
-0.023134199902415276,
0.05311604589223862,
-0.0596524216234684,
0.0023451128508895636,
-0.021519044414162636,
0.03712137043476105,
-0.0004394892603158951,
0.08822762966156006,
0.0030604186467826366,
-0.010891607031226158,
0.03826163709163666,
0.077595554292202,
-0.09591000527143478,
-0.030241087079048157,
-0.04850165918469429,
-0.022318385541439056,
-0.03005819581449032,
0.046743281185626984,
-0.04306606575846672,
0.11582714319229126,
-0.09614847600460052,
0.01967361569404602,
-0.08051791042089462,
0.08928767591714859,
-0.06138811260461807,
0.034357450902462006,
0.05469084158539772,
0.06837134808301926,
0.020150400698184967,
-0.06718488782644272,
0.08256920427083969,
-0.05657769739627838,
-0.02687268704175949,
0.06471855193376541,
0.012974271550774574,
-0.04663655534386635,
0.09626085311174393,
0.05072709918022156,
0.04152630642056465,
-0.022046318277716637,
-0.005011413712054491,
-0.05670361965894699,
0.0640096664428711,
-0.05622486770153046,
0.06668494641780853,
0.01940244808793068,
-0.10366847366094589,
0.11546345800161362,
-0.008315992541611195,
-0.08920900523662567,
-0.04087112471461296,
0.0530732199549675,
0.0064836242236196995,
0.046118829399347305,
0.010840690694749355,
0.03389740735292435,
-0.00638476712629199,
-0.07979106158018112,
0.044777482748031616,
0.027259640395641327,
0.02698151022195816,
-0.06162754446268082,
0.000539081753231585,
0.019632546231150627,
0.003483641194179654,
-0.047788940370082855,
-0.006058827508240938,
0.04301958158612251,
0.022550828754901886,
-0.01814316026866436,
0.04383768513798714,
0.04908980429172516,
-0.06595758348703384,
-0.010047019459307194,
-0.010784892365336418,
0.05105089768767357,
-0.006466941442340612,
-0.006261683534830809,
-0.03698945790529251,
1.8905852169809098e-33,
0.012202455662190914,
0.049342382699251175,
0.06451721489429474,
-0.06828569620847702,
-0.00792071595788002,
-0.03404608368873596,
-0.037872374057769775,
0.013097691349685192,
0.09354423731565475,
0.0742596685886383,
0.024097202345728874,
0.0651845708489418,
-0.05892423540353775,
0.02569630928337574,
-0.06587602198123932,
-0.05076003447175026,
0.004476398695260286,
0.11002004891633987,
0.0393524095416069,
-0.0018962030299007893,
0.04213982820510864,
0.05074143782258034,
-0.002285976428538561,
-0.06370467692613602,
0.028715746477246284,
0.02998693287372589,
0.10260842740535736,
0.022090598940849304,
0.017638545483350754,
0.01578688994050026,
-0.0006458539282903075,
-0.02622578665614128,
0.008795513771474361,
0.011842025443911552,
-0.009610246866941452,
-0.09266749024391174,
-0.041191503405570984,
0.006150068715214729,
-0.03891023248434067,
0.011631953530013561,
-0.008020477369427681,
0.015514192171394825,
-0.03138435259461403,
-0.11277393251657486,
-0.022146303206682205,
-0.0037815929390490055,
0.03609912469983101,
-0.08434122800827026,
0.023535003885626793,
-0.00273696961812675,
0.047764603048563004,
-0.06255379319190979,
-0.03756774961948395,
-0.04315195977687836,
-0.07814650237560272,
0.04907890781760216,
0.036349061876535416,
-0.06547237187623978,
0.023564189672470093,
0.027816621586680412,
-0.0959429144859314,
0.07401156425476074,
0.041276104748249054,
-0.030969474464654922,
-0.05062530189752579,
0.04509751871228218,
0.06688939034938812,
0.035355664789676666,
0.027547374367713928,
0.003649069694802165,
0.08140187710523605,
0.0291209165006876,
-0.03560950234532356,
-0.0609714649617672,
0.08448328077793121,
0.021101249381899834,
0.021186459809541702,
-0.015869053080677986,
-0.03992050886154175,
0.06690214574337006,
0.0003823334991466254,
0.016284417361021042,
0.01947716251015663,
-0.009605885483324528,
0.0287851020693779,
0.03663276135921478,
0.05756416544318199,
-0.029902877286076546,
-0.01152415294200182,
-0.099809929728508,
-0.031043585389852524,
-0.0012141751358285546,
-0.04701602831482887,
0.04226227104663849,
0.016028467565774918,
-2.4855295402259553e-33,
0.042175471782684326,
0.01613830402493477,
0.04317007586359978,
-0.053417474031448364,
0.012090764939785004,
-0.037749405950307846,
-0.024135500192642212,
0.06073956936597824,
0.02343481406569481,
-0.0978635847568512,
-0.02233162149786949,
-0.10086075216531754,
-0.010680479928851128,
0.04023556783795357,
0.035554107278585434,
-0.08057046681642532,
-0.005173000972718,
-0.01147734560072422,
0.0056124161928892136,
-0.007045502308756113,
-0.03817455843091011,
-0.01988285593688488,
-0.06955636292695999,
0.055624984204769135,
-0.012462804093956947,
-0.08477049320936203,
0.028608880937099457,
0.12117769569158554,
-0.03974846377968788,
-0.01916559785604477,
-0.05926983803510666,
-0.045670609921216965,
0.05615736544132233,
-0.028194429352879524,
-0.05798054859042168,
0.07541458308696747,
-0.007878012023866177,
-0.005151920020580292,
-0.04480115696787834,
-0.03383741155266762,
0.04565485566854477,
-0.024644486606121063,
-0.018520668148994446,
-0.04768042266368866,
-0.05505841597914696,
-0.020729651674628258,
-0.032285794615745544,
0.053574200719594955,
0.05719795078039169,
0.07826108485460281,
-0.050657667219638824,
-0.011334574781358242,
-0.07172061502933502,
-0.06925146281719208,
-0.05788542702794075,
-0.03949279710650444,
0.046973876655101776,
0.02321663312613964,
0.02970152720808983,
0.059400152415037155,
-0.03377436101436615,
0.04799072816967964,
0.047542769461870193,
-0.03155362233519554,
0.05312719568610191,
0.02445880137383938,
0.004377968609333038,
-0.06649672240018845,
0.049184322357177734,
0.010521437972784042,
0.11219949275255203,
0.03234051540493965,
-0.05979851260781288,
0.03182819113135338,
-0.08110698312520981,
-0.021344177424907684,
0.035076890140771866,
-0.03081517107784748,
-0.0780835822224617,
0.03406989946961403,
0.008958984166383743,
0.008421172387897968,
-0.0032364886719733477,
0.08866626769304276,
-0.06282148510217667,
0.006784629542380571,
0.09809672832489014,
-0.04145197942852974,
-0.002563545713201165,
-0.06015121564269066,
-0.06791269779205322,
0.0003113973361905664,
-0.044962525367736816,
0.07853961735963821,
-0.01571301929652691,
-3.978612284072369e-8,
-0.03291312977671623,
0.02488347887992859,
0.006795031949877739,
0.058313921093940735,
0.007397362031042576,
-0.019594348967075348,
0.06023279204964638,
-0.007480298634618521,
-0.026059582829475403,
-0.01114917267113924,
0.09714944660663605,
-0.053372789174318314,
-0.08365859091281891,
0.09215319156646729,
-0.04092009738087654,
-0.010374905541539192,
0.045958176255226135,
0.015001662075519562,
0.018413247540593147,
-0.07180207222700119,
0.025088289752602577,
0.06279736757278442,
0.09566576778888702,
-0.09342221915721893,
0.12376619130373001,
-0.08394525200128555,
-0.021105928346514702,
0.10505736619234085,
0.10786611586809158,
0.06729162484407425,
0.012878984212875366,
-0.012341002002358437,
-0.07114645093679428,
-0.05441221967339516,
-0.0093533955514431,
-0.001965988427400589,
-0.020020345225930214,
-0.05639713630080223,
-0.01387988030910492,
-0.020625442266464233,
0.005700135603547096,
0.06870415806770325,
-0.10030048340559006,
0.019465086981654167,
-0.01136358454823494,
-0.0012733601033687592,
0.021013790741562843,
-0.07841163873672485,
0.06996186822652817,
-0.06093508377671242,
0.06089561805129051,
-0.024923082441091537,
0.09292774647474289,
-0.00740379374474287,
0.08171014487743378,
0.011572425253689289,
-0.005302322097122669,
-0.07244839519262314,
-0.030667511746287346,
0.07577428966760635,
0.018739890307188034,
0.0031801522709429264,
-0.057153940200805664,
0.06181278079748154
] | 0 | 0 | 0 | 0 |
|
SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration | https://openreview.net/forum?id=OL44KtasKc | [
"Jintao Zhang",
"Jia wei",
"Pengle Zhang",
"Jun Zhu",
"Jianfei Chen"
] | Poster | The transformer architecture predominates across various models. As the heart of the transformer, attention has a computational complexity of $O(N^2)$, compared to $O(N)$ for linear transformations. When handling large sequence lengths, attention becomes the primary time-consuming component. Although quantization has proven to be an effective method for accelerating model inference, existing quantization methods primarily focus on optimizing the linear layer.
In response, we first analyze the feasibility of quantization in attention detailedly. Following that, we propose SageAttention, a highly efficient and accurate quantization method for attention. The OPS (operations per second) of our approach outperforms FlashAttention2 and xformers by about 2.1x and 2.7x, respectively. SageAttention also achieves superior accuracy performance over FlashAttention3. Comprehensive experiments confirm that our approach incurs almost no end-to-end metrics loss across diverse models—including those for large language processing, image generation, and video generation. The code is available at https://github.com/thu-ml/SageAttention. | Attention, Quantization, quantized attention, inference acceleration | We propose a quantization method for Attention that achieves speedups of 2.1x and 2.7x compared to FlashAttention2 and xformers, respectively, without lossing end-to-end metrics across various models. | 13,596 | null | [
-0.02581658400595188,
-0.07964684814214706,
-0.003355717984959483,
-0.012068760581314564,
0.06455054879188538,
0.0218642707914114,
-0.050053536891937256,
-0.00723774079233408,
0.09669104963541031,
-0.06277777254581451,
-0.0008834511390887201,
-0.041179243475198746,
-0.07722111791372299,
0.025787046179175377,
0.01257818378508091,
0.014099220745265484,
0.030443521216511726,
0.0700969249010086,
-0.10545045137405396,
-0.06553176045417786,
0.09602098166942596,
-0.0626964271068573,
0.018749110400676727,
-0.04418772831559181,
0.07703808695077896,
0.0037420045118778944,
-0.019985144957900047,
-0.05182725191116333,
0.03508458286523819,
-0.04045289382338524,
0.03899617865681648,
-0.024509454146027565,
0.07158057391643524,
0.016342658549547195,
-0.06373334676027298,
0.09528736025094986,
-0.023114575073122978,
-0.039802730083465576,
-0.02324223332107067,
-0.049864206463098526,
0.008856598287820816,
-0.01987633667886257,
0.039589036256074905,
-0.011435492895543575,
0.09913002699613571,
0.007768815848976374,
0.07108049094676971,
-0.024275202304124832,
0.052033502608537674,
0.00529377069324255,
-0.06603279709815979,
-0.03023793175816536,
-0.05628553405404091,
0.07196903228759766,
-0.07137428224086761,
0.042578112334012985,
0.061965227127075195,
0.011109371669590473,
0.008986424654722214,
-0.018495751544833183,
-0.08491695672273636,
-0.01631058380007744,
0.04969717189669609,
-0.008548756130039692,
-0.04139164835214615,
-0.052589062601327896,
0.08036647737026215,
-0.07243799418210983,
-0.04567323997616768,
0.0019913949072360992,
-0.055393148213624954,
0.052916381508111954,
-0.05443210154771805,
0.07032491266727448,
0.028080737218260765,
0.0633525401353836,
0.03725534304976463,
-0.07740260660648346,
0.0659092590212822,
-0.07382435351610184,
0.06332024186849594,
-0.04112514853477478,
0.03407108411192894,
0.0062667978927493095,
0.1325731724500656,
-0.0013404152123257518,
-0.024129217490553856,
0.04983123019337654,
-0.015092871151864529,
0.050131361931562424,
-0.12923340499401093,
-0.03147796541452408,
0.06495703756809235,
-0.004616954363882542,
0.07051341235637665,
0.05220235884189606,
0.02276221476495266,
-0.05175120756030083,
-0.047098878771066666,
0.07898373156785965,
0.04174969717860222,
0.015613878145813942,
0.004228305537253618,
-0.029287278652191162,
0.0025371431838721037,
-0.025532804429531097,
0.05514359101653099,
0.06931255012750626,
-0.06622885167598724,
-0.08656928688287735,
0.0688568726181984,
0.01913338527083397,
-0.0071767354384064674,
-0.05920584872364998,
0.0829465389251709,
0.012830057181417942,
-0.0071531059220433235,
0.0035772838164120913,
-0.03220658376812935,
0.017358440905809402,
0.07198544591665268,
0.07562489807605743,
-0.02260459028184414,
0.09757963567972183,
0.0768241211771965,
-0.07425703853368759,
-0.02964170277118683,
5.018197873287823e-33,
0.01789318211376667,
0.05646365508437157,
0.06726863235235214,
0.04052915424108505,
0.020214708521962166,
0.002466042758896947,
0.05555448308587074,
0.026844650506973267,
-0.03905512019991875,
0.02227642573416233,
0.026902513578534126,
-0.01619080640375614,
-0.031023913994431496,
0.03916264325380325,
0.059598322957754135,
-0.01931685395538807,
-0.006355183199048042,
0.0736939013004303,
-0.057628557085990906,
-0.01332841720432043,
0.03031485341489315,
-0.07027579098939896,
0.012645813636481762,
-0.0012367734452709556,
-0.0597059465944767,
0.02695964276790619,
0.05123644694685936,
-0.05443323403596878,
-0.04552711546421051,
-0.018435657024383545,
-0.13755592703819275,
-0.05858570337295532,
-0.017960619181394577,
-0.02972353622317314,
0.07890003174543381,
-0.08493517339229584,
-0.003032147418707609,
-0.07904351502656937,
0.02622649446129799,
-0.009173153899610043,
-0.044094257056713104,
0.06964915990829468,
0.012599433772265911,
-0.018834080547094345,
-0.10851850360631943,
-0.006317782215774059,
-0.030136331915855408,
0.041239477694034576,
0.011077380739152431,
0.021607255563139915,
0.1153472289443016,
0.0062887356616556644,
-0.0315299853682518,
0.04490819573402405,
0.018566979095339775,
-0.019460570067167282,
0.08136726170778275,
0.026820752769708633,
0.015974225476384163,
0.04803035780787468,
0.011600636877119541,
0.009861204773187637,
-0.03199176490306854,
0.04195176810026169,
-0.03181562200188637,
-0.007170988246798515,
0.01061840821057558,
0.07368999719619751,
-0.010024956427514553,
0.0896553248167038,
-0.053573571145534515,
-0.004179186653345823,
-0.005943838506937027,
-0.0414142943918705,
0.061013538390398026,
0.00849450659006834,
0.04728131741285324,
-0.0638408213853836,
0.013315510004758835,
-0.00919919740408659,
-0.046243470162153244,
0.06856390833854675,
-0.006530565209686756,
0.010205553844571114,
-0.04190211370587349,
-0.03603729233145714,
0.020519614219665527,
-0.06433923542499542,
-0.044163838028907776,
-0.025313593447208405,
-0.018721679225564003,
0.0021719099022448063,
-0.05211737006902695,
-0.06023214012384415,
-0.057148389518260956,
-2.871347724658954e-33,
0.010050952434539795,
0.0530182421207428,
-0.029934609308838844,
0.07918946444988251,
-0.07407670468091965,
-0.09942415356636047,
0.005027574021369219,
0.011132541112601757,
0.03291293606162071,
-0.08002479374408722,
-0.03357814624905586,
-0.048343755304813385,
0.005539859179407358,
-0.041263651102781296,
0.05812397599220276,
-0.005843228194862604,
-0.016357723623514175,
0.015802815556526184,
-0.06756258010864258,
0.02569388411939144,
-0.01723518595099449,
0.08151528239250183,
0.01642526127398014,
-0.06174030527472496,
-0.06981584429740906,
0.03353632614016533,
-0.05967959761619568,
0.04756970703601837,
0.04363805055618286,
-0.08170562237501144,
-0.0295634213835001,
0.003126318333670497,
-0.05104498937726021,
-0.061308182775974274,
-0.05966794118285179,
0.03855694457888603,
0.08039738982915878,
0.02872665785253048,
-0.021983040496706963,
0.06641072779893875,
0.1094493716955185,
-0.09334009885787964,
-0.08912044018507004,
0.044878825545310974,
-0.053574685007333755,
0.034072622656822205,
-0.08982113003730774,
0.02459740824997425,
-0.013994773849844933,
0.08191897720098495,
-0.019389284774661064,
0.028870731592178345,
-0.047573916614055634,
0.034956201910972595,
-0.001783051062375307,
-0.06875844299793243,
-0.0875309556722641,
-0.03145008161664009,
0.07873383164405823,
-0.054858893156051636,
-0.03722076117992401,
0.027221858501434326,
-0.017308557406067848,
-0.10499384254217148,
0.049565840512514114,
0.007286124862730503,
-0.0030413216445595026,
-0.01111997663974762,
0.021300550550222397,
0.029678789898753166,
0.1027231365442276,
-0.029300518333911896,
-0.01593983732163906,
0.04258191958069801,
-0.04979310929775238,
-0.015636784955859184,
0.016324356198310852,
-0.03361748531460762,
0.04170123115181923,
-0.023302530869841576,
-0.002459730487316847,
0.009837049059569836,
0.042749565094709396,
-0.04829790070652962,
0.013881203718483448,
0.10178902745246887,
0.11502976715564728,
0.04652443900704384,
0.018582649528980255,
0.034410495311021805,
-0.08130089938640594,
0.12290757149457932,
0.05435251072049141,
0.04046226292848587,
-0.053263138979673386,
-4.486202698217312e-8,
-0.08742862939834595,
0.09865877777338028,
-0.019504092633724213,
-0.010284926742315292,
-0.024459144100546837,
-0.04607265442609787,
-0.005078890826553106,
0.13113071024417877,
0.017312098294496536,
-0.03511614352464676,
0.11329246312379837,
-0.02969217672944069,
0.028012016788125038,
-0.024711938574910164,
-0.038228195160627365,
0.05930550396442413,
0.04334036260843277,
-0.01717112585902214,
-0.033928755670785904,
-0.046008627861738205,
-0.05348321050405502,
0.11142555624246597,
-0.02584872767329216,
-0.05180499702692032,
-0.011422498151659966,
-0.07170975208282471,
0.0063238381408154964,
0.12036330252885818,
0.10391218215227127,
-0.013869422487914562,
0.019803883507847786,
0.039687611162662506,
0.04210124537348747,
-0.04651203379034996,
0.03263949230313301,
0.03155452758073807,
-0.0025324481539428234,
-0.004140487406402826,
0.045021746307611465,
-0.033320147544145584,
-0.024310119450092316,
0.03029068000614643,
-0.05932815745472908,
0.03832674026489258,
-0.005174391902983189,
-0.03507322072982788,
-0.003404834307730198,
-0.06591077893972397,
0.02787191793322563,
-0.038306258618831635,
0.019186243414878845,
-0.007963971234858036,
-0.06670448929071426,
0.026571014896035194,
-0.03266887366771698,
0.007309803739190102,
0.04820400848984718,
-0.026787161827087402,
0.007135105784982443,
0.034873027354478836,
0.005068080499768257,
-0.023235177621245384,
-0.01935107819736004,
-0.004956691991537809
] | 0 | 0 | 0 | 0 |
|
Neural Multi-Objective Combinatorial Optimization via Graph-Image Multimodal Fusion | https://openreview.net/forum?id=4sJ2FYE65U | [
"Jinbiao Chen",
"Jiahai Wang",
"Zhiguang Cao",
"Yaoxin Wu"
] | Poster | Existing neural multi-objective combinatorial optimization (MOCO) methods still exhibit an optimality gap since they fail to fully exploit the intrinsic features of problem instances. A significant factor contributing to this shortfall is their reliance solely on graph-modal information. To overcome this, we propose a novel graph-image multimodal fusion (GIMF) framework that enhances neural MOCO methods by integrating graph and image information of the problem instances. Our GIMF framework comprises three key components: (1) a constructed coordinate image to better represent the spatial structure of the problem instance, (2) a problem-size adaptive resolution strategy during the image construction process to improve the cross-size generalization of the model, and (3) a multimodal fusion mechanism with modality-specific bottlenecks to efficiently couple graph and image information. We demonstrate the versatility of our GIMF by implementing it with two state-of-the-art neural MOCO backbones. Experimental results on classic MOCO problems show that our GIMF significantly outperforms state-of-the-art neural MOCO methods and exhibits superior generalization capability. | Neural Multi-Objective Combinatorial Optimization, Multimodal Fusion, Deep Reinforcement Learning | This paper proposes a graph-image multimodal fusion framework for neural multi-objective combinatorial optimization. | 13,570 | null | [
-0.01708954945206642,
-0.12935195863246918,
0.06167563423514366,
0.020073797553777695,
0.0499006025493145,
-0.09705570340156555,
0.003896881127730012,
-0.0736173689365387,
0.02101214975118637,
-0.025695182383060455,
-0.0008232670370489359,
-0.04045902565121651,
0.07225503027439117,
0.11616574227809906,
0.010839423164725304,
0.04871020466089249,
0.018411073833703995,
0.09446155279874802,
-0.08505607396364212,
-0.01266881637275219,
0.029109884053468704,
-0.08958116918802261,
0.014531652443110943,
-0.06868532299995422,
0.05266691371798515,
0.042620278894901276,
-0.06556067615747452,
-0.04651956632733345,
0.054336581379175186,
-0.004320845007896423,
0.08675646036863327,
0.049889370799064636,
-0.06288660317659378,
0.04945821315050125,
0.0019681144040077925,
0.08196288347244263,
-0.04685189574956894,
0.03184107691049576,
0.0136038176715374,
0.024680910632014275,
-0.03038599155843258,
0.12292616814374924,
0.04657034948468208,
-0.01651737466454506,
0.0670984610915184,
0.0975806713104248,
-0.027905989438295364,
-0.059247441589832306,
0.06749435514211655,
0.0005509507027454674,
-0.10234246402978897,
-0.06823480129241943,
-0.1351093053817749,
0.03780356049537659,
0.10881169885396957,
0.010558433830738068,
0.018773691728711128,
-0.05663037672638893,
0.036016177386045456,
0.07143683731555939,
-0.031863853335380554,
0.0036244895309209824,
-0.023637663573026657,
0.018389558419585228,
0.09390581399202347,
-0.06336453557014465,
0.030411342158913612,
-0.051911547780036926,
-0.03318718075752258,
0.0032083585392683744,
0.027025457471609116,
-0.02352004125714302,
0.032489631325006485,
0.0015238204505294561,
0.015540174208581448,
0.057179391384124756,
0.08142047375440598,
-0.0042150262743234634,
0.01668281853199005,
-0.02078697271645069,
0.06674381345510483,
0.06743090599775314,
0.03648640960454941,
0.0038144588470458984,
0.11554062366485596,
-0.005837193224579096,
-0.03753204643726349,
0.0024108539801090956,
-0.01701268181204796,
0.030931826680898666,
-0.07493345439434052,
-0.06554831564426422,
0.004512053448706865,
0.01334545947611332,
0.04372018948197365,
-0.03372807800769806,
0.05946743115782738,
-0.02277933806180954,
-0.002095583127811551,
0.06179564818739891,
-0.011901697143912315,
-0.05094094201922417,
0.08327696472406387,
-0.031106198206543922,
0.0703003853559494,
-0.012927901931107044,
0.0416160449385643,
0.06054089963436127,
0.035301610827445984,
-0.029778212308883667,
0.04466972500085831,
0.09231579303741455,
0.027700208127498627,
0.03922702372074127,
-0.018261993303894997,
-0.016713585704565048,
0.0061536082066595554,
0.026130452752113342,
0.08301644772291183,
0.06730524450540543,
-0.03440601006150246,
-0.0070695034228265285,
-0.06187871843576431,
-0.03491048142313957,
-0.032149914652109146,
0.043330393731594086,
-0.04720258340239525,
1.4919108896052576e-33,
0.04937687888741493,
0.033274903893470764,
0.09217432141304016,
0.02124263159930706,
0.09081006050109863,
0.008024078793823719,
-0.031087007373571396,
-0.003393970662727952,
-0.02981949970126152,
0.03501152619719505,
-0.04352201521396637,
-0.013474080711603165,
-0.06083144620060921,
0.11377847194671631,
0.03857593983411789,
-0.051414791494607925,
-0.07133296132087708,
-0.01499141938984394,
-0.0038946501445025206,
0.016142506152391434,
-0.02450893260538578,
-0.04073435440659523,
0.034986671060323715,
0.05231231078505516,
0.04034235328435898,
0.013233632780611515,
0.06304120272397995,
-0.08318430930376053,
-0.027704527601599693,
0.00869405921548605,
-0.1399981528520584,
0.007940839044749737,
0.006801339332014322,
0.06036803126335144,
-0.007225074339658022,
-0.0788155049085617,
-0.06392450630664825,
-0.06726758927106857,
-0.0396311916410923,
0.0341644212603569,
-0.006337004713714123,
-0.03736993297934532,
-0.09297244995832443,
0.0039808922447264194,
-0.07883502542972565,
-0.03539599850773811,
0.019196443259716034,
0.025218170136213303,
-0.02997910976409912,
-0.023577099665999413,
-0.015988774597644806,
0.02778918482363224,
-0.011211472563445568,
-0.04191555455327034,
-0.011535988189280033,
-0.03666491433978081,
0.01853511855006218,
-0.0723491907119751,
0.08070676028728485,
-0.015688233077526093,
-0.06094036251306534,
0.002370018744841218,
-0.03343772515654564,
0.05362270027399063,
0.051184289157390594,
0.010309808887541294,
-0.059672918170690536,
-0.03548989072442055,
0.10409728437662125,
0.009076290763914585,
-0.018210994079709053,
0.09133566915988922,
0.020850500091910362,
-0.0808638334274292,
0.018324969336390495,
0.0005268172826617956,
0.09343071281909943,
-0.05914019048213959,
0.0013782931491732597,
0.01785055547952652,
-0.03321201726794243,
0.07501296699047089,
0.04096668213605881,
-0.050437964498996735,
-0.011273158714175224,
0.02705085091292858,
0.07869011908769608,
0.03533528372645378,
-0.057763706892728806,
-0.04557061567902565,
-0.03897949308156967,
-0.020506424829363823,
-0.04059883579611778,
-0.045434340834617615,
-0.0010062928777188063,
-5.078609925889938e-34,
0.0648457482457161,
-0.033586595207452774,
0.015913892537355423,
-0.03527878597378731,
0.015468845143914223,
-0.051783107221126556,
-0.008553152903914452,
-0.08518997579813004,
-0.03181561082601547,
-0.020347241312265396,
-0.05051850527524948,
-0.02393488585948944,
0.007462408859282732,
-0.05792354419827461,
-0.05100708827376366,
-0.026288947090506554,
-0.0449821762740612,
-0.009502139873802662,
-0.013281844556331635,
0.060994479805231094,
0.05919818952679634,
0.09777908772230148,
0.003543686354532838,
-0.004401565063744783,
-0.06758773326873779,
0.038953863084316254,
-0.061696529388427734,
-0.03704221919178963,
0.05045027658343315,
0.041885536164045334,
0.04316556453704834,
-0.08057475090026855,
-0.03872008994221687,
-0.020884184166789055,
0.023887354880571365,
0.07017768174409866,
-0.035058289766311646,
0.0028765485621988773,
-0.018452895805239677,
0.08430405706167221,
-0.015993259847164154,
-0.055220216512680054,
-0.09003408998250961,
0.033532217144966125,
0.016788160428404808,
0.002062846440821886,
-0.041277870535850525,
-0.025467410683631897,
0.03487519919872284,
0.06173054873943329,
-0.03145422413945198,
-0.02414124086499214,
-0.08293422311544418,
-0.0043939827010035515,
-0.03461003303527832,
-0.11180268973112106,
-0.008937607519328594,
0.0032791844569146633,
0.09483499079942703,
-0.03790125623345375,
-0.02642364427447319,
-0.09893621504306793,
-0.0018548138905316591,
-0.07306066900491714,
-0.007844928652048111,
0.04163699969649315,
-0.042445193976163864,
-0.06208396703004837,
0.015311924740672112,
0.024923764169216156,
-0.051006730645895004,
0.011400224640965462,
-0.029836738482117653,
0.0033171135000884533,
-0.02234414778649807,
0.032252371311187744,
0.00882222130894661,
0.061450108885765076,
0.051630910485982895,
-0.01566684991121292,
0.025260834023356438,
-0.0519605427980423,
-0.005946981720626354,
0.11000142246484756,
0.07667449861764908,
0.10889648646116257,
0.02729843556880951,
0.0997249037027359,
0.09588943421840668,
-0.0001438600884284824,
0.004511740989983082,
-0.031058717519044876,
-0.009202839806675911,
0.014484639279544353,
0.05920199304819107,
-3.9953398811576335e-8,
-0.09604475647211075,
0.046632975339889526,
0.004340990446507931,
-0.05114099755883217,
0.02798478491604328,
-0.03383563458919525,
0.08775350451469421,
0.03324084356427193,
0.028073644265532494,
-0.05227869749069214,
0.14733557403087616,
0.022759221494197845,
0.01343170553445816,
0.007316390983760357,
-0.0059647043235599995,
-0.0329761765897274,
-0.013216212391853333,
-0.01977040059864521,
0.06649163365364075,
0.005256549455225468,
-0.038403384387493134,
-0.04077184572815895,
0.013797641731798649,
0.07450727373361588,
-0.030021904036402702,
-0.0898350402712822,
-0.03116024099290371,
0.05008753016591072,
-0.03148828446865082,
-0.03278392180800438,
-0.044412657618522644,
0.008464113809168339,
0.038500551134347916,
-0.02040797844529152,
-0.03316335380077362,
0.02703719213604927,
0.030847297981381416,
0.023059358820319176,
-0.096078060567379,
-0.07713604718446732,
0.029429666697978973,
0.044494930654764175,
-0.010160977952182293,
0.0372336208820343,
-0.029804211109876633,
0.06815595179796219,
0.03528861328959465,
-0.051185112446546555,
-0.02985277958214283,
0.04076693207025528,
0.04102463275194168,
0.00564562389627099,
-0.026689086109399796,
0.07383278757333755,
0.05280815437436104,
-0.03433063626289368,
0.03068423829972744,
-0.07327698171138763,
0.05631539970636368,
0.07408953458070755,
-0.03437730669975281,
-0.010409396141767502,
-0.07845411449670792,
-0.07033038884401321
] | 0 | 0 | 0 | 0 |
|
Have the VLMs Lost Confidence? A Study of Sycophancy in VLMs | https://openreview.net/forum?id=E2PFv7ad3p | [
"Shuo Li",
"Tao Ji",
"Xiaoran Fan",
"Linsheng Lu",
"Leyi Yang",
"Yuming Yang",
"Zhiheng Xi",
"Rui Zheng",
"Yuran Wang",
"xh.zhao",
"Tao Gui",
"Qi Zhang",
"Xuanjing Huang"
] | Poster | In the study of LLMs, sycophancy represents a prevalent hallucination that poses significant challenges to these models. Specifically, LLMs often fail to adhere to original correct responses, instead blindly agreeing with users' opinions, even when those opinions are incorrect or malicious. However, research on sycophancy in visual language models (VLMs) has been scarce. In this work, we extend the exploration of sycophancy from LLMs to VLMs, introducing the MM-SY benchmark to evaluate this phenomenon. We present evaluation results from multiple representative models, addressing the gap in sycophancy research for VLMs. To mitigate sycophancy, we propose a synthetic dataset for training and employ methods based on prompts, supervised fine-tuning, and DPO. Our experiments demonstrate that these methods effectively alleviate sycophancy in VLMs. Additionally, we probe VLMs to assess the semantic impact of sycophancy and analyze the attention distribution of visual tokens. Our findings indicate that the ability to prevent sycophancy is predominantly observed in higher layers of the model. The lack of attention to image knowledge in these higher layers may contribute to sycophancy, and enhancing image attention at high layers proves beneficial in mitigating this issue. | Multi-modal Model, Visual-Language Model, Sycophancy, Hallucination | Have the VLMs Lost Confidence? A Study of Sycophancy in VLMs | 13,565 | 2410.11302 | [
0.06219777837395668,
-0.12652911245822906,
0.014988565817475319,
0.019824836403131485,
0.060573939234018326,
0.005442068446427584,
0.03139607980847359,
0.030946092680096626,
0.048057492822408676,
0.014879281632602215,
-0.04676186665892601,
-0.10459130257368088,
0.04097375273704529,
0.0017534679500386119,
0.023511502891778946,
-0.006848474033176899,
0.14019235968589783,
0.023669302463531494,
-0.07063697278499603,
0.008593158796429634,
0.05074761435389519,
0.047483671456575394,
0.0512457937002182,
-0.007966478355228901,
-0.003564943093806505,
0.006200146861374378,
-0.032983019948005676,
-0.01526598073542118,
-0.02018885314464569,
-0.05192592740058899,
-0.03700808435678482,
0.04109577089548111,
-0.004131266847252846,
0.06877405196428299,
-0.0597229041159153,
0.014603819698095322,
-0.008901710622012615,
0.05349767580628395,
0.05113433301448822,
-0.12693314254283905,
-0.10072188824415207,
-0.061497773975133896,
0.031172381713986397,
0.010936434380710125,
0.08256072551012039,
0.0017535366350784898,
0.00020931677136104554,
-0.004820383153855801,
-0.03152528032660484,
-0.04201212897896767,
-0.0908256247639656,
-0.07311353087425232,
0.02233506552875042,
0.04499843344092369,
-0.10662291944026947,
0.036759138107299805,
0.0374218188226223,
0.0435166098177433,
0.0235606599599123,
-0.004869117867201567,
-0.10331827402114868,
-0.07993614673614502,
-0.0798536017537117,
0.01235350500792265,
-0.024410588666796684,
0.02695338986814022,
0.006166634149849415,
0.013597577810287476,
-0.027251655235886574,
0.044255580753088,
-0.023891257122159004,
0.04923282936215401,
0.0012250555446371436,
0.024209288880228996,
0.021641157567501068,
0.08392705768346786,
-0.006159990094602108,
-0.07099617272615433,
0.11070128530263901,
-0.09218940138816833,
0.0235016830265522,
0.07708238065242767,
0.007045484613627195,
0.01976342312991619,
0.1498194932937622,
0.0009716737549751997,
0.02626863867044449,
0.049503084272146225,
-0.0649532899260521,
0.016332874074578285,
-0.07681212574243546,
-0.06397239863872528,
0.09265510737895966,
-0.002400076948106289,
0.06656362861394882,
0.05653105676174164,
0.00017603598826099187,
-0.07409650087356567,
0.03795015811920166,
0.04827561974525452,
-0.013509457930922508,
0.07674098014831543,
0.003092352766543627,
-0.02617575041949749,
0.018267201259732246,
0.01464463584125042,
0.07980720698833466,
0.023875879123806953,
0.012875877320766449,
-0.07085065543651581,
-0.05223139747977257,
0.09166394174098969,
0.009390274062752724,
-0.023415599018335342,
0.041991107165813446,
-0.058592360466718674,
-0.010680900886654854,
-0.0585523322224617,
0.03519865497946739,
0.012642085552215576,
0.030597643926739693,
-0.024829944595694542,
0.04017747566103935,
-0.029890887439250946,
0.08823441714048386,
-0.06196229159832001,
-0.06749168783426285,
2.1509456444155306e-33,
0.010296941734850407,
0.0146751394495368,
0.011703096330165863,
0.05865875631570816,
0.08365390449762344,
0.016105156391859055,
-0.04742499813437462,
-0.018576517701148987,
-0.015237116254866123,
0.014771359972655773,
-0.06752292066812515,
-0.07304082810878754,
-0.00039840611862018704,
0.09857822954654694,
-0.005315741058439016,
0.03624927997589111,
-0.03204776346683502,
0.04980500414967537,
-0.12064666301012039,
-0.03918341174721718,
0.013906450010836124,
-0.040312886238098145,
0.02619507722556591,
-0.01855410821735859,
0.003221361432224512,
0.01649770326912403,
0.0600312277674675,
-0.01566329039633274,
-0.02935338392853737,
0.04616511985659599,
-0.099159374833107,
0.04555072262883186,
0.02375742979347706,
0.0442882776260376,
0.04407951608300209,
-0.004607933573424816,
0.005386258475482464,
-0.020236918702721596,
0.0544600747525692,
-0.017612701281905174,
0.011978354305028915,
0.04830082505941391,
-0.05715752765536308,
0.022774580866098404,
-0.0798206701874733,
0.02573459967970848,
0.00764329032972455,
0.007400161121040583,
-0.07267408072948456,
-0.01060683187097311,
0.03726769983768463,
0.023278582841157913,
-0.023979980498552322,
0.01571640558540821,
-0.04378047585487366,
0.022892145439982414,
0.005950447171926498,
0.014758582226932049,
-0.036004528403282166,
-0.007898567244410515,
0.0024541555903851986,
0.08988484740257263,
-0.012348001822829247,
0.04425346106290817,
0.013895921409130096,
-0.012217785231769085,
-0.043231621384620667,
0.018971331417560577,
0.0006079942686483264,
-0.035452693700790405,
0.010271665640175343,
0.03371633216738701,
-0.06335403025150299,
-0.04384788125753403,
-0.012688590213656425,
-0.028590435162186623,
0.06553377211093903,
-0.017093850299715996,
0.031211785972118378,
0.02904088795185089,
0.11479434370994568,
-0.05941831320524216,
-0.07027789205312729,
-0.0974595695734024,
-0.029326824471354485,
-0.1076238825917244,
0.022003641352057457,
-0.10038179904222488,
-0.04388701543211937,
-0.08167718350887299,
-0.01900431513786316,
-0.06054762378334999,
0.00764525355771184,
-0.012081051245331764,
-0.07959575951099396,
-2.880004505868791e-33,
-0.04594625532627106,
0.011120048351585865,
0.009184632450342178,
0.0778287723660469,
-0.10720837861299515,
-0.021752018481492996,
0.06247471272945404,
0.03676579147577286,
-0.03718898445367813,
-0.06663740426301956,
-0.00135141983628273,
-0.06227913498878479,
-0.004792325664311647,
0.03930935263633728,
0.022103728726506233,
-0.0036803504917770624,
0.008882173337042332,
-0.034738291054964066,
-0.006920387968420982,
0.04362567141652107,
0.010100256651639938,
0.08342879265546799,
-0.12534016370773315,
-0.04718469828367233,
-0.022099768742918968,
0.03138391673564911,
-0.07966899871826172,
0.052642010152339935,
-0.03511404991149902,
-0.01653609424829483,
-0.0003528915112838149,
0.0835023820400238,
-0.016026075929403305,
-0.05295436084270477,
0.005109656136482954,
0.07667145133018494,
0.04768769443035126,
-0.08597398549318314,
-0.053153492510318756,
0.06790684908628464,
0.04443389177322388,
-0.05048425495624542,
-0.06147991120815277,
-0.05735042318701744,
-0.040168650448322296,
-0.06737080961465836,
-0.023327697068452835,
-0.03757372871041298,
-0.005410671234130859,
0.01573624089360237,
-0.004126433748751879,
0.0239486712962389,
-0.018865222111344337,
0.03384023532271385,
-0.058901552110910416,
-0.07426919043064117,
-0.051859300583601,
0.002075182506814599,
0.049675844609737396,
-0.027665842324495316,
-0.0490262508392334,
-0.016551529988646507,
0.004497222602367401,
-0.05211225897073746,
0.05434497073292732,
-0.038120467215776443,
-0.03726020082831383,
-0.026189979165792465,
0.08186818659305573,
-0.05392658710479736,
0.04237647354602814,
-0.019097458571195602,
-0.07188614457845688,
0.008125082589685917,
0.03158242627978325,
-0.03468826040625572,
-0.02230994589626789,
-0.01896839775145054,
-0.02918040007352829,
-0.023059142753481865,
0.039770834147930145,
-0.05277097970247269,
0.008722303435206413,
0.09307827800512314,
-0.002022217260673642,
0.08705408126115799,
0.04051774740219116,
0.05300535634160042,
-0.013893126510083675,
0.0034630452282726765,
-0.045895956456661224,
0.0013764186296612024,
-0.022684866562485695,
0.13358792662620544,
-0.03403186425566673,
-4.389578478480871e-8,
-0.059014055877923965,
-0.027674078941345215,
0.03459859639406204,
0.032704416662454605,
0.06006215512752533,
-0.0014359038323163986,
-0.04636291414499283,
0.14080865681171417,
-0.06597279012203217,
0.017877278849482536,
0.02101065032184124,
0.004962878301739693,
-0.061962470412254333,
-0.0531979538500309,
0.06636425107717514,
0.05952491983771324,
0.007833967916667461,
0.08876010030508041,
-0.04045833274722099,
-0.02761128544807434,
0.07984102517366409,
0.0772477313876152,
0.022155284881591797,
-0.04575316607952118,
0.060862500220537186,
-0.02880287542939186,
-0.023236380890011787,
0.033463362604379654,
-0.030947545543313026,
-0.03472701087594032,
0.07054349780082703,
0.05222562327980995,
-0.041913777589797974,
-0.07567524164915085,
0.0077628991566598415,
0.08525346964597702,
0.024607859551906586,
-0.04430234804749489,
0.022296559065580368,
-0.005497680976986885,
0.03375307843089104,
0.10017424821853638,
-0.01611309126019478,
0.05685676634311676,
0.035094063729047775,
0.048089154064655304,
-0.02061399444937706,
-0.08846882730722427,
0.009461687877774239,
-0.09193235635757446,
-0.025893431156873703,
0.03039068728685379,
-0.028772039338946342,
0.13090203702449799,
-0.022175656631588936,
0.06060405820608139,
0.03067902661859989,
-0.0011971152853220701,
0.009063588455319405,
0.08670303225517273,
0.06259980797767639,
0.06654202193021774,
-0.042860157787799835,
-0.01756257936358452
] | 0 | 0 | 0 | 0 |
|
Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass | https://openreview.net/forum?id=bc3sUsS6ck | [
"Tong Chen",
"Hao Fang",
"Patrick Xia",
"Xiaodong Liu",
"Benjamin Van Durme",
"Luke Zettlemoyer",
"Jianfeng Gao",
"Hao Cheng"
] | Poster | Large language models (LLMs) acquire substantial knowledge during pretraining but often need adaptation to new contexts, tasks, or domains, typically achieved through fine-tuning or prompting. However, fine-tuning incurs significant training costs, while prompting increases inference overhead. Inspired by fast weight memory, we introduce GenerativeAdapter, an effective and efficient adaptation method that encode test-time context into language model parameters with a single forward pass.
GenerativeAdapter augments a frozen pretrained LM with a lightweight adapter generator, trained via self-supervised learning, to produce parameter-efficient adapters.
Notably, our generator is general-purpose, i.e., one generator can adapt the corresponding base model for all langauge processing scenarios.
We apply GenerativeAdapter to two pretrained LMs (Mistral-7B-Instruct and Llama2-7B-Chat) and evaluate the adapted models across knowledge acquisition from documents, learning from demonstrations, and personalization for users.
In StreamingQA, our approach is effective in injecting knowledge into the LM's parameters, achieving a 63.5\% improvement in F1 score over the model with supervised fine-tuning (from $19.5$ to $31.5$) for contexts as long as 32K tokens.
In the MetaICL in-context learning evaluation, our method achieves an average accuracy of $44.9$ across 26 tasks, outperforming the base model.
On MSC, our method proves to be highly competitive in memorizing user information from conversations with a 4x reduction in computation and memory costs compared to
prompting with full conversation history.
Overall, GenerativeAdapter provides a viable solution for adapting large LMs to evolving information and providing tailored user experience, while reducing training and inference costs relative to traditional fine-tuning and prompting techniques. | language model; efficient adaptation; fine-tuning; prompting | null | 13,563 | 2411.05877 | [
-0.029294373467564583,
-0.08528949320316315,
0.05706203356385231,
0.05936508998274803,
0.031653594225645065,
0.04363562539219856,
0.030000828206539154,
-0.0005516217788681388,
-0.050364717841148376,
-0.047026634216308594,
-0.004359576851129532,
-0.12838751077651978,
0.051820024847984314,
-0.004663923755288124,
0.04357433691620827,
-0.007435317151248455,
0.11216521263122559,
0.033798255026340485,
-0.058850694447755814,
-0.03667831793427467,
0.03937436640262604,
0.042892877012491226,
0.037511494010686874,
0.006015487015247345,
0.040588896721601486,
-0.007991064339876175,
-0.010850323364138603,
-0.049618568271398544,
0.06917581707239151,
-0.043747641146183014,
0.09496916085481644,
0.05815619230270386,
-0.06551641970872879,
0.07517648488283157,
-0.06852296739816666,
0.044590894132852554,
-0.09094072878360748,
0.03653230518102646,
-0.01007478404790163,
-0.05249950289726257,
-0.0351652167737484,
-0.018955878913402557,
0.02487850748002529,
-0.004037878010421991,
0.02389259822666645,
0.021421628072857857,
-0.0008334357407875359,
0.024556070566177368,
-0.010793059132993221,
0.08438844978809357,
-0.0638660341501236,
-0.04572797194123268,
-0.016108151525259018,
0.03542090579867363,
-0.044274743646383286,
0.06282959878444672,
0.0640207827091217,
0.04327176511287689,
-0.01019041333347559,
0.015005291439592838,
-0.09349711984395981,
-0.09865684807300568,
-0.015139504335820675,
-0.017835525795817375,
-0.08434663712978363,
-0.04441763833165169,
0.034633103758096695,
0.06650148332118988,
0.027389366179704666,
-0.03374916687607765,
-0.018568269908428192,
0.07270564138889313,
-0.008947090245783329,
0.04807329922914505,
0.04124922677874565,
0.05156147480010986,
-0.033632792532444,
-0.053782522678375244,
0.023464754223823547,
-0.07924221456050873,
0.03337991610169411,
0.016122855246067047,
0.014840290881693363,
-0.04095514118671417,
0.10492341220378876,
0.05086646229028702,
0.04185391962528229,
0.08978034555912018,
0.009647962637245655,
0.015839776024222374,
-0.003940490540117025,
-0.05404030904173851,
-0.01937027834355831,
0.018499549478292465,
0.010601981543004513,
0.077491894364357,
0.03572695329785347,
-0.07465176284313202,
-0.055303942412137985,
0.06991705298423767,
0.05081554502248764,
0.10941728204488754,
-0.005541618447750807,
-0.016027169302105904,
-0.03126023709774017,
-0.02117634192109108,
0.032511014491319656,
0.018545564264059067,
0.020184341818094254,
-0.06807754933834076,
0.052151042968034744,
0.0465874969959259,
0.03772597387433052,
-0.01884588412940502,
0.007401847746223211,
0.010113903321325779,
0.031246496364474297,
-0.06382788717746735,
-0.028351860120892525,
0.0619220994412899,
-0.012642347253859043,
0.022161871194839478,
0.03298654034733772,
0.026421992108225822,
-0.017471130937337875,
-0.04623929038643837,
-0.05896013230085373,
1.2336696580799787e-33,
0.08363556116819382,
0.03477717563509941,
-0.020392917096614838,
0.0868910700082779,
0.023346729576587677,
-0.0019605811685323715,
0.048805274069309235,
0.059526123106479645,
-0.034052956849336624,
-0.10425502061843872,
0.0009740981622599065,
-0.02055918611586094,
-0.07011999189853668,
0.06937015056610107,
0.05821862071752548,
0.027759667485952377,
-0.09168106317520142,
0.051691826432943344,
0.016561808064579964,
-0.05865912139415741,
0.034970708191394806,
-0.0001236217503901571,
-0.025108758360147476,
-0.01655280776321888,
0.05578114092350006,
0.04515464976429939,
0.11425255239009857,
-0.016267022117972374,
-0.024363085627555847,
0.002862872090190649,
-0.09905248880386353,
-0.04069104045629501,
-0.012875804677605629,
0.01682129129767418,
0.028785236179828644,
-0.08826186507940292,
-0.038026753813028336,
-0.03552524000406265,
0.007214638404548168,
-0.0312683992087841,
-0.025084195658564568,
0.004352137446403503,
0.04155968502163887,
0.035273585468530655,
-0.11386564373970032,
-0.1467766910791397,
0.02697933092713356,
0.03144766017794609,
-0.01743077114224434,
-0.03489139676094055,
0.046245262026786804,
0.02446364425122738,
-0.03569961339235306,
-0.0804806724190712,
-0.01931488886475563,
-0.017682163044810295,
0.03255285322666168,
0.08203434944152832,
0.03362654522061348,
0.05546271800994873,
0.00916454941034317,
0.07403997331857681,
-0.0011527449823915958,
0.14375418424606323,
0.07210182398557663,
-0.016336198896169662,
0.014134310185909271,
0.03723745420575142,
0.07902155071496964,
-0.0504169762134552,
-0.0410892590880394,
-0.08488477766513824,
0.002440993906930089,
-0.030436309054493904,
0.03302236273884773,
0.019885927438735962,
-0.014036682434380054,
-0.08101162314414978,
0.010533181019127369,
-0.0015925491461530328,
0.03911499306559563,
0.033787306398153305,
-0.028836561366915703,
-0.006932367570698261,
0.01724981889128685,
-0.07840698957443237,
0.05881593003869057,
-0.13021020591259003,
-0.0002174405672121793,
-0.07240980118513107,
0.014252050779759884,
-0.013655001297593117,
-0.03010665625333786,
-0.024254247546195984,
0.018755650147795677,
-3.154475365452591e-34,
0.025301668792963028,
-0.008664187975227833,
-0.04601050540804863,
0.09510501474142075,
0.028104837983846664,
-0.05023844912648201,
0.03163805231451988,
0.09408098459243774,
-0.09057892858982086,
-0.06517024338245392,
-0.11087709665298462,
-0.011875251308083534,
0.015274900943040848,
-0.008768544532358646,
0.02516988106071949,
-0.016139734536409378,
0.005005771294236183,
0.007340948097407818,
-0.01785040833055973,
0.021047238260507584,
-0.04145105928182602,
0.1116257905960083,
-0.07530996948480606,
0.004464141558855772,
-0.07490037381649017,
0.01596568338572979,
-0.028523825109004974,
0.08457964658737183,
0.028048621490597725,
0.05008919537067413,
-0.01929495483636856,
0.02052321657538414,
0.010687295347452164,
-0.03175118565559387,
-0.06357383728027344,
0.08240316808223724,
0.08062298595905304,
-0.025307876989245415,
-0.04083053395152092,
0.11010279506444931,
0.09790439903736115,
-0.009584563784301281,
-0.055338069796562195,
-0.0156323853880167,
-0.06747736781835556,
-0.010581615380942822,
-0.11633605509996414,
-0.06487199664115906,
0.02874058298766613,
0.04988965019583702,
-0.04456465691328049,
0.011715344153344631,
-0.0589151531457901,
-0.05126045271754265,
-0.03926056995987892,
-0.13379760086536407,
-0.0003320797986816615,
-0.08217016607522964,
0.05268920958042145,
-0.06104988232254982,
-0.0402509942650795,
-0.05585972219705582,
0.07845915853977203,
-0.060526203364133835,
0.0031815676484256983,
-0.05028845742344856,
-0.017355598509311676,
-0.017792019993066788,
0.00022465306392405182,
0.003165783127769828,
0.06995165348052979,
-0.014473364688456059,
0.003889797953888774,
-0.012606339529156685,
0.05458144098520279,
-0.024909792467951775,
0.03147182986140251,
-0.0015979629242792726,
-0.010889973491430283,
-0.11376657336950302,
-0.0030897229444235563,
-0.004634552169591188,
0.00017471578030381352,
0.026240121573209763,
0.044878844171762466,
0.04597436264157295,
0.03906436264514923,
0.13784551620483398,
0.014986688271164894,
0.03085954487323761,
0.005653460044413805,
0.04420243203639984,
0.004367283545434475,
0.04757337272167206,
-0.02975606545805931,
-3.972233031390715e-8,
-0.1000640019774437,
0.03160003572702408,
0.007372016087174416,
0.12082125246524811,
-0.0012488000793382525,
-0.030870554968714714,
-0.07914848625659943,
0.0813981294631958,
0.03135901689529419,
-0.015438809990882874,
0.03518298268318176,
0.008929274044930935,
0.01356364693492651,
-0.015163101255893707,
-0.003415700513869524,
0.03714887797832489,
0.027312543243169785,
0.04697960615158081,
-0.04651212319731712,
-0.014594871550798416,
0.06560182571411133,
0.053094398230314255,
0.001968208933249116,
-0.03252514824271202,
0.03686979413032532,
-0.033205412328243256,
0.017921073362231255,
0.08950508385896683,
0.03393468260765076,
-0.026721255853772163,
-0.025866160169243813,
-0.018889106810092926,
-0.06136054918169975,
-0.06616716831922531,
-0.015983087942004204,
0.06245671585202217,
-0.009632919915020466,
-0.035658784210681915,
0.007748722098767757,
-0.038570255041122437,
0.035008978098630905,
0.06587719917297363,
-0.07264482229948044,
0.04085919260978699,
-0.009712062776088715,
-0.057745311409235,
-0.05840824916958809,
-0.14496663212776184,
-0.013767465017735958,
-0.028649870306253433,
0.0066662561148405075,
-0.05158660188317299,
0.01252706814557314,
0.037539172917604446,
0.05417404696345329,
0.02339046448469162,
-0.015812058001756668,
0.03537841886281967,
0.032116975635290146,
0.04468401521444321,
-0.004112196154892445,
0.05166005343198776,
-0.017371950671076775,
0.03183983638882637
] | 0 | 0 | 0 | 0 |
|
ICLR: In-Context Learning of Representations | https://openreview.net/forum?id=pXlmOmlHJZ | [
"Core Francisco Park",
"Andrew Lee",
"Ekdeep Singh Lubana",
"Yongyi Yang",
"Maya Okawa",
"Kento Nishi",
"Martin Wattenberg",
"Hidenori Tanaka"
] | Poster | Recent work demonstrates that structured patterns in pretraining data influence how representations of different concepts are organized in a large language model’s (LLM) internals, with such representations then driving downstream abilities. Given the open-ended nature of LLMs, e.g., their ability to in-context learn novel tasks, we ask whether models can flexibly alter their semantically grounded organization of concepts. Specifically, if we provide in-context exemplars wherein a concept plays a different role than what the pretraining data suggests, can models infer these novel semantics and reorganize representations in accordance with them? To answer this question, we define a toy “graph tracing” task wherein the nodes of the graph are referenced via concepts seen during training (e.g., apple, bird, etc.), and the connectivity of the graph is defined via some predefined structure (e.g., a square grid). Given exemplars that indicate traces of random walks on the graph, we analyze intermediate representations of the model and find that as the amount of context is scaled, there is a sudden re-organization of representations according to the graph’s structure. Further, we find that when reference concepts have correlations in their semantics (e.g., Monday, Tuesday, etc.), the context-specified graph structure is still present in the representations, but is unable to dominate the pretrained structure. To explain these results, we analogize our task to energy minimization for a predefined graph topology, which shows getting non-trivial performance on the task requires for the model to infer a connected component. Overall, our findings indicate context-size may be an underappreciated scaling axis that can flexibly re-organize model representations, unlocking novel capabilities. | In-Context Learning, Representational Geometry, World Models, Emergence, Percolation | Large language models develop emergent task specific representations given enough in-context exemplars. | 13,560 | 2501.00070 | [
0.018754323944449425,
-0.10803944617509842,
0.007552081253379583,
0.05884907394647598,
0.08313634246587753,
-0.048245999962091446,
0.031229140236973763,
-0.0197514109313488,
0.09743563085794449,
0.0009918059222400188,
-0.03152618929743767,
-0.007904277183115482,
0.05376630648970604,
0.08643527328968048,
0.06052834913134575,
-0.022422727197408676,
0.04956189915537834,
0.09871545433998108,
-0.11037404835224152,
-0.032904911786317825,
0.06350868940353394,
-0.036677125841379166,
-0.010274727828800678,
-0.005182118620723486,
-0.005104805808514357,
0.056979455053806305,
-0.061947401612997055,
-0.010226765647530556,
0.05635730177164078,
-0.020469238981604576,
0.08933314681053162,
0.09543953835964203,
-0.010172264650464058,
0.07433101534843445,
-0.018426747992634773,
0.09202103316783905,
-0.005041488911956549,
0.0724542960524559,
0.057013534009456635,
-0.0338919498026371,
-0.0005420989473350346,
0.03520822152495384,
0.01984500326216221,
0.03882613033056259,
0.06826817244291306,
0.040982723236083984,
-0.01036011055111885,
-0.01627858355641365,
-0.11340750753879547,
0.031230906024575233,
-0.13628628849983215,
-0.0601198673248291,
-0.0533839613199234,
0.026022525504231453,
0.013086547143757343,
0.041160136461257935,
0.059584226459264755,
-0.004234889056533575,
-0.040574584156274796,
-0.07178377360105515,
-0.020408043637871742,
-0.11032039672136307,
-0.04852912202477455,
0.01902957074344158,
-0.060559485107660294,
-0.031818874180316925,
0.021953318268060684,
0.07166122645139694,
0.04584934189915657,
0.006634324789047241,
0.06334786117076874,
0.03354964777827263,
-0.014992591924965382,
0.018937302753329277,
0.07395116239786148,
0.031052814796566963,
0.029279261827468872,
0.011740724556148052,
0.025836477056145668,
-0.12300925701856613,
0.04880871623754501,
0.07365811616182327,
0.0800560936331749,
-0.0002272614510729909,
0.09071464836597443,
0.048846788704395294,
0.02389627881348133,
-0.02982032112777233,
0.01096674706786871,
0.014510555192828178,
-0.05717199295759201,
-0.09858249872922897,
-0.009770318865776062,
-0.005925596226006746,
0.05814191326498985,
0.02011132799088955,
0.05593814328312874,
-0.07548633217811584,
0.014305291697382927,
0.10403914004564285,
0.028229285031557083,
0.029731567949056625,
-0.0021420293487608433,
-0.01453488226979971,
-0.000340349564794451,
-0.0461469329893589,
-0.060907866805791855,
0.03272242844104767,
0.011628434993326664,
-0.023994343355298042,
-0.027864260599017143,
0.022234594449400902,
-0.011641845107078552,
0.07106190919876099,
0.022977715358138084,
-0.04059519246220589,
0.041367270052433014,
-0.08085115253925323,
0.029149292036890984,
0.08191660791635513,
-0.042058467864990234,
0.07342705875635147,
0.015779247507452965,
-0.013957403600215912,
-0.0021574038546532393,
-0.054242219775915146,
-0.041224926710128784,
-1.7788621931169286e-33,
0.04807538911700249,
0.022862102836370468,
0.03028450347483158,
0.04499175772070885,
0.10208551585674286,
-0.04311637952923775,
-0.001158252009190619,
0.007619215175509453,
0.0664864033460617,
0.0049094404093921185,
-0.02189706638455391,
0.09000853449106216,
0.0016899474430829287,
0.07923193275928497,
0.04752439633011818,
0.019419483840465546,
-0.09106922149658203,
0.0033774408511817455,
-0.02277383953332901,
-0.08690479397773743,
-0.026531634852290154,
0.001806368469260633,
0.009556401520967484,
-0.05268196016550064,
0.03316674008965492,
0.07605662196874619,
0.019667772576212883,
-0.06574981659650803,
-0.018611837178468704,
0.018197884783148766,
-0.04922838136553764,
0.016833320260047913,
-0.025857845321297646,
0.061630379408597946,
-0.030095035210251808,
-0.03102940320968628,
0.03022424317896366,
-0.05673147737979889,
0.023440353572368622,
-0.04143877699971199,
-0.01349597703665495,
0.017583075910806656,
-0.011032059788703918,
-0.020866934210062027,
-0.08707746118307114,
-0.027810735628008842,
-0.022773077711462975,
-0.01937166415154934,
-0.08214955776929855,
-0.0696367546916008,
0.06067747250199318,
0.021252404898405075,
0.04628371074795723,
-0.060711100697517395,
0.02153143286705017,
0.026584504172205925,
-0.016743438318371773,
0.03374137356877327,
0.04869871959090233,
-0.005899036768823862,
0.025245903059840202,
0.018481001257896423,
0.017175031825900078,
0.11850787699222565,
0.08226737380027771,
0.01800527423620224,
-0.04038532078266144,
0.04308418557047844,
0.050139155238866806,
-0.014559351839125156,
-0.07819874584674835,
0.07151302695274353,
-0.037358202040195465,
-0.0727359727025032,
0.012442043051123619,
-0.025410329923033714,
-0.04219542816281319,
-0.10357525944709778,
-0.10206285864114761,
0.07138989120721817,
-0.034540534019470215,
-0.07164846360683441,
0.02114308625459671,
-0.030159497633576393,
-0.07837550342082977,
-0.07559163868427277,
0.049184370785951614,
-0.14533652365207672,
-0.0014224076876416802,
-0.08583708852529526,
-0.05210544914007187,
-0.04261624068021774,
-0.039307162165641785,
0.019569529220461845,
-0.010037697851657867,
-5.446994243403296e-34,
-0.030360540375113487,
0.01667090691626072,
-0.019446304067969322,
0.009570905938744545,
-0.009486470371484756,
-0.08717626333236694,
0.008021003566682339,
0.021145496517419815,
-0.08556509763002396,
-0.021058019250631332,
-0.07200250774621964,
-0.06275352835655212,
0.024878548458218575,
0.05825189873576164,
0.033982083201408386,
-0.06923030316829681,
-0.01519632525742054,
0.044820189476013184,
0.032187771052122116,
0.041729867458343506,
0.04395192861557007,
0.021284393966197968,
-0.09137970954179764,
0.0035173457581549883,
0.03527737781405449,
0.03512788936495781,
-0.0018865049351006746,
0.042856499552726746,
-0.01926795393228531,
0.0591682605445385,
-0.04599076882004738,
-0.03299455717206001,
0.03746876120567322,
-0.029403051361441612,
-0.09191916137933731,
0.0741150975227356,
-0.004058484919369221,
-0.03460575267672539,
-0.05880427733063698,
0.03103870153427124,
0.05779411271214485,
-0.028190098702907562,
-0.07172458618879318,
-0.015064180828630924,
-0.0007772610988467932,
-0.03465238958597183,
-0.07766631990671158,
0.041879937052726746,
-0.024018146097660065,
-0.04742950573563576,
0.004177313297986984,
0.030885471031069756,
-0.03889744356274605,
-0.047737061977386475,
0.0043496787548065186,
-0.010332256555557251,
0.06052524968981743,
-0.03682464361190796,
-0.02420441433787346,
-0.06631972640752792,
-0.06247444823384285,
-0.05392109602689743,
-0.008025460876524448,
-0.056473810225725174,
0.014748260378837585,
-0.051520735025405884,
-0.05698489025235176,
-0.10825259238481522,
0.019370917230844498,
0.005480451509356499,
-0.011252986267209053,
0.017324235290288925,
-0.10109220445156097,
-0.044935837388038635,
0.01916077360510826,
-0.024063831195235252,
-0.02654608152806759,
-0.026238301768898964,
-0.0386192612349987,
-0.08110913634300232,
-0.04807859659194946,
-0.0771293118596077,
0.053426310420036316,
0.0862097516655922,
0.07932431250810623,
0.031212329864501953,
-0.018980352208018303,
0.1228540912270546,
0.04841157793998718,
0.04714736342430115,
0.006556494161486626,
0.012576872482895851,
-0.03484831750392914,
0.08120975643396378,
-0.09706518054008484,
-4.78531170244878e-8,
-0.13996076583862305,
0.02876145951449871,
0.02200555056333542,
0.0668451189994812,
0.0781974196434021,
0.02029050886631012,
0.017631834372878075,
0.08553966879844666,
-0.04879835247993469,
0.027594827115535736,
0.07471135258674622,
0.020111598074436188,
-0.02586950734257698,
-0.016811789944767952,
0.026850903406739235,
0.07036087661981583,
0.023286830633878708,
-0.046554822474718094,
-0.01623416692018509,
-0.01286264043301344,
0.04718297719955444,
0.011290020309388638,
-0.04768836870789528,
0.05118430405855179,
0.03551222383975983,
-0.09985646605491638,
-0.051089756190776825,
0.055968839675188065,
0.02681608498096466,
0.009579095058143139,
-0.05467427894473076,
0.05907728895545006,
0.023248886689543724,
-0.00414779270067811,
-0.05696603283286095,
0.05493423342704773,
0.07182833552360535,
-0.05335216596722603,
-0.019131463021039963,
-0.02896854467689991,
0.029386386275291443,
0.03743406757712364,
-0.05293036997318268,
0.019324863329529762,
0.04626215621829033,
0.029014350846409798,
-0.017559051513671875,
-0.004542374052107334,
0.03833382576704025,
-0.03510945662856102,
-0.0814654603600502,
0.07774975150823593,
-0.03849051147699356,
0.01013718917965889,
0.04537159204483032,
0.040709491819143295,
-0.05043594539165497,
-0.03277040272951126,
-0.00483127124607563,
0.059911493211984634,
-0.0003552185953594744,
0.03250127285718918,
-0.0030197156593203545,
-0.025311144068837166
] | 0 | 0 | 0 | 0 |
|
Ensembling Diffusion Models via Adaptive Feature Aggregation | https://openreview.net/forum?id=e32cI4r8Eo | [
"Cong Wang",
"Kuan Tian",
"Yonghang Guan",
"Fei Shen",
"Zhiwei Jiang",
"Qing Gu",
"Jun Zhang"
] | Poster | The success of the text-guided diffusion model has inspired the development and release of numerous powerful diffusion models within the open-source community. These models are typically fine-tuned on various expert datasets, showcasing diverse denoising capabilities. Leveraging multiple high-quality models to produce stronger generation ability is valuable, but has not been extensively studied. Existing methods primarily adopt parameter merging strategies to produce a new static model. However, they overlook the fact that the divergent denoising capabilities of the models may dynamically change across different states, such as when experiencing different prompts, initial noises, denoising steps, and spatial locations. In this paper, we propose a novel ensembling method, Adaptive Feature Aggregation (AFA), which dynamically adjusts the contributions of multiple models at the feature level according to various states (i.e., prompts, initial noises, denoising steps, and spatial locations), thereby keeping the advantages of multiple diffusion models, while suppressing their disadvantages. Specifically, we design a lightweight Spatial-Aware Block-Wise (SABW) feature aggregator that adaptive aggregates the block-wise intermediate features from multiple U-Net denoisers into a unified one. The core idea lies in dynamically producing an individual attention map for each model's features by comprehensively considering various states. It is worth noting that only SABW is trainable with about 50 million parameters, while other models are frozen. Both the quantitative and qualitative experiments demonstrate the effectiveness of our proposed method. | Image Generation, Diffusion Models, Model Ensembling | We propose Adaptive Feature Aggregation (AFA) to ensemble multiple diffusion models dynamically based on different states. | 13,559 | 2405.17082 | [
-0.0022625650744885206,
-0.15442867577075958,
0.069721519947052,
0.10095153748989105,
0.057552769780159,
-0.007404733914881945,
-0.1031423956155777,
-0.07334215193986893,
0.000729983439669013,
-0.04994568973779678,
-0.027226928621530533,
0.004402380436658859,
0.04600503295660019,
0.022943956777453423,
-0.029518375173211098,
0.03936542570590973,
0.010505198501050472,
0.05395474284887314,
-0.08720944821834564,
-0.04110762104392052,
-0.054619207978248596,
0.04931653290987015,
-0.07549869269132614,
0.016099227592349052,
0.13552117347717285,
0.06221475079655647,
-0.00716719776391983,
-0.036793310195207596,
0.031071143224835396,
0.012655835598707199,
0.04712684080004692,
-0.0023915076162666082,
-0.061297137290239334,
-0.005346612073481083,
0.04293457791209221,
-0.04947679862380028,
-0.045670285820961,
0.0913621261715889,
0.03809163719415665,
-0.027506481856107712,
0.039528101682662964,
-0.007117729168385267,
-0.01377417054027319,
0.007861103862524033,
0.050814829766750336,
-0.010314378887414932,
-0.06535256654024124,
0.008449390530586243,
0.019695255905389786,
-0.007943044416606426,
0.015730412676930428,
-0.0064839646220207214,
-0.09736113995313644,
0.10398810356855392,
0.04827186092734337,
-0.016894517466425896,
0.030119257047772408,
0.03803621605038643,
0.025188380852341652,
-0.023949528113007545,
0.007143799215555191,
-0.07275473326444626,
-0.022166360169649124,
0.059184785932302475,
0.040072668343782425,
-0.021017208695411682,
0.043941598385572433,
0.1502765715122223,
-0.030463112518191338,
0.0656883716583252,
-0.004481100011616945,
0.06639327853918076,
-0.07427703589200974,
0.012505228631198406,
0.009630603715777397,
0.06792858242988586,
0.010741941630840302,
0.005053068045526743,
0.018735338002443314,
-0.0559484027326107,
0.030442148447036743,
0.04849240183830261,
0.024259604513645172,
-0.09463836997747421,
0.10554414987564087,
-0.04525599628686905,
0.029327211901545525,
-0.013197539374232292,
-0.0544842854142189,
-0.012795546092092991,
-0.06128253415226936,
0.008220857940614223,
-0.006675510201603174,
-0.05301893502473831,
-0.01499455701559782,
0.0705936998128891,
0.010118131525814533,
-0.03412482142448425,
-0.00259898672811687,
0.11417064815759659,
-0.061298515647649765,
-0.034046631306409836,
0.048039257526397705,
-0.08197629451751709,
0.03271421417593956,
-0.09904242306947708,
0.06480361521244049,
0.029373718425631523,
-0.0011878478107973933,
0.023041563108563423,
0.020518308505415916,
0.022951913997530937,
0.030469391494989395,
0.016356710344552994,
0.07254111021757126,
-0.046940870583057404,
0.07986157387495041,
0.02629009447991848,
0.08593869209289551,
0.01498863659799099,
-0.004958854056894779,
-0.06422507017850876,
-0.034025948494672775,
0.015588732436299324,
0.07921741157770157,
0.08471373468637466,
-0.04562983289361,
2.8633620772550398e-33,
0.030247429385781288,
-0.02781224437057972,
0.0027452458161860704,
0.023663686588406563,
0.07092798501253128,
-0.04597613587975502,
-0.0281540434807539,
-0.00044604751747101545,
-0.02236458659172058,
0.011966677382588387,
-0.03336174413561821,
0.09661109745502472,
-0.027449050918221474,
0.06831176578998566,
0.051921550184488297,
-0.03103461302816868,
-0.03358778357505798,
0.07776685804128647,
-0.00979677401483059,
-0.04246105998754501,
-0.007102403789758682,
-0.09190957248210907,
-0.0187733992934227,
-0.034988656640052795,
-0.08826762437820435,
-0.036784324795007706,
0.03433448076248169,
0.0010380423627793789,
-0.10466916114091873,
0.016811050474643707,
-0.11491715162992477,
0.04986901953816414,
-0.018249133601784706,
0.057794634252786636,
-0.007701195310801268,
-0.08603239804506302,
0.0014302709605544806,
-0.08656957000494003,
0.022324858233332634,
-0.011746815405786037,
-0.029024703428149223,
0.0374755784869194,
-0.12691634893417358,
-0.04781513288617134,
-0.046052929013967514,
0.036117009818553925,
-0.054941676557064056,
-0.09137317538261414,
-0.04678758606314659,
-0.03298824280500412,
0.027453437447547913,
-0.060424625873565674,
-0.024582576006650925,
-0.018703363835811615,
-0.03371407091617584,
-0.01031872071325779,
-0.009263710118830204,
-0.060152359306812286,
0.06870445609092712,
-0.0044020297937095165,
-0.014535866677761078,
0.05020878463983536,
-0.049609504640102386,
0.019946590065956116,
0.12536852061748505,
0.004725039470940828,
0.04521476477384567,
0.1496371477842331,
-0.018267543986439705,
-0.04910041764378548,
-0.01836739107966423,
-0.003558609401807189,
-0.03070918098092079,
-0.03120170719921589,
0.039151713252067566,
-0.06631387770175934,
0.021498847752809525,
0.004301657900214195,
0.015283866785466671,
-0.0022770778741687536,
0.032468054443597794,
-0.022636540234088898,
-0.023870516568422318,
-0.07506530731916428,
-0.0012594949221238494,
0.016196873039007187,
-0.018954981118440628,
-0.033903371542692184,
-0.059957098215818405,
-0.0844619870185852,
-0.042798448354005814,
0.015430969186127186,
0.012643306516110897,
-0.008236922323703766,
0.04470612853765488,
-2.1398770794055925e-33,
-0.011920250952243805,
0.015069840475916862,
0.02378247305750847,
-0.010954676195979118,
-0.020532285794615746,
0.01723545789718628,
-0.04761093109846115,
-0.013224859721958637,
0.1047835573554039,
-0.06390809267759323,
-0.060105208307504654,
-0.03934171423316002,
0.053357258439064026,
-0.03527947515249252,
0.04758257046341896,
-0.05601764842867851,
-0.006829326506704092,
-0.0417037196457386,
-0.03461838513612747,
0.03937750309705734,
-0.03212766349315643,
0.018066570162773132,
-0.04276352375745773,
0.06673505157232285,
-0.06042433902621269,
0.026475828140974045,
-0.06790778040885925,
0.07630468904972076,
0.019966376945376396,
-0.04759849235415459,
-0.05056174844503403,
0.0328611396253109,
-0.018190110102295876,
-0.06493351608514786,
-0.07228954881429672,
0.08907930552959442,
-0.002686028368771076,
-0.05471951887011528,
-0.00999235175549984,
0.08578310906887054,
0.056183911859989166,
-0.0010336651466786861,
-0.01883939653635025,
0.03899204730987549,
0.011388122104108334,
-0.028494933620095253,
-0.034342389553785324,
0.04838000610470772,
-0.030122527852654457,
-0.005844568833708763,
0.0030993048567324877,
0.08527620881795883,
-0.12186231464147568,
0.10040470212697983,
-0.05293343588709831,
0.008632042445242405,
0.0040590460412204266,
-0.04252983257174492,
0.033443909138441086,
0.000007032482699287357,
-0.030134396627545357,
0.006528448313474655,
-0.015885908156633377,
-0.0510084442794323,
-0.010581554844975471,
-0.025891903787851334,
0.013883828185498714,
0.049903906881809235,
-0.03160634636878967,
0.06579860299825668,
-0.0241729523986578,
-0.059836842119693756,
-0.03994256630539894,
-0.03396216779947281,
0.01986163854598999,
-0.07922865450382233,
0.028147626668214798,
-0.08750785142183304,
-0.02055562660098076,
-0.044926147907972336,
0.020641833543777466,
0.038126785308122635,
-0.020702801644802094,
0.06846149265766144,
0.06929212808609009,
0.0289275161921978,
-0.028339197859168053,
0.01625845767557621,
0.01713321916759014,
-0.02619689330458641,
-0.012276298366487026,
-0.022872691974043846,
0.007576562464237213,
0.10797794163227081,
-0.02175000123679638,
-4.912032025572444e-8,
-0.0943395271897316,
0.07129628211259842,
-0.017005663365125656,
0.017391538247466087,
0.022702831774950027,
-0.07370717823505402,
0.008200307376682758,
0.11599257588386536,
0.030427729710936546,
-0.059565406292676926,
0.060936540365219116,
-0.039680007845163345,
-0.04475679621100426,
0.012586246244609356,
0.029110828414559364,
0.08598736673593521,
0.007523634936660528,
0.044977735728025436,
-0.03468731418251991,
-0.08060557395219803,
0.08923929929733276,
0.030661724507808685,
0.06694786995649338,
-0.003177311271429062,
0.07929351180791855,
-0.015762319788336754,
-0.03219705820083618,
0.05793575197458267,
0.0895867794752121,
-0.0033327455166727304,
0.048611681908369064,
0.030797308310866356,
-0.02323276363313198,
0.01945464126765728,
0.04640847072005272,
0.13586926460266113,
0.09429654479026794,
-0.004393925424665213,
-0.06735092401504517,
-0.017297111451625824,
-0.0009306010906584561,
0.04198375716805458,
-0.0572488009929657,
0.02589571662247181,
-0.02699904888868332,
0.02808469347655773,
0.06236385926604271,
-0.09410663694143295,
0.05552174523472786,
0.0027556021232157946,
0.002978844800963998,
-0.03306833654642105,
0.020114367827773094,
0.04394935443997383,
0.043958887457847595,
0.044790610671043396,
0.002587889786809683,
0.0036368120927363634,
0.11388421058654785,
0.03511985391378403,
0.019779128953814507,
-0.009509709663689137,
-0.0072363377548754215,
0.061343155801296234
] | https://github.com/tenvence/afa | 5 | 0 | 0 | 0 |
Solving New Tasks by Adapting Internet Video Knowledge | https://openreview.net/forum?id=p01BR4njlY | [
"Calvin Luo",
"Zilai Zeng",
"Yilun Du",
"Chen Sun"
] | Poster | Video generative models demonstrate great promise in robotics by serving as visual planners or as policy supervisors. When pretrained on internet-scale data, such video models intimately understand alignment with natural language, and can thus facilitate generalization to novel downstream behavior through text-conditioning. However, they may not be sensitive to the specificities of the particular environment the agent inhabits. On the other hand, training video models on in-domain examples of robotic behavior naturally encodes environment-specific intricacies, but the scale of available demonstrations may not be sufficient to support generalization to unseen tasks via natural language specification. In this work, we investigate different adaptation techniques that integrate in-domain information with large-scale pretrained video models, and explore the extent to which they enable novel text-conditioned generalization for robotic tasks, while also considering their independent data and resource considerations. We successfully demonstrate across robotic environments that adapting powerful video models with small scales of example data can successfully facilitate generalization to novel behaviors. In particular, we present a novel adaptation strategy, termed *Inverse Probabilistic Adaptation*, that not only consistently achieves strong generalization performance across robotic tasks and settings, but also exhibits robustness to the quality of adaptation data, successfully solving novel tasks even when only suboptimal in-domain demonstrations are available. | Text-Conditioned Generalization, Video Diffusion, Adaptation, Planning, Policy Learning | We compare techniques for adapting large-scale video generative models to in-domain robotic data, and demonstrate that it facilitates text-conditioned generalization to novel tasks. | 13,550 | null | [
-0.0017583360895514488,
-0.14566107094287872,
0.06462559103965759,
-0.003760233521461487,
0.11513674259185791,
-0.0010549152502790093,
-0.01592065952718258,
-0.0378309041261673,
0.004995429888367653,
-0.033636365085840225,
-0.01917988620698452,
-0.06902475655078888,
0.0337831974029541,
0.09424847364425659,
0.07347377389669418,
-0.015919705852866173,
0.15537656843662262,
0.05085448920726776,
-0.12913091480731964,
-0.013552672229707241,
0.02045460417866707,
0.052247364073991776,
0.10030547529459,
0.025100290775299072,
0.019002312794327736,
0.03621194511651993,
0.08942658454179764,
-0.024738997220993042,
-0.0009364503785036504,
-0.037381354719400406,
0.01169521827250719,
-0.041376370936632156,
-0.020499426871538162,
-0.01473350077867508,
-0.03496962785720825,
0.09659641236066818,
-0.12610727548599243,
0.04786494001746178,
-0.049907904118299484,
-0.06940759718418121,
0.031288065016269684,
-0.005018855910748243,
0.09377457201480865,
-0.008145065978169441,
0.022547367960214615,
0.07561732083559036,
-0.00022323887969832867,
-0.09244295954704285,
0.01953776367008686,
-0.016486667096614838,
-0.049542684108018875,
-0.04498875141143799,
0.002307109534740448,
0.010278003290295601,
-0.05301080271601677,
0.054685961455106735,
0.057185858488082886,
-0.02887975610792637,
0.07191552966833115,
-0.011631111614406109,
0.03901129215955734,
-0.11960244178771973,
0.01072729378938675,
-0.005955749657005072,
-0.021567363291978836,
-0.04187900200486183,
0.01736658252775669,
0.03846072405576706,
-0.034836962819099426,
0.030769728124141693,
-0.03580827638506889,
0.045965492725372314,
0.054054781794548035,
0.08820934593677521,
0.03118857555091381,
-0.006146008148789406,
0.017348503693938255,
-0.05898801609873772,
-0.006296859122812748,
-0.05497540533542633,
0.022451814264059067,
-0.0457669273018837,
0.0634893849492073,
0.10207826644182205,
0.12612856924533844,
0.018411362543702126,
0.025275904685258865,
0.03917019069194794,
0.0009871673537418246,
-0.0065054274164140224,
-0.06487476080656052,
-0.09154346585273743,
0.026871951296925545,
-0.03571729362010956,
-0.022259438410401344,
0.0068333507515490055,
-0.043628908693790436,
-0.1775485873222351,
-0.019645411521196365,
0.04260031506419182,
0.023102298378944397,
0.04208214953541756,
0.018469998612999916,
0.013542456552386284,
-0.020661093294620514,
-0.04545602947473526,
0.05278734117746353,
0.05684433877468109,
0.02547641098499298,
-0.046969812363386154,
0.06275246292352676,
-0.03371497616171837,
0.06235276162624359,
0.010431446135044098,
0.01358911395072937,
-0.03159096837043762,
0.006830956321209669,
-0.004827270284295082,
-0.09086614847183228,
0.07956120371818542,
0.028151068836450577,
-0.019977182149887085,
0.02568715065717697,
-0.021019509062170982,
0.06478945165872574,
-0.0052544488571584225,
0.04720311239361763,
-6.750706558896486e-34,
0.0908089131116867,
0.013809514231979847,
0.0848059356212616,
0.0853343978524208,
0.09167292714118958,
0.007806810550391674,
0.005371020641177893,
-0.0024128886871039867,
-0.037230756133794785,
-0.004769227467477322,
0.017385531216859818,
0.0063523282296955585,
-0.058420680463314056,
0.09228669106960297,
-0.025656284764409065,
0.0697319284081459,
-0.06679222732782364,
0.06506479531526566,
0.01713385246694088,
0.041366055607795715,
0.025960423052310944,
0.0535392165184021,
-0.07391983270645142,
-0.0007235775119625032,
0.00773746008053422,
-0.05584832280874252,
0.025389645248651505,
-0.0516120120882988,
0.008936554193496704,
-0.04177188500761986,
-0.11633658409118652,
0.032871413975954056,
0.02052665501832962,
0.051842350512742996,
0.09040714055299759,
-0.014127692207694054,
-0.0374123677611351,
-0.07415701448917389,
-0.013476348482072353,
-0.0047026388347148895,
0.0265729408711195,
0.03881419450044632,
-0.000015223345144477207,
-0.05468342453241348,
-0.09562762081623077,
-0.0940207988023758,
0.0446370430290699,
0.043741725385189056,
-0.07712802290916443,
-0.06712193787097931,
0.09449426084756851,
0.06888236105442047,
0.0039052122738212347,
-0.07025895267724991,
-0.0322117954492569,
-0.036956507712602615,
0.04441516101360321,
0.02959390915930271,
0.031554777175188065,
-0.0394306518137455,
0.04251205921173096,
0.0039531756192445755,
-0.03498097509145737,
0.07497145235538483,
0.021695038303732872,
0.011157926172018051,
0.04379521682858467,
0.03978327289223671,
0.061419613659381866,
0.026093274354934692,
-0.04853739216923714,
0.03473992645740509,
-0.03613054379820824,
-0.04431438073515892,
0.049645766615867615,
0.040397487580776215,
-0.022332027554512024,
-0.10023466497659683,
0.013403850607573986,
0.05758674442768097,
-0.019335223361849785,
-0.06303625553846359,
-0.02089536190032959,
-0.0023367132525891066,
-0.02353348582983017,
-0.05633753165602684,
0.036618612706661224,
-0.09656930714845657,
0.0008973752846941352,
-0.03478112816810608,
-0.020479653030633926,
-0.0992245227098465,
-0.09541982412338257,
0.0326315313577652,
0.06025606393814087,
-5.925251007697556e-34,
-0.021486053243279457,
-0.023425016552209854,
0.02148936130106449,
-0.00006518868758575991,
-0.010793270543217659,
-0.009927586652338505,
0.012160204350948334,
0.02112290821969509,
0.014943942427635193,
-0.11237728595733643,
-0.07763303071260452,
-0.03046809695661068,
-0.018236426636576653,
-0.005188460927456617,
-0.025768069550395012,
-0.029111163690686226,
-0.04549003019928932,
0.03059622459113598,
-0.006518552079796791,
-0.03825036808848381,
0.02357667312026024,
0.08408959209918976,
-0.033872220665216446,
0.009553118608891964,
-0.005046948790550232,
0.053952962160110474,
-0.002018782775849104,
0.09407418966293335,
0.020321160554885864,
0.019607041031122208,
-0.058044034987688065,
-0.018381137400865555,
-0.00984980445355177,
0.017581751570105553,
-0.1073884665966034,
0.07660325616598129,
0.05007954686880112,
-0.020257970318198204,
-0.09196583926677704,
0.12461353838443756,
0.05276954546570778,
-0.06426346302032471,
-0.04881894588470459,
-0.008482178673148155,
-0.009601675905287266,
-0.0630398765206337,
-0.0537722185254097,
-0.02766580507159233,
-0.016093866899609566,
0.020708268508315086,
-0.027250703424215317,
0.07232430577278137,
-0.08101579546928406,
-0.03064139373600483,
-0.024245575070381165,
-0.09260627627372742,
-0.042484384030103683,
0.00517403706908226,
0.06271979212760925,
-0.0029643585439771414,
-0.0754804015159607,
-0.005661215167492628,
0.001998731167986989,
0.036158107221126556,
-0.04761064797639847,
-0.0005860383389517665,
-0.04348962754011154,
0.0401141420006752,
0.037741366773843765,
-0.032740116119384766,
0.020045718178153038,
-0.063109390437603,
-0.0006194072775542736,
-0.02650897018611431,
0.02822168730199337,
-0.04250755161046982,
0.06350422650575638,
-0.005615985952317715,
-0.005674207117408514,
-0.03442731872200966,
0.01761847920715809,
-0.0004566629359032959,
0.03532038629055023,
0.06420687586069107,
0.05889088287949562,
0.1583881378173828,
0.03943434730172157,
0.052215248346328735,
-0.0009340318501926959,
0.007135680411010981,
-0.05059676244854927,
-0.009484962560236454,
-0.0351262241601944,
0.03979691118001938,
-0.08537881076335907,
-3.371514267769271e-8,
-0.12745743989944458,
0.05068109184503555,
0.000027571299142437056,
0.04806579276919365,
-0.030885063111782074,
-0.010578680783510208,
-0.02556312084197998,
0.001861322089098394,
0.0011212044628337026,
-0.013566221110522747,
0.052232176065444946,
0.01634877547621727,
0.05973071604967117,
-0.04245685040950775,
-0.019186383113265038,
0.057769108563661575,
-0.0001941676455317065,
0.01619817689061165,
0.003499195910990238,
0.02836993709206581,
0.03291231021285057,
-0.013274573720991611,
0.022403964772820473,
-0.04376008361577988,
-0.00527593307197094,
-0.030581830069422722,
-0.03317762166261673,
0.07287292182445526,
0.041570995002985,
0.023847848176956177,
-0.004903771448880434,
-0.03587692975997925,
0.0029571272898465395,
-0.044491760432720184,
0.00809005182236433,
-0.005574177950620651,
0.01034393347799778,
-0.05657399445772171,
0.02378581278026104,
-0.05315984785556793,
0.001983360853046179,
0.07469704747200012,
-0.013594838790595531,
0.020398542284965515,
-0.010976859368383884,
0.05991469696164131,
-0.014802255667746067,
-0.106024831533432,
0.014877662993967533,
-0.03522707521915436,
-0.0438389889895916,
-0.02372637763619423,
0.001140225795097649,
0.08876834809780121,
0.11138710379600525,
0.03305695950984955,
0.053409941494464874,
-0.03368031978607178,
0.0023534568026661873,
0.08332432061433792,
-0.02613730914890766,
0.00611314969137311,
-0.033049844205379486,
-0.032939717173576355
] | 0 | 0 | 0 | 0 |
|
SoftMatcha: A Soft and Fast Pattern Matcher for Billion-Scale Corpus Searches | https://openreview.net/forum?id=Q6PAnqYVpo | [
"Hiroyuki Deguchi",
"Go Kamoda",
"Yusuke Matsushita",
"Chihiro Taguchi",
"Kohei Suenaga",
"Masaki Waga",
"Sho Yokoi"
] | Poster | Researchers and practitioners in natural language processing and computational linguistics frequently observe and analyze the real language usage in large-scale corpora.
For that purpose, they often employ off-the-shelf pattern-matching tools, such as grep, and keyword-in-context concordancers, which is widely used in corpus linguistics for gathering examples.
Nonetheless, these existing techniques rely on surface-level string matching, and thus they suffer from the major limitation of not being able to handle orthographic variations and paraphrasing---notable and common phenomena in any natural language.
In addition, existing continuous approaches such as dense vector search tend to be overly coarse, often retrieving texts that are unrelated but share similar topics.
Given these challenges, we propose a novel algorithm that achieves soft (or semantic) yet efficient pattern matching by relaxing a surface-level matching with word embeddings.
Our algorithm is highly scalable with respect to the size of the corpus text utilizing inverted indexes.
We have prepared an efficient implementation, and we provide an accessible web tool.
Our experiments demonstrate that the proposed method
(i) can execute searches on billion-scale corpora in less than a second, which is comparable in speed to surface-level string matching and dense vector search;
(ii) can extract harmful instances that semantically match queries from a large set of English and Japanese Wikipedia articles;
and (iii) can be effectively applied to corpus-linguistic analyses of Latin, a language with highly diverse inflections. | natural language processing, full-text search, word embeddings, inverted index, pattern match | Achieving soft yet efficient full-text search, we introduce a novel method that extends inverted indexes with word embeddings to enable sub-second searches of billion-scale corpora without GPUs. | 13,545 | 2503.03703 | [
-0.050917427986860275,
-0.060618530958890915,
0.056628547608852386,
-0.023328382521867752,
0.10042053461074829,
-0.0013572301249951124,
0.012212351895868778,
0.03761141747236252,
-0.012630777433514595,
-0.04816611483693123,
0.00902603380382061,
-0.04820046201348305,
0.02415563352406025,
0.04661765694618225,
0.009511254727840424,
0.07559630274772644,
0.06659337878227234,
0.05786623805761337,
-0.014651025645434856,
0.050878290086984634,
0.05277225747704506,
0.048012759536504745,
0.02277524583041668,
-0.07471038401126862,
0.001752881915308535,
0.05066331848502159,
0.007879138924181461,
-0.09481682628393173,
0.034625642001628876,
0.010662026703357697,
0.03512527048587799,
0.07740306109189987,
-0.002647627145051956,
0.10661442577838898,
-0.019468087702989578,
0.042992476373910904,
-0.05175165832042694,
0.08997730910778046,
-0.0012146126246079803,
0.0058793481439352036,
-0.04248075187206268,
-0.05168699845671654,
-0.02194312959909439,
0.058319177478551865,
0.11169657856225967,
0.05740281194448471,
-0.07590179145336151,
0.05116019397974014,
0.03525694087147713,
-0.005256469827145338,
-0.1039748266339302,
-0.014238138683140278,
0.0033932009246200323,
0.06012330949306488,
-0.01128618884831667,
-0.021396586671471596,
-0.004194643814116716,
0.006865666247904301,
0.0631680116057396,
-0.05864401161670685,
0.0007812460535205901,
-0.08514797687530518,
0.026944972574710846,
-0.059120312333106995,
-0.042439382523298264,
-0.07576337456703186,
0.029312653467059135,
0.02312176488339901,
0.012054412625730038,
0.0661299079656601,
-0.03439641371369362,
0.06852707266807556,
-0.07340437173843384,
0.0701521784067154,
-0.08113402873277664,
0.12099506705999374,
-0.028757402673363686,
-0.09916984289884567,
0.02597210183739662,
-0.06946505606174469,
-0.010174289345741272,
-0.013759960420429707,
0.05020477622747421,
-0.00012640914064832032,
0.05832216516137123,
-0.04047932103276253,
0.03802203759551048,
-0.03674723580479622,
0.023888694122433662,
0.03696700185537338,
-0.05316072702407837,
-0.06658458709716797,
0.0787847563624382,
-0.06864834576845169,
0.03328869864344597,
0.022480018436908722,
0.013079953379929066,
-0.002178006339818239,
0.05291832983493805,
0.046675991266965866,
0.01981130801141262,
0.01588893122971058,
0.010296732187271118,
-0.08295474946498871,
-0.12778159976005554,
0.012813765555620193,
-0.05755225941538811,
0.01704629696905613,
0.07401833683252335,
-0.07209064066410065,
-0.04546261578798294,
0.027653727680444717,
0.005971493665128946,
0.026064226403832436,
0.0669960305094719,
-0.0330238938331604,
0.06025109440088272,
-0.05429947003722191,
0.032437510788440704,
0.07493586838245392,
-0.07445918023586273,
0.07008875906467438,
-0.044790271669626236,
-0.03201435133814812,
-0.05825906991958618,
0.06995546072721481,
0.0016641458496451378,
-1.2194761148067765e-33,
0.05311470851302147,
0.05128249153494835,
-0.09067871421575546,
0.00505225220695138,
-0.016510114073753357,
-0.023789118975400925,
-0.02787509560585022,
0.10434319823980331,
-0.07609744369983673,
0.012714304961264133,
-0.028842132538557053,
0.10023695230484009,
0.016864627599716187,
-0.003516906639561057,
-0.021859390661120415,
-0.03325566276907921,
0.01746455952525139,
-0.010612464509904385,
-0.05880489572882652,
-0.001438324456103146,
-0.02028919942677021,
-0.006475129164755344,
0.016039608046412468,
0.02101750485599041,
-0.0035757068544626236,
-0.06261919438838959,
0.08935055136680603,
-0.06245657056570053,
0.030788002535700798,
-0.019756831228733063,
-0.03995866701006889,
-0.028468606993556023,
0.05659692734479904,
0.022951627150177956,
0.04994314908981323,
-0.03908224403858185,
0.021875113248825073,
-0.04952429234981537,
-0.05509277060627937,
-0.05205891653895378,
-0.14560216665267944,
-0.025289146229624748,
0.02995438501238823,
0.015188375487923622,
-0.04461471363902092,
-0.04194745421409607,
-0.05542255565524101,
0.038909345865249634,
0.06109258905053139,
-0.02427060902118683,
0.129503533244133,
0.02577866055071354,
-0.04113331437110901,
0.05024385079741478,
0.07371813803911209,
0.022047709673643112,
-0.008166131563484669,
0.013590840622782707,
0.017724255099892616,
0.029116306453943253,
0.032906994223594666,
0.021491598337888718,
0.041410788893699646,
0.04683484882116318,
-0.010115664452314377,
-0.009142466820776463,
0.023459915071725845,
0.06208926439285278,
0.016602544113993645,
0.058408334851264954,
0.024950241670012474,
0.05836204066872597,
0.02686896175146103,
-0.08260256797075272,
0.06007321923971176,
0.04767382889986038,
0.029333598911762238,
-0.08006563782691956,
-0.056968871504068375,
0.009250739589333534,
-0.03638218343257904,
-0.03489391505718231,
0.07382164895534515,
-0.07287997752428055,
-0.042439524084329605,
-0.04956751689314842,
0.012413881719112396,
-0.05530116707086563,
0.011207537725567818,
-0.022061876952648163,
-0.028203148394823074,
-0.04784407839179039,
-0.0670543983578682,
-0.05361197143793106,
0.032336290925741196,
1.849628606075361e-33,
-0.02011967822909355,
-0.11046824604272842,
0.013862395659089088,
0.0017942284466698766,
0.01977139711380005,
-0.05091170221567154,
0.002043065382167697,
-0.06946714967489243,
-0.04391169920563698,
-0.04584198817610741,
-0.0642685517668724,
-0.04434992000460625,
0.07087821513414383,
-0.013824891299009323,
-0.0045067849569022655,
0.03976375237107277,
-0.004174822010099888,
0.07586725056171417,
-0.06404133141040802,
0.10922380536794662,
0.011269216425716877,
0.06510353833436966,
-0.07308101654052734,
0.10204553604125977,
-0.022428270429372787,
0.0331975594162941,
0.0029752159025520086,
-0.0023256801068782806,
-0.05736641213297844,
-0.08612961322069168,
-0.05655353516340256,
0.01602647639811039,
-0.042414404451847076,
-0.05148567259311676,
-0.061763666570186615,
0.04931216314435005,
0.03433264046907425,
0.039428289979696274,
-0.05624721944332123,
-0.005194497294723988,
0.03930540010333061,
0.053801361471414566,
-0.08672486245632172,
-0.014210999011993408,
-0.04880810156464577,
-0.031948525458574295,
-0.1269424706697464,
0.006473521236330271,
0.020042264834046364,
0.010936073027551174,
-0.004236593842506409,
-0.033631667494773865,
-0.02346956357359886,
0.024299200624227524,
0.03781731799244881,
-0.07584784179925919,
-0.13278138637542725,
0.002538427012041211,
-0.10396318137645721,
-0.0011193878017365932,
-0.010730556212365627,
0.02186266891658306,
-0.033134084194898605,
0.017272645607590675,
0.07849492877721786,
-0.052530668675899506,
-0.02155260182917118,
0.0034125498495996,
-0.023808419704437256,
0.011819046922028065,
0.06642641872167587,
-0.07509542256593704,
-0.008182462304830551,
0.04894838482141495,
-0.005012275651097298,
0.034000031650066376,
0.09113076329231262,
-0.07039918750524521,
-0.009057129733264446,
0.0021963033359497786,
0.06062036752700806,
-0.023617517203092575,
0.06045353040099144,
0.0428888276219368,
0.03387754037976265,
0.05116979032754898,
-0.04083690792322159,
0.08608659356832504,
0.05561412125825882,
0.0841895118355751,
-0.0060277655720710754,
-0.04994915798306465,
-0.017359070479869843,
0.04328826069831848,
0.03527044132351875,
-3.7284699772044405e-8,
-0.004227106459438801,
0.06107938662171364,
-0.06825805455446243,
-0.005409259349107742,
0.012201724573969841,
-0.0015653616283088923,
-0.0037314221262931824,
0.035798121243715286,
-0.08520225435495377,
-0.019403768703341484,
-0.049938250333070755,
-0.021175125613808632,
-0.024509264156222343,
-0.008177630603313446,
0.05978037789463997,
0.07364529371261597,
-0.026885811239480972,
-0.06722038984298706,
0.008908501826226711,
0.036139100790023804,
-0.030570946633815765,
0.10829845815896988,
-0.011123756878077984,
0.10110607743263245,
0.028529314324259758,
0.018037879839539528,
0.005700418725609779,
0.07462123036384583,
0.04158319905400276,
-0.050054460763931274,
-0.008339906111359596,
0.03322896361351013,
-0.0830029845237732,
-0.08628077059984207,
0.05158475786447525,
0.07283727824687958,
0.012797579169273376,
0.01445530541241169,
-0.022580614313483238,
0.03996354714035988,
0.1033855676651001,
0.03620057553052902,
-0.03273981809616089,
0.024594295769929886,
0.07661055028438568,
0.002014637691900134,
-0.03840963542461395,
-0.04586804658174515,
0.0724300742149353,
-0.017214424908161163,
-0.01192386168986559,
-0.07382187247276306,
-0.033912044018507004,
0.021983711048960686,
-0.035016510635614395,
0.02616874687373638,
-0.0843176394701004,
0.04676899313926697,
0.08228281885385513,
-0.0037374889943748713,
0.005248363595455885,
-0.01704869605600834,
0.07495476305484772,
0.04272971302270889
] | 0 | 0 | 0 | 0 |
|
Model merging with SVD to tie the Knots | https://openreview.net/forum?id=67X93aZHII | [
"George Stoica",
"Pratik Ramesh",
"Boglarka Ecsedi",
"Leshem Choshen",
"Judy Hoffman"
] | Poster | Recent model merging methods demonstrate that the parameters of fully-finetuned models specializing in distinct tasks can be combined into one model capable of solving all tasks without retraining. Yet, this success does not transfer well when merging LoRA finetuned models. We study this phenomenon and observe that the weights of LoRA finetuned models showcase a lower degree of alignment compared to their fully-finetuned counterparts. We hypothesize that improving this alignment is key to obtaining better LoRA model merges, and propose KnOTS to address this problem. KnOTS uses the SVD to jointly transform the weights of different LoRA models into an aligned space, where existing merging methods can be applied. In addition, we introduce a new benchmark that explicitly evaluates whether merged models are general models. Notably, KnOTS consistently improves LoRA merging by up to 4.3% across several vision and language benchmarks, including our new setting. We release our code at: https://github.com/gstoica27/KnOTS. | model merging; lora PEFT; computer vision; | We align LoRA PEFT models to improve the application of existing merging methods. | 13,539 | 2410.19735 | [
-0.08733753859996796,
-0.08792471140623093,
0.03268388658761978,
-0.0248191449791193,
0.03528715670108795,
-0.07477720826864243,
-0.04063256457448006,
-0.03960993513464928,
0.032048314809799194,
-0.050252970308065414,
0.006478842347860336,
0.01514237280935049,
-0.022893555462360382,
0.029056377708911896,
-0.03505229577422142,
0.005045830272138119,
0.0015379434917122126,
0.08121655881404877,
-0.1475951075553894,
0.024788351729512215,
-0.03586015850305557,
0.049324844032526016,
0.027118977159261703,
0.008211823180317879,
0.06305240839719772,
0.02434738725423813,
0.017422012984752655,
-0.04746781289577484,
-0.003757609985768795,
-0.08212976157665253,
0.012539496645331383,
0.022530218586325645,
-0.05391550436615944,
0.052583325654268265,
-0.02277088537812233,
0.04630059748888016,
0.024403853341937065,
0.0012122651096433401,
-0.013032681308686733,
-0.0031109319534152746,
-0.007494584657251835,
0.019582530483603477,
0.016877304762601852,
-0.018582429736852646,
0.07653522491455078,
0.028149204328656197,
-0.059939928352832794,
-0.07874448597431183,
-0.01966813951730728,
0.04457475244998932,
-0.08645709604024887,
-0.05405288562178612,
-0.0379265695810318,
0.030186235904693604,
-0.06177361309528351,
-0.008855256251990795,
0.041853323578834534,
-0.045074213296175,
0.00814581848680973,
-0.0185411274433136,
0.02361513115465641,
-0.026132993400096893,
-0.04306943342089653,
0.036695029586553574,
0.04067268595099449,
-0.1321050077676773,
0.03994211554527283,
0.09383156895637512,
-0.06772679835557938,
0.01683417521417141,
-0.027082722634077072,
0.07787753641605377,
-0.050198186188936234,
0.05238467827439308,
0.028553932905197144,
0.04531797766685486,
0.09318192303180695,
0.046155259013175964,
0.04908491298556328,
-0.058523017913103104,
-0.02687937021255493,
0.0032478899229317904,
0.0018382353009656072,
0.019207092002034187,
0.0637846514582634,
0.016635240986943245,
-0.0395938903093338,
0.04022253304719925,
0.029742231592535973,
-0.058122552931308746,
-0.017193777486681938,
0.001393655315041542,
0.054269563406705856,
-0.03059137612581253,
0.03981484845280647,
0.10396384447813034,
0.08450231701135635,
-0.01041445042937994,
-0.028677422553300858,
0.0998741164803505,
-0.01223048847168684,
-0.009501000866293907,
0.03980964049696922,
-0.09422619640827179,
0.0131450230255723,
-0.023631855845451355,
0.0034047397784888744,
-0.0022170040756464005,
-0.0029174196533858776,
-0.13366495072841644,
0.058325789868831635,
-0.031877655535936356,
0.01247048657387495,
0.03592332825064659,
-0.003222103463485837,
-0.009109266102313995,
-0.014524636790156364,
0.00463046645745635,
0.034754276275634766,
-0.02872418612241745,
-0.02679586596786976,
-0.02276577614247799,
0.01950218714773655,
0.05969776585698128,
-0.020264100283384323,
0.013073875568807125,
-0.09342125058174133,
4.302484178794257e-33,
0.0960334837436676,
0.020531876012682915,
0.010872681625187397,
-0.007232666481286287,
0.023438800126314163,
0.004178524017333984,
-0.044897254556417465,
0.021620213985443115,
-0.07920099794864655,
0.037025678902864456,
-0.10985061526298523,
0.0035954462364315987,
-0.04611653462052345,
-0.006525459233671427,
-0.016650114208459854,
-0.018101884052157402,
-0.03161600977182388,
0.060390815138816833,
-0.12045063823461533,
0.022434215992689133,
0.08016527444124222,
0.022766435518860817,
0.045301079750061035,
-0.0005422718822956085,
0.05798155814409256,
0.04288258031010628,
0.07070822268724442,
-0.03789670020341873,
-0.06000010669231415,
0.009939980693161488,
-0.030248718336224556,
-0.054534442722797394,
0.06740202009677887,
0.10086788982152939,
0.012389590963721275,
0.018173104152083397,
0.03648877143859863,
-0.0465262196958065,
-0.04070131108164787,
-0.04841563105583191,
0.0016547228442505002,
0.0312989167869091,
-0.08019509166479111,
-0.0422113873064518,
-0.008172372356057167,
-0.034830886870622635,
0.016703931614756584,
0.025187047198414803,
-0.003686793614178896,
0.016481932252645493,
0.013266115449368954,
0.02032153680920601,
-0.013975332491099834,
-0.04093742370605469,
-0.12219399958848953,
0.01329940464347601,
-0.04429549723863602,
-0.007198028732091188,
-0.008774596266448498,
0.05748473480343819,
0.00882596056908369,
0.042039431631565094,
-0.03207363188266754,
0.010437850840389729,
0.06024366617202759,
-0.02266082353889942,
0.09293430298566818,
0.10136979073286057,
0.08967822045087814,
0.0019099719356745481,
-0.047494906932115555,
-0.03695224970579147,
-0.017808012664318085,
-0.020639147609472275,
0.07565241307020187,
-0.08231613785028458,
-0.006379263009876013,
-0.10261591523885727,
0.042346734553575516,
-0.0025929417461156845,
0.021239830181002617,
0.07689780741930008,
0.053249381482601166,
-0.06447052955627441,
-0.006064003799110651,
-0.07374325394630432,
-0.008670425973832607,
-0.02633616514503956,
-0.09200693666934967,
0.02194027602672577,
-0.006875202525407076,
0.010262664407491684,
0.06605268269777298,
-0.031805213540792465,
0.053459107875823975,
-3.0764769990072127e-33,
0.026098351925611496,
-0.018510542809963226,
-0.0016018797177821398,
0.09706667810678482,
0.007203724235296249,
-0.08415684849023819,
-0.019885599613189697,
-0.03700543940067291,
-0.07050525397062302,
-0.03765811771154404,
0.04917771741747856,
-0.07326612621545792,
-0.02340974658727646,
-0.011659931391477585,
0.06867776066064835,
-0.05320315062999725,
-0.034252967685461044,
-0.052733518183231354,
-0.02057792991399765,
0.030042331665754318,
0.04754067212343216,
0.04233579337596893,
0.01190022099763155,
0.01821272075176239,
-0.009765408001840115,
0.015923067927360535,
-0.03402116149663925,
-0.014007090590894222,
-0.0070985970087349415,
-0.013944079168140888,
-0.01804262399673462,
-0.06154102459549904,
-0.07475051283836365,
-0.04765941575169563,
-0.058226607739925385,
-0.0028150754515081644,
0.029842820018529892,
-0.005685599520802498,
0.026311416178941727,
0.044991154223680496,
-0.006805293262004852,
-0.022654712200164795,
-0.08079777657985687,
0.04721158370375633,
-0.0008229087688960135,
-0.0010259250411763787,
-0.10788742452859879,
0.023002926260232925,
-0.05125122517347336,
-0.05251552537083626,
-0.011379405856132507,
0.006347305607050657,
-0.03576525300741196,
-0.03439167141914368,
0.029067181050777435,
-0.060950867831707,
-0.03437818959355354,
-0.023048684000968933,
-0.08693597465753555,
-0.0066810715943574905,
-0.03290281817317009,
-0.024083063006401062,
0.003258928656578064,
0.027957670390605927,
0.09832491725683212,
0.010084104724228382,
0.020358780398964882,
-0.06829556077718735,
-0.06719722598791122,
0.01169421523809433,
-0.024280523881316185,
-0.018798474222421646,
-0.059265222400426865,
0.03355306386947632,
0.01411817129701376,
-0.058513443917036057,
0.04357798025012016,
-0.03703337162733078,
0.11719229817390442,
-0.03913334384560585,
-0.0718209445476532,
-0.06142347306013107,
0.05603181570768356,
0.07290126383304596,
0.01581350527703762,
0.06694382429122925,
0.01906551793217659,
0.10577410459518433,
0.041138771921396255,
0.04196593165397644,
-0.05627886578440666,
0.04905648157000542,
0.07927542924880981,
0.06160736829042435,
-0.06628882884979248,
-4.066374259537042e-8,
-0.07552653551101685,
0.0665460079908371,
-0.08938069641590118,
0.050822913646698,
-0.013821299187839031,
0.005912199150770903,
-0.013060220517218113,
0.1213955506682396,
0.003392014652490616,
0.012904676608741283,
0.02024935372173786,
0.03732734173536301,
-0.08299961686134338,
0.024022094905376434,
0.048928502947092056,
0.00041627386235632,
-0.006605379283428192,
0.08226414024829865,
-0.02550702914595604,
0.014071379788219929,
0.08625908195972443,
0.09202708303928375,
-0.021221186965703964,
0.08951389044523239,
0.07551583647727966,
-0.09457669407129288,
-0.09568615257740021,
0.134712353348732,
0.16491034626960754,
0.017767824232578278,
-0.0036161032039672136,
-0.028188811615109444,
0.008939792402088642,
-0.03011731058359146,
0.029252663254737854,
0.03727279603481293,
0.015201154164969921,
0.04628495126962662,
-0.008962608873844147,
0.03410780429840088,
0.07360003888607025,
0.13605917990207672,
-0.05178826302289963,
0.03173274174332619,
0.0720202773809433,
-0.02586594969034195,
-0.003797823330387473,
-0.10274355858564377,
-0.004696710966527462,
-0.05421094968914986,
-0.02262216806411743,
-0.02358926273882389,
0.01957574486732483,
0.016776498407125473,
-0.012574712745845318,
0.04663171246647835,
0.05746394768357277,
-0.0438656248152256,
0.07587822526693344,
0.040211133658885956,
0.015686700120568275,
-0.06144742667675018,
-0.07832043617963791,
0.06597990542650223
] | https://github.com/gstoica27/knots | 52 | 0 | 0 | 0 |
PolyhedronNet: Representation Learning for Polyhedra with Surface-attributed Graph | https://openreview.net/forum?id=BpyHIrpUOL | [
"Dazhou Yu",
"Genpei Zhang",
"Liang Zhao"
] | Poster | Ubiquitous geometric objects can be precisely and efficiently represented as polyhedra. The transformation of a polyhedron into a vector, known as polyhedra representation learning, is crucial for manipulating these shapes with mathematical and statistical tools for tasks like classification, clustering, and generation. Recent years have witnessed significant strides in this domain, yet most efforts focus on the vertex sequence of a polyhedron, neglecting the complex surface modeling crucial in real-world polyhedral objects.
This study proposes \textbf{PolyhedronNet}, a general framework tailored for learning representations of 3D polyhedral objects. We propose the concept of the surface-attributed graph to seamlessly model the vertices, edges, faces, and their geometric interrelationships within a polyhedron.
To effectively learn the representation of the entire surface-attributed graph, we first propose to break it down into local rigid representations to effectively learn each local region's relative positions against the remaining regions without geometric information loss. Subsequently, we propose PolyhedronGNN to hierarchically aggregate the local rigid representation via intra-face and inter-face geometric message passing modules, to obtain a global representation that minimizes information loss while maintaining rotation and translation invariance.
Our experimental evaluations on four distinct datasets, encompassing both classification and retrieval tasks, substantiate PolyhedronNet's efficacy in capturing comprehensive and informative representations of 3D polyhedral objects. | polygon, polyhedron, polygonal representation, representation learning, graph neural networks | A framework for encoding polyhedral objects using surface-attributed graph | 13,538 | 2502.01814 | [
-0.019062072038650513,
-0.03767762705683708,
-0.010781919583678246,
0.022100873291492462,
0.04173795133829117,
0.0017882331740111113,
-0.012184656225144863,
-0.054519470781087875,
-0.01535666361451149,
0.0068200016394257545,
-0.0849490836262703,
-0.033201396465301514,
0.03274646773934364,
0.11463028937578201,
-0.021960243582725525,
0.025721313431859016,
0.10177711397409439,
0.04310464859008789,
-0.03271745517849922,
0.0921899825334549,
0.0281374454498291,
0.009384391829371452,
-0.0010629752650856972,
0.0343073233962059,
0.0864722728729248,
0.06579146534204483,
0.05302917957305908,
-0.013132857158780098,
0.026405271142721176,
-0.006671831477433443,
0.0597413145005703,
0.011406900361180305,
-0.06412016600370407,
-0.003937100991606712,
-0.041638344526290894,
0.08236131817102432,
-0.07565716654062271,
0.07858840376138687,
-0.038237571716308594,
-0.06106158718466759,
0.026503073051571846,
0.08089207112789154,
0.033433347940444946,
0.05737690255045891,
0.08269470930099487,
0.06017753481864929,
-0.09457992762327194,
-0.033792655915021896,
0.00797203741967678,
-0.006278380285948515,
-0.04509502649307251,
-0.09319783747196198,
-0.016842177137732506,
-0.009647078812122345,
-0.006533553823828697,
0.04326203837990761,
0.014750311151146889,
-0.07580804079771042,
0.005492497235536575,
-0.02718917466700077,
0.05511389672756195,
-0.05678306147456169,
0.04097793623805046,
-0.03230408951640129,
-0.07879342138767242,
-0.025541294366121292,
0.0434192530810833,
0.03535929694771767,
-0.05070239305496216,
0.059546928852796555,
-0.04868881404399872,
0.0430934801697731,
-0.07189731299877167,
0.013301737606525421,
0.06344566494226456,
-0.040732331573963165,
0.05573620647192001,
-0.04985112324357033,
-0.0009855331154540181,
-0.03943656384944916,
-0.04883301630616188,
0.03293609619140625,
0.029233142733573914,
0.02756359614431858,
0.05847156420350075,
-0.0005578227574005723,
-0.026815133169293404,
0.1015537679195404,
-0.02940608188509941,
0.0015723052201792598,
-0.06463538855314255,
0.03178633004426956,
0.02662019245326519,
-0.043481990694999695,
-0.029286973178386688,
-0.01908189430832863,
0.03159669041633606,
-0.04525874927639961,
0.022641433402895927,
0.06965792924165726,
-0.0494326613843441,
-0.02468164637684822,
0.011004934087395668,
-0.04385305568575859,
-0.024028204381465912,
0.0036002807319164276,
-0.058790989220142365,
-0.02016572281718254,
0.008805138990283012,
-0.07390015572309494,
-0.03796491026878357,
0.01375525165349245,
-0.06193642318248749,
0.018203606829047203,
-0.06525828689336777,
0.0051733809523284435,
-0.011126123368740082,
-0.11403477191925049,
0.07438506931066513,
0.01845894195139408,
-0.0329250693321228,
0.006486305966973305,
-0.021646054461598396,
0.09129288792610168,
-0.016130711883306503,
-0.06033362075686455,
-0.06883490830659866,
-1.0192564579186202e-33,
0.04035145416855812,
0.03929208964109421,
0.059922970831394196,
0.0790800005197525,
0.08947017043828964,
0.052672941237688065,
0.010294616222381592,
0.020219283178448677,
0.044007379561662674,
-0.000890546478331089,
-0.13206692039966583,
-0.04103056713938713,
0.06937005370855331,
0.10809765011072159,
0.02482829988002777,
0.006838948465883732,
-0.02774825692176819,
-0.007189816329628229,
-0.045451514422893524,
-0.009159895591437817,
0.013972842134535313,
-0.0194782055914402,
0.010942613705992699,
0.03668069466948509,
-0.0668911412358284,
0.06551449000835419,
0.05613422393798828,
0.017433764412999153,
-0.02619681879878044,
-0.012830227613449097,
-0.08643164485692978,
-0.006808930076658726,
0.005820177029818296,
0.026486171409487724,
0.04106718674302101,
-0.0244669858366251,
-0.016059163957834244,
-0.08094160258769989,
0.001941726659424603,
-0.046045225113630295,
0.010998963378369808,
0.03442488610744476,
-0.01749345101416111,
-0.052223607897758484,
-0.008145860396325588,
0.04313315451145172,
0.0070196483284235,
0.0734613761305809,
-0.048003315925598145,
-0.01845650188624859,
0.021774819120764732,
0.04706193506717682,
0.0072523015551269054,
0.03474648296833038,
-0.061401527374982834,
0.019551699981093407,
-0.009908144362270832,
0.008861183188855648,
0.04747864603996277,
-0.010027536191046238,
-0.015545695088803768,
0.025069892406463623,
-0.025287488475441933,
0.047754429280757904,
-0.05108153820037842,
0.010894536972045898,
0.06943697482347488,
0.09873522073030472,
0.07885202765464783,
-0.026329148560762405,
-0.0155593017116189,
0.05582670867443085,
-0.04993065074086189,
0.00877306330949068,
0.010844106785953045,
0.031246785074472427,
0.0042482721619307995,
-0.09359030425548553,
-0.012161601334810257,
0.08975277096033096,
-0.03276554122567177,
-0.021739035844802856,
-0.007865858264267445,
-0.11062093824148178,
-0.06326432526111603,
-0.05422739312052727,
0.10985256731510162,
-0.07711897790431976,
0.03600836172699928,
0.0057733794674277306,
-0.05811773240566254,
-0.13345696032047272,
0.03811684995889664,
-0.013178771361708641,
-0.024384042248129845,
-7.374626817889301e-34,
-0.013553460128605366,
-0.011950337328016758,
-0.02960030548274517,
-0.03755991905927658,
0.0019859960302710533,
-0.06526518613100052,
-0.08249419182538986,
-0.03259917348623276,
-0.07140412926673889,
-0.03708617389202118,
-0.0025123252999037504,
-0.03022061288356781,
0.01025616005063057,
-0.0030861152336001396,
0.0036619058810174465,
0.0618608258664608,
-0.0734969824552536,
-0.031583890318870544,
-0.061986085027456284,
-0.0453747995197773,
0.01666322350502014,
0.044209953397512436,
-0.10390107333660126,
0.008887020871043205,
-0.017258623614907265,
0.043991416692733765,
0.011243962682783604,
0.01956351101398468,
0.024889567866921425,
0.10077884048223495,
-0.06818384677171707,
-0.06056387722492218,
-0.004581065848469734,
0.03475853428244591,
-0.028254348784685135,
0.03280102089047432,
-0.008437045849859715,
-0.0400015227496624,
0.04419626668095589,
0.022817688062787056,
0.016604235395789146,
-0.019623668864369392,
-0.09105623513460159,
0.10561062395572662,
-0.009833100251853466,
0.04181521013379097,
-0.0691153034567833,
0.006434793118387461,
-0.051050715148448944,
0.04756425693631172,
-0.09630470722913742,
0.022663312032818794,
0.013238489627838135,
-0.026262473315000534,
-0.02197347581386566,
-0.06201925128698349,
-0.007042366079986095,
0.041674885898828506,
0.0745483785867691,
-0.041851822286844254,
-0.018513444811105728,
-0.09017176926136017,
-0.015106670558452606,
0.03972255811095238,
-0.008996397256851196,
-0.013063670136034489,
0.029926510527729988,
-0.12914632260799408,
0.00011994584201602265,
0.06457636505365372,
0.042985524982213974,
-0.00021851136989425868,
-0.05602069944143295,
0.02437063865363598,
0.04401633143424988,
-0.05914207547903061,
0.026426907628774643,
0.08080936223268509,
-0.04200024530291557,
-0.03633740916848183,
0.0015826302114874125,
-0.05916285887360573,
0.05182046443223953,
0.07241714745759964,
0.08967787772417068,
-0.03424582630395889,
-0.02452254481613636,
0.0739220678806305,
0.01688103750348091,
0.03385290503501892,
-0.05081859230995178,
0.07401368767023087,
-0.04078660160303116,
0.040761999785900116,
0.019552597776055336,
-3.5408316279017527e-8,
-0.09681177139282227,
0.024127405136823654,
-0.0020209664944559336,
-0.07990890741348267,
0.0200442336499691,
0.06100153550505638,
0.060351163148880005,
0.08448182791471481,
-0.09434771537780762,
0.03112959861755371,
0.07442722469568253,
-0.0515645332634449,
-0.030222568660974503,
-0.05485925450921059,
0.004907345864921808,
0.10372624546289444,
0.012114559300243855,
-0.02930501475930214,
0.02895232103765011,
0.009449041448533535,
-0.01430346630513668,
0.03013775125145912,
0.005553922615945339,
0.015972459688782692,
0.023722900077700615,
-0.0941910669207573,
-0.0755772516131401,
0.15078003704547882,
0.08994506299495697,
-0.06225508451461792,
-0.04641842097043991,
0.03396683558821678,
0.08354765921831131,
-0.027439136058092117,
-0.035960663110017776,
0.12662753462791443,
0.05327916145324707,
-0.05729351192712784,
0.012192563153803349,
-0.034808844327926636,
0.03158934414386749,
0.11208543181419373,
-0.030005596578121185,
-0.015939099714159966,
0.08899254351854324,
0.0676952600479126,
0.04484637454152107,
-0.022771013900637627,
0.05048162862658501,
-0.024825941771268845,
-0.042109403759241104,
0.0033515174873173237,
-0.055495407432317734,
0.05760245770215988,
-0.022653751075267792,
0.09081798791885376,
-0.07534381747245789,
-0.06878934055566788,
0.08828838914632797,
0.016587531194090843,
-0.016515230759978294,
0.02471165731549263,
-0.047989003360271454,
-0.029143240302801132
] | https://github.com/dyu62/3d_polyhedron | 1 | 0 | 0 | 0 |
SafeDiffuser: Safe Planning with Diffusion Probabilistic Models | https://openreview.net/forum?id=ig2wk7kK9J | [
"Wei Xiao",
"Tsun-Hsuan Wang",
"Chuang Gan",
"Ramin Hasani",
"Mathias Lechner",
"Daniela Rus"
] | Poster | Diffusion models have shown promise in data-driven planning. While these planners are commonly employed in applications where decisions are critical, they still lack established safety guarantees. In this paper, we address this limitation by introducing SafeDiffuser, a method to equip diffusion models with safety guarantees via control barrier functions. The key idea of our approach is to embed finite-time diffusion invariance, i.e., a form of specification consisting of safety constraints, into the denoising diffusion procedure. This way we enable data generation under safety constraints. We show that SafeDiffusers maintain the generative performance of diffusion models while also providing robustness in safe data generation. We evaluate our method on a series of tasks, including maze path generation, legged robot locomotion, and 3D space manipulation, and demonstrate the advantages of robustness over vanilla diffusion models. | Diffusion model, Safety guarantees, Planning and control | We propose a new method to ensure diffusion probabilistic models satisfy specifications by using a control theoretic method | 13,531 | 2306.00148 | [
-0.030558649450540543,
-0.08846736699342728,
-0.006954406388103962,
-0.003988290671259165,
0.04523804038763046,
-0.04370588809251785,
-0.04887841269373894,
-0.03219446539878845,
-0.014083869755268097,
0.0762937143445015,
-0.019537797197699547,
-0.026656344532966614,
0.10004467517137527,
0.03768104314804077,
-0.0473049096763134,
-0.011471560224890709,
0.04237717017531395,
0.02069847658276558,
-0.12064041942358017,
-0.014445976354181767,
-0.021792666986584663,
0.011924264952540398,
0.010021599009633064,
-0.0031848556827753782,
-0.03992975875735283,
-0.02813020534813404,
0.059254731982946396,
-0.03720453381538391,
0.015144150704145432,
-0.036211248487234116,
0.03410112485289574,
-0.0970950797200203,
-0.07360109686851501,
-0.05787689983844757,
0.016748130321502686,
0.0541410855948925,
-0.09992828220129013,
-0.026065321639180183,
-0.03380909562110901,
-0.02461826056241989,
-0.0698971226811409,
0.024038322269916534,
-0.010162578895688057,
0.04644477367401123,
0.006394893396645784,
0.006731506437063217,
-0.01843011938035488,
-0.07267702370882034,
0.010584319941699505,
0.011249693110585213,
-0.027293778955936432,
-0.02174995094537735,
0.030593983829021454,
0.005330657586455345,
-0.007046625949442387,
-0.03809908404946327,
0.0821065604686737,
-0.005270360503345728,
-0.0023604300804436207,
-0.042957939207553864,
-0.018988389521837234,
-0.06111305207014084,
0.0007396175642497838,
-0.05529696121811867,
-0.030004872009158134,
-0.029762564226984978,
-0.01684209518134594,
0.001387926982715726,
0.0432034395635128,
0.08675695955753326,
-0.0010509591083973646,
0.0009440950816497207,
-0.06613893061876297,
0.01968703791499138,
0.024839147925376892,
-0.013955408707261086,
0.0017506350995972753,
-0.026603147387504578,
-0.011806951835751534,
-0.03986091539263725,
0.07351574301719666,
0.0011110203340649605,
0.03510137274861336,
0.07334356755018234,
0.034571293741464615,
0.022594358772039413,
-0.014367262832820415,
0.09321858733892441,
0.10937640815973282,
-0.04079751297831535,
-0.016803767532110214,
0.08570164442062378,
-0.02030067890882492,
0.027417533099651337,
-0.03087616339325905,
-0.04017709940671921,
-0.015795286744832993,
-0.036426249891519547,
-0.008901008404791355,
0.013262266293168068,
-0.048588208854198456,
-0.025030221790075302,
0.010002010501921177,
0.04427314177155495,
0.07196617126464844,
-0.0815521702170372,
0.07512068003416061,
-0.030461510643363,
-0.020257966592907906,
0.05384218692779541,
0.06010367348790169,
-0.03634074702858925,
0.08606535196304321,
0.05955060198903084,
-0.020519420504570007,
-0.07659025490283966,
-0.07667561620473862,
0.027942923828959465,
0.020944179967045784,
0.06068806350231171,
0.0575610026717186,
-0.021133719012141228,
0.05350278690457344,
0.0007311781519092619,
0.025601662695407867,
-0.0799206793308258,
0.06398062407970428,
4.924889703113442e-34,
0.004243003204464912,
-0.05569735914468765,
0.0074311369098722935,
0.03985792398452759,
0.06516263633966446,
-0.025218136608600616,
-0.021433714777231216,
-0.020849790424108505,
-0.02190566249191761,
-0.025843601673841476,
-0.028987212106585503,
-0.035952258855104446,
-0.12892615795135498,
0.03917356953024864,
0.13245949149131775,
-0.017093965783715248,
0.01145191490650177,
0.06258590519428253,
-0.05232781544327736,
0.014221473596990108,
0.12301023304462433,
-0.07961877435445786,
-0.05989560857415199,
-0.0041237627156078815,
0.03369569033384323,
0.03339803218841553,
-0.030981671065092087,
-0.04351748898625374,
0.06503858417272568,
0.030622901394963264,
-0.18587985634803772,
-0.0010023587383329868,
-0.058598410338163376,
0.04819861054420471,
0.01355677843093872,
0.11522117257118225,
-0.1082700714468956,
-0.03931954503059387,
-0.04511908441781998,
0.027401985600590706,
0.03648093342781067,
-0.01988980360329151,
0.05519887059926987,
-0.002495319815352559,
-0.09417204558849335,
-0.0323357991874218,
0.02632325142621994,
0.0028355284593999386,
-0.07908986508846283,
-0.06103441119194031,
0.013382554054260254,
0.05166544392704964,
0.02610000967979431,
-0.10182545334100723,
-0.048324525356292725,
-0.05938471108675003,
0.04029865562915802,
-0.05235324800014496,
-0.017991621047258377,
0.061817560344934464,
0.017876889556646347,
0.04579184949398041,
0.011467145755887032,
0.014864679425954819,
0.01830957643687725,
-0.012301328592002392,
-0.02300911210477352,
0.049255117774009705,
0.12277563661336899,
-0.013109389692544937,
-0.09056741744279861,
0.016208985820412636,
0.0058664013631641865,
-0.06447318196296692,
-0.028819257393479347,
-0.03541455790400505,
0.026813309639692307,
-0.013149997219443321,
0.026859914883971214,
-0.08426608890295029,
-0.04566631838679314,
-0.06321515142917633,
-0.06035171076655388,
-0.012619251385331154,
0.05726482346653938,
-0.07358382642269135,
-0.02630486525595188,
0.03932241350412369,
-0.10998651385307312,
-0.02933247573673725,
-0.02187202498316765,
-0.044063568115234375,
-0.020976033061742783,
0.09246791899204254,
-0.02601350098848343,
-2.756130727508146e-33,
-0.029859542846679688,
-0.012925415299832821,
0.06964016705751419,
0.02755962498486042,
-0.04736744612455368,
-0.04455098137259483,
-0.03840697556734085,
-0.044944047927856445,
0.09755222499370575,
-0.035281769931316376,
-0.15011714398860931,
0.006932997144758701,
0.04510100930929184,
-0.0003841864818241447,
0.034044019877910614,
-0.052231501787900925,
0.05400240421295166,
-0.04126443713903427,
-0.06783895939588547,
-0.012015698477625847,
-0.02113247476518154,
0.053351886570453644,
-0.1296168565750122,
0.04148012772202492,
0.012064743787050247,
0.030377112329006195,
-0.03716571256518364,
-0.0071814823895692825,
0.023171501234173775,
0.08386843651533127,
-0.06544657051563263,
-0.02183535322546959,
0.00023111743212211877,
-0.0041078669019043446,
-0.04159820079803467,
-0.013693045824766159,
0.044421400874853134,
0.06426626443862915,
-0.0038244540337473154,
-0.00010354831465519965,
0.043848875910043716,
-0.0042733424343168736,
-0.05132054165005684,
0.0367683582007885,
0.05610784515738487,
-0.008502393960952759,
-0.03229738026857376,
0.024052629247307777,
-0.019042765721678734,
-0.013980447314679623,
0.0003264149709139019,
0.085633285343647,
-0.03876524418592453,
-0.04516417533159256,
0.015041391365230083,
-0.029325131326913834,
-0.04791160300374031,
-0.05107155442237854,
0.11971325427293777,
0.00544344075024128,
-0.0703451856970787,
0.02063904143869877,
-0.03418435528874397,
0.10604037344455719,
-0.04844711348414421,
0.043963320553302765,
-0.09463849663734436,
-0.028988562524318695,
0.037242524325847626,
0.017540857195854187,
0.04005860164761543,
0.0529298298060894,
0.01502973586320877,
0.01047599408775568,
0.0416443794965744,
-0.08181946724653244,
0.041984934359788895,
0.004536516964435577,
0.05146404728293419,
0.02466743066906929,
0.06346838176250458,
0.040333785116672516,
0.09975956380367279,
0.03663954511284828,
-0.019067564979195595,
0.013984913937747478,
-0.029303627088665962,
-0.02807798609137535,
-0.031264808028936386,
-0.02647198922932148,
-0.01735050231218338,
-0.025042319670319557,
-0.008512280881404877,
0.03275709226727486,
-0.10014700889587402,
-3.739400966651374e-8,
-0.03476087376475334,
0.07109172642230988,
0.06278891861438751,
0.01954244263470173,
0.08035751432180405,
0.01608910970389843,
0.02453254722058773,
-0.021052932366728783,
0.03408501669764519,
0.030548149719834328,
0.08909010142087936,
0.017390312626957893,
0.0027063400484621525,
-0.0040084063075482845,
-0.0699692815542221,
0.11922530829906464,
0.023582573980093002,
0.01725243590772152,
0.02087855525314808,
0.03391609340906143,
-0.025815187022089958,
0.024594560265541077,
-0.04734570160508156,
0.03941211849451065,
0.04775972291827202,
0.0018509146757423878,
-0.07934784889221191,
-0.019614888355135918,
0.09354978054761887,
0.06180734932422638,
0.009719327092170715,
-0.03813955560326576,
0.06389761716127396,
0.12683674693107605,
-0.014883688651025295,
0.04781638830900192,
0.0061918036080896854,
0.024631906300783157,
-0.01632423885166645,
-0.056517668068408966,
0.06090173497796059,
0.10919101536273956,
-0.03366146609187126,
-0.0019633451011031866,
-0.02773469313979149,
0.05517282709479332,
0.019410204142332077,
-0.0842926874756813,
-0.0000297419373964658,
-0.021145718172192574,
-0.05394280329346657,
-0.042396217584609985,
0.03147782012820244,
0.08996023237705231,
0.08314299583435059,
0.06626298278570175,
0.014177592471241951,
-0.08388078212738037,
0.05729058012366295,
0.030818335711956024,
-0.06058818846940994,
0.04883565381169319,
0.0629878044128418,
0.013551772572100163
] | 0 | 0 | 0 | 0 |
|
BrainACTIV: Identifying visuo-semantic properties driving cortical selectivity using diffusion-based image manipulation | https://openreview.net/forum?id=CGON8Btleu | [
"Diego Garcia Cerdas",
"Christina Sartzetaki",
"Magnus Petersen",
"Gemma Roig",
"Pascal Mettes",
"Iris Groen"
] | Poster | The human brain efficiently represents visual inputs through specialized neural populations that selectively respond to specific categories. Advancements in generative modeling have enabled data-driven discovery of neural selectivity using brain-optimized image synthesis. However, current methods independently generate one sample at a time, without enforcing structural constraints on the generations; thus, these individual images have no explicit point of comparison, making it hard to discern which image features drive neural response selectivity. To address this issue, we introduce Brain Activation Control Through Image Variation (BrainACTIV), a method for manipulating a reference image to enhance or decrease activity in a target cortical region using pretrained diffusion models. Starting from a reference image allows for fine-grained and reliable offline identification of optimal visuo-semantic properties, as well as producing controlled stimuli for novel neuroimaging studies. We show that our manipulations effectively modulate predicted fMRI responses and agree with hypothesized preferred categories in established regions of interest, while remaining structurally close to the reference image. Moreover, we demonstrate how our method accentuates differences between brain regions that are selective to the same category, and how it could be used to explore neural representation of brain regions with unknown selectivities. Hence, BrainACTIV holds the potential to formulate robust hypotheses about brain representation and to facilitate the production of naturalistic stimuli for neuroscientific experiments. | brain, selectivity, visual cortex, fMRI, manipulation, variation, diffusion, neuroscience | BrainACTIV manipulates a reference image to drive or suppress activity in a target brain region using pretrained diffusion models, providing insights into neural selectivity for experimental purposes. | 13,524 | null | [
0.05428843945264816,
-0.07961981743574142,
0.10735435038805008,
0.10717463493347168,
0.05813846364617348,
0.0003543643979355693,
-0.01104348711669445,
-0.0410759299993515,
0.021912267431616783,
-0.035055261105298996,
-0.05646728724241257,
-0.11539868265390396,
-0.0020961137488484383,
0.039390064775943756,
0.016768788918852806,
0.030289363116025925,
0.05632564797997475,
0.038078825920820236,
-0.016446560621261597,
0.035177603363990784,
-0.007191915530711412,
-0.039466191083192825,
0.008351330645382404,
-0.06315623968839645,
0.04283539950847626,
0.04746517539024353,
0.035656820982694626,
-0.012546587735414505,
0.02696087770164013,
-0.05375858023762703,
0.05777408555150032,
0.026096010580658913,
-0.04165686294436455,
0.012618140317499638,
-0.10067449510097504,
0.060282234102487564,
-0.045051004737615585,
0.011236583814024925,
0.03970417380332947,
-0.04930474981665611,
0.034965332597494125,
0.000786173390224576,
-0.0270156878978014,
0.028046496212482452,
-0.011175282299518585,
0.020854011178016663,
-0.006921215448528528,
-0.0519704595208168,
-0.006881562527269125,
-0.010090989992022514,
-0.04149197041988373,
0.025958839803934097,
-0.04235891252756119,
0.06858044117689133,
-0.014263942837715149,
0.08436461538076401,
0.06795268505811691,
0.030720902606844902,
-0.0019103247905150056,
0.006455794908106327,
-0.066068135201931,
-0.029640087857842445,
0.03692912310361862,
-0.06955323368310928,
0.025424329563975334,
-0.0024880997370928526,
-0.010691188275814056,
-0.06308580935001373,
-0.07459075003862381,
-0.031950242817401886,
0.02115698531270027,
0.06650493294000626,
-0.038927990943193436,
-0.01878361403942108,
0.021069947630167007,
0.04298369586467743,
0.019463544711470604,
-0.03808131441473961,
0.06462346017360687,
-0.11738883703947067,
0.027751488611102104,
0.06764578819274902,
-0.002576331375166774,
0.03827127069234848,
0.08889231830835342,
0.048135921359062195,
-0.020478224381804466,
0.06999564915895462,
0.10127212852239609,
0.02798965387046337,
-0.06051803380250931,
-0.03195708617568016,
-0.050800975412130356,
-0.043772436678409576,
0.06458450108766556,
0.025309374555945396,
-0.0409417562186718,
-0.09504327923059464,
0.028099874034523964,
0.025025686249136925,
0.036690693348646164,
0.007140676025301218,
-0.010557767003774643,
0.007167850621044636,
0.025053346529603004,
-0.050483282655477524,
0.1094704195857048,
0.020500941202044487,
0.015269805677235126,
-0.007998846471309662,
0.01606636680662632,
0.01673762872815132,
-0.002931887051090598,
0.017818843945860863,
-0.02034737728536129,
-0.06251726299524307,
0.015787234529852867,
0.02419212833046913,
0.045793164521455765,
-0.020482953637838364,
0.048163559287786484,
0.013413944281637669,
-0.05160455405712128,
0.016942717134952545,
0.006180481519550085,
-0.13094088435173035,
-0.11261827498674393,
1.4202776413985898e-33,
0.0013934122398495674,
-0.10247766226530075,
0.0590687096118927,
0.0673949271440506,
-0.0016124555841088295,
-0.05134982615709305,
-0.06510043144226074,
-0.06696397066116333,
-0.025956233963370323,
-0.025758665055036545,
-0.05035325884819031,
0.03857703506946564,
-0.07256283611059189,
0.18682311475276947,
0.06469418853521347,
-0.027304809540510178,
-0.07235082983970642,
0.08471665531396866,
0.05502935126423836,
-0.05028539150953293,
-0.00891486182808876,
0.005038642790168524,
-0.06398214399814606,
-0.05324361473321915,
-0.06455443799495697,
-0.02327832207083702,
0.006380717270076275,
-0.04749191552400589,
0.0018970549572259188,
-0.009461643174290657,
-0.14499710500240326,
-0.0365782156586647,
0.022630227729678154,
-0.0008175913244485855,
-0.016369741410017014,
0.03857173025608063,
-0.009942884556949139,
-0.011260533705353737,
-0.0071552530862390995,
0.057028140872716904,
0.08138065785169601,
0.0469798743724823,
-0.017268426716327667,
0.051712073385715485,
-0.06590645015239716,
0.05264727398753166,
-0.013056139461696148,
0.12161429226398468,
-0.07667778432369232,
-0.02674802578985691,
0.01586243510246277,
-0.024763019755482674,
0.025513947010040283,
-0.05783630907535553,
-0.06482953578233719,
-0.026829062029719353,
-0.04406963661313057,
-0.058225370943546295,
0.08277323842048645,
-0.0033890961203724146,
0.03629983589053154,
0.03374055027961731,
-0.023940281942486763,
0.06899121403694153,
0.02144535444676876,
0.03782352805137634,
-0.008895759470760822,
0.019255589693784714,
-0.02327917329967022,
0.06077801436185837,
-0.030202994123101234,
0.06810219585895538,
-0.020533982664346695,
-0.08513320982456207,
0.032675690948963165,
-0.04780999943614006,
0.0890212133526802,
-0.007979845628142357,
0.003495489712804556,
-0.02680348977446556,
-0.034413378685712814,
-0.0052459691651165485,
-0.056911274790763855,
-0.08029986917972565,
-0.03436996415257454,
0.005991020239889622,
0.014882916584610939,
-0.04815502464771271,
-0.10056576132774353,
-0.10384482890367508,
0.010137492790818214,
0.016537776216864586,
0.016256529837846756,
0.010438773781061172,
-0.010585217736661434,
-1.7700705973987318e-33,
-0.013464986346662045,
-0.031168930232524872,
0.07540857046842575,
0.015293603762984276,
-0.06506159901618958,
0.010011302307248116,
0.08637256920337677,
-0.03326762095093727,
0.039119794964790344,
-0.03057650290429592,
-0.01704344153404236,
-0.050987668335437775,
-0.02811959944665432,
-0.08506518602371216,
-0.01238307449966669,
-0.1009073555469513,
-0.029806459322571754,
0.05604243651032448,
0.011458438821136951,
0.06353159993886948,
-0.024388037621974945,
0.10836242884397507,
-0.04974360764026642,
0.033835891634225845,
-0.09989195317029953,
0.10443994402885437,
-0.005380571354180574,
0.06576363742351532,
-0.020457573235034943,
-0.005088902544230223,
-0.09803475439548492,
0.03349107876420021,
0.0012794721405953169,
-0.04530727118253708,
-0.01657070219516754,
0.11188048869371414,
0.020360402762889862,
-0.011587273329496384,
-0.07800117135047913,
0.07135947048664093,
0.004855175968259573,
-0.04955730959773064,
-0.045994244515895844,
0.08306828141212463,
-0.019781680777668953,
-0.044824179261922836,
-0.02915947325527668,
0.006895445287227631,
0.010534788481891155,
0.10293178260326385,
0.011694465763866901,
0.015555205754935741,
-0.13415539264678955,
-0.010862495750188828,
-0.020567774772644043,
-0.04844412952661514,
-0.027905503287911415,
-0.06536034494638443,
0.12030698359012604,
-0.006821127608418465,
-0.05193566530942917,
-0.04286396503448486,
-0.0009225762914866209,
-0.058491889387369156,
-0.041851628571748734,
0.013836965896189213,
-0.048105474561452866,
-0.018514078110456467,
0.055715326219797134,
0.0559823177754879,
-0.06275945156812668,
0.0035656967666000128,
0.12034936249256134,
-0.03035513311624527,
0.07582099735736847,
-0.06839434802532196,
0.050827331840991974,
0.009246512316167355,
0.0034726003650575876,
-0.027660949155688286,
0.038063473999500275,
-0.03945598006248474,
-0.04093252867460251,
-0.007709199097007513,
0.02945675514638424,
0.07294322550296783,
-0.07233374565839767,
0.007770525757223368,
0.0561586394906044,
0.006836968008428812,
-0.04422018676996231,
-0.05130217224359512,
0.0281971525400877,
0.028204521164298058,
-0.06719741225242615,
-4.2671356226264834e-8,
-0.05773831531405449,
0.023651648312807083,
0.09372761845588684,
0.05279400944709778,
0.04472114518284798,
-0.03083907440304756,
0.0006156311719678342,
0.031773198395967484,
0.028301026672124863,
-0.06999567896127701,
0.030901331454515457,
0.011165192350745201,
0.02377859316766262,
-0.08365203440189362,
0.02455105632543564,
0.05694183334708214,
-0.007036102470010519,
0.05509016662836075,
0.005400066263973713,
-0.03488291800022125,
-0.028922665864229202,
-0.005408720578998327,
0.03720032796263695,
-0.06567802280187607,
0.0972680076956749,
-0.04213690757751465,
-0.06589192152023315,
0.0331009179353714,
0.009697986766695976,
-0.034740813076496124,
0.08685488998889923,
0.023628881201148033,
0.032386768609285355,
-0.008749588392674923,
0.053930144757032394,
0.03127089515328407,
-0.008708938024938107,
0.012330560013651848,
-0.08408825844526291,
-0.0075972010381519794,
-0.029865503311157227,
0.051016367971897125,
-0.005804500542581081,
-0.009124010801315308,
0.00704932538792491,
-0.011408129706978798,
0.06098347157239914,
-0.0024055717512965202,
0.07777441293001175,
0.05730322375893593,
0.00476278318092227,
-0.008861050941050053,
0.03583567216992378,
0.09851378202438354,
0.049846477806568146,
-0.049779705703258514,
0.007626708131283522,
0.004265157040208578,
0.03488058224320412,
0.02075866237282753,
0.08840183913707733,
0.07215837389230728,
-0.035827722400426865,
-0.004452441819012165
] | 0 | 0 | 0 | 0 |
|
Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data | https://openreview.net/forum?id=HN0CYZbAPw | [
"Zhiyuan Zhou",
"Andy Peng",
"Qiyang Li",
"Sergey Levine",
"Aviral Kumar"
] | Poster | The modern paradigm in machine learning involves pre-training on diverse data, followed by task-specific fine-tuning. In reinforcement learning (RL), this translates to learning via offline RL on a diverse historical dataset, followed by rapid online RL fine-tuning using interaction data. Most RL fine-tuning methods require continued training on offline data for stability and performance. However, this is undesirable because training on diverse offline data is slow and expensive for large datasets, and should, in principle, also limit the performance improvement possible because of constraints or pessimism on offline data. In this paper, we show that retaining offline data is unnecessary as long as we use a properly-designed online RL approach for fine-tuning offline RL initializations. To build this approach, we start by analyzing the role of retaining offline data in online fine-tuning. We find that continued training on offline data is mostly useful for preventing a sudden divergence in the value function at the onset of fine-tuning, caused by a distribution mismatch between the offline data and online rollouts. This divergence typically results in unlearning and forgetting the benefits of offline pre-training. Our approach, Warm-start RL (WSRL), mitigates the catastrophic forgetting of pre-trained initializations using a very simple idea. WSRL employs a warmup phase that seeds the online RL run with a very small number of rollouts from the pre-trained policy to do fast online RL. The data collected during warmup bridges the distribution mismatch, and helps ``recalibrate'' the offline Q-function to the online distribution, allowing us to completely discard offline data without destabilizing the online RL fine-tuning. We show that WSRL is able to fine-tune without retaining any offline data, and is able to learn faster and attains higher performance than existing algorithms irrespective of whether they do or do not retain offline data. | Reinforcement learning, fast fine-tuning | We find that previous RL fine-tuning methods fail because of Q-divergence when no offline data is kept, and propose a new method WSRL that can finetune more efficiently with no data retention. | 13,523 | 2412.07762 | [
-0.08821123838424683,
-0.05282408744096756,
0.03373328596353531,
0.07820477336645126,
0.03150755167007446,
0.014197918586432934,
-0.026179978623986244,
-0.003128543496131897,
-0.054902635514736176,
-0.019673462957143784,
-0.03488728776574135,
0.03413037955760956,
-0.005493734497576952,
-0.019606564193964005,
0.017158977687358856,
0.020591186359524727,
0.07495272159576416,
0.03683280199766159,
-0.14663788676261902,
-0.04606853425502777,
-0.030650904402136803,
0.022830702364444733,
0.03241570666432381,
0.02949320524930954,
0.012858197093009949,
-0.02222502790391445,
0.011699085123836994,
-0.013337826356291771,
0.048207838088274,
-0.044324710965156555,
-0.008181028068065643,
0.016405152156949043,
-0.06422954052686691,
-0.07299449294805527,
-0.09485440701246262,
0.050810087472200394,
-0.11448635160923004,
0.0347861722111702,
0.004415839444845915,
0.013236433267593384,
-0.02857515960931778,
0.011283383704721928,
0.006096191238611937,
0.020873956382274628,
0.06762520968914032,
-0.011636273004114628,
-0.028059422969818115,
-0.056498486548662186,
0.027718814089894295,
0.047714587301015854,
-0.07546854019165039,
0.03160221874713898,
0.00960919726639986,
0.09096194803714752,
0.0403515100479126,
0.04446578025817871,
0.06856881827116013,
0.0291457362473011,
0.01623118668794632,
-0.005558500997722149,
-0.03264792636036873,
-0.10188402235507965,
0.015362442471086979,
-0.02915710024535656,
-0.11207414418458939,
-0.023713285103440285,
0.03181677311658859,
0.04547238349914551,
0.09130482375621796,
0.029709894210100174,
-0.06789350509643555,
0.07889524847269058,
-0.057802196592092514,
0.015652401372790337,
0.025708910077810287,
0.007758917286992073,
0.013685145415365696,
-0.058241311460733414,
0.016524553298950195,
0.017359722405672073,
-0.002660006983205676,
-0.006298601161688566,
-0.039769336581230164,
-0.02188631147146225,
0.086545929312706,
-0.03373290225863457,
0.03690639138221741,
0.011998133733868599,
0.16300424933433533,
-0.010349764488637447,
0.061662327498197556,
0.1127668172121048,
0.07652920484542847,
-0.00420926371589303,
-0.041342366486787796,
0.018967682495713234,
0.030499737709760666,
-0.07726376503705978,
-0.010864130221307278,
0.08661942183971405,
0.02085096761584282,
0.10090182721614838,
0.013643537648022175,
0.05887233838438988,
-0.04561799392104149,
-0.08402551710605621,
0.04201555252075195,
0.06808049976825714,
0.010750317014753819,
-0.07277705520391464,
0.01500744465738535,
0.016936838626861572,
0.020639535039663315,
-0.0017612206283956766,
-0.04369863122701645,
0.05340100824832916,
-0.028654104098677635,
-0.06891553848981857,
-0.0594383180141449,
0.08157910406589508,
0.030279414728283882,
-0.05915604159235954,
0.00847939308732748,
0.010545155964791775,
0.015391082502901554,
-0.07123579829931259,
-0.05174000933766365,
4.738791682304996e-33,
0.045739538967609406,
-0.06620080769062042,
0.015504433773458004,
-0.008887223899364471,
0.01661696471273899,
-0.057487793266773224,
0.05664872005581856,
0.021213935688138008,
-0.05981728062033653,
-0.023411249741911888,
0.014717858284711838,
0.007932109758257866,
-0.0434766449034214,
0.026974570006132126,
0.031884998083114624,
-0.005514618940651417,
-0.006654673721641302,
0.04106568172574043,
0.0504768081009388,
-0.03718465566635132,
0.09573021531105042,
-0.08764161914587021,
-0.005449860356748104,
-0.03608322888612747,
0.029922688379883766,
0.039358627051115036,
0.04272741451859474,
0.006276588421314955,
0.020654235035181046,
-0.015772324055433273,
0.03371492400765419,
-0.03829527646303177,
0.005199586506932974,
0.0739179328083992,
0.028116166591644287,
-0.009364008903503418,
-0.10984684526920319,
-0.03274600952863693,
-0.0270611010491848,
0.03058427758514881,
0.021929768845438957,
0.03938993439078331,
-0.0018954048864543438,
0.014109151437878609,
-0.007825493812561035,
-0.17592564225196838,
0.04145737737417221,
-0.02362058311700821,
-0.04181212931871414,
-0.024195637553930283,
-0.02339857630431652,
0.0759379044175148,
-0.015306128188967705,
-0.03130631148815155,
-0.07507982105016708,
0.045001234859228134,
-0.017935344949364662,
0.05577931925654411,
-0.05901462584733963,
0.040169090032577515,
-0.0066653043031692505,
-0.003981344867497683,
0.009909571148455143,
0.0695720762014389,
0.022684605792164803,
0.06912993639707565,
-0.014397962018847466,
0.01999160833656788,
-0.013398806564509869,
-0.02471223659813404,
0.025691408663988113,
0.0030278542544692755,
0.041942037642002106,
-0.07945374399423599,
0.11564917117357254,
-0.00862739235162735,
0.0074965571984648705,
-0.05761830508708954,
0.035141922533512115,
-0.006717678625136614,
0.06681565195322037,
0.016396578401327133,
-0.01923208124935627,
-0.025720927864313126,
0.0031514770817011595,
-0.14255063235759735,
-0.013377783820033073,
-0.014006715267896652,
-0.0500514879822731,
0.021106472238898277,
-0.11420432478189468,
-0.005639501381665468,
-0.060021910816431046,
-0.07240596413612366,
0.04067940637469292,
-3.643336007146229e-33,
-0.03182077407836914,
-0.10259842872619629,
0.005523685831576586,
0.06545277684926987,
-0.01672258973121643,
0.021302873268723488,
0.022137951105833054,
-0.022089466452598572,
0.010534136556088924,
-0.039374593645334244,
-0.04861752316355705,
0.011493035592138767,
-0.004371787887066603,
-0.01388698723167181,
-0.007011412642896175,
-0.05058629810810089,
-0.07111801952123642,
0.04414936900138855,
0.030401630327105522,
0.05867430195212364,
0.057591427117586136,
-0.001432639081031084,
-0.0893281102180481,
0.0767132043838501,
-0.02369139902293682,
0.008456342853605747,
-0.04821881279349327,
0.05679506063461304,
-0.024425173178315163,
0.06959550827741623,
-0.045509904623031616,
-0.04420576989650726,
-0.022135552018880844,
-0.005760814994573593,
-0.045139510184526443,
0.05099508911371231,
0.045391831547021866,
0.017733987420797348,
-0.07741034030914307,
0.13865245878696442,
0.07683596760034561,
0.004155413713306189,
-0.09556402266025543,
-0.040081266313791275,
0.022188767790794373,
-0.013917922042310238,
-0.12399424612522125,
0.04927654564380646,
-0.02810802310705185,
0.025662878528237343,
-0.01612081378698349,
0.08237431198358536,
-0.0475873164832592,
-0.020610589534044266,
-0.02477704919874668,
-0.03810913488268852,
-0.0403127446770668,
-0.0009701655944809318,
-0.011677674949169159,
0.004993070382624865,
-0.019453786313533783,
-0.04517624154686928,
-0.05269622802734375,
0.033506568521261215,
0.02961292676627636,
-0.025993039831519127,
0.017744848504662514,
0.02434862218797207,
0.0069036888889968395,
0.03379879146814346,
-0.011008121073246002,
0.0026845999527722597,
0.04550425335764885,
-0.05092174932360649,
0.005983444396406412,
-0.028489967808127403,
-0.012637612409889698,
-0.04003596305847168,
0.011026980355381966,
0.022679327055811882,
0.004686530213803053,
0.01763634756207466,
0.07034287601709366,
0.053743280470371246,
0.02412397973239422,
0.08041895180940628,
-0.02986452542245388,
0.12599146366119385,
0.043991219252347946,
-0.053538329899311066,
-0.005198023747652769,
-0.013137638568878174,
-0.03608880937099457,
0.058860793709754944,
-0.04499537870287895,
-4.379760554229506e-8,
-0.019566768780350685,
0.05587194859981537,
0.0887024849653244,
0.08333123475313187,
0.1393318474292755,
0.007888502441346645,
0.010402660816907883,
0.051236942410469055,
-0.025671087205410004,
-0.003800703678280115,
0.023306088522076607,
0.00008315778541145846,
-0.013267376460134983,
-0.09783415496349335,
0.03222700580954552,
0.05992890149354935,
-0.010941354557871819,
-0.06593118607997894,
-0.0011528576724231243,
0.022513242438435555,
0.07178414613008499,
0.006512449122965336,
-0.038184598088264465,
-0.05241799354553223,
0.06526169180870056,
-0.03917139023542404,
0.029465429484844208,
0.1258946657180786,
0.039373207837343216,
-0.06830765306949615,
0.05031123757362366,
-0.07322419434785843,
0.1003589853644371,
-0.04963592439889908,
0.0513032004237175,
0.035332534462213516,
0.08003387600183487,
-0.05583382770419121,
-0.00741838151589036,
-0.030930591747164726,
0.0022014491260051727,
0.08868227154016495,
-0.019573524594306946,
-0.0280343908816576,
0.05581560730934143,
0.0005396869964897633,
-0.1052166223526001,
-0.05690121650695801,
0.08534958958625793,
-0.0880681574344635,
-0.008894511498510838,
-0.03582736477255821,
-0.02072717249393463,
0.026183096691966057,
0.10455945879220963,
-0.03555143252015114,
0.0016345089534297585,
-0.031644806265830994,
0.04237890988588333,
0.09526875615119934,
-0.008680245839059353,
-0.058771710842847824,
-0.09439792484045029,
0.017727769911289215
] | https://github.com/zhouzypaul/wsrl | 41 | 0 | 0 | 0 |
Making Text Embedders Few-Shot Learners | https://openreview.net/forum?id=wfLuiDjQ0u | [
"Chaofan Li",
"Minghao Qin",
"Shitao Xiao",
"Jianlyu Chen",
"Kun Luo",
"Defu Lian",
"Yingxia Shao",
"Zheng Liu"
] | Poster | Large language models (LLMs) with decoder-only architectures have demonstrated exceptional text-generation capabilities across a variety of tasks. Some researchers have also adapted these models for text representation tasks. However, in text representation tasks, these models often face performance degradation on unseen tasks. In-context learning (ICL), which leverages examples provided in the input context, enables LLMs to handle unseen tasks effectively. Inspired by this, we aim to fully utilize the inherent properties of LLMs to enhance text representation performance across different tasks through the ICL approach.
In this paper, we introduce a simple yet effective training strategy, which significantly improves text representation capabilities. Unlike previous models that prepend task instructions to the text, our method randomly samples a varying number of examples during training, endowing the embedding model with in-context learning abilities while maintaining its zero-shot capabilities. This approach does not require additional data construction or modifications to the model architecture. On the contrary, we find that some popular modifications to the model, such as bidirectional attention, can degrade performance, undermining the inherent characteristics of LLMs. We have publicly released our method at this \href{https://github.com/FlagOpen/FlagEmbedding}{repo}. | large language model, embedding model, in-context learning | null | 13,517 | 2409.15700 | [
-0.031241336837410927,
-0.04190240800380707,
0.035359710454940796,
0.10312189906835556,
0.10391794890165329,
0.012599312700331211,
0.01333900447934866,
-0.021919481456279755,
0.08514091372489929,
-0.03530709072947502,
-0.05414973944425583,
-0.005165801849216223,
0.07354623824357986,
0.02994299866259098,
-0.005409381352365017,
0.0036271719727665186,
0.14155523478984833,
0.05697740241885185,
-0.09858927130699158,
-0.029927289113402367,
0.012179132550954819,
-0.047712501138448715,
0.10207667946815491,
-0.002367864828556776,
0.030390728265047073,
-0.028343984857201576,
-0.061732739210128784,
-0.028152652084827423,
0.023072650656104088,
-0.002803351264446974,
0.06567391008138657,
-0.025079211220145226,
-0.043409109115600586,
0.024084849283099174,
-0.03513842076063156,
0.08600166440010071,
-0.04745256155729294,
0.10060437023639679,
-0.016784051433205605,
-0.07096198946237564,
0.01619158871471882,
0.014755739830434322,
0.0010605992283672094,
0.07869580388069153,
0.07345986366271973,
-0.03619716316461563,
0.016916098073124886,
-0.08804906159639359,
-0.006166588049381971,
0.006412307731807232,
-0.09637880325317383,
-0.03524813428521156,
-0.02788740210235119,
0.0026494197081774473,
-0.055907897651195526,
-0.009755471721291542,
-0.01904159225523472,
0.017177755013108253,
-0.03273129090666771,
-0.03279060125350952,
-0.058791372925043106,
-0.07447738945484161,
-0.034048475325107574,
-0.007504064589738846,
0.009454527869820595,
-0.03914232552051544,
0.025185538455843925,
0.05951370671391487,
-0.043097034096717834,
0.025777040049433708,
-0.008357826620340347,
0.05889084190130234,
-0.04364551231265068,
0.08396812528371811,
0.009883822873234749,
0.02854338102042675,
0.0048852297477424145,
-0.046409863978624344,
0.08426371216773987,
-0.05771249905228615,
0.051022719591856,
-0.007682176772505045,
0.04006011411547661,
-0.026243574917316437,
0.12970858812332153,
0.035576339811086655,
0.06571508198976517,
0.0213027261197567,
-0.0156848207116127,
-0.010655734688043594,
-0.025190453976392746,
-0.07087987661361694,
0.09793281555175781,
-0.0022673557978123426,
0.05708395689725876,
0.024426469579339027,
0.021976184099912643,
-0.018611200153827667,
-0.09426338225603104,
0.08728855848312378,
0.03301548585295677,
0.04809888079762459,
0.006895710248500109,
-0.04462607577443123,
-0.02522539719939232,
-0.0029076295904815197,
0.02569693699479103,
0.0674426481127739,
-0.01960691623389721,
-0.09107492864131927,
0.044152650982141495,
0.045771460980176926,
-0.040761224925518036,
0.02153940126299858,
0.045985735952854156,
-0.0552377812564373,
0.015230061486363411,
-0.11292938143014908,
0.04694458469748497,
0.08052284270524979,
-0.03071383200585842,
0.045779068022966385,
-0.0747465267777443,
0.00007401566108455881,
-0.026736313477158546,
-0.07466151565313339,
-0.01312381960451603,
9.215835302828425e-34,
0.07606757432222366,
0.035423289984464645,
0.03137166053056717,
0.05292292684316635,
0.04824751615524292,
-0.015562783926725388,
0.027237895876169205,
0.07374414801597595,
-0.038097262382507324,
-0.017912114039063454,
0.03578587621450424,
-0.07087202370166779,
0.001268698601052165,
0.07117308676242828,
0.023712076246738434,
-0.023807233199477196,
-0.08300132304430008,
0.009393575601279736,
-0.008840027265250683,
0.028242209926247597,
-0.0281362347304821,
-0.004245253745466471,
0.007875575684010983,
-0.062092408537864685,
-0.01169324479997158,
0.0584721677005291,
0.08394163846969604,
-0.09570839256048203,
-0.07619369775056839,
0.010850819759070873,
-0.12150222063064575,
-0.03558533266186714,
0.005820108577609062,
0.05727064982056618,
-0.01649629883468151,
-0.06510525196790695,
-0.00710661057382822,
-0.007430919911712408,
0.036961548030376434,
-0.02754271775484085,
-0.012261058203876019,
0.030025817453861237,
-0.00023869355209171772,
-0.014148060232400894,
-0.10757081210613251,
-0.04359688609838486,
-0.021835921332240105,
0.030814291909337044,
-0.07542004436254501,
0.020629605278372765,
0.07662078738212585,
0.06808370351791382,
-0.04284526780247688,
-0.05641717091202736,
-0.000413930305512622,
-0.009803008288145065,
0.014599885791540146,
0.026740651577711105,
0.023658616468310356,
0.007596238050609827,
0.033366382122039795,
-0.004493526183068752,
-0.033223509788513184,
0.11497639119625092,
0.05124959722161293,
0.03101428784430027,
0.006254854146391153,
0.039796534925699234,
0.03269512578845024,
-0.06941290944814682,
-0.04519660770893097,
0.014576046727597713,
-0.04857194423675537,
-0.03340984880924225,
0.030959490686655045,
-0.01667613349854946,
0.01631396822631359,
-0.1153118833899498,
0.027179144322872162,
0.08219704777002335,
0.038172610104084015,
-0.05428572744131088,
0.022918587550520897,
-0.03746620565652847,
-0.07046999037265778,
-0.09881505370140076,
0.027968373149633408,
-0.1414567083120346,
0.027047639712691307,
-0.03301135450601578,
-0.04339073598384857,
-0.04922545328736305,
-0.08923152834177017,
-0.027337370440363884,
-0.00238905125297606,
-4.157323575272663e-34,
0.05087760090827942,
-0.005024838726967573,
-0.087004154920578,
0.018395569175481796,
-0.021009886637330055,
0.026915697380900383,
0.045571401715278625,
0.03973684087395668,
-0.011854088865220547,
-0.07482573390007019,
-0.027185063809156418,
-0.004739005584269762,
0.040299106389284134,
-0.03491664677858353,
-0.009390493854880333,
-0.004033880308270454,
0.002100388053804636,
0.029048576951026917,
-0.047934405505657196,
0.03025856427848339,
0.019579041749238968,
0.06612440198659897,
-0.014995787292718887,
0.04169287160038948,
0.00775490328669548,
0.05563762038946152,
-0.031144309788942337,
0.06482598185539246,
-0.013462423346936703,
-0.037566889077425,
-0.013795064762234688,
0.02903163805603981,
0.04795943200588226,
-0.0860455334186554,
-0.11664267629384995,
0.10236810892820358,
0.05760473012924194,
0.029032627120614052,
-0.055056337267160416,
0.08466995507478714,
0.14861854910850525,
-0.04100902006030083,
-0.03990958631038666,
0.028220923617482185,
-0.046867307275533676,
-0.01430220901966095,
-0.11780928820371628,
-0.059500932693481445,
0.04393725097179413,
0.06460164487361908,
-0.05338576063513756,
-0.026674184948205948,
-0.08978518098592758,
-0.018114466220140457,
-0.0678575336933136,
-0.10642950981855392,
-0.019710075110197067,
-0.03529341146349907,
0.03767523914575577,
-0.03030386194586754,
0.0007429396500810981,
-0.006421871483325958,
0.027988458052277565,
-0.06972426176071167,
0.04483916983008385,
-0.0785209909081459,
-0.04432600736618042,
0.015362895093858242,
0.013245106674730778,
-0.013557958416640759,
0.05390378460288048,
-0.053559381514787674,
-0.023073507472872734,
-0.006215349305421114,
0.025391003116965294,
0.047910552471876144,
0.05499936640262604,
-0.04151274263858795,
-0.045638371258974075,
-0.07840416580438614,
0.008902350440621376,
-0.035689521580934525,
-0.01611883006989956,
0.030471323058009148,
0.10491487383842468,
0.08557707071304321,
0.018325673416256905,
0.03890696540474892,
-0.009440837427973747,
0.031581323593854904,
-0.029120255261659622,
0.08503790944814682,
-0.019670426845550537,
0.10737717151641846,
-0.029456419870257378,
-3.828178662956816e-8,
-0.09465736150741577,
0.026669681072235107,
-0.03820458799600601,
0.06118205189704895,
-0.04766331985592842,
0.0006471964297816157,
-0.054021768271923065,
0.08310496807098389,
0.008463079109787941,
-0.04583403095602989,
0.07093440741300583,
-0.022764522582292557,
-0.012384181842207909,
-0.06431964784860611,
0.0031484642531722784,
0.07829304039478302,
0.02497476153075695,
0.008046252653002739,
0.006479287054389715,
-0.02756410278379917,
0.0415106862783432,
0.07558328658342361,
-0.014255904592573643,
-0.049290940165519714,
-0.0038798535242676735,
-0.05327743664383888,
-0.014279446564614773,
0.0866679698228836,
0.05931411683559418,
-0.031396131962537766,
-0.0024632178246974945,
0.03591243922710419,
-0.010243323631584644,
-0.03609722852706909,
0.005321964621543884,
0.11989489942789078,
0.04854637384414673,
-0.04755651578307152,
-0.011123710311949253,
-0.029359295964241028,
0.04648619517683983,
0.04046082869172096,
-0.025998713448643684,
0.020547211170196533,
0.03032599948346615,
0.03917551040649414,
0.02170940302312374,
-0.09969163686037064,
0.039547014981508255,
-0.0677608996629715,
-0.000484775984659791,
-0.022995267063379288,
-0.02002820000052452,
0.02319461479783058,
0.06808681786060333,
0.048131488263607025,
0.05362639203667641,
0.012074170634150505,
0.021737772971391678,
0.07641127705574036,
0.040081936866045,
0.023432428017258644,
-0.04432082176208496,
0.037357185035943985
] | https://github.com/flagopen/flagembedding | 9,473 | 0 | 0 | 0 |
Diverse Preference Learning for Capabilities and Alignment | https://openreview.net/forum?id=pOq9vDIYev | [
"Stewart Slocum",
"Asher Parker-Sartori",
"Dylan Hadfield-Menell"
] | Poster | As LLMs increasingly impact society, their ability to represent diverse perspectives is critical. However, recent studies reveal that alignment algorithms such as RLHF and DPO significantly reduce the diversity of LLM outputs. Not only do aligned LLMs generate text with repetitive structure and word choice, they also approach problems in more uniform ways, and their responses reflect a narrower range of societal perspectives. We attribute this problem to the KL divergence regularizer employed in preference learning algorithms. This causes the model to overweight majority opinions and sacrifice diversity in exchange for optimal reward. To address this, we propose Soft Preference Learning, which decouples the entropy and cross-entropy terms in the KL penalty — allowing for fine-grained control over LLM generation diversity. From a capabilities perspective, LLMs trained using Soft Preference Learning attain higher accuracy on difficult repeated sampling tasks and produce outputs with greater semantic and lexical diversity. From an alignment perspective, they are capable of representing a wider range of societal viewpoints and display improved logit calibration. Notably, Soft Preference Learning resembles, but is a Pareto improvement over, standard temperature scaling. | alignment, diversity, natural language processing, reinforcement learning | We identify a cause of diversity loss from RLHF/DPO and propose a simple solution. | 13,501 | null | [
-0.03360642492771149,
-0.0958733782172203,
-0.019402451813220978,
-0.027962880209088326,
0.04514132812619209,
0.02638869546353817,
-0.0023823122028261423,
-0.06907378137111664,
0.08266276121139526,
0.025701330974698067,
-0.034423597157001495,
-0.09199264645576477,
0.05625636503100395,
0.044136375188827515,
0.02810288965702057,
0.03890763223171234,
0.061889324337244034,
0.059170112013816833,
-0.12345579266548157,
-0.019639339298009872,
0.016293412074446678,
0.02763512171804905,
0.03360699117183685,
0.02432088367640972,
-0.017061132937669754,
0.030316438525915146,
0.04295625537633896,
0.04638008400797844,
0.0016677196836099029,
-0.05296870693564415,
0.06647656112909317,
0.019299890846014023,
0.020199840888381004,
0.034104328602552414,
-0.04234936460852623,
0.0691082775592804,
-0.1544116884469986,
0.0010931978467851877,
-0.01838737539947033,
-0.002631271257996559,
-0.05992349237203598,
-0.016245566308498383,
-0.01606280356645584,
0.02960735186934471,
0.0711185559630394,
0.07305986434221268,
-0.06826196610927582,
-0.048418473452329636,
0.03337031602859497,
0.04369974508881569,
0.030062399804592133,
-0.014548401348292828,
0.0445922389626503,
0.012275590561330318,
-0.010511117987334728,
0.08105117827653885,
-0.05296991020441055,
0.0418766550719738,
-0.0031155627220869064,
-0.06191609054803848,
-0.04584858566522598,
-0.1289043128490448,
-0.03339145705103874,
-0.015178394503891468,
0.004404190927743912,
-0.058338411152362823,
-0.008562777191400528,
0.08506862074136734,
-0.05611925944685936,
-0.022558873519301414,
-0.03910208120942116,
0.06003229320049286,
0.0026001352816820145,
0.05968980863690376,
0.06360387802124023,
0.048058491200208664,
0.00797625444829464,
-0.04907653480768204,
0.04994983598589897,
-0.05043409392237663,
-0.012735399417579174,
0.0653076320886612,
0.022794215008616447,
0.03868716210126877,
0.11389689892530441,
-0.06935572624206543,
-0.003872469998896122,
0.04996655881404877,
0.06624943017959595,
0.02376045659184456,
-0.08926619589328766,
-0.037180397659540176,
0.11818796396255493,
-0.04199870303273201,
0.02745872363448143,
0.09641353785991669,
-0.009396051056683064,
-0.07425334304571152,
-0.0007001489866524935,
0.06326454877853394,
0.018302984535694122,
0.07338757812976837,
-0.02591634728014469,
-0.022699793800711632,
0.011927747167646885,
-0.10741455107927322,
-0.04813506081700325,
0.05798564478754997,
0.044823795557022095,
-0.11610011756420135,
0.016676098108291626,
0.0715375617146492,
-0.015215526334941387,
0.0676589235663414,
0.037015318870544434,
-0.07987485080957413,
0.0326833613216877,
-0.060549039393663406,
0.030878528952598572,
0.0699843317270279,
-0.08970008790493011,
-0.01593848504126072,
-0.06357212364673615,
0.026500800624489784,
0.04608474671840668,
0.01929570361971855,
0.010061816312372684,
3.7300972707172547e-34,
0.03593455255031586,
0.09214777499437332,
0.004016323946416378,
-0.011659079231321812,
0.0221612099558115,
0.013721784576773643,
-0.030575523152947426,
0.05926289036870003,
-0.1358601152896881,
0.004063583444803953,
-0.033780161291360855,
0.06296783685684204,
0.005293165799230337,
0.03363537788391113,
0.052616238594055176,
-0.02567432075738907,
-0.08923403918743134,
0.08211087435483932,
-0.041304104030132294,
-0.029328763484954834,
-0.011944545432925224,
0.12226122617721558,
0.06806273758411407,
-0.06101171299815178,
-0.005419724155217409,
-0.03529266640543938,
0.06255774199962616,
-0.0005961788119748235,
-0.0320303849875927,
-0.007050831336528063,
-0.008409660309553146,
-0.06492707133293152,
0.04523163288831711,
-0.02338274195790291,
-0.0034610142465680838,
-0.0023288049269467592,
-0.029783101752400398,
-0.012909004464745522,
-0.005943624768406153,
-0.036867108196020126,
-0.03601435571908951,
0.033881958574056625,
0.010479921475052834,
0.0016344661125913262,
-0.06639891862869263,
0.055876120924949646,
0.03799447417259216,
0.03037344291806221,
-0.03587424382567406,
-0.017034148797392845,
0.03566465899348259,
0.012323349714279175,
0.041616447269916534,
-0.019474614411592484,
0.005485397297888994,
0.0071730343624949455,
0.00002002983819693327,
0.030632900074124336,
0.02062978222966194,
-0.06257305294275284,
-0.050644319504499435,
0.048645876348018646,
0.0225668977946043,
0.04641060531139374,
0.10197464376688004,
0.034340519458055496,
0.02550889551639557,
0.015613380819559097,
0.04379839077591896,
-0.015786847099661827,
0.052355553954839706,
0.020538058131933212,
-0.06031245365738869,
-0.08421755582094193,
0.005359995644539595,
0.03375744819641113,
0.10129237920045853,
-0.042274683713912964,
0.030733412131667137,
-0.024727800861001015,
-0.01735590025782585,
-0.0038649479392915964,
0.018740439787507057,
-0.07787448912858963,
-0.06708397716283798,
-0.08253301680088043,
0.07167496532201767,
-0.08833618462085724,
0.028469225391745567,
0.028376508504152298,
-0.08383239060640335,
-0.06281454116106033,
-0.021533817052841187,
0.011649467051029205,
-0.01207469031214714,
2.6153720748238928e-34,
-0.07369627058506012,
-0.09784739464521408,
0.024343067780137062,
0.08309672027826309,
0.004916164558380842,
0.0001970785961020738,
-0.01157132163643837,
-0.02082483284175396,
-0.04793684929609299,
-0.03513101115822792,
-0.030087728053331375,
-0.008837863802909851,
0.06373720616102219,
-0.03785872831940651,
0.023088980466127396,
0.05863789841532707,
-0.05658337101340294,
0.0972348153591156,
-0.00787911843508482,
0.013538645580410957,
0.053623877465724945,
0.12453275918960571,
-0.05528813973069191,
0.10883833467960358,
0.021755483001470566,
0.03854856267571449,
-0.006465852726250887,
0.07211671024560928,
0.009684580378234386,
-0.024357162415981293,
-0.031487032771110535,
-0.05361004173755646,
-0.063242606818676,
-0.043940089643001556,
-0.024306440725922585,
0.008323241956532001,
-0.019015097990632057,
0.023182449862360954,
-0.04068189114332199,
0.10501058399677277,
-0.08040794730186462,
-0.031230762600898743,
-0.0781896561384201,
0.0015651046996936202,
-0.052911270409822464,
-0.03020578809082508,
-0.08207571506500244,
-0.055371806025505066,
0.08262920379638672,
-0.00007708086195634678,
0.016001496464014053,
0.0009271507151424885,
-0.03537891432642937,
0.0925799012184143,
-0.06529378145933151,
-0.07940772920846939,
-0.06119395047426224,
-0.013429172337055206,
0.020780019462108612,
0.0063495640642941,
-0.03506024554371834,
-0.0438385084271431,
0.03789820894598961,
-0.005294483155012131,
0.01785610429942608,
-0.08065538853406906,
0.01964034140110016,
-0.00344221293926239,
-0.06651105731725693,
-0.02467614971101284,
0.06641162931919098,
-0.08354926109313965,
0.02079368568956852,
-0.003871631110087037,
0.031347595155239105,
-0.044256631284952164,
0.03718778118491173,
-0.010849659331142902,
-0.05930013954639435,
-0.012804646976292133,
0.05124673619866371,
0.03672298043966293,
0.0457916222512722,
0.07976342737674713,
0.04849909245967865,
0.020963644608855247,
0.04773698002099991,
0.07197784632444382,
0.048126742243766785,
-0.0029301282484084368,
0.02564682438969612,
-0.0038291295059025288,
0.04761573672294617,
0.03693336248397827,
-0.06749725341796875,
-4.1590343613506775e-8,
-0.054615553468465805,
-0.033941563218832016,
-0.0069849141873419285,
0.12779097259044647,
-0.007675850763916969,
-0.02004426345229149,
-0.03512314334511757,
0.024765754118561745,
-0.05116530880331993,
-0.02188704162836075,
0.09618581831455231,
0.043442919850349426,
-0.05896565318107605,
-0.03458182141184807,
0.050757165998220444,
0.07058566063642502,
-0.026470331475138664,
0.0007323840982280672,
0.02178781107068062,
0.005869749933481216,
0.038995642215013504,
0.06719523668289185,
0.02910769172012806,
-0.05106177181005478,
0.043596308678388596,
-0.007650059647858143,
-0.11565523594617844,
0.058997977524995804,
0.04375297203660011,
-0.010981406085193157,
0.02995656244456768,
-0.019677413627505302,
-0.043895021080970764,
-0.007153811398893595,
-0.02539660781621933,
0.0818227156996727,
-0.031852059066295624,
0.022969381883740425,
-0.002312732394784689,
0.008230162784457207,
0.005330732092261314,
0.02873239479959011,
-0.020719802007079124,
0.06476268172264099,
-0.08003592491149902,
-0.060780711472034454,
-0.04570275545120239,
-0.06182226538658142,
0.08340448141098022,
-0.06553347408771515,
0.029427288100123405,
-0.04999976232647896,
0.027943503111600876,
-0.007862191647291183,
0.06347966939210892,
-0.033810701221227646,
-0.012184763327240944,
-0.05887073650956154,
0.06734877079725266,
0.061846569180488586,
0.08554240316152573,
-0.01609465666115284,
-0.08338765054941177,
-0.01573423109948635
] | 0 | 0 | 0 | 0 |
|
Efficient Reinforcement Learning with Large Language Model Priors | https://openreview.net/forum?id=e2NRNQ0sZe | [
"Xue Yan",
"Yan Song",
"Xidong Feng",
"Mengyue Yang",
"Haifeng Zhang",
"Haitham Bou Ammar",
"Jun Wang"
] | Poster | In sequential decision-making (SDM) tasks, methods like reinforcement learning (RL) and heuristic search have made notable advances in specific cases. However, they often require extensive exploration and face challenges in generalizing across diverse environments due to their limited grasp of the underlying decision dynamics. In contrast, large language models (LLMs) have recently emerged as powerful general-purpose tools, due to their capacity to maintain vast amounts of domain-specific knowledge. To harness this rich prior knowledge for efficiently solving complex SDM tasks, we propose treating LLMs as prior action distributions and integrating them into RL frameworks through Bayesian inference methods, making use of variational inference and direct posterior sampling. The proposed approaches facilitate the seamless incorporation of fixed LLM priors into both policy-based and value-based RL frameworks. Our experiments show that incorporating LLM-based action priors significantly reduces exploration and optimization complexity, substantially improving sample efficiency compared to traditional RL techniques, e.g., using LLM priors decreases the number of required samples by over 90\% in offline learning scenarios. | Reinforcement Learning; Probabilistic Inference; Language Prior; | To harness LLMs' rich prior knowledge for efficiently solving complex SDM tasks, we propose treating LLMs as prior action distributions and integrating them into RL frameworks through Bayesian inference methods. | 13,493 | 2410.07927 | [
-0.022834274917840958,
-0.07238935679197311,
0.07547641545534134,
0.0044219885021448135,
0.05242225527763367,
0.05927948281168938,
-0.009222512133419514,
-0.014932435937225819,
0.029935631901025772,
0.019262654706835747,
-0.04234293848276138,
-0.02006220631301403,
0.005743733607232571,
0.054476048797369,
0.04300614073872566,
0.018862145021557808,
0.1049535870552063,
0.04137876629829407,
-0.05500585213303566,
-0.05567304417490959,
0.01951962150633335,
0.051084548234939575,
0.0882035419344902,
-0.05308026075363159,
0.03074515610933304,
-0.004770528059452772,
0.0522456057369709,
-0.01820666901767254,
0.03319668769836426,
0.013332411646842957,
0.07857096940279007,
0.022482747212052345,
0.02296706847846508,
-0.07461140304803848,
-0.044022638350725174,
0.044970132410526276,
-0.1514468938112259,
-0.03455163910984993,
-0.03531946614384651,
-0.014433557167649269,
-0.039280105382204056,
-0.0032946839928627014,
0.056913502514362335,
0.03793211281299591,
0.05043250322341919,
-0.032819002866744995,
0.014215027913451195,
-0.017136387526988983,
0.050574515014886856,
0.020359447225928307,
-0.006025987211614847,
-0.06693211942911148,
0.01610635407269001,
-0.01247958280146122,
-0.008569027297198772,
0.0026048854924738407,
0.00725693441927433,
-0.04947000369429588,
0.03469518944621086,
-0.06727253645658493,
-0.03845462203025818,
-0.17969346046447754,
-0.00960853137075901,
-0.02646760083734989,
-0.07858247309923172,
-0.004364730324596167,
0.0418868362903595,
0.06710653007030487,
0.05236783251166344,
0.018214261159300804,
-0.026488568633794785,
0.07531243562698364,
0.007117489818483591,
0.04627615958452225,
0.008230394683778286,
-0.021681679412722588,
-0.0024774072226136923,
-0.0546601302921772,
0.04495083540678024,
-0.03664792701601982,
-0.04253372922539711,
-0.02595292590558529,
-0.042192503809928894,
0.05256364122033119,
0.14001017808914185,
0.030732132494449615,
-0.0033534185495227575,
0.05652843415737152,
0.09122025221586227,
0.030835481360554695,
-0.019318711012601852,
-0.0777728334069252,
-0.01424160785973072,
-0.0432548001408577,
-0.06778102368116379,
0.02096404880285263,
-0.01757417432963848,
-0.055453646928071976,
-0.016283903270959854,
0.0338166244328022,
0.07708585262298584,
0.11748690903186798,
-0.010946209542453289,
-0.028183002024888992,
-0.0397726371884346,
-0.07204661518335342,
-0.06207950413227081,
0.078276127576828,
0.02392646297812462,
-0.07244615256786346,
-0.010624926537275314,
0.004548129625618458,
0.08981754630804062,
0.07589039951562881,
-0.018717274069786072,
-0.016756894066929817,
0.1231645792722702,
-0.0917695164680481,
-0.0794583335518837,
0.05733267590403557,
-0.006793531123548746,
0.004455745220184326,
0.015536709688603878,
-0.003961322363466024,
-0.014726650901138783,
-0.033577438443899155,
-0.029901612550020218,
3.096283344134599e-33,
0.09962736070156097,
-0.012700676918029785,
0.051627710461616516,
0.08272077888250351,
0.0235537588596344,
0.013634340837597847,
0.07366512715816498,
0.04325529560446739,
-0.029975123703479767,
0.016353659331798553,
-0.01090039499104023,
0.013952495530247688,
-0.05278770998120308,
0.020901722833514214,
0.0066655888222157955,
-0.03467365726828575,
-0.05515101179480553,
0.0617876797914505,
-0.019213849678635597,
-0.01596676930785179,
0.05762740224599838,
-0.0068071940913796425,
0.013974240981042385,
-0.07212810218334198,
0.04627905413508415,
0.027841415256261826,
0.07984448969364166,
-0.0325147844851017,
-0.003644524607807398,
-0.0184578038752079,
-0.09653878957033157,
-0.04384815692901611,
-0.024858908727765083,
0.08339264988899231,
0.026549264788627625,
0.025484096258878708,
-0.0527014397084713,
-0.030068442225456238,
-0.006415151059627533,
-0.010307862423360348,
-0.02962234802544117,
0.009714586660265923,
0.041778985410928726,
0.0396152064204216,
-0.10421741753816605,
-0.1219964325428009,
-0.009009134024381638,
-0.03262101113796234,
0.01024482399225235,
-0.021069839596748352,
0.07287666201591492,
0.010443049483001232,
0.032461609691381454,
-0.05584269389510155,
0.01916540041565895,
0.04102278873324394,
-0.03549902141094208,
0.05950602889060974,
-0.030399808660149574,
0.07425161451101303,
-0.004559818655252457,
0.03532341495156288,
-0.03147271275520325,
0.116862952709198,
0.054648544639348984,
0.07122556865215302,
-0.04329710081219673,
0.058674152940511703,
0.054041944444179535,
-0.05307639762759209,
0.05370920151472092,
-0.014277220703661442,
0.05490553379058838,
-0.07929591834545135,
0.06685639172792435,
0.01584383100271225,
0.032535210251808167,
-0.08030348271131516,
0.01853444240987301,
-0.013046124950051308,
-0.0046119652688503265,
-0.051648955792188644,
-0.03592487797141075,
0.014416970312595367,
0.06365019828081131,
-0.0270674005150795,
0.03765125572681427,
-0.12574468553066254,
0.009536261670291424,
-0.0760173574090004,
-0.0841154083609581,
-0.01307897549122572,
-0.047807011753320694,
-0.0048106079921126366,
0.06061803922057152,
-3.0189325097726308e-33,
-0.025565529242157936,
-0.08267244696617126,
0.049159687012434006,
0.01779031939804554,
-0.005795828532427549,
-0.0733155757188797,
-0.011046744883060455,
-0.026832588016986847,
0.03479114547371864,
-0.12097334861755371,
-0.09476491808891296,
0.01825365237891674,
0.027297653257846832,
-0.010548233985900879,
0.01450473815202713,
0.0020684897899627686,
0.01359705999493599,
0.03885367140173912,
0.00940597616136074,
0.026032980531454086,
-0.06431402266025543,
0.06708911806344986,
-0.08847744017839432,
0.034952562302351,
-0.0549926720559597,
0.024454060941934586,
-0.027807509526610374,
0.025136664509773254,
0.0024446432944387197,
0.04004902392625809,
-0.03438396751880646,
-0.04177005961537361,
-0.024829771369695663,
-0.020970407873392105,
-0.08181576430797577,
0.06465674936771393,
0.10017291456460953,
0.05167186260223389,
-0.06458287686109543,
0.11752714961767197,
0.04357165843248367,
-0.07135623693466187,
-0.053002290427684784,
-0.0063252197578549385,
-0.0008670652168802917,
0.006965750828385353,
-0.06651575118303299,
0.005135133862495422,
-0.024613870307803154,
-0.024736933410167694,
-0.04044026881456375,
0.055493731051683426,
-0.0655551329255104,
0.0198829248547554,
-0.03712858632206917,
-0.08102308958768845,
-0.04053603857755661,
-0.039187848567962646,
0.07648233324289322,
-0.002848984906449914,
-0.018059037625789642,
0.07309018820524216,
0.0334697961807251,
-0.03111199103295803,
-0.023065486922860146,
0.00029338637250475585,
-0.00305158831179142,
0.011765318922698498,
-0.030038179829716682,
-0.054850101470947266,
0.010894215665757656,
-0.04493572190403938,
0.04116882011294365,
0.00031338256667368114,
0.0990760326385498,
-0.05342448130249977,
0.016399160027503967,
-0.029609227553009987,
0.0036745532415807247,
-0.018399199470877647,
0.07260830700397491,
0.008571768179535866,
0.0034786902833729982,
0.04430333524942398,
0.02642485313117504,
0.07681126892566681,
-0.02374759130179882,
0.06664571166038513,
0.0409059152007103,
-0.029908927157521248,
-0.030077960342168808,
0.022513795644044876,
0.055979374796152115,
0.0521652027964592,
-0.07328213006258011,
-4.275783282992052e-8,
-0.05347069725394249,
0.05194525048136711,
0.050968460738658905,
0.04218646138906479,
0.08103270083665848,
0.006073904223740101,
-0.09516580402851105,
0.06956622749567032,
0.027092790231108665,
-0.05401281639933586,
0.0537053681910038,
0.026216940954327583,
0.04046833887696266,
-0.11352024972438812,
-0.04371718317270279,
0.06796298176050186,
0.06822721660137177,
-0.05231931433081627,
-0.014306609518826008,
-0.006719654425978661,
0.05014486610889435,
0.01217991765588522,
0.03458857163786888,
-0.04400922358036041,
0.055928438901901245,
-0.06439829617738724,
-0.009341804310679436,
0.09222890436649323,
0.09135908633470535,
-0.010433662682771683,
-0.003661159425973892,
-0.029195310547947884,
-0.05479743331670761,
0.0520966462790966,
0.05072365701198578,
-0.00597597798332572,
0.0291125550866127,
-0.094149649143219,
0.010604893788695335,
-0.06886625289916992,
-0.048328954726457596,
0.04846322908997536,
-0.022307317703962326,
0.061018966138362885,
-0.024055395275354385,
0.04876663163304329,
-0.09907231479883194,
-0.11315437406301498,
0.07076053321361542,
-0.034624338150024414,
-0.007672314532101154,
0.000922693288885057,
-0.013970168307423592,
0.06327986717224121,
0.04027107357978821,
0.03965083137154579,
0.050757184624671936,
-0.03034214675426483,
0.01962055079638958,
0.0100473091006279,
0.030118878930807114,
-0.028065934777259827,
-0.08196529000997543,
0.04824985936284065
] | 0 | 0 | 0 | 0 |
|
Geometry-Aware Approaches for Balancing Performance and Theoretical Guarantees in Linear Bandits | https://openreview.net/forum?id=Oeb0I3JcVc | [
"Yuwei Luo",
"Mohsen Bayati"
] | Poster | This paper is motivated by recent research in the $d$-dimensional stochastic linear bandit literature, which has revealed an unsettling discrepancy: algorithms like Thompson sampling and Greedy demonstrate promising empirical performance, yet this contrasts with their pessimistic theoretical regret bounds. The challenge arises from the fact that while these algorithms may perform poorly in certain problem instances, they generally excel in typical instances. To address this, we propose a new data-driven technique that tracks the geometric properties of the uncertainty ellipsoid around the main problem parameter. This methodology enables us to formulate a data-driven frequentist regret bound, which incorporates the geometric information, for a broad class of base algorithms, including Greedy, OFUL, and Thompson sampling. This result allows us to identify and ``course-correct" problem instances in which the base algorithms perform poorly. The course-corrected algorithms achieve the minimax optimal regret of order $\tilde{\mathcal{O}}(d\sqrt{T})$ for a $T$-period decision-making scenario, effectively maintaining the desirable attributes of the base algorithms, including their empirical efficacy. We present simulation results to validate our findings using synthetic and real data. | Linear bandit, Thompson sampling, Greedy, Data-driven exploration | null | 13,487 | 2306.14872 | [
-0.0889001339673996,
-0.00671895919367671,
-0.02146807126700878,
-0.021011684089899063,
0.02316664531826973,
-0.03643560782074928,
0.0019381634192541242,
0.06065848842263222,
0.030539458617568016,
0.008280560374259949,
-0.0829850286245346,
0.027037091553211212,
0.11819002032279968,
-0.038540564477443695,
0.00879447627812624,
0.04673328623175621,
0.05496245250105858,
0.05570047348737717,
0.019298896193504333,
0.0306222103536129,
-0.006240051239728928,
0.020360251888632774,
0.03150833770632744,
-0.03939048945903778,
0.03767821937799454,
-0.01766457036137581,
0.028501495718955994,
-0.012115786783397198,
-0.005158859770745039,
0.0159969013184309,
-0.012350521981716156,
0.035402581095695496,
0.05158858373761177,
-0.05755779892206192,
-0.040430981665849686,
-0.010649889707565308,
-0.12511584162712097,
-0.015534719452261925,
0.04397880658507347,
0.024712782353162766,
-0.003054766682907939,
0.052301857620477676,
0.050849027931690216,
0.07436983287334442,
0.015023397281765938,
0.012623186223208904,
-0.01912085898220539,
-0.019593168050050735,
0.03423528000712395,
-0.02299530990421772,
-0.059796202927827835,
-0.06813167780637741,
-0.04127020761370659,
-0.059362996369600296,
-0.0769999697804451,
0.01447866577655077,
0.04639082029461861,
0.04218040034174919,
0.038792990148067474,
0.006552761886268854,
0.052939217537641525,
-0.156281977891922,
-0.015208136290311813,
-0.015525656752288342,
-0.05246651545166969,
-0.07356396317481995,
0.013942885212600231,
0.04285001754760742,
-0.02528197504580021,
0.08590773493051529,
0.015828272327780724,
0.12364042550325394,
-0.11303845047950745,
0.08670662343502045,
0.07079380750656128,
0.018788306042551994,
0.010287932120263577,
-0.0777779147028923,
-0.024880867451429367,
0.015832016244530678,
-0.058823708444833755,
0.052762094885110855,
0.04136879742145538,
-0.009472194127738476,
0.09480082243680954,
-0.004571036901324987,
0.0642259418964386,
0.0197428110986948,
0.14378872513771057,
-0.07478643208742142,
0.008936364203691483,
-0.01963449828326702,
-0.09329797327518463,
0.012206028215587139,
-0.05291247367858887,
0.04051537439227104,
-0.021593745797872543,
-0.014103367924690247,
-0.015640534460544586,
0.0771985650062561,
0.06396930664777756,
0.05562422797083855,
0.006984202656894922,
0.0008071624906733632,
0.018854454159736633,
-0.04759242385625839,
0.05264345929026604,
0.056946080178022385,
0.03264211118221283,
-0.1301221400499344,
0.03445986285805702,
0.004208235070109367,
0.05199218541383743,
0.043738268315792084,
-0.0401567667722702,
-0.02587629295885563,
-0.023984236642718315,
-0.0019835487473756075,
-0.013331633992493153,
0.0941278412938118,
0.014729737304151058,
0.04266877472400665,
0.07810544967651367,
0.052855685353279114,
0.006689997389912605,
0.02915661409497261,
-0.024310914799571037,
9.202715683847294e-34,
0.04788542538881302,
0.010882607661187649,
0.039888348430395126,
-0.05865032598376274,
0.00488025089725852,
0.0011900709941983223,
0.03150147199630737,
-0.03945024311542511,
-0.057270631194114685,
0.06863575428724289,
-0.0075643425807356834,
-0.015601062215864658,
-0.02454274520277977,
0.02226770855486393,
0.15826143324375153,
0.0024874184746295214,
-0.024599280208349228,
0.029953472316265106,
0.02735188603401184,
-0.0975167527794838,
0.02885550633072853,
-0.05244637280702591,
-0.020244114100933075,
-0.04298494756221771,
-0.0038926952984184027,
-0.009444641880691051,
-0.06478802114725113,
0.05919625610113144,
0.008922964334487915,
0.018129657953977585,
-0.05294572189450264,
0.04708842933177948,
-0.06848649680614471,
0.049090426415205,
0.027017947286367416,
0.0007065193494781852,
-0.048246145248413086,
-0.000565787311643362,
-0.06672608852386475,
0.0029535626526921988,
-0.044010795652866364,
0.05308009684085846,
0.06451308727264404,
0.00564196240156889,
-0.0973140299320221,
-0.12421490997076035,
-0.03239312395453453,
0.008615512400865555,
0.006045072805136442,
-0.04724198579788208,
-0.010417520999908447,
0.005922166630625725,
0.03335804492235184,
0.014627067372202873,
-0.03193292394280434,
-0.010158059187233448,
-0.03084232658147812,
0.007831037975847721,
-0.032635822892189026,
0.040858712047338486,
0.03430655226111412,
-0.057874396443367004,
0.005086975172162056,
0.018879055976867676,
-0.07542506605386734,
0.11331553757190704,
0.036796532571315765,
-0.024396702647209167,
0.015883758664131165,
0.03163269907236099,
0.0006872343365103006,
-0.047163158655166626,
-0.0036147329956293106,
-0.06093365699052811,
0.059960588812828064,
-0.04635455086827278,
0.09482306241989136,
0.00481055723503232,
0.022856349125504494,
-0.07406941056251526,
0.023990385234355927,
-0.04903576895594597,
-0.006755767855793238,
-0.07940452545881271,
-0.052224256098270416,
-0.04270023852586746,
0.055171381682157516,
-0.11902571469545364,
-0.022647112607955933,
0.027340585365891457,
-0.09023338556289673,
-0.01087268814444542,
-0.09113027900457382,
0.034192778170108795,
0.0211194921284914,
-1.1609667100140626e-33,
-0.02904057316482067,
-0.03493231534957886,
0.0875299945473671,
0.024412978440523148,
0.035189565271139145,
-0.049031492322683334,
-0.05680931732058525,
-0.07672148942947388,
0.015193259343504906,
-0.08575498312711716,
-0.11372129619121552,
-0.002853594720363617,
0.004294766578823328,
-0.019350100308656693,
0.018490014597773552,
-0.010469973087310791,
0.01739911176264286,
0.004359276033937931,
-0.042842257767915726,
-0.01967422291636467,
-0.0059303101152181625,
0.011620155535638332,
-0.07062941789627075,
-0.017851728945970535,
0.015412348322570324,
0.014210950583219528,
-0.03385855630040169,
0.027460265904664993,
-0.04838915914297104,
-0.018436742946505547,
-0.0390060693025589,
-0.09713765978813171,
-0.023812230676412582,
0.022335028275847435,
-0.05718890205025673,
0.04467107728123665,
0.05386979505419731,
0.02887144684791565,
-0.03037298284471035,
0.12125509232282639,
0.014463581144809723,
-0.029932286590337753,
0.021356983110308647,
-0.036143507808446884,
0.037195563316345215,
0.02924647182226181,
-0.005373439285904169,
-0.000985089223831892,
0.03775361180305481,
-0.04746943712234497,
-0.00360268191434443,
0.1420411318540573,
-0.035089507699012756,
0.12415211647748947,
-0.020526865497231483,
-0.03571111708879471,
-0.05719463899731636,
0.028093893080949783,
0.06484640389680862,
0.0025433478876948357,
-0.08117465674877167,
0.0228738933801651,
-0.1040351390838623,
0.05297698825597763,
-0.005346819758415222,
0.01929161697626114,
-0.07910753786563873,
0.015437138266861439,
0.0000286241320281988,
0.05330231413245201,
-0.07768578827381134,
0.016149984672665596,
-0.034533943980932236,
-0.02970898523926735,
-0.04410089924931526,
0.03752902150154114,
0.042365506291389465,
0.05835844203829765,
-0.012666006572544575,
0.041789304465055466,
0.05539001524448395,
0.03932710364460945,
0.10438906401395798,
0.03264356777071953,
0.031789883971214294,
0.0883340984582901,
0.054625481367111206,
0.023837454617023468,
-0.019774988293647766,
-0.03143465146422386,
0.036742132157087326,
-0.053008899092674255,
0.05327044054865837,
0.025179322808980942,
0.017208915203809738,
-4.718459933883423e-8,
-0.0775582566857338,
0.01141071505844593,
0.04439608380198479,
0.029241185635328293,
0.12050507962703705,
0.04198700934648514,
-0.021064195781946182,
0.07029827684164047,
-0.13319659233093262,
-0.0005573987145908177,
0.015473373234272003,
0.0769558921456337,
-0.04071228206157684,
-0.03284949064254761,
-0.011158359237015247,
0.05585623160004616,
-0.011850968934595585,
-0.057955384254455566,
-0.022325022146105766,
0.030737238004803658,
-0.022888308390975,
0.02487589232623577,
0.046172406524419785,
-0.11699187010526657,
0.00047276256373152137,
-0.054343730211257935,
-0.0023571543861180544,
0.09947853535413742,
0.09564696252346039,
0.10759685188531876,
-0.01473882794380188,
-0.048644740134477615,
0.01568298600614071,
0.002799004316329956,
0.06883655488491058,
0.002618700498715043,
-0.00781218009069562,
-0.048086974769830704,
-0.011173532344400883,
-0.03495999425649643,
0.023168331012129784,
0.0616057924926281,
-0.00417905580252409,
0.02801191620528698,
0.09691978245973587,
0.06071385741233826,
-0.011472375132143497,
-0.01773347705602646,
0.016653845086693764,
-0.06610769778490067,
0.025554979220032692,
-0.031621746718883514,
-0.03193153813481331,
0.08660551905632019,
0.07501208782196045,
0.02928963117301464,
-0.020878151059150696,
-0.06875031441450119,
-0.057127490639686584,
0.027325687929987907,
0.004679033532738686,
-0.011980577372014523,
-0.0396956242620945,
-0.005472939927130938
] | 0 | 0 | 0 | 0 |
|
Langevin Soft Actor-Critic: Efficient Exploration through Uncertainty-Driven Critic Learning | https://openreview.net/forum?id=FvQsk3la17 | [
"Haque Ishfaq",
"Guangyuan Wang",
"Sami Nur Islam",
"Doina Precup"
] | Poster | Existing actor-critic algorithms, which are popular for continuous control reinforcement learning (RL) tasks, suffer from poor sample efficiency due to lack of principled exploration mechanism within them. Motivated by the success of Thompson sampling for efficient exploration in RL, we propose a novel model-free RL algorithm, \emph{Langevin Soft Actor Critic} (LSAC), which prioritizes enhancing critic learning through uncertainty estimation over policy optimization. LSAC employs three key innovations: approximate Thompson sampling through distributional Langevin Monte Carlo (LMC) based $Q$ updates, parallel tempering for exploring multiple modes of the posterior of the $Q$ function, and diffusion synthesized state-action samples regularized with $Q$ action gradients. Our extensive experiments demonstrate that LSAC outperforms or matches the performance of mainstream model-free RL algorithms for continuous control tasks.
Notably, LSAC marks the first successful application of an LMC based Thompson sampling in continuous control tasks with continuous action spaces. | Actor-Critic, Exploration, Reinforcement Learning, Thompson Sampling, Langevin Monte Carlo, Deep Reinforcement learning, Continuous Control | We propose a principled exploration approach for actor-critic algorithms through critic learning using approximate Thompson sampling | 13,486 | null | [
-0.07571890950202942,
-0.07110023498535156,
0.045012593269348145,
0.07097061723470688,
-0.005605100654065609,
0.040089838206768036,
-0.008929907344281673,
-0.00874107051640749,
0.061469074338674545,
0.07417973130941391,
-0.062419161200523376,
-0.02375079318881035,
0.05214483663439751,
0.05435078963637352,
-0.03137936815619469,
0.03489656373858452,
0.05872591957449913,
0.06226034462451935,
-0.03299548476934433,
-0.01990547776222229,
-0.0020998737309128046,
0.029040968045592308,
0.09020738303661346,
-0.028355151414871216,
0.02556511014699936,
-0.0031139289494603872,
0.029237965121865273,
-0.04449489712715149,
0.003484138287603855,
-0.013834046199917793,
0.032876402139663696,
-0.067493736743927,
-0.02774258516728878,
-0.12932716310024261,
-0.029209593310952187,
0.04977376013994217,
-0.16980589926242828,
0.011291363276541233,
0.035071324557065964,
-0.007399475667625666,
-0.0218436811119318,
0.03455764427781105,
0.042302582412958145,
-0.0029295608401298523,
0.1222623884677887,
0.0006216388428583741,
-0.03110543079674244,
-0.03199262544512749,
0.08714642375707626,
0.024928033351898193,
0.054903607815504074,
-0.04325111210346222,
0.018113620579242706,
-0.08569684624671936,
0.011146844364702702,
0.011180136352777481,
-0.013213087804615498,
0.01762296073138714,
0.0527590773999691,
0.009752627462148666,
-0.014856860041618347,
-0.11934196949005127,
-0.0014740368351340294,
-0.026591463014483452,
-0.06409952789545059,
-0.011525520123541355,
-0.02390240877866745,
0.07119787484407425,
0.07208853960037231,
0.0027626713272184134,
-0.025200920179486275,
0.07505786418914795,
-0.007379383780062199,
-0.007694610394537449,
0.05291800945997238,
-0.07257118076086044,
-0.0015280063962563872,
-0.01799321547150612,
-0.008128520101308823,
-0.008241115137934685,
0.04583831503987312,
-0.030810723081231117,
0.0075883022509515285,
0.010121311992406845,
-0.00016683996364008635,
0.013568183407187462,
0.01939203217625618,
0.051312364637851715,
0.10909043997526169,
-0.02915702946484089,
0.041486915200948715,
0.04198193922638893,
-0.08639276027679443,
-0.06810950487852097,
-0.02657388523221016,
-0.02708582393825054,
0.0026448899880051613,
-0.0343608520925045,
-0.02622220851480961,
0.02923424169421196,
0.0785185769200325,
0.08324509859085083,
-0.0009083379409275949,
0.02082609012722969,
0.0031898815650492907,
-0.13784955441951752,
0.052374765276908875,
0.017841583117842674,
0.03253218159079552,
-0.06544061750173569,
-0.00624033622443676,
-0.02629139833152294,
0.04185684770345688,
0.04298661649227142,
0.007887844927608967,
-0.004452035762369633,
0.021032409742474556,
-0.017624957486987114,
-0.03628615289926529,
0.10101816058158875,
-0.044590726494789124,
-0.09910868853330612,
-0.034673549234867096,
0.010057896375656128,
0.016428308561444283,
-0.007788459304720163,
0.022807471454143524,
2.3650634097857487e-33,
0.04961054027080536,
-0.02051914855837822,
0.06499793380498886,
0.04875391721725464,
0.03019718825817108,
-0.0047924332320690155,
0.06851665675640106,
-0.10225677490234375,
-0.09079228341579437,
-0.04270782694220543,
0.049789588898420334,
0.0026127530727535486,
-0.12084098160266876,
0.0467163510620594,
0.03924407809972763,
0.01072542741894722,
-0.08327533304691315,
0.018625814467668533,
0.007293147034943104,
-0.04965680465102196,
0.028737377375364304,
-0.003630491206422448,
-0.026632126420736313,
-0.041949279606342316,
0.05953162536025047,
0.024177532643079758,
0.047530580312013626,
0.07618440687656403,
-0.04149767383933067,
0.012879865244030952,
-0.08138780295848846,
0.012833599001169205,
-0.07570967078208923,
0.04099123552441597,
0.036573030054569244,
0.052710894495248795,
-0.058771684765815735,
-0.0011337475152686238,
-0.010609968565404415,
-0.054318707436323166,
0.03078269399702549,
0.029745781794190407,
0.040568508207798004,
0.02080831117928028,
-0.12350164353847504,
-0.07781602442264557,
-0.03063562698662281,
-0.030556702986359596,
-0.027995718643069267,
0.0083203399553895,
0.004517086781561375,
0.03866402432322502,
0.026356050744652748,
-0.04417130723595619,
-0.007462628185749054,
-0.013798888772726059,
0.03842858597636223,
-0.001103438320569694,
-0.07460494339466095,
0.03839222341775894,
0.060574524104595184,
0.02842177078127861,
-0.036363594233989716,
0.0739113911986351,
0.04803727939724922,
0.13272495567798615,
-0.10347536206245422,
0.09667357057332993,
0.005261982791125774,
0.003506444161757827,
0.006873652804642916,
-0.015192613936960697,
0.04398985952138901,
-0.12377206236124039,
0.10622633993625641,
-0.06233547627925873,
0.024332422763109207,
-0.004223801661282778,
-0.04411404952406883,
-0.028005359694361687,
-0.03365176171064377,
-0.06433895975351334,
-0.09078380465507507,
0.020629722625017166,
0.01294085942208767,
-0.07657317072153091,
-0.013906496576964855,
-0.07599234580993652,
0.01957915909588337,
-0.005398663692176342,
-0.07070524990558624,
-0.043022193014621735,
0.013393583707511425,
0.08117084950208664,
-0.08261709660291672,
-2.7751087164601872e-33,
0.015683777630329132,
-0.052537109702825546,
0.06885264813899994,
0.05312957242131233,
-0.04859067499637604,
-0.05174184590578079,
-0.04276754707098007,
-0.07233573496341705,
0.07765522599220276,
-0.07142119109630585,
-0.11091141402721405,
0.024348927661776543,
-0.03241496905684471,
0.022445008158683777,
0.019791634753346443,
-0.0027225289959460497,
0.02132568322122097,
0.021863119676709175,
0.007703116629272699,
0.01698673702776432,
-0.004020723979920149,
0.012479826807975769,
0.0026219673454761505,
0.006721036043018103,
-0.012041845358908176,
0.034620534628629684,
-0.0003747393493540585,
0.11768588423728943,
-0.005001664627343416,
-0.00992109626531601,
-0.06106460839509964,
-0.07764720916748047,
-0.00647216010838747,
-0.004107102286070585,
-0.10193633288145065,
0.019637515768408775,
0.11303803324699402,
0.10426453500986099,
0.0050214920192956924,
0.0733317881822586,
0.017929477617144585,
-0.0767478421330452,
-0.05738931521773338,
0.018165281042456627,
0.08506902307271957,
-0.02055233158171177,
0.038578324019908905,
-0.06497573852539062,
-0.055695123970508575,
-0.009813527576625347,
-0.009545853361487389,
0.05111744627356529,
-0.05388227477669716,
0.036989953368902206,
-0.06581360101699829,
-0.05022323504090309,
-0.06553096324205399,
0.06110992655158043,
0.07997338473796844,
0.026012007147073746,
-0.08215559273958206,
0.04355751350522041,
-0.044293783605098724,
0.03150869160890579,
-0.0440099872648716,
0.09155545383691788,
-0.016860848292708397,
0.026047267019748688,
0.06999658793210983,
-0.03869827091693878,
-0.0376226007938385,
0.0009363209828734398,
0.059488050639629364,
0.022641507908701897,
0.06592992693185806,
-0.054705504328012466,
0.006600218825042248,
-0.03706812113523483,
-0.00666336202993989,
-0.005331545136868954,
0.052706070244312286,
0.003243003971874714,
0.04516959562897682,
0.00973817054182291,
0.019469715654850006,
0.08037691563367844,
-0.013923448510468006,
0.06013172119855881,
0.0018164823995903134,
-0.042960647493600845,
0.02076008729636669,
0.009527782909572124,
0.12280262261629105,
-0.0000991875131148845,
-0.05815976858139038,
-3.479048871213308e-8,
-0.04836980253458023,
0.07381396740674973,
0.07085992395877838,
0.011988508515059948,
0.08152226358652115,
0.01867465302348137,
-0.014301920309662819,
-0.09096100181341171,
-0.01942327246069908,
0.005050985608249903,
0.09042884409427643,
0.030132047832012177,
0.029713667929172516,
-0.10679933428764343,
-0.023040587082505226,
0.015240458771586418,
0.048299603164196014,
0.02038263902068138,
0.010602118447422981,
-0.0019551473669707775,
0.018458418548107147,
0.022048084065318108,
0.021635806187987328,
-0.055426761507987976,
0.04600130394101143,
-0.0348534882068634,
-0.006787707097828388,
0.018664950504899025,
0.08916658163070679,
-0.022804662585258484,
0.013476906344294548,
-0.0012375134974718094,
-0.003074520267546177,
0.06956670433282852,
-0.024283625185489655,
0.02914591319859028,
0.027664300054311752,
-0.042456138879060745,
0.027186842635273933,
-0.021432165056467056,
-0.05913223326206207,
0.08250383287668228,
-0.06585489958524704,
-0.003793239826336503,
-0.09126784652471542,
0.048753805458545685,
-0.05961001291871071,
-0.09515462070703506,
0.09497140347957611,
-0.019352326169610023,
0.00325977336615324,
-0.029834290966391563,
-0.03834646940231323,
0.027091559022665024,
0.03846829757094383,
0.004307400900870562,
0.030605904757976532,
-0.04289395362138748,
0.002520430600270629,
0.012009551748633385,
0.0013103565433993936,
-0.03637561947107315,
-0.088435098528862,
0.024111799895763397
] | 0 | 0 | 0 | 0 |
|
Searching for Optimal Solutions with LLMs via Bayesian Optimization | https://openreview.net/forum?id=aVfDrl7xDV | [
"Dhruv Agarwal",
"Manoj Ghuhan Arivazhagan",
"Rajarshi Das",
"Sandesh Swamy",
"Sopan Khosla",
"Rashmi Gangadharaiah"
] | Poster | Scaling test-time compute to search for optimal solutions is an important step towards building generally-capable language models that can reason. Recent work, however, shows that tasks of varying complexity require distinct search strategies to solve optimally, thus making it challenging to design a one-size-fits-all approach. Prior solutions either attempt to predict task difficulty to select the optimal search strategy, often infeasible in practice, or use a static, pre-defined strategy, e.g., repeated parallel sampling or greedy sequential search, which is sub-optimal. In this work, we argue for an alternative view using the probabilistic framework of Bayesian optimization (BO), where the search strategy is adapted dynamically based on the evolving uncertainty estimates of solutions as search progresses. To this end, we introduce Bayesian-OPRO (BOPRO)––a generalization of a recent method for in-context optimization, which iteratively samples from new proposal distributions by modifying the prompt to the LLM with a subset of its previous generations selected to explore or exploit different parts of the search space. We evaluate our method on word search, molecule optimization, and a joint hypothesis+program search task using a 1-D version of the challenging Abstraction and Reasoning Corpus (1D-ARC). Our results show that BOPRO outperforms all baselines in word search (≥10 points) and molecule optimization (higher quality and 17% fewer invalid molecules), but trails a best-k prompting strategy in program search. Our analysis reveals that despite the ability to balance exploration and exploitation using BOPRO, failure is likely due to the inability of code representation models in distinguishing sequences with low edit-distances. | search, optimization, LLMs, test-time steering, bayesian optimization, reasoning | We present Bayesian-OPRO (BOPRO), a method for generating optimal solutions with LLMs via Bayesian optimization that explores and exploits the search space based on evolving uncertainty estimates of the solution space as search progresses. | 13,475 | null | [
-0.05560266226530075,
-0.07190003991127014,
0.025769950821995735,
-0.0038000622298568487,
0.05576161667704582,
-0.016393175348639488,
0.014039091765880585,
0.052685193717479706,
0.04164282977581024,
-0.013458805158734322,
-0.036944177001714706,
-0.06709744781255722,
0.020003823563456535,
0.014119305647909641,
0.02254265546798706,
-0.007237943820655346,
0.10035127401351929,
0.055446602404117584,
-0.004661025013774633,
-0.04243360087275505,
0.036893635988235474,
0.0536113865673542,
0.09404926002025604,
-0.07317046076059341,
-0.0061522903852164745,
0.004728302825242281,
0.045344747602939606,
-0.006091381888836622,
0.04191069304943085,
0.06702835112810135,
0.07837098836898804,
0.05928288400173187,
0.06651921570301056,
-0.01858481392264366,
0.025635486468672752,
0.0368540920317173,
-0.16881510615348816,
0.02436644956469536,
0.018516624346375465,
0.0007096281042322516,
-0.07825694233179092,
0.006755443289875984,
0.020237846300005913,
0.022672517225146294,
0.026822488754987717,
-0.05552683770656586,
0.009976700879633427,
-0.018358264118433,
0.04293004795908928,
-0.031072624027729034,
-0.0482129342854023,
-0.078084297478199,
-0.060202062129974365,
-0.029447445645928383,
0.012044323608279228,
0.010715167969465256,
0.030439356341958046,
-0.025238996371626854,
0.038082022219896317,
-0.06831339001655579,
-0.04731331393122673,
-0.13631241023540497,
-0.0286501906812191,
-0.0332334041595459,
0.04214369133114815,
0.010337567888200283,
0.050128091126680374,
0.03711283579468727,
-0.017811136320233345,
0.08531515300273895,
-0.040480948984622955,
0.10133101046085358,
-0.009086272679269314,
0.07590969651937485,
-0.008859527297317982,
0.036982838064432144,
0.03317451849579811,
-0.022900190204381943,
0.003193106735125184,
-0.057596635073423386,
-0.004300773609429598,
-0.024419724941253662,
-0.00011800174979725853,
0.034247953444719315,
0.09167864173650742,
0.053275082260370255,
-0.030682608485221863,
0.023496538400650024,
0.15709424018859863,
-0.010402780957520008,
-0.06043814867734909,
-0.07694493234157562,
-0.08498258888721466,
-0.058184754103422165,
-0.028176113963127136,
0.001052789157256484,
-0.00857432372868061,
-0.056465230882167816,
0.048768654465675354,
0.04025474563241005,
0.03978390619158745,
0.057324472814798355,
0.09018648415803909,
-0.0837159976363182,
-0.05414655804634094,
-0.031595416367053986,
-0.04773743823170662,
0.057800862938165665,
0.0340743213891983,
-0.046599820256233215,
-0.0020154959056526423,
0.03046243265271187,
0.12054429948329926,
0.022860107943415642,
-0.05650648847222328,
-0.0034143272787332535,
0.0520135834813118,
-0.04079751297831535,
-0.04694555327296257,
0.13231246173381805,
0.0022290381602942944,
0.023229334503412247,
-0.00013849510287400335,
-0.06543109565973282,
-0.06618255376815796,
0.021223539486527443,
-0.022456811740994453,
-4.504469556767043e-34,
0.10307018458843231,
0.022143088281154633,
0.06815130263566971,
0.09226276725530624,
0.025976568460464478,
0.023889722302556038,
0.02897747978568077,
0.06886924803256989,
-0.07589757442474365,
-0.027307355776429176,
-0.022201959043741226,
0.01728122867643833,
-0.06681276112794876,
0.007581600919365883,
0.08960826694965363,
-0.033372536301612854,
-0.0326344259083271,
0.09012190252542496,
-0.05393000319600105,
-0.02183791808784008,
0.008841152302920818,
-0.01656297780573368,
0.023248879238963127,
-0.05247598886489868,
0.05756177008152008,
-0.030537445098161697,
0.07266432791948318,
-0.1006682738661766,
-0.08129379898309708,
-0.010061661712825298,
-0.15148420631885529,
-0.027562305331230164,
-0.07018765062093735,
0.08826885372400284,
-0.015970079228281975,
-0.012272205203771591,
-0.007207524497061968,
-0.03587275743484497,
-0.026620732620358467,
-0.02049187570810318,
-0.032380226999521255,
0.00201404863037169,
0.12739147245883942,
0.052471816539764404,
-0.044970110058784485,
-0.11293525248765945,
-0.005528854206204414,
-0.05858762934803963,
0.013204867951571941,
-0.07951780408620834,
0.09526511281728745,
0.006648405455052853,
0.006782230921089649,
-0.030971333384513855,
0.05575936287641525,
0.03964614495635033,
-0.031142817810177803,
0.017284218221902847,
0.04040765389800072,
0.05124206095933914,
0.03241366147994995,
0.039418045431375504,
-0.0462835393846035,
0.06168540194630623,
0.028923751786351204,
0.061304617673158646,
-0.029056241735816002,
0.04322978854179382,
0.08282126486301422,
-0.009146456606686115,
0.08772542327642441,
-0.0703476071357727,
0.012766406871378422,
-0.0738791823387146,
0.07952255755662918,
0.04757118225097656,
0.09699149429798126,
0.004353278316557407,
-0.011162658222019672,
-0.05649572238326073,
0.01172495074570179,
-0.05355190113186836,
-0.02121647447347641,
0.026475215330719948,
0.000563491543289274,
-0.02737257443368435,
0.04754991829395294,
-0.104835145175457,
-0.051943667232990265,
-0.0525256022810936,
-0.054749418050050735,
-0.008234984241425991,
-0.0715685784816742,
0.009401260875165462,
0.03808752819895744,
2.291473101238206e-34,
-0.05301010608673096,
-0.09926457703113556,
0.06540538370609283,
0.035550836473703384,
0.04851113259792328,
-0.06313034892082214,
-0.009183577261865139,
-0.12269576638936996,
0.0016341288574039936,
-0.10922259092330933,
-0.10109604895114899,
-0.021470263600349426,
0.059628285467624664,
-0.04006008431315422,
-0.06650438904762268,
0.02602849155664444,
0.002217922592535615,
0.05927709490060806,
-0.004719601944088936,
0.04978560656309128,
0.0003676501219160855,
0.06194812059402466,
-0.022858629003167152,
0.006747459527105093,
0.024722710251808167,
0.013557917438447475,
0.036958154290914536,
0.02930702082812786,
-0.003369915299117565,
-0.0067572216503322124,
-0.08189968019723892,
0.0027963244356215,
-0.05776510760188103,
0.05282052978873253,
-0.028648026287555695,
0.09268952161073685,
0.03388263285160065,
0.024602366611361504,
-0.05220850184559822,
0.09901434183120728,
0.025444040074944496,
-0.03116287849843502,
-0.016237037256360054,
-0.05358889698982239,
-0.021175483241677284,
0.01406838558614254,
-0.05600332096219063,
-0.057554811239242554,
0.05456339567899704,
-0.038821786642074585,
0.012036179192364216,
0.09857277572154999,
-0.01486223191022873,
0.03763497620820999,
-0.05377181991934776,
-0.04795467481017113,
-0.06307554244995117,
-0.08384212851524353,
0.07372859120368958,
0.043349072337150574,
0.011967124417424202,
0.00043788610491901636,
0.05020918697118759,
-0.013085131533443928,
-0.019160475581884384,
-0.004778592847287655,
-0.00220898212864995,
0.014795378781855106,
-0.025909103453159332,
-0.024405665695667267,
0.0008248522644862533,
-0.041956640779972076,
0.052293308079242706,
0.011549308896064758,
0.006735640577971935,
0.028419621288776398,
0.04357060790061951,
-0.05059049278497696,
-0.017743278294801712,
-0.05212812498211861,
0.029395692050457,
0.007656478323042393,
-0.03366927057504654,
0.07540614902973175,
-0.039412252604961395,
0.01791849173605442,
0.06509499996900558,
0.06050446256995201,
0.0541851781308651,
-0.03611629828810692,
-0.05639435723423958,
-0.05861454829573631,
0.005960182752460241,
0.025011183694005013,
-0.03421364352107048,
-4.4707523016995765e-8,
-0.013307086192071438,
0.05381077527999878,
0.0332249291241169,
0.011374170891940594,
0.10232260078191757,
0.012658459134399891,
-0.04575876519083977,
0.0005250205867923796,
0.025555668398737907,
-0.08925603330135345,
0.06467931717634201,
0.03187499940395355,
0.049561142921447754,
-0.02036522701382637,
-0.028679395094513893,
0.03345504403114319,
-0.007276261691004038,
-0.055861253291368484,
-0.020015660673379898,
-0.04892078414559364,
0.024371130391955376,
0.11110839992761612,
-0.0012558384332805872,
0.01847267523407936,
-0.00264801736921072,
-0.05788205564022064,
-0.012690499424934387,
0.03415248170495033,
0.10142546147108078,
-0.022809559479355812,
-0.04399726539850235,
-0.005551537498831749,
-0.09380455315113068,
0.06676534563302994,
-0.018320459872484207,
-0.015474733896553516,
-0.05242154002189636,
-0.04058823734521866,
-0.014705674722790718,
-0.05207071825861931,
0.03866000100970268,
0.028568146750330925,
-0.06353117525577545,
0.05626296624541283,
0.023304050788283348,
0.03623877093195915,
-0.028712552040815353,
-0.000952197820879519,
0.03778815641999245,
-0.020031895488500595,
0.03214425593614578,
0.01868698000907898,
-0.025872109457850456,
0.031387437134981155,
0.034022171050310135,
0.0636916533112526,
0.0037166504189372063,
-0.009688252583146095,
-0.017586711794137955,
0.008869173936545849,
0.02071266993880272,
0.025256816297769547,
-0.057767197489738464,
0.0403631329536438
] | 0 | 0 | 0 | 0 |
|
Context Steering: Controllable Personalization at Inference Time | https://openreview.net/forum?id=xQCXInDq0m | [
"Jerry Zhi-Yang He",
"Sashrika Pandey",
"Mariah L Schrum",
"Anca Dragan"
] | Poster | To deliver high-quality, personalized responses, large language models (LLMs) must effectively incorporate context — personal, demographic, and cultural information specific to an end-user. For example, asking the model to explain Newton's second law with the context "I am a toddler'' should produce a response different from when the context is "I am a physics professor''. However, leveraging the context in practice is a nuanced and challenging task, and is often dependent on the specific situation or user base. The model must strike a balance between providing specific, personalized responses and maintaining general applicability. Current solutions, such as prompt-engineering and fine-tuning, require collection of contextually appropriate responses as examples, making them time-consuming and less flexible to use across different contexts. In this work, we introduce Context Steering (CoS) —a simple, training-free decoding approach that amplifies the influence of the context in next token predictions. CoS computes contextual influence by comparing the output probabilities from two LLM forward passes: one that includes the context and one that does not. By linearly scaling the contextual influence, CoS allows practitioners to flexibly control the degree of personalization for different use cases. We show that CoS can be applied to autoregressive LLMs, and demonstrates strong performance in personalized recommendations. Additionally, we show that CoS can function as a Bayesian Generative model to infer and quantify correlations between open-ended texts, broadening its potential applications. | personalization, context, large language model, inference, controllable generation | We propose Context Steering (CoS), an inference-time technique that enables generating outputs more relevant to the user-provided contexts, leading to better personalization and various applications for large language models. | 13,474 | 2405.01768 | [
-0.02911425568163395,
-0.042179327458143234,
-0.024457963183522224,
0.04607464745640755,
0.12254580855369568,
-0.01160582434386015,
0.08645400404930115,
0.05979349836707115,
0.015422438271343708,
-0.03565061092376709,
-0.032433196902275085,
-0.13659052550792694,
0.04438496381044388,
0.04745723679661751,
0.07356802374124527,
-0.010057765990495682,
0.12973491847515106,
0.0694052129983902,
-0.08066461980342865,
-0.03072573058307171,
0.033465247601270676,
0.01952146366238594,
0.03825509548187256,
-0.008824167773127556,
-0.02788607031106949,
-0.00025431832182221115,
0.018466921523213387,
0.013650337234139442,
0.040735941380262375,
0.042172323912382126,
0.0615634061396122,
0.09403743594884872,
-0.020314689725637436,
-0.0017104730941355228,
-0.07663524895906448,
0.05283992365002632,
-0.05359407141804695,
0.04054419323801994,
-0.020967161282896996,
0.002862423425540328,
-0.07095178216695786,
-0.017733385786414146,
0.06496107578277588,
0.04652479290962219,
0.08842123299837112,
-0.022907117381691933,
0.0527377724647522,
0.05241019278764725,
-0.05215921998023987,
0.038724809885025024,
-0.08582545071840286,
-0.03628962114453316,
0.014681895263493061,
-0.03785601258277893,
-0.02725405804812908,
-0.00492922542616725,
0.07752028107643127,
0.06953925639390945,
-0.022194791585206985,
-0.03442223742604256,
-0.041247908025979996,
-0.1343558430671692,
-0.09149203449487686,
0.04163368418812752,
-0.045458126813173294,
0.010348360054194927,
-0.014592470601201057,
0.08773328363895416,
-0.023928046226501465,
0.04138236120343208,
-0.009816805832087994,
0.03986421227455139,
0.013668659143149853,
0.06461767107248306,
0.06293428689241409,
0.017986953258514404,
-0.004162624012678862,
-0.009602971374988556,
0.07188718765974045,
-0.06657017022371292,
0.022488050162792206,
0.0072598326951265335,
0.0553949773311615,
-0.051633525639772415,
0.05427170172333717,
0.006123534869402647,
0.04291098192334175,
0.0692620798945427,
0.03785541653633118,
0.0849112793803215,
-0.022524621337652206,
-0.09762710332870483,
0.07874275743961334,
0.004589397925883532,
0.011083829216659069,
0.024487193673849106,
0.013505980372428894,
-0.0925445705652237,
-0.02543814666569233,
0.08573151379823685,
-0.028433451429009438,
0.05898332595825195,
-0.006781446747481823,
-0.022167691960930824,
-0.00210238597355783,
0.02555224671959877,
-0.06348740309476852,
0.08011982589960098,
0.031145595014095306,
-0.012429168447852135,
-0.007168487645685673,
0.10658019036054611,
-0.053665608167648315,
-0.024867556989192963,
0.04823540896177292,
-0.04189153388142586,
0.11836831271648407,
-0.06189731881022453,
0.04421311244368553,
0.05206785351037979,
-0.031342361122369766,
-0.0008775267633609474,
-0.04045474901795387,
0.003620224306359887,
-0.03754212707281113,
-0.07034925371408463,
-0.03346925973892212,
8.567528127830855e-34,
0.0014606626937165856,
0.03573450446128845,
0.038189202547073364,
0.08797992020845413,
0.03268543258309364,
0.01643187366425991,
-0.059834904968738556,
-0.013687128201127052,
0.016789721325039864,
-0.04458159953355789,
0.07452593743801117,
0.11088313162326813,
0.023283623158931732,
0.03987208753824234,
-0.008090118877589703,
0.054294660687446594,
-0.1037093847990036,
0.06326627731323242,
-0.01763281598687172,
-0.0037794054951518774,
-0.019932067021727562,
-0.017647523432970047,
-0.02577296271920204,
-0.031999796628952026,
0.053052548319101334,
-0.034151919186115265,
0.09250615537166595,
0.013487354852259159,
0.028554219752550125,
-0.018117358908057213,
-0.11249247938394547,
-0.04090607911348343,
-0.030396537855267525,
-0.012851620092988014,
0.04030274227261543,
-0.02753056399524212,
-0.003975882660597563,
-0.07711373269557953,
-0.02253853902220726,
-0.0062242294661700726,
-0.022717971354722977,
-0.006717566400766373,
0.022099900990724564,
0.0691579207777977,
-0.12707509100437164,
-0.1009211614727974,
-0.009222700260579586,
0.0014199168654158711,
-0.06787341088056564,
0.003167870920151472,
0.03801647946238518,
0.021387014538049698,
-0.0040532853454351425,
-0.059669990092515945,
0.0028148747514933348,
-0.021358288824558258,
-0.012816491536796093,
0.04642530158162117,
0.0018059558933600783,
-0.04449846222996712,
0.04849304258823395,
0.010303189978003502,
0.05898399278521538,
0.1095876544713974,
0.013529042713344097,
0.004360797815024853,
-0.056625429540872574,
0.014867071993649006,
0.058142345398664474,
-0.05527959018945694,
-0.061422474682331085,
0.011361116543412209,
-0.03507564961910248,
-0.003237491473555565,
0.03153803199529648,
0.02892252430319786,
0.04040715470910072,
-0.04714568704366684,
0.008857808075845242,
0.02282717451453209,
0.048357043415308,
-0.0059907082468271255,
-0.005976516287773848,
0.019169606268405914,
-0.04959637671709061,
-0.01744370348751545,
-0.025956720113754272,
-0.11171179264783859,
-0.03701671585440636,
-0.10994742810726166,
-0.06484012305736542,
-0.046098493039608,
-0.012494643218815327,
0.06183233484625816,
-0.0024206768721342087,
-2.157896076189004e-34,
-0.03887972980737686,
-0.004974172916263342,
0.01073321234434843,
-0.019131723791360855,
-0.0011537830578163266,
-0.03401269018650055,
0.045095622539520264,
0.011472940444946289,
-0.008330831304192543,
-0.07812020182609558,
-0.06000400707125664,
-0.059490103274583817,
0.06002992391586304,
0.02676391787827015,
0.03489193320274353,
-0.01500907726585865,
0.0167737677693367,
0.02461189031600952,
-0.010857042856514454,
0.029274925589561462,
0.03309765085577965,
0.0845598354935646,
-0.12409118562936783,
-0.045396845787763596,
-0.02178417146205902,
-0.022363649681210518,
-0.08128166198730469,
0.016603093594312668,
0.03428276628255844,
0.002747897058725357,
-0.0582895502448082,
0.008110152557492256,
0.034964486956596375,
-0.042529478669166565,
-0.08125709742307663,
0.054170943796634674,
0.01530594751238823,
-0.02675243467092514,
-0.050779860466718674,
0.06478362530469894,
0.04260603338479996,
-0.01640567183494568,
-0.01943979412317276,
-0.011101826094090939,
-0.015467826277017593,
0.0030860784463584423,
-0.04877409711480141,
-0.027859121561050415,
0.03055415116250515,
0.01458375621587038,
-0.06496921181678772,
-0.008757119067013264,
-0.0920347347855568,
-0.022378381341695786,
-0.09304926544427872,
-0.08805617690086365,
0.09060491621494293,
-0.0891600102186203,
0.08603201061487198,
-0.05528761073946953,
-0.05436179041862488,
-0.03127795830368996,
0.05456499010324478,
-0.03158718720078468,
0.0028973666485399008,
-0.08655881881713867,
-0.010894539766013622,
0.0011821530060842633,
-0.0038313264958560467,
-0.027838844805955887,
0.05614803731441498,
0.0003573443682398647,
-0.04024335741996765,
-0.017907077446579933,
0.04237080365419388,
-0.018458494916558266,
0.07236730307340622,
-0.09580519795417786,
-0.04172306880354881,
-0.06697723269462585,
0.031408440321683884,
-0.013209732249379158,
0.04607430100440979,
0.030540967360138893,
0.05508141219615936,
0.031245674937963486,
0.012508085928857327,
0.07783595472574234,
0.052706342190504074,
0.022117076441645622,
-0.00737419817596674,
0.02263057604432106,
-0.060434065759181976,
0.10495807230472565,
-0.09338083118200302,
-5.014551973658854e-8,
-0.08101238310337067,
0.08239570260047913,
0.00884281937032938,
0.11362539231777191,
-0.04063604399561882,
0.06589504331350327,
-0.017056947574019432,
-0.00365682621486485,
-0.00971182156354189,
-0.004384197294712067,
-0.009784340858459473,
0.00676635792478919,
-0.006231626030057669,
-0.03791071102023125,
0.010768115520477295,
0.0923117846250534,
-0.002101264661177993,
0.009003765881061554,
-0.013303940184414387,
-0.003617282724007964,
0.09228885918855667,
0.03394736722111702,
-0.009261608123779297,
0.05838274210691452,
0.08272866904735565,
-0.005095634143799543,
-0.009965319186449051,
0.11245781928300858,
0.0323529914021492,
-0.08594373613595963,
-0.03554125502705574,
0.01782185025513172,
-0.07075625658035278,
-0.02710309624671936,
-0.08197008073329926,
0.03419487550854683,
0.004170299042016268,
-0.03897124528884888,
0.06114935874938965,
0.01648947410285473,
0.021377623081207275,
0.044554274529218674,
-0.08800757676362991,
0.08819714933633804,
0.028655389323830605,
0.037368934601545334,
0.026828859001398087,
-0.08227021992206573,
0.007279324345290661,
0.040062155574560165,
0.0030902910511940718,
-0.07562103867530823,
-0.014104506932199001,
0.04908785596489906,
0.04365424066781998,
0.04840662330389023,
-0.026659036055207253,
0.04401254653930664,
0.01155389565974474,
0.08862286806106567,
-0.022201338782906532,
0.09586493670940399,
-0.06424128264188766,
0.00804218091070652
] | https://github.com/sashrikap/context-steering | 12 | 0 | 0 | 0 |
Efficient Policy Evaluation with Safety Constraint for Reinforcement Learning | https://openreview.net/forum?id=Dem5LyVk8R | [
"Claire Chen",
"Shuze Liu",
"Shangtong Zhang"
] | Poster | In reinforcement learning, classic on-policy evaluation methods often suffer from high variance and require massive online data to attain the desired accuracy. Previous studies attempt to reduce evaluation variance by searching for or designing proper behavior policies to collect data. However, these approaches ignore the safety of such behavior policies---the designed behavior policies have no safety guarantee and may lead to severe damage during online executions. In this paper, to address the challenge of reducing variance while ensuring safety simultaneously, we propose an optimal variance-minimizing behavior policy under safety constraints. Theoretically, while ensuring safety constraints, our evaluation method is unbiased and has lower variance than on-policy evaluation. Empirically, our method is the only existing method to achieve both substantial variance reduction and safety constraint satisfaction. Furthermore, we show our method is even superior to previous methods in both variance reduction and execution safety. | Reinforcement Learning | null | 13,457 | 2410.05655 | [
-0.06123380362987518,
0.027162402868270874,
0.0008491823682561517,
-0.0017402753001078963,
0.04123548045754433,
0.07519236207008362,
0.021422991529107094,
-0.000685275299474597,
-0.04674874246120453,
0.04141882061958313,
-0.049946852028369904,
-0.004099811427295208,
0.0531405545771122,
0.01743038184940815,
-0.02823827415704727,
-0.005263064987957478,
0.11116494238376617,
0.04933353140950203,
-0.0867779478430748,
-0.024577748030424118,
0.041785724461078644,
0.01934153214097023,
0.10124947130680084,
0.053895510733127594,
0.028764896094799042,
-0.09105676412582397,
-0.016279924660921097,
0.0036891489289700985,
0.014031266793608665,
-0.005339059513062239,
0.028784317895770073,
-0.15204143524169922,
-0.0035753794945776463,
-0.07519500702619553,
-0.011240222491323948,
-0.006093333475291729,
-0.11175976693630219,
-0.05240988731384277,
-0.05224098637700081,
0.00327299814671278,
-0.09858240187168121,
0.006125994957983494,
0.01202953141182661,
0.030572686344385147,
0.058698371052742004,
-0.013878824189305305,
0.011360357515513897,
-0.13379736244678497,
0.016579095274209976,
0.006138769444078207,
-0.009002848528325558,
-0.015741664916276932,
0.02901093102991581,
0.032268233597278595,
-0.0057098581455647945,
0.004659249447286129,
0.02771846018731594,
-0.002640909980982542,
0.02038332261145115,
0.05834604427218437,
-0.03840012848377228,
-0.12020894140005112,
-0.023843057453632355,
-0.029445137828588486,
-0.08794838935136795,
-0.010835906490683556,
-0.050246573984622955,
-0.004097888711839914,
0.007095992565155029,
0.04543652385473251,
-0.06187398359179497,
0.034981656819581985,
-0.01819092407822609,
0.023821860551834106,
0.004804098978638649,
0.030102364718914032,
-0.022324424237012863,
-0.08209289610385895,
0.016510264948010445,
0.014729117974638939,
-0.047754574567079544,
-0.037582848221063614,
-0.020617619156837463,
0.05697152763605118,
0.10315163433551788,
-0.03726736828684807,
0.025827977806329727,
0.010927305556833744,
0.1374618113040924,
0.004573766607791185,
0.005166241433471441,
0.10176832228899002,
0.08808092772960663,
0.06388282030820847,
0.005918955430388451,
-0.014157501980662346,
-0.028598565608263016,
-0.10842692852020264,
-0.03479864448308945,
0.09152538329362869,
0.03597864508628845,
0.08662215620279312,
-0.07830670475959778,
0.010109877213835716,
0.026327520608901978,
-0.02416633814573288,
0.018120456486940384,
0.006104426924139261,
-0.015163064934313297,
-0.04767727851867676,
0.028024468570947647,
0.024964185431599617,
0.01872357353568077,
-0.008215558715164661,
-0.07834316790103912,
0.08013010770082474,
-0.05544406920671463,
0.015332975424826145,
-0.021706756204366684,
0.11688496917486191,
0.07266045361757278,
-0.07448823750019073,
0.09770407527685165,
0.03834175691008568,
0.040633514523506165,
-0.07287035882472992,
0.0009581308113411069,
3.1811477914502294e-33,
0.07145631313323975,
0.00631709722802043,
0.0348920114338398,
-0.044705066829919815,
0.011999505572021008,
0.02512301504611969,
0.01731535606086254,
-0.012550399638712406,
-0.05293043702840805,
-0.04468433931469917,
0.012321154586970806,
-0.03711245208978653,
-0.0541909895837307,
0.0004957554629072547,
0.08225492388010025,
0.007647023070603609,
-0.08486965298652649,
0.1279863864183426,
-0.009176863357424736,
-0.08146016299724579,
0.08055658638477325,
-0.11508715152740479,
0.019048532471060753,
0.024195896461606026,
0.02118753083050251,
0.0298811886459589,
-0.021271316334605217,
0.023447023704648018,
0.027228137478232384,
0.016695337370038033,
-0.05134081840515137,
-0.0025301300920546055,
0.008074020966887474,
0.08635291457176208,
0.031416960060596466,
0.02873195707798004,
-0.1297331154346466,
0.06117653474211693,
-0.003453075885772705,
0.022209778428077698,
-0.039691489189863205,
-0.016450099647045135,
0.10835826396942139,
0.004268508870154619,
-0.031914956867694855,
-0.08276727795600891,
0.005959963425993919,
-0.070368692278862,
-0.021179914474487305,
-0.05840733274817467,
-0.0589144341647625,
0.043607693165540695,
0.06112698093056679,
-0.010003665462136269,
-0.07269024103879929,
0.0423668697476387,
0.026492157950997353,
0.08074767142534256,
-0.06380648165941238,
0.020791513845324516,
-0.017908399924635887,
0.02490006946027279,
0.013006214052438736,
0.04010431468486786,
-0.008237441070377827,
0.04375484958291054,
0.018946783617138863,
-0.01662558875977993,
0.058956440538167953,
-0.07833509147167206,
-0.018312804400920868,
0.06232494115829468,
0.020090864971280098,
-0.08927004039287567,
-0.016319261863827705,
0.016548380255699158,
0.07080185413360596,
0.00546650867909193,
0.004243894945830107,
-0.0935552567243576,
0.04615272209048271,
0.01900634542107582,
0.014520739205181599,
-0.0312713086605072,
0.003529439913108945,
-0.04403647780418396,
-0.06092972680926323,
-0.002971658017486334,
-0.02721954695880413,
0.07880885899066925,
-0.06673150509595871,
-0.02425083890557289,
-0.022036772221326828,
0.023317310959100723,
0.023090694099664688,
-2.7736578992918832e-33,
-0.0320708230137825,
-0.03313252329826355,
-0.017565660178661346,
0.021464629098773003,
-0.03231615200638771,
-0.034486595541238785,
-0.02254912070930004,
-0.060283806174993515,
0.019871555268764496,
-0.02486426755785942,
-0.07528918236494064,
0.049684882164001465,
-0.016158150508999825,
-0.01187317818403244,
-0.04727105796337128,
-0.012559713795781136,
-0.020463693886995316,
0.04710241034626961,
-0.017170390114188194,
-0.00931514985859394,
0.06218082830309868,
0.03923676535487175,
-0.12982165813446045,
0.09475354850292206,
0.013940497301518917,
0.02216028794646263,
-0.0559573695063591,
0.0003935772692784667,
0.020982442423701286,
0.09156642109155655,
-0.04271639883518219,
-0.053030386567115784,
-0.026102349162101746,
0.014700940810143948,
-0.001955956220626831,
-0.016583865508437157,
0.086023710668087,
0.015690254047513008,
-0.038175467401742935,
0.07689264416694641,
0.07719823718070984,
-0.020239390432834625,
-0.017219454050064087,
-0.04203014075756073,
0.04256051778793335,
-0.0008956934907473624,
-0.046225883066654205,
0.022512394934892654,
-0.0130686704069376,
-0.010135220363736153,
-0.046638134866952896,
-0.014375390484929085,
-0.0844588354229927,
0.0009605919476598501,
0.0004100839141756296,
-0.08154013752937317,
0.0026701258029788733,
0.007009054999798536,
0.03322577849030495,
0.026131510734558105,
-0.027098404243588448,
0.04870210215449333,
-0.08698802441358566,
0.09757993370294571,
0.00122494762763381,
0.06108207255601883,
-0.047133270651102066,
-0.004031844902783632,
0.014846516773104668,
0.0049073351547122,
0.006842353381216526,
0.025964591652154922,
-0.0036126996856182814,
-0.02931971102952957,
-0.007304737344384193,
-0.03806167468428612,
0.027021408081054688,
-0.00810575857758522,
-0.05606824532151222,
0.027198493480682373,
0.09196750074625015,
0.024018703028559685,
0.10836245864629745,
0.004448378924280405,
0.035775553435087204,
0.08518029749393463,
-0.015001476742327213,
0.11812132596969604,
0.03665197268128395,
0.005685607902705669,
-0.06994783878326416,
-0.020832298323512077,
-0.006045198533684015,
-0.014057398773729801,
-0.06838920712471008,
-4.229458383520068e-8,
0.01695088855922222,
0.05484475940465927,
0.0635799765586853,
0.05577565357089043,
0.09954271465539932,
0.040642354637384415,
-0.041231222450733185,
-0.016926126554608345,
-0.007974646054208279,
-0.06317131966352463,
0.057266466319561005,
0.03755234554409981,
0.03884882107377052,
-0.07010906934738159,
-0.07352814823389053,
0.016210483387112617,
-0.0495886392891407,
0.018485410138964653,
-0.00906811747699976,
0.06278366595506668,
0.01464180089533329,
0.048458125442266464,
-0.03202497586607933,
-0.1029777079820633,
0.07748909294605255,
-0.020002370700240135,
0.009132162667810917,
0.11241371929645538,
0.0780346542596817,
0.05431875213980675,
0.03357652947306633,
-0.06129557266831398,
0.012327225878834724,
-0.011490450240671635,
-0.004211952909827232,
0.03637024760246277,
0.0001607796730240807,
-0.044640034437179565,
0.015632765367627144,
-0.09720993787050247,
0.013698960654437542,
0.07342854142189026,
0.010719827376306057,
0.03325225040316582,
-0.00036356880445964634,
0.06733066588640213,
-0.0617188960313797,
-0.09614331275224686,
0.08260121196508408,
-0.06287158280611038,
0.03739377483725548,
-0.07309439778327942,
-0.02522937022149563,
0.10236357152462006,
0.03760211169719696,
-0.02493000216782093,
0.028792211785912514,
-0.09173838049173355,
0.054285693913698196,
0.0681486651301384,
0.04093664884567261,
-0.041970521211624146,
-0.017543034628033638,
0.023983823135495186
] | 0 | 0 | 0 | 0 |
|
Agent S: An Open Agentic Framework that Uses Computers Like a Human | https://openreview.net/forum?id=lIVRgt4nLv | [
"Saaket Agashe",
"Jiuzhou Han",
"Shuyu Gan",
"Jiachen Yang",
"Ang Li",
"Xin Eric Wang"
] | Poster | We present Agent S, an open agentic framework that enables autonomous interaction with computers through Graphical User Interface (GUI), aimed at transforming human-computer interaction by automating complex, multi-step tasks. Agent S addresses three key challenges in automating computer tasks: acquiring domain-specific knowledge, planning over long task horizons, and handling dynamic, non-uniform interfaces. To this end, Agent S introduces experience-augmented hierarchical planning, which learns from external knowledge search and internal experience retrieval at multiple levels, facilitating efficient task planning and subtask execution.
In addition, it employs an Agent-Computer Interface (ACI) to better elicit the reasoning and control capabilities of GUI agents based on Multimodal Large Language Models (MLLMs). Evaluation on the OSWorld benchmark shows that Agent S outperforms the baseline by 9.37\% on success rate (an 83.6\% relative improvement) and achieves a new state-of-the-art. Comprehensive analysis highlights the effectiveness of individual components and provides insights for future improvements. Furthermore, Agent S demonstrates broad generalizability to different operating systems on a newly-released WindowsAgentArena benchmark. Code available at https://github.com/simular-ai/Agent-S. | Large Vision and Language Model, Agents, Retrieval Augmented Generation, GUI, Large Language Models, Agent Computer Interface | null | 13,452 | 2410.08164 | [
0.009398548863828182,
-0.07868135720491409,
-0.016693539917469025,
0.016539590433239937,
0.06234857812523842,
-0.052446674555540085,
-0.03356855735182762,
0.03478221222758293,
-0.016116410493850708,
-0.022512471303343773,
-0.0675072893500328,
-0.08871668577194214,
0.07436422258615494,
0.05833135545253754,
0.04758153483271599,
-0.0030769440345466137,
0.09166033565998077,
-0.03444672003388405,
-0.09451726824045181,
-0.03592730686068535,
0.020320752635598183,
0.012026969343423843,
0.030345134437084198,
-0.043433718383312225,
-0.0036955010145902634,
0.01979152113199234,
0.000434191373642534,
-0.10528780519962311,
-0.023568399250507355,
0.019259680062532425,
0.06112334877252579,
0.004708930850028992,
0.030299777165055275,
0.01955495774745941,
0.005925990175455809,
0.05483057349920273,
-0.0786786898970604,
0.03390217572450638,
-0.0121931042522192,
-0.06513126939535141,
-0.07618039101362228,
-0.0394461452960968,
-0.008186348713934422,
0.0035159264225512743,
0.06108836457133293,
0.000713499030098319,
-0.06739270687103271,
-0.05503081530332565,
0.07217877358198166,
0.04213324561715126,
-0.10889821499586105,
-0.0956987664103508,
0.09033573418855667,
0.04023303464055061,
-0.051006972789764404,
0.024180667474865913,
-0.012900509871542454,
-0.0013395734131336212,
0.018786726519465446,
-0.06014349311590195,
-0.030052388086915016,
-0.06438326090574265,
0.008241803385317326,
0.009106690995395184,
-0.041814420372247696,
0.0008954060031101108,
-0.013241762295365334,
-0.08792385458946228,
0.021975664421916008,
-0.07620195299386978,
-0.07328521460294724,
0.016069410368800163,
0.00613027811050415,
0.03170096501708031,
-0.005117244552820921,
-0.003861704608425498,
-0.019727710634469986,
-0.054809801280498505,
0.09454632550477982,
-0.07071372866630554,
-0.006451669614762068,
-0.010175274685025215,
-0.03745302930474281,
0.05464356020092964,
0.07622648775577545,
0.05634887143969536,
-0.003979234024882317,
0.04705074429512024,
0.04040253907442093,
0.04349945858120918,
0.016507834196090698,
-0.10523726791143417,
0.0675252377986908,
-0.01737327128648758,
0.03580429032444954,
0.04303167760372162,
-0.004595705773681402,
-0.09957542270421982,
-0.08041580766439438,
0.08135296404361725,
-0.024026857689023018,
0.017749927937984467,
-0.016712408512830734,
-0.017477957531809807,
-0.019580604508519173,
0.0035945475101470947,
0.05750303715467453,
0.035435356199741364,
0.015297395177185535,
-0.111935093998909,
-0.045829497277736664,
-0.05729486420750618,
0.06996320188045502,
-0.015510892495512962,
0.06850402802228928,
0.017450222745537758,
0.035953447222709656,
0.059009432792663574,
0.022033166140317917,
-0.021100658923387527,
0.03310463950037956,
-0.0005010714521631598,
0.005035886075347662,
0.006573867052793503,
0.09475847333669662,
0.046536728739738464,
0.017642907798290253,
7.41197861006416e-34,
0.0871240571141243,
0.038314007222652435,
0.08545441925525665,
0.10220199823379517,
0.050630588084459305,
-0.028581038117408752,
0.011496685445308685,
0.06448907405138016,
-0.011815773323178291,
-0.003177586942911148,
-0.0027632247656583786,
0.06748203933238983,
-0.0775086060166359,
0.07609251141548157,
0.04468848928809166,
0.00037198513746261597,
-0.04984870180487633,
0.07590146362781525,
-0.05837228521704674,
-0.07479550689458847,
0.05275269225239754,
0.01411348395049572,
0.03597860783338547,
-0.026599640026688576,
0.02500457875430584,
0.050946492701768875,
0.05467263609170914,
-0.06329754739999771,
0.056640803813934326,
-0.021606145426630974,
-0.14673759043216705,
0.045820172876119614,
-0.051111143082380295,
0.048675041645765305,
-0.04161573573946953,
-0.026849202811717987,
-0.029381616041064262,
-0.08919015526771545,
0.09824389219284058,
0.05761249363422394,
-0.07175559550523758,
-0.003819054923951626,
0.08794713765382767,
-0.05035572126507759,
-0.022126879543066025,
-0.0351833812892437,
-0.02174997515976429,
0.05539634823799133,
-0.002841621171683073,
-0.021448038518428802,
0.04884098097681999,
0.09288632869720459,
0.04158112406730652,
0.03640564903616905,
0.012832692824304104,
0.015171215869486332,
0.06312421709299088,
0.07879441231489182,
0.03827778622508049,
0.02391558140516281,
0.026285888627171516,
0.014838315546512604,
0.011744444258511066,
0.12847906351089478,
0.040399275720119476,
-0.009410961531102657,
0.016328472644090652,
0.14155921339988708,
0.04511672630906105,
0.037126705050468445,
-0.019558770582079887,
0.04251638054847717,
0.06918571889400482,
-0.006739611737430096,
0.0033541047014296055,
-0.026942623779177666,
0.00825460534542799,
-0.07400181144475937,
0.03793438524007797,
-0.015811119228601456,
-0.0630178451538086,
-0.019121063873171806,
-0.030176730826497078,
-0.0727156400680542,
0.03705457225441933,
-0.036385852843523026,
-0.005528311710804701,
-0.07532793283462524,
0.025508340448141098,
-0.021706845611333847,
-0.05654361844062805,
-0.00502464035525918,
-0.0783383846282959,
0.0681111142039299,
-0.0012825613375753164,
-8.428338576433957e-34,
0.027022304013371468,
-0.1054450273513794,
-0.07190585136413574,
-0.0138206472620368,
-0.011234181933104992,
-0.07076190412044525,
-0.01779015362262726,
-0.001157455611974001,
-0.057593703269958496,
-0.06078806892037392,
-0.0338163748383522,
0.047994498163461685,
0.07701544463634491,
0.0015808971365913749,
0.020303141325712204,
-0.030911311507225037,
0.0009552448173053563,
0.008297820575535297,
-0.002120635472238064,
-0.008444982580840588,
-0.0570853054523468,
0.05802878737449646,
-0.07363311201334,
-0.017027491703629494,
-0.04112924635410309,
-0.008300695568323135,
0.0018847695318982005,
0.012592841871082783,
0.01568802259862423,
0.04753715172410011,
0.017915964126586914,
0.016376346349716187,
-0.011037378571927547,
-0.02290220558643341,
0.01918191649019718,
0.09641638398170471,
0.012592584826052189,
-0.052633222192525864,
-0.08510620146989822,
0.07006125152111053,
0.09385319799184799,
-0.0760694071650505,
-0.07271068543195724,
-0.059269413352012634,
-0.009475682862102985,
0.06442604959011078,
-0.16260972619056702,
0.0125189870595932,
0.008525287732481956,
0.018421677872538567,
0.003097101580351591,
-0.018045559525489807,
-0.04825827106833458,
-0.0839146226644516,
-0.01829834282398224,
-0.03953162208199501,
-0.007868200540542603,
-0.027178578078746796,
0.014831626787781715,
-0.0587150901556015,
-0.06366802006959915,
0.0057752481661736965,
0.02276957407593727,
-0.0056915851309895515,
-0.02644611895084381,
0.0504407063126564,
-0.05152655392885208,
-0.004782129544764757,
-0.011224268935620785,
-0.041141409426927567,
0.032216981053352356,
-0.0924619734287262,
-0.04277496412396431,
0.008863747119903564,
0.08368790149688721,
0.03107559122145176,
0.03771872818470001,
-0.03757020831108093,
-0.0010157667566090822,
-0.07823731005191803,
-0.016736164689064026,
0.01226734183728695,
0.055506862699985504,
0.03960305079817772,
-0.02632969617843628,
0.07638398557901382,
-0.03508150205016136,
0.11072100698947906,
0.013929478824138641,
0.036534905433654785,
-0.01641579158604145,
0.02259220741689205,
0.042155921459198,
0.0861889198422432,
-0.06505722552537918,
-4.294791366987738e-8,
-0.061913713812828064,
0.0682363510131836,
0.0529620461165905,
-0.018744317814707756,
-0.012535661458969116,
-0.04770389199256897,
-0.046560224145650864,
0.03619076684117317,
0.023521384224295616,
0.020630793645977974,
0.052555765956640244,
-0.05445115640759468,
0.05945679545402527,
-0.05555249750614166,
0.03775354474782944,
0.04178999736905098,
-0.006067593581974506,
-0.004727587569504976,
0.02017846517264843,
-0.07781201601028442,
0.10375228524208069,
0.016305195167660713,
-0.008625668473541737,
0.00619002990424633,
-0.0061828563921153545,
-0.03383725881576538,
-0.07353632897138596,
0.08795539289712906,
-0.009926469065248966,
0.030221745371818542,
0.03445947542786598,
-0.03325653448700905,
0.03573758527636528,
0.01980026625096798,
0.058612968772649765,
0.04478037357330322,
-0.01759766787290573,
-0.05768588185310364,
0.020240897312760353,
-0.07279393821954727,
0.004379679448902607,
0.03548423945903778,
-0.022573798894882202,
-0.03839432820677757,
0.0182252936065197,
0.04435271769762039,
-0.09283438324928284,
-0.08227529376745224,
0.060788173228502274,
-0.0470161959528923,
-0.09992694854736328,
-0.030515803024172783,
0.03632618114352226,
0.07772081345319748,
0.09973081946372986,
-0.00586519343778491,
0.02439536340534687,
-0.017678072676062584,
0.1369665414094925,
0.01846073567867279,
0.011869734153151512,
0.022495705634355545,
-0.02469099499285221,
0.03031766228377819
] | https://github.com/simular-ai/agent-s | 2,745 | 0 | 0 | 0 |
Semantic Loss Guided Data Efficient Supervised Fine Tuning for Safe Responses in LLMs | https://openreview.net/forum?id=kO0DgO07hW | [
"Yuxiao Lu",
"Arunesh Sinha",
"Pradeep Varakantham"
] | Poster | Large Language Models (LLMs) generating unsafe responses to toxic prompts is a significant issue in their applications. While various efforts aim to address this safety concern, previous approaches often demand substantial human data collection or rely on the less dependable option of using another LLM to generate corrective data. In this paper, we aim to take this problem and overcome limitations of requiring significant high-quality human data. Our method requires only a small set of unsafe responses to toxic prompts, easily obtained from the unsafe LLM itself. By employing a semantic cost combined with a negative Earth Mover Distance (EMD) loss, we guide the LLM away from generating unsafe responses. Additionally, we propose a novel lower bound for EMD loss, enabling more efficient optimization. Our results demonstrate superior performance and data efficiency compared to baselines, and we further examine the nuanced effects of over-alignment and potential degradation of language capabilities when using contrastive data. | Large Language Model, Safe Large Language Model, Earth Mover Distance, Supervised Fine-tuning | null | 13,451 | 2412.06843 | [
0.03545327112078667,
-0.05466867610812187,
0.06231367588043213,
-0.0038924901746213436,
0.09776432067155838,
-0.025918645784258842,
-0.027202704921364784,
-0.02872324548661709,
0.0404885970056057,
-0.01222759485244751,
-0.012897822074592113,
-0.09382408112287521,
0.1116049513220787,
0.012656676582992077,
-0.021330038085579872,
0.019481493160128593,
0.10920269787311554,
0.05082559213042259,
-0.13143236935138702,
0.01490740291774273,
0.01025652140378952,
0.13610008358955383,
0.02226829156279564,
0.018505608662962914,
-0.02261453866958618,
-0.027292771264910698,
-0.03382578119635582,
0.019453244283795357,
0.008728687651455402,
-0.03650614619255066,
0.02234000153839588,
0.0005077687092125416,
-0.039894334971904755,
0.027503343299031258,
0.010639162734150887,
0.01203269511461258,
-0.04133369401097298,
0.0024551020469516516,
-0.022022882476449013,
-0.01299713458865881,
-0.059101879596710205,
-0.014035291969776154,
-0.02025008574128151,
0.052081149071455,
-0.0013841058826074004,
0.016530035063624382,
-0.014050670899450779,
-0.07887683063745499,
0.009782249107956886,
-0.0009886282496154308,
-0.0663834884762764,
0.04740053787827492,
-0.02971322275698185,
0.06623302400112152,
-0.06621013581752777,
-0.0008364671957679093,
0.030900510028004646,
0.01309188175946474,
-0.06656462699174881,
-0.010828240774571896,
-0.023580482229590416,
-0.12386816740036011,
-0.017383188009262085,
-0.036029789596796036,
-0.027418911457061768,
-0.036857154220342636,
-0.02869967743754387,
0.018068958073854446,
-0.002229641191661358,
0.0832999125123024,
-0.030472224578261375,
-0.021252991631627083,
-0.018911708146333694,
0.02911677211523056,
-0.02158905379474163,
0.10191559791564941,
-0.0011775102466344833,
-0.0562945231795311,
0.0501931831240654,
-0.07059498876333237,
0.04068583622574806,
-0.030320122838020325,
0.07392999529838562,
0.0536087304353714,
0.11241315305233002,
0.04997929558157921,
0.02636011689901352,
0.04465605318546295,
0.04383082315325737,
0.05354723706841469,
-0.049437858164310455,
-0.09599683433771133,
0.10493629425764084,
0.0906316265463829,
-0.01799866557121277,
-0.008607638999819756,
-0.0193716362118721,
-0.11065156757831573,
-0.01914091780781746,
0.0888533741235733,
0.006112118251621723,
0.08083262294530869,
-0.04423654079437256,
-0.07397735118865967,
-0.01467869896441698,
-0.08094366639852524,
0.005563962738960981,
0.02158655785024166,
0.023856064304709435,
0.00842053722590208,
0.023636575788259506,
0.07252926379442215,
-0.018499860540032387,
-0.05014920234680176,
0.009228399954736233,
-0.050805699080228806,
-0.029352203011512756,
-0.06568382680416107,
-0.0027648494578897953,
-0.040884166955947876,
-0.02465755119919777,
-0.025715798139572144,
0.00905259232968092,
0.05374482274055481,
0.05432506278157234,
-0.07538429647684097,
-0.0028936981689184904,
3.128938944542427e-33,
0.052411388605833054,
0.029790526255965233,
-0.002511498751118779,
0.017649056389927864,
0.044309597462415695,
0.003109075129032135,
-0.051114145666360855,
0.08856712281703949,
-0.05011225864291191,
0.012474569492042065,
-0.011108423583209515,
-0.075715571641922,
0.015539462678134441,
0.044484693557024,
0.026024721562862396,
0.00870571006089449,
-0.0013265600427985191,
0.06552916765213013,
-0.13057273626327515,
-0.03467434272170067,
0.04675626754760742,
-0.05435630679130554,
0.021600697189569473,
-0.04571593552827835,
0.01264459453523159,
0.0037443607579916716,
-0.02561516873538494,
-0.018345406278967857,
0.007171595003455877,
-0.022668948397040367,
-0.17266687750816345,
-0.009955031797289848,
0.021206321194767952,
0.027871832251548767,
0.07268206030130386,
-0.02542778290808201,
-0.07757791876792908,
-0.043409183621406555,
-0.0428505539894104,
-0.0031484251376241446,
-0.04723682254552841,
0.11275285482406616,
0.007207920774817467,
-0.021898355334997177,
0.024301400408148766,
-0.014851749874651432,
0.0133419344201684,
-0.020509785041213036,
-0.037719596177339554,
-0.05830775573849678,
0.0489150770008564,
0.0062698922120034695,
-0.027858639135956764,
0.024642154574394226,
-0.05475681275129318,
-0.016891853883862495,
0.02251880057156086,
0.02488771267235279,
-0.018726594746112823,
0.027463719248771667,
-0.010515819303691387,
0.009612754918634892,
0.028325412422418594,
0.03432788699865341,
0.09642375260591507,
-0.03840749338269234,
0.04630603641271591,
0.03925609961152077,
0.011475907638669014,
-0.009591097012162209,
-0.012873835861682892,
-0.005016290582716465,
0.03389320522546768,
-0.01622423529624939,
-0.03489798679947853,
-0.062416814267635345,
0.0715552493929863,
-0.050956133753061295,
0.07812722027301788,
0.002875678474083543,
0.04906056076288223,
-0.024942880496382713,
-0.023650215938687325,
-0.055288784205913544,
-0.02518114261329174,
-0.09956145286560059,
0.024966195225715637,
-0.09190470725297928,
-0.03359494358301163,
0.06112317740917206,
-0.041316788643598557,
-0.052915845066308975,
-0.05553503334522247,
0.0243470948189497,
-0.039264142513275146,
-2.921525170391742e-33,
-0.1018768921494484,
-0.05315244942903519,
-0.01822272688150406,
0.06115510314702988,
-0.10840673744678497,
-0.09069632738828659,
0.04190334677696228,
0.11541563272476196,
0.04060765728354454,
0.008758081123232841,
-0.031463880091905594,
-0.05338228866457939,
0.0636911690235138,
-0.0187129657715559,
0.0515577457845211,
0.003992494195699692,
-0.010536560788750648,
0.02558242902159691,
0.008777091279625893,
0.03706442937254906,
0.0346839465200901,
0.1439836472272873,
-0.1660814881324768,
0.07969986647367477,
-0.027950305491685867,
0.037366218864917755,
-0.019955413416028023,
0.02610861510038376,
-0.03580029681324959,
-0.03893434256315231,
-0.056074514985084534,
0.055403463542461395,
-0.027628231793642044,
-0.0016712135402485728,
-0.07962825894355774,
0.00038769652019254863,
0.02510053664445877,
-0.02241000533103943,
-0.015948252752423286,
0.04802819341421127,
0.032558538019657135,
0.0889863446354866,
-0.05056004971265793,
0.012363484129309654,
-0.032370101660490036,
-0.01692204736173153,
-0.12371102720499039,
0.011034023948013783,
0.04531266167759895,
-0.003504389664158225,
0.0248054601252079,
-0.03844493627548218,
-0.0714111179113388,
0.03455394506454468,
0.020600929856300354,
-0.10945980250835419,
0.025196295231580734,
-0.07892397791147232,
0.02863488718867302,
-0.0010016955202445388,
-0.046044640243053436,
-0.01946116052567959,
0.020159730687737465,
-0.03066118061542511,
0.015158675611019135,
0.003027210244908929,
-0.02017848566174507,
0.00047905786777846515,
-0.033256400376558304,
-0.016485104337334633,
0.07032258808612823,
0.02667083404958248,
-0.02771054022014141,
0.0022872828412801027,
0.01816990226507187,
-0.07958245277404785,
0.020853174850344658,
-0.040785133838653564,
-0.03175053372979164,
-0.018403485417366028,
0.08275527507066727,
0.03068637102842331,
0.02793770283460617,
0.1368444263935089,
0.04428926855325699,
0.08103218674659729,
0.010224196128547192,
0.06370651721954346,
0.013053162954747677,
-0.005658276844769716,
-0.060098931193351746,
0.013496733270585537,
0.018502293154597282,
0.04419396072626114,
-0.06564009934663773,
-4.565139022361109e-8,
-0.10865344852209091,
0.04750433936715126,
-0.0070833745412528515,
0.07172729820013046,
-0.005708304233849049,
-0.02668645791709423,
-0.008136908523738384,
0.11722981184720993,
0.026487981900572777,
-0.0010808256920427084,
0.05329768359661102,
-0.013164982199668884,
-0.04749467968940735,
0.0128004290163517,
-0.004212574567645788,
0.06654409319162369,
0.015899280086159706,
0.03644426167011261,
-0.014483666978776455,
0.07269565761089325,
0.010388562455773354,
0.13395285606384277,
0.0009159636101685464,
0.025943847373127937,
0.13008174300193787,
-0.05286024510860443,
-0.0020087065640836954,
0.08669377118349075,
0.07119851559400558,
-0.04912951588630676,
-0.0032208559568971395,
-0.056812677532434464,
-0.01142803393304348,
0.05476565286517143,
-0.010308611206710339,
0.08816082030534744,
-0.022458607330918312,
-0.00728320749476552,
-0.04576636478304863,
0.02868461236357689,
0.03308387100696564,
0.06339066475629807,
-0.07358020544052124,
0.04162554070353508,
0.04395408555865288,
0.013530542142689228,
-0.023599015548825264,
-0.09068866819143295,
0.005259404424577951,
-0.07052215188741684,
0.03645392507314682,
-0.03809727728366852,
0.025492222979664803,
0.07123022526502609,
0.03868236392736435,
0.007793566212058067,
-0.032029371708631516,
-0.008413640782237053,
0.07155155390501022,
0.0625900849699974,
0.04779503494501114,
-0.010267610661685467,
-0.03259497508406639,
0.010617702268064022
] | 0 | 0 | 0 | 0 |
|
Adversarial Generative Flow Network for Solving Vehicle Routing Problems | https://openreview.net/forum?id=tBom4xOW1H | [
"Ni Zhang",
"Jingfeng Yang",
"Zhiguang Cao",
"Xu Chi"
] | Poster | Recent research into solving vehicle routing problems (VRPs) has gained significant traction, particularly through the application of deep (reinforcement) learning for end-to-end solution construction. However, many current construction-based neural solvers predominantly utilize Transformer architectures, which can face scalability challenges and struggle to produce diverse solutions. To address these limitations, we introduce a novel framework beyond Transformer-based approaches, i.e., Adversarial Generative Flow Networks (AGFN). This framework integrates the generative flow network (GFlowNet)—a probabilistic model inherently adept at generating diverse solutions (routes)—with a complementary model for discriminating (or evaluating) the solutions. These models are trained alternately in an adversarial manner to improve the overall solution quality, followed by a proposed hybrid decoding method to construct the solution. We apply the AGFN framework to solve the capacitated vehicle routing problem (CVRP) and travelling salesman problem (TSP), and our experimental results demonstrate that AGFN surpasses the popular construction-based neural solvers, showcasing strong generalization capabilities on synthetic and real-world benchmark instances. | Generative Flow Network, Adversarial Training, Vehicle Routing Problem | null | 13,449 | 2503.01931 | [
-0.13143044710159302,
-0.062401726841926575,
0.04668041691184044,
0.02460525929927826,
0.016891004517674446,
0.005789215210825205,
-0.009485458955168724,
-0.024468090385198593,
-0.010172157548367977,
-0.04736277833580971,
-0.0743754580616951,
-0.05093683674931526,
-0.019377097487449646,
0.07351336628198624,
-0.06475301831960678,
0.0472189337015152,
0.07252126932144165,
0.08544409275054932,
-0.05645429342985153,
-0.04758330434560776,
0.011173268780112267,
0.058126747608184814,
0.02431309223175049,
-0.002464874414727092,
0.041687943041324615,
0.007535790093243122,
0.03292205184698105,
-0.05230167508125305,
-0.0010504797101020813,
0.0052557215094566345,
0.07810822874307632,
-0.0043625785037875175,
-0.09741248190402985,
0.045285914093256,
-0.07781321555376053,
0.1027773842215538,
-0.14542008936405182,
0.03181372210383415,
-0.004679626785218716,
-0.11131797730922699,
0.023465868085622787,
-0.010497733019292355,
-0.028186479583382607,
-0.0019944007508456707,
0.08096595108509064,
-0.039522476494312286,
0.024762634187936783,
-0.0892874076962471,
0.06373507529497147,
-0.018855247646570206,
0.023478766903281212,
-0.01208546943962574,
-0.03001343086361885,
0.017794420942664146,
-0.0057013449259102345,
0.002933854004368186,
0.026425037533044815,
-0.027433358132839203,
-0.0427037812769413,
0.04117189347743988,
-0.0028029608074575663,
-0.12090269476175308,
-0.005177625920623541,
-0.06295955181121826,
-0.05667988955974579,
-0.035997424274683,
0.017388131469488144,
0.08445362746715546,
-0.03513181209564209,
0.01460140012204647,
0.05570308119058609,
0.0650366023182869,
-0.011888018809258938,
0.055075060576200485,
0.08977464586496353,
0.07435557246208191,
0.06319979578256607,
-0.03516293689608574,
0.023663824424147606,
0.0016713719815015793,
0.034723278135061264,
0.012800092808902264,
0.03574081510305405,
0.06315282732248306,
0.09120435267686844,
-0.003099314169958234,
-0.017421646043658257,
0.11950583010911942,
0.014021670445799828,
0.009608243592083454,
-0.09267760068178177,
0.014482480473816395,
0.02220209315419197,
-0.041667040437459946,
0.03733737766742706,
-0.00043248635483905673,
-0.073074109852314,
-0.05556256324052811,
-0.0200662761926651,
0.0795241966843605,
-0.0249653123319149,
0.026662981137633324,
0.0888398215174675,
-0.06398461014032364,
0.01142555009573698,
0.03821771591901779,
0.027977578341960907,
0.05087181180715561,
0.008491733111441135,
-0.06867537647485733,
0.05945068597793579,
0.016298523172736168,
0.08698224276304245,
-0.0381711982190609,
-0.0639071986079216,
0.006411380134522915,
-0.10366061329841614,
0.0027570317033678293,
-0.05890674889087677,
0.10495419800281525,
-0.08889708667993546,
-0.014863464049994946,
0.01786893419921398,
0.029699115082621574,
0.013473362661898136,
-0.012830816209316254,
0.05129329115152359,
9.89090793318124e-35,
-0.006211067084223032,
0.023916348814964294,
0.06611412018537521,
0.07900193333625793,
0.13457684218883514,
0.06220021843910217,
-0.009615188464522362,
0.08700061589479446,
-0.012667044997215271,
-0.01787874847650528,
-0.06199657917022705,
0.009516309946775436,
-0.09045121818780899,
0.05523009970784187,
0.025324594229459763,
-0.0318121574819088,
-0.11386015266180038,
-0.033296260982751846,
-0.010853164829313755,
-0.08715540915727615,
0.03983229771256447,
-0.07073759287595749,
0.010208559222519398,
-0.00094836694188416,
0.07469278573989868,
-0.04145852476358414,
0.044113028794527054,
-0.0020351216662675142,
0.019326668232679367,
-0.024292489513754845,
-0.11115983128547668,
0.028727827593684196,
0.002572976052761078,
0.0829690471291542,
0.003820577636361122,
-0.001400110311806202,
-0.04648541659116745,
-0.07119325548410416,
-0.023060087114572525,
0.02529183402657509,
0.014837623573839664,
0.03234393894672394,
-0.04475004971027374,
0.08419220894575119,
-0.0780668780207634,
-0.058693647384643555,
0.033432383090257645,
-0.01733982376754284,
-0.07554223388433456,
0.04072076827287674,
-0.019569184631109238,
0.01755322329699993,
0.034621115773916245,
-0.02825927548110485,
-0.022074928507208824,
-0.022805802524089813,
0.047123778611421585,
0.07805297523736954,
0.022462746128439903,
0.0059968214482069016,
0.04294035956263542,
0.021076910197734833,
-0.04809628054499626,
0.05682987719774246,
0.003237021854147315,
0.01760193705558777,
-0.0609719380736351,
0.08836834132671356,
0.09031333029270172,
-0.02618696540594101,
0.01613517850637436,
0.021211568266153336,
0.007705298252403736,
0.024038933217525482,
0.07549451291561127,
0.024177508428692818,
-0.0027426492888480425,
0.021419718861579895,
-0.009211276657879353,
-0.025604890659451485,
-0.08092781156301498,
0.08141269534826279,
-0.03782626613974571,
0.012349025346338749,
0.04497186094522476,
0.048994030803442,
0.01730906032025814,
-0.06917654722929001,
0.030821355059742928,
0.02541213668882847,
-0.06131843850016594,
-0.027359968051314354,
-0.06075301393866539,
0.030935999006032944,
0.07562095671892166,
-4.440255422788491e-34,
-0.029874349012970924,
0.052089981734752655,
-0.08083827793598175,
0.020869750529527664,
0.027319444343447685,
-0.03817976638674736,
-0.04944000393152237,
-0.028293035924434662,
0.07357604801654816,
-0.01051407866179943,
-0.07254324853420258,
0.03845377266407013,
0.09686695039272308,
-0.0619596503674984,
0.05037955939769745,
-0.09830673038959503,
-0.002326466143131256,
-0.04844607785344124,
-0.010726174339652061,
0.02391708455979824,
-0.0008516263915225863,
0.13361811637878418,
-0.11104695498943329,
-0.06859501451253891,
0.022367415949702263,
-0.022196045145392418,
-0.01479251217097044,
0.04519615322351456,
0.02357633225619793,
0.07147849351167679,
0.035583969205617905,
-0.061164695769548416,
-0.028494849801063538,
0.003565956139937043,
-0.02195589803159237,
0.06927226483821869,
0.04474702849984169,
0.026421748101711273,
-0.019594890996813774,
0.09211338311433792,
0.009654445573687553,
-0.12274374812841415,
-0.03353101760149002,
0.06365031003952026,
-0.04485301673412323,
-0.000571860873606056,
-0.001218329300172627,
-0.07495507597923279,
0.025567544624209404,
-0.002366219414398074,
0.01328206155449152,
0.058577582240104675,
-0.11913681030273438,
0.031272776424884796,
-0.02539258450269699,
-0.03809206932783127,
0.05090504139661789,
-0.00977338943630457,
0.05374961718916893,
-0.007153395563364029,
0.015182668343186378,
-0.04226727783679962,
0.06954288482666016,
-0.04011203348636627,
-0.030842531472444534,
-0.057801343500614166,
-0.0020121412817388773,
0.010695894248783588,
-0.009923533536493778,
-0.0017997027607634664,
-0.00472290487959981,
-0.019480710849165916,
0.07385416328907013,
0.07012313604354858,
0.00657969806343317,
-0.04176044836640358,
0.05089516565203667,
0.017857203260064125,
0.01560046337544918,
-0.11419640481472015,
0.006432239431887865,
-0.07895774394273758,
0.07242338359355927,
0.03577195480465889,
0.05139237269759178,
0.02514997124671936,
0.032164085656404495,
0.00742965517565608,
0.01408617477864027,
0.06842011958360672,
-0.02897520922124386,
0.03266399726271629,
-0.050320979207754135,
-0.019635168835520744,
-0.0992804765701294,
-3.967870298993148e-8,
-0.05564134940505028,
0.07139632105827332,
-0.004609858151525259,
-0.053557537496089935,
0.05015059933066368,
0.005933952052146196,
-0.002312909346073866,
0.09088943898677826,
-0.03852986544370651,
0.00783429853618145,
0.062188323587179184,
0.03255882486701012,
-0.01970265433192253,
-0.06743263453245163,
-0.04537425935268402,
0.0320311114192009,
0.028243189677596092,
0.02557966485619545,
-0.007436772808432579,
-0.013871932402253151,
-0.0350811742246151,
0.025122296065092087,
-0.04348248988389969,
-0.011328360065817833,
-0.002558312378823757,
-0.058311305940151215,
-0.053356919437646866,
-0.011312373913824558,
0.07841239869594574,
0.000025136179829132743,
-0.039763979613780975,
0.029196519404649734,
0.036702461540699005,
0.0038464840035885572,
0.025581667199730873,
0.09386616200208664,
0.009905689395964146,
-0.012067441828548908,
-0.014749589376151562,
-0.05861193314194679,
0.053603820502758026,
0.03413199260830879,
-0.01841978169977665,
-0.013012145645916462,
-0.016013748943805695,
0.03923431783914566,
-0.0035577653907239437,
-0.03788699582219124,
0.01663077250123024,
0.02182186022400856,
0.04944479465484619,
-0.043552082031965256,
-0.03227376565337181,
0.0913378968834877,
0.06220577657222748,
-0.05332136154174805,
0.008328975178301334,
-0.11181093007326126,
0.048935409635305405,
0.0682963877916336,
-0.08004261553287506,
0.025298694148659706,
0.005940989125519991,
-0.022970665246248245
] | 0 | 0 | 0 | 0 |
|
Tracking the Copyright of Large Vision-Language Models through Parameter Learning Adversarial Images | https://openreview.net/forum?id=K7xpl3LZQp | [
"Yubo Wang",
"Jianting Tang",
"Chaohu Liu",
"Linli Xu"
] | Poster | Large vision-language models (LVLMs) have demonstrated remarkable image understanding and dialogue capabilities, allowing them to handle a variety of visual question answering tasks. However, their widespread availability raises concerns about unauthorized usage and copyright infringement, where users or individuals can develop their own LVLMs by fine-tuning published models. In this paper, we propose a novel method called Parameter Learning Attack (PLA) for tracking the copyright of LVLMs without modifying the original model. Specifically, we construct adversarial images through targeted attacks against the original model, enabling it to generate specific outputs. To ensure these attacks remain effective on potential fine-tuned models to trigger copyright tracking, we allow the original model to learn the trigger images by updating parameters in the opposite direction during the adversarial attack process. Notably, the proposed method can be applied after the release of the original model, thus not affecting the model’s performance and behavior. To simulate real-world applications, we fine-tune the original model using various strategies across diverse datasets, creating a range of models for copyright verification. Extensive experiments demonstrate that our method can more effectively identify the original copyright of fine-tuned models compared to baseline methods. Therefore, this work provides a powerful tool for tracking copyrights and detecting unlicensed usage of LVLMs. | Copyright Tracking, Large Vision-Language Models, Adversarial Attacks, Fine-tuning | By allowing model parameter updates during adversarial attacks, the adversarial images can more effectively trigger fine-tuned models to output specific texts, thereby achieving the copyright tracking. | 13,438 | 2502.16593 | [
-0.11608260869979858,
-0.06525403261184692,
-0.04284624755382538,
0.007266966626048088,
0.1082671582698822,
-0.008549371734261513,
0.0012578986352309585,
-0.05725633725523949,
0.07073408365249634,
-0.03126389533281326,
-0.02158733457326889,
-0.04144937917590141,
0.062356069684028625,
-0.009961755014955997,
-0.03623851388692856,
-0.00005733411671826616,
0.058093637228012085,
0.09464330971240997,
-0.10280190408229828,
-0.030732622370123863,
0.01735644042491913,
0.0069205667823553085,
0.04322255402803421,
0.012510061264038086,
0.014644252136349678,
-0.039754245430231094,
-0.00554110249504447,
0.0014342352515086532,
0.014815215021371841,
-0.04929826036095619,
0.0294323917478323,
-0.04634450003504753,
-0.009249481372535229,
0.054848480969667435,
-0.06974288076162338,
0.02747992053627968,
-0.023303832858800888,
-0.016197126358747482,
0.05319177359342575,
-0.0632961094379425,
-0.04089422523975372,
-0.0416669063270092,
-0.019315993413329124,
-0.02034302055835724,
0.06120386719703674,
-0.037830792367458344,
0.07383347302675247,
0.05430319905281067,
0.012030631303787231,
0.031117482110857964,
-0.03171711415052414,
-0.09067980200052261,
-0.046891383826732635,
0.015818234533071518,
-0.047576963901519775,
-0.06731803715229034,
0.009914970025420189,
0.025883030146360397,
0.0048439051024615765,
0.021143531426787376,
-0.08610720187425613,
-0.03650263324379921,
-0.04334157332777977,
0.05932118743658066,
-0.08139805495738983,
-0.011344140395522118,
0.011653373949229717,
0.05489460378885269,
-0.06418478488922119,
0.040654536336660385,
0.026251958683133125,
0.07791111618280411,
0.00631834426894784,
0.004800053779035807,
0.01773887127637863,
0.06442615389823914,
-0.016945449635386467,
-0.026282276958227158,
0.019171424210071564,
-0.15222769975662231,
0.05172779783606529,
-0.06467598676681519,
0.057872045785188675,
0.04664165526628494,
0.05745188146829605,
0.039475396275520325,
0.011486472561955452,
0.05849751457571983,
0.1490357220172882,
-0.01378556713461876,
-0.08585464209318161,
-0.0037495093420147896,
0.10761938244104385,
0.012318110093474388,
0.05641007795929909,
0.05461778864264488,
-0.03862990066409111,
-0.06919287890195847,
-0.020856618881225586,
0.08311637490987778,
-0.03335408866405487,
0.025001471862196922,
0.04242418706417084,
-0.06027105450630188,
0.1170564666390419,
0.04492640495300293,
0.08542558550834656,
0.0018565442878752947,
-0.0038228894118219614,
-0.0246074628084898,
0.04111642763018608,
0.08807610720396042,
-0.01755566895008087,
-0.0034450970124453306,
0.1224796324968338,
-0.03953631594777107,
0.007010961417108774,
-0.010381769388914108,
-0.048781462013721466,
-0.008371254429221153,
0.02207699790596962,
0.02401427924633026,
-0.013495301827788353,
-0.012694960460066795,
-0.040197402238845825,
-0.06432182341814041,
-0.06463287025690079,
-7.715681350923522e-34,
0.0590205043554306,
0.025616826489567757,
0.030866490676999092,
-0.020661834627389908,
0.06492547690868378,
0.0077789658680558205,
-0.015992166474461555,
0.09824036061763763,
-0.05660582333803177,
-0.01460623275488615,
-0.01680341176688671,
-0.03934761881828308,
-0.045031413435935974,
0.11127128452062607,
-0.00023281345784198493,
0.05355433002114296,
-0.03596532344818115,
0.0297970250248909,
-0.002325282897800207,
0.03527441248297691,
0.06620297580957413,
-0.06813407689332962,
0.022353794425725937,
0.04758455604314804,
-0.020860953256487846,
-0.03610064834356308,
0.04667433723807335,
-0.0570613332092762,
-0.015350241214036942,
0.03080877847969532,
-0.11660269647836685,
-0.022666214033961296,
0.03748444840312004,
0.034067969769239426,
0.0232530664652586,
-0.020463479682803154,
0.007963735610246658,
0.006475202739238739,
0.03430800139904022,
0.049831684678792953,
0.020338062196969986,
-0.009765209630131721,
0.0005950300255790353,
0.008392172865569592,
-0.0957280769944191,
0.023597706109285355,
-0.0543183758854866,
0.046931687742471695,
-0.07191874831914902,
-0.005980268120765686,
0.033808112144470215,
0.015796354040503502,
-0.05926152691245079,
-0.04991339519619942,
-0.08054253458976746,
-0.05086158215999603,
-0.05667441710829735,
-0.041975781321525574,
-0.0024220433551818132,
-0.011200210079550743,
-0.01818804442882538,
0.07362014800310135,
0.006619366817176342,
0.12120357155799866,
-0.019989123567938805,
-0.004864810965955257,
0.030791662633419037,
-0.014631915837526321,
0.010953848250210285,
0.005168888717889786,
-0.004867075942456722,
-0.027242785319685936,
-0.0463876910507679,
-0.01759985461831093,
-0.028841082006692886,
-0.04668077453970909,
0.03509809076786041,
0.04541029408574104,
0.07287652790546417,
0.07232259213924408,
-0.002761109499260783,
-0.07378842681646347,
0.07285410910844803,
-0.06689950823783875,
-0.09192834049463272,
-0.06285912543535233,
0.009709110483527184,
-0.09141997992992401,
-0.019515911117196083,
-0.008009528741240501,
-0.03344142436981201,
-0.04075169935822487,
-0.06066424399614334,
-0.026290422305464745,
0.033695004880428314,
-1.9952198262384402e-34,
-0.022552821785211563,
0.07917960733175278,
-0.023212948814034462,
0.03784724324941635,
-0.0720861405134201,
-0.05005090683698654,
-0.05868198350071907,
0.10331068933010101,
-0.004065643530339003,
-0.0651167705655098,
-0.06290844082832336,
-0.03191256523132324,
-0.027699023485183716,
-0.004597240127623081,
0.012680038809776306,
-0.10949964821338654,
0.004117847420275211,
0.010494607500731945,
-0.08820681273937225,
-0.007071053143590689,
0.008963419124484062,
0.06220133230090141,
-0.0891544371843338,
0.043399613350629807,
0.011754908598959446,
-0.015329877845942974,
-0.055217694491147995,
0.037173401564359665,
0.03191386163234711,
0.047891996800899506,
-0.05419134721159935,
0.031903158873319626,
-0.04229871928691864,
-0.058566126972436905,
-0.060272399336099625,
0.04074454307556152,
0.09805889427661896,
0.018615851178765297,
0.02324564754962921,
0.09746123105287552,
0.06496712565422058,
-0.057171717286109924,
-0.05419036000967026,
0.001809823326766491,
-0.07419252395629883,
-0.09135042130947113,
0.014219765551388264,
-0.03330250456929207,
0.015001986175775528,
-0.030270250514149666,
0.004965875763446093,
-0.03865808621048927,
0.004577485378831625,
-0.006899725645780563,
-0.033603254705667496,
-0.05799845978617668,
-0.06454049050807953,
0.0442800335586071,
0.10454997420310974,
0.011928838677704334,
0.017401378601789474,
-0.06967485696077347,
-0.04512583091855049,
-0.03134627640247345,
-0.0028768745250999928,
-0.02338324673473835,
-0.02191942371428013,
0.04361583665013313,
0.003994268830865622,
0.020351329818367958,
0.06425240635871887,
-0.040461260825395584,
0.001009347033686936,
0.050763342529535294,
-0.005923835560679436,
-0.06802598387002945,
0.012635274790227413,
-0.038701582700014114,
0.01820366457104683,
0.006642351858317852,
0.11106990277767181,
-0.07791954278945923,
0.09333442151546478,
0.056045062839984894,
0.06960711628198624,
0.013894283212721348,
0.05243990942835808,
0.016660962253808975,
-0.002821087371557951,
0.048493802547454834,
-0.003517830278724432,
-0.0031123992521315813,
-0.04449225962162018,
0.09359130263328552,
0.017776980996131897,
-3.540724335948653e-8,
-0.09347706288099289,
0.040470391511917114,
0.006733275484293699,
0.08881768584251404,
0.0023987300228327513,
0.04548558220267296,
-0.0842127874493599,
0.054845768958330154,
0.02687288448214531,
-0.07442347705364227,
0.009393136948347092,
0.003637149930000305,
0.0235603004693985,
-0.0679042860865593,
-0.10861018300056458,
0.08302526921033859,
0.018592331558465958,
0.04887530580163002,
-0.005176535341888666,
-0.008846440352499485,
-0.019870519638061523,
-0.0010191674809902906,
0.051833074539899826,
-0.07194657623767853,
-0.0021022085566073656,
-0.04487480968236923,
-0.028761211782693863,
0.010735535062849522,
0.06186193600296974,
0.10261744260787964,
0.013813216239213943,
-0.014589846134185791,
0.006067870184779167,
-0.005342183634638786,
0.01693415269255638,
0.07690184563398361,
-0.04177568852901459,
-0.05029106140136719,
0.015015630051493645,
-0.016339603811502457,
0.01805899292230606,
0.0539470873773098,
0.03733456879854202,
-0.013600506819784641,
0.01836930401623249,
-0.007576731964945793,
0.014024763368070126,
-0.1265692412853241,
0.06319517642259598,
-0.0025577470660209656,
0.04408221319317818,
-0.028473153710365295,
0.055098600685596466,
0.12380340695381165,
0.0002965512976516038,
-0.07194672524929047,
0.0392204150557518,
0.011325164698064327,
0.04096442833542824,
0.13920238614082336,
0.06648105382919312,
0.021464303135871887,
0.047824595123529434,
0.014736335724592209
] | 0 | 0 | 0 | 0 |
|
Self-Supervised Diffusion MRI Denoising via Iterative and Stable Refinement | https://openreview.net/forum?id=wxPnuFp8fZ | [
"Chenxu Wu",
"Qingpeng Kong",
"Zihang Jiang",
"S Kevin Zhou"
] | Poster | Magnetic Resonance Imaging (MRI), including diffusion MRI (dMRI), serves as a ``microscope'' for anatomical structures and routinely mitigates the influence of low signal-to-noise ratio scans by compromising temporal or spatial resolution. However, these compromises fail to meet clinical demands for both efficiency and precision. Consequently, denoising is a vital preprocessing step, particularly for dMRI, where clean data is unavailable. In this paper, we introduce Di-Fusion, a fully self-supervised denoising method that leverages the latter diffusion steps and an adaptive sampling process. Unlike previous approaches, our single-stage framework achieves efficient and stable training without extra noise model training and offers adaptive and controllable results in the sampling process. Our thorough experiments on real and simulated data demonstrate that Di-Fusion achieves state-of-the-art performance in microstructure modeling, tractography tracking, and other downstream tasks. Code is available at https://github.com/FouierL/Di-Fusion. | Diffusion based models, Self-supervised MRI denoising | null | 13,426 | 2501.13514 | [
-0.02278451807796955,
-0.1367643028497696,
0.06230827420949936,
-0.02914028987288475,
-0.02806015871465206,
-0.11548520624637604,
0.012167824432253838,
-0.11122000217437744,
-0.02939831092953682,
-0.04082712158560753,
-0.05821089446544647,
0.008486696518957615,
-0.016721369698643684,
0.0866793841123581,
-0.05504810810089111,
-0.014283315278589725,
-0.0004835818544961512,
0.022788627073168755,
-0.06038559600710869,
-0.0001148838855442591,
-0.07927309721708298,
-0.014339637011289597,
-0.03242570906877518,
0.0196806900203228,
0.049230705946683884,
0.023758521303534508,
0.046390414237976074,
-0.11073865741491318,
-0.014399551786482334,
-0.07027904689311981,
0.13866007328033447,
-0.027531959116458893,
-0.05449169874191284,
-0.013918178156018257,
-0.018211523070931435,
-0.06189543008804321,
-0.007552549242973328,
0.1095903217792511,
0.017073383554816246,
-0.001673499122262001,
0.012934020720422268,
0.08313196152448654,
-0.0011381799122318625,
-0.009850907139480114,
0.06081625446677208,
0.0367514044046402,
0.027044977992773056,
-0.0667923092842102,
0.09891851991415024,
0.05039336159825325,
-0.020479971542954445,
0.014815343543887138,
-0.06606151163578033,
0.07157646119594574,
0.0035824552178382874,
0.05531441792845726,
0.06052205339074135,
-0.024157080799341202,
-0.03285406529903412,
0.0297846682369709,
0.04749337583780289,
-0.03640967607498169,
0.0011146910255774856,
0.017179934307932854,
-0.009896166622638702,
0.008313064463436604,
0.04643310233950615,
0.04128091782331467,
-0.007993936538696289,
0.05538139119744301,
0.003283858299255371,
-0.011998944915831089,
-0.01378573663532734,
0.030870774760842323,
-0.015374748036265373,
-0.036168236285448074,
-0.021118316799402237,
-0.0063822162337601185,
0.08321364969015121,
-0.004422259517014027,
0.00927992258220911,
0.08269637078046799,
0.07168231904506683,
-0.0389837883412838,
0.023293474689126015,
-0.004971638787537813,
-0.013488787226378918,
0.03411697968840599,
0.0031884037889540195,
-0.005357422400265932,
0.014408731833100319,
0.0461852140724659,
-0.16741296648979187,
-0.017591876909136772,
-0.0385875441133976,
-0.009440598078072071,
0.019230538979172707,
0.06721112132072449,
0.05904288962483406,
0.061933115124702454,
-0.03597700968384743,
-0.01233825832605362,
-0.05738554149866104,
0.05543448030948639,
0.0038429819978773594,
-0.09041354060173035,
0.10371160507202148,
0.012494251132011414,
0.028621738776564598,
0.054161347448825836,
0.07557544112205505,
-0.003379363566637039,
0.004345031920820475,
0.02190382033586502,
0.10703197866678238,
-0.08129923790693283,
0.0873110443353653,
0.02345513366162777,
0.04496330767869949,
0.01921674981713295,
-0.01809971034526825,
-0.05376916378736496,
-0.05783601477742195,
-0.029674552381038666,
0.06839033961296082,
0.0296656284481287,
-0.06356213241815567,
2.715666308944256e-34,
0.005475039593875408,
0.003710281802341342,
0.03438045084476471,
0.00421521719545126,
0.019072098657488823,
-0.003272436559200287,
-0.042658090591430664,
-0.053135041147470474,
0.03962837532162666,
0.06252650171518326,
0.010903402231633663,
0.006503124255686998,
-0.046880465000867844,
0.03622856363654137,
-0.00719440495595336,
-0.04459366574883461,
0.004258213099092245,
0.07790248841047287,
-0.045425958931446075,
-0.08379678428173065,
0.035983506590127945,
0.03201259672641754,
-0.05192851275205612,
0.004611540120095015,
0.011536793783307076,
0.007020770106464624,
-0.004614936653524637,
-0.01083784643560648,
0.06254325062036514,
0.01831521838903427,
-0.12857747077941895,
-0.041871026158332825,
-0.04889432340860367,
0.040136076509952545,
0.012225118465721607,
0.009854336269199848,
-0.020238732919096947,
0.016190966591238976,
-0.03216303512454033,
0.0306109506636858,
0.05913157761096954,
0.00454720901325345,
0.02418370172381401,
-0.05453524738550186,
-0.07571418583393097,
-0.003436256665736437,
-0.022827623412013054,
-0.013108924962580204,
-0.007561883423477411,
-0.08869403600692749,
0.001041680108755827,
-0.0382046103477478,
-0.031223295256495476,
-0.09599554538726807,
-0.016797397285699844,
0.027569923549890518,
0.029946837574243546,
-0.05161045864224434,
0.0666184276342392,
0.007484796456992626,
0.048972390592098236,
-0.06895166635513306,
-0.08140621334314346,
0.06885841488838196,
0.08712831139564514,
-0.014445047825574875,
-0.042144402861595154,
0.04228544607758522,
0.02381931245326996,
0.010301783680915833,
-0.035909004509449005,
0.03424980491399765,
0.00014152390940580517,
-0.07942087203264236,
0.07087568938732147,
0.030648136511445045,
0.05904918536543846,
0.06767480075359344,
0.0033202271442860365,
-0.05266839638352394,
-0.0626237764954567,
0.00941783469170332,
-0.08026029914617538,
-0.029800619930028915,
-0.018649067729711533,
0.07837376743555069,
0.026430362835526466,
-0.018663788214325905,
-0.03826488181948662,
-0.1014721617102623,
-0.06608881801366806,
-0.014487780630588531,
0.0011389428982511163,
-0.01466779038310051,
-0.04757135733962059,
-7.622647401540669e-34,
0.0003732255136128515,
0.06607841700315475,
0.04333650693297386,
0.032063983380794525,
-0.0060975500382483006,
0.026327447965741158,
0.0033863787539303303,
0.09490232914686203,
0.09273095428943634,
-0.00005820730802952312,
-0.07463473081588745,
-0.04689320921897888,
0.00649528531357646,
-0.00556194456294179,
-0.020040059462189674,
-0.028844444081187248,
-0.03885415196418762,
-0.0540604367852211,
-0.06367748975753784,
0.016088038682937622,
-0.03720071166753769,
0.13574786484241486,
0.02780895307660103,
0.03193792328238487,
-0.08733977377414703,
0.04775206744670868,
-0.044730111956596375,
0.029753005132079124,
0.08251719921827316,
-0.005633978638797998,
-0.07045424729585648,
-0.01167482603341341,
0.0412060022354126,
-0.09513791650533676,
-0.05817878246307373,
0.09545383602380753,
-0.01912952959537506,
-0.0666058138012886,
0.019724102690815926,
0.0009840233251452446,
-0.0035229355562478304,
-0.03199995681643486,
-0.06400715559720993,
0.025976527482271194,
0.017223091796040535,
-0.09299656748771667,
-0.051910050213336945,
0.050289444625377655,
0.026385895907878876,
0.05038632079958916,
-0.04882258176803589,
0.05646313354372978,
-0.10780560225248337,
0.04348931089043617,
-0.03843552619218826,
-0.06456425040960312,
-0.005447037052363157,
-0.024329127743840218,
0.06884288042783737,
0.006881908979266882,
-0.12480904906988144,
-0.011643927544355392,
-0.03399505093693733,
-0.0069113243371248245,
-0.05430959165096283,
0.06829146295785904,
-0.01569533534348011,
0.02335650846362114,
-0.015154819004237652,
0.12540289759635925,
-0.05514954403042793,
0.04775690287351608,
-0.0016627738950774074,
-0.00719706853851676,
0.03251711651682854,
-0.08957894146442413,
0.011889621615409851,
-0.05424733832478523,
0.008880922570824623,
0.03491959348320961,
0.06645894795656204,
-0.0509641207754612,
-0.019994523376226425,
0.0649307519197464,
0.10682413727045059,
0.0019000619649887085,
-0.019822286441922188,
-0.06452067196369171,
-0.014703907072544098,
-0.013289352878928185,
0.0893036499619484,
-0.0194978266954422,
0.042582787573337555,
0.008034851402044296,
-0.0799146369099617,
-4.019563704105167e-8,
-0.0812954306602478,
0.06312210857868195,
0.01136149000376463,
0.0022530832793563604,
0.015089738182723522,
-0.10759169608354568,
0.05741160735487938,
0.04574209824204445,
0.045198261737823486,
-0.037534527480602264,
0.12882210314273834,
-0.09965799003839493,
-0.0897059440612793,
-0.037759218364953995,
0.022251572459936142,
0.055119309574365616,
0.008577943779528141,
0.03463657945394516,
-0.005412389524281025,
0.000380921148462221,
-0.03811458870768547,
-0.0003105686919298023,
0.0086827352643013,
-0.02876964397728443,
0.0555984303355217,
0.06658867001533508,
0.016995349898934364,
-0.00954508688300848,
0.016355669125914574,
-0.009998612105846405,
0.033335305750370026,
0.01636406034231186,
0.04979966580867767,
0.060114651918411255,
-0.046389006078243256,
0.012639258988201618,
0.10012778639793396,
0.05360151082277298,
-0.09900461882352829,
-0.04355054721236229,
-0.056744787842035294,
0.014793110080063343,
-0.041792355477809906,
-0.006755944807082415,
-0.06142013892531395,
0.04221010208129883,
0.010885708965361118,
-0.02617708407342434,
0.021556882187724113,
0.00891572143882513,
0.036566510796546936,
-0.020266372710466385,
0.12404848635196686,
-0.005280076991766691,
0.06007436662912369,
0.05050127953290939,
0.032482050359249115,
0.043511778116226196,
0.03504762426018715,
0.08930910378694534,
0.004465488251298666,
-0.0012813621433451772,
-0.014934911392629147,
0.015529036521911621
] | https://github.com/fouierl/di-fusion | 15 | 0 | 0 | 0 |
LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement | https://openreview.net/forum?id=fMbLszVO1H | [
"Zhifan Ye",
"Kejing Xia",
"Yonggan Fu",
"Xin Dong",
"Jihoon Hong",
"Xiangchi Yuan",
"Shizhe Diao",
"Jan Kautz",
"Pavlo Molchanov",
"Yingyan Celine Lin"
] | Poster | State space models (SSMs) have emerged as an efficient alternative to Transformer models for language modeling, offering linear computational complexity and constant memory usage as context length increases. However, despite their efficiency in handling long contexts, recent studies have shown that SSMs, such as Mamba models, generally underperform compared to Transformers in long-context understanding tasks. To address this significant shortfall and achieve both efficient and accurate long-context understanding, we propose LongMamba, a training-free technique that significantly enhances the long-context capabilities of Mamba models. LongMamba builds on our discovery that the hidden channels in Mamba can be categorized into local and global channels based on their receptive field lengths, with global channels primarily responsible for long-context capability. These global channels can become the key bottleneck as the input context lengthens. Specifically, when input lengths largely exceed the training sequence length, global channels exhibit limitations in adaptively extend their receptive fields, leading to Mamba’s poor long-context performance. The key idea of LongMamba is to mitigate the hidden state memory decay in these global channels by preventing the accumulation of unimportant tokens in their memory. This is achieved by first identifying critical tokens in the global channels and then applying token filtering to accumulate only those critical tokens. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for Mamba’s long-context performance, significantly extending its operational range without requiring additional training. Our code is available at https://github.com/GATECH-EIC/LongMamba. | Large Language Models, State Space Models, Long Context Understanding | null | 13,423 | null | [
0.02180374227464199,
-0.08886180073022842,
0.02573208510875702,
0.018441082909703255,
0.041097600013017654,
0.022094326093792915,
-0.05275632068514824,
-0.05302976444363594,
0.027233360335230827,
-0.03947758302092552,
-0.030868228524923325,
-0.06144265830516815,
0.03868710622191429,
-0.01190491858869791,
0.08534320443868637,
-0.05377955362200737,
0.047201287001371384,
0.07033024728298187,
-0.07119053602218628,
-0.02791403792798519,
0.1229785606265068,
-0.001985653769224882,
0.031314898282289505,
0.010532191023230553,
-0.02435408905148506,
0.048265885561704636,
-0.005428329110145569,
-0.06424468755722046,
0.05244119465351105,
-0.05766937509179115,
0.09276732057332993,
0.10140970349311829,
0.05915059521794319,
0.00823344849050045,
-0.07199303060770035,
0.08637861162424088,
-0.12433996051549911,
0.03944263607263565,
-0.053339291363954544,
-0.05315013974905014,
-0.0008864905103109777,
0.03987264633178711,
0.05775928124785423,
0.06590083241462708,
0.03893190249800682,
-0.061608247458934784,
0.06077863648533821,
-0.006536904256790876,
0.02164623886346817,
0.028231648728251457,
-0.08934655785560608,
-0.0013298699632287025,
-0.012887123972177505,
0.11050842702388763,
-0.031202957034111023,
0.0008103853324428201,
-0.00815534871071577,
0.11287781596183777,
-0.05709519609808922,
0.039048220962285995,
-0.05856798216700554,
-0.08800166845321655,
-0.02561119943857193,
-0.027734601870179176,
-0.0997157171368599,
-0.01444422546774149,
0.029485858976840973,
0.06256308406591415,
-0.06163179501891136,
0.011421549133956432,
-0.019297439604997635,
0.019404448568820953,
-0.019750982522964478,
0.035933561623096466,
0.026865746825933456,
-0.00027072170632891357,
0.035045091062784195,
-0.02360176108777523,
0.07026262581348419,
-0.045538149774074554,
0.06563805788755417,
-0.01097848266363144,
0.03425716981291771,
-0.057424478232860565,
0.10671460628509521,
0.03443249687552452,
-0.007994935847818851,
-0.002782404189929366,
0.0034951530396938324,
0.003764895023778081,
-0.02313908189535141,
-0.15805189311504364,
0.000796047504991293,
-0.011067516170442104,
0.03274906054139137,
-0.033204637467861176,
-0.027318043634295464,
-0.0038178341928869486,
-0.07731415331363678,
0.05659489333629608,
0.0445714108645916,
0.01670728623867035,
-0.011124455370008945,
0.034278374165296555,
-0.024849804118275642,
-0.014149286784231663,
0.003878457937389612,
0.08624506741762161,
0.048338375985622406,
-0.01621408388018608,
-0.010047062300145626,
0.07955534756183624,
-0.012471618130803108,
0.05529756098985672,
0.03928528353571892,
0.01893593557178974,
0.039072487503290176,
-0.10420343279838562,
-0.03157632425427437,
0.03870726004242897,
-0.018499860540032387,
-0.049375224858522415,
-0.017544476315379143,
0.01955941691994667,
-0.015388011932373047,
-0.009635520167648792,
-0.06445492058992386,
2.108071508350203e-33,
-0.016414016485214233,
0.045458316802978516,
-0.0028677689842879772,
0.01675443910062313,
0.04646474868059158,
0.023683758452534676,
0.014428943395614624,
0.029879527166485786,
-0.014847319573163986,
0.04060325399041176,
0.010274561122059822,
0.0969570204615593,
-0.0583200640976429,
0.07902222871780396,
0.034477151930332184,
-0.01967841573059559,
-0.08633943647146225,
0.0513790138065815,
-0.033856794238090515,
-0.07778489589691162,
0.002776969224214554,
-0.0055620321072638035,
0.031590525060892105,
-0.028117870911955833,
0.05147644877433777,
0.015195580199360847,
0.05425780639052391,
-0.04754061996936798,
-0.06799622625112534,
0.00018197890312876552,
-0.10160684585571289,
0.02537398599088192,
-0.023646289482712746,
0.010008018463850021,
0.0444766990840435,
-0.11149027943611145,
0.03875518590211868,
-0.046614568680524826,
0.02178996428847313,
-0.06702647358179092,
-0.014776957221329212,
-0.010205251164734364,
0.04208918288350105,
-0.004253783263266087,
-0.06646841764450073,
-0.12253539264202118,
0.018673274666070938,
-0.005643583368510008,
-0.08899480104446411,
0.08594851940870285,
0.08825292438268661,
0.010943423956632614,
-0.03825259953737259,
-0.04431358352303505,
-0.01867195963859558,
0.031578078866004944,
0.024226408451795578,
0.04659661278128624,
0.01394819375127554,
0.03759244084358215,
-0.017748940736055374,
-0.013339399360120296,
0.021397123113274574,
0.09310203790664673,
0.02115483209490776,
0.03716481104493141,
-0.046164434403181076,
0.07985755801200867,
0.0773102194070816,
-0.067497618496418,
-0.004595546517521143,
0.0022690033074468374,
-0.029557326808571815,
0.010816114023327827,
0.00859068613499403,
-0.07562188804149628,
0.038379088044166565,
-0.0356626994907856,
-0.03161291405558586,
0.07997819781303406,
0.028570348396897316,
-0.01674349419772625,
-0.0090017756447196,
0.013770799152553082,
-0.023600168526172638,
-0.006541814189404249,
0.030966531485319138,
-0.1209191232919693,
-0.026270318776369095,
-0.07690900564193726,
-0.02311626449227333,
-0.010054036974906921,
0.009590325877070427,
-0.0503356009721756,
-0.009373430162668228,
-2.402635698986883e-33,
-0.025352347642183304,
-0.013168665580451488,
-0.04862289875745773,
0.03655695915222168,
-0.009282116778194904,
-0.09414496272802353,
0.019851211458444595,
0.1297634392976761,
-0.10105696320533752,
-0.09257421642541885,
-0.056607529520988464,
-0.03579464182257652,
0.04529604688286781,
0.004727891646325588,
0.0659937858581543,
-0.02503698132932186,
0.02536214329302311,
0.013975808396935463,
0.014852749183773994,
0.018351642414927483,
0.0681794136762619,
0.06158308684825897,
-0.11446989327669144,
0.020828327164053917,
-0.05917860567569733,
0.02144506201148033,
-0.06039990484714508,
0.03918115794658661,
0.019081691280007362,
-0.006059595849364996,
-0.06758376955986023,
0.02563897892832756,
-0.06134023517370224,
-0.044527817517519,
-0.08526943624019623,
0.014107489958405495,
-0.038934625685214996,
-0.014792620204389095,
-0.0403149351477623,
0.07694955915212631,
0.09982991963624954,
-0.04785553738474846,
-0.01767188124358654,
0.021261751651763916,
-0.04477392137050629,
-0.001403596717864275,
-0.04466037452220917,
0.023459820076823235,
0.011512752622365952,
0.057238753885030746,
-0.0026471028104424477,
0.0495133250951767,
-0.05553126335144043,
-0.031900666654109955,
-0.04588937759399414,
-0.06758931279182434,
-0.03434941917657852,
-0.003278205171227455,
0.0321987122297287,
-0.06115112826228142,
-0.022632911801338196,
-0.04137198626995087,
0.0016970105934888124,
-0.09099046885967255,
0.08350928127765656,
0.015226752497255802,
0.0391777902841568,
0.04055148735642433,
0.05114074796438217,
-0.031710073351860046,
0.040860194712877274,
0.05008099228143692,
-0.0997045636177063,
0.01988912560045719,
0.00008776001777732745,
0.03126313537359238,
-0.10619236528873444,
-0.09559723734855652,
0.006317188497632742,
-0.03088299371302128,
-0.039416395127773285,
0.013621081598103046,
-0.042537540197372437,
0.021543102338910103,
0.058119047433137894,
0.12099606543779373,
-0.006173120811581612,
0.06929001957178116,
0.007012849673628807,
0.016354793682694435,
0.016994763165712357,
0.05233300104737282,
0.009401161223649979,
0.059017933905124664,
-0.08544624596834183,
-3.8269686086778165e-8,
-0.08589891344308853,
0.0617198683321476,
-0.027993276715278625,
-0.0013406850630417466,
0.05479241535067558,
-0.011779069900512695,
-0.03391367197036743,
0.08993726968765259,
0.013192296959459782,
0.006903023459017277,
0.08080806583166122,
-0.011831365525722504,
0.02361409366130829,
0.012806898914277554,
0.029190532863140106,
0.07280688732862473,
0.007076167967170477,
0.03140741214156151,
-0.020002568140625954,
0.006709489971399307,
0.08802489936351776,
0.04729267582297325,
-0.029716873541474342,
0.06488526612520218,
-0.010916313156485558,
-0.061338070780038834,
-0.05302318558096886,
0.13878144323825836,
0.07216612994670868,
-0.03797456994652748,
-0.024967540055513382,
0.023742128163576126,
-0.030882567167282104,
-0.010280190967023373,
-0.06898223608732224,
0.021366417407989502,
0.04892671853303909,
-0.03191220760345459,
0.020325902849435806,
0.027082093060016632,
0.0049366336315870285,
0.03893144428730011,
-0.015341893769800663,
0.04424722120165825,
-0.07467211782932281,
0.06257107108831406,
0.019659217447042465,
-0.10418537259101868,
-0.01609947718679905,
-0.03971176967024803,
-0.014999949373304844,
0.03683434799313545,
-0.046837933361530304,
0.02721528150141239,
0.042258311063051224,
0.02606535144150257,
0.006195087917149067,
0.003796859411522746,
0.0408051535487175,
0.0692838653922081,
-0.08977735787630081,
0.07083217799663544,
-0.10583650320768356,
0.03557970002293587
] | 0 | 0 | 0 | 0 |
|
Robust Weight Initialization for Tanh Neural Networks with Fixed Point Analysis | https://openreview.net/forum?id=Es4RPNDtmq | [
"Hyun woo Lee",
"Hayoung Choi",
"Hyunju Kim"
] | Poster | As a neural network's depth increases, it can improve generalization performance. However, training deep networks is challenging due to gradient and signal propagation issues. To address these challenges, extensive theoretical research and various methods have been introduced. Despite these advances, effective weight initialization methods for tanh neural networks remain insufficiently investigated. This paper presents a novel weight initialization method for neural networks with tanh activation function. Based on an analysis of the fixed points of the function $\tanh(ax)$, the proposed method aims to determine values of $a$ that mitigate activation saturation. A series of experiments on various classification datasets and physics-informed neural networks demonstrates that the proposed method outperforms Xavier initialization methods (with or without normalization) in terms of robustness across different network sizes, data efficiency, and convergence speed. Code is available at https://github.com/1HyunwooLee/Tanh-Init. | Weight initialization, Signal propagation, Physics informed neural networks | null | 13,420 | 2410.02242 | [
-0.0737900510430336,
-0.05177491530776024,
0.07725290954113007,
0.040384307503700256,
-0.0635378360748291,
0.022937051951885223,
0.02698138728737831,
-0.011036388576030731,
-0.03607568144798279,
-0.05996018648147583,
0.03234207257628441,
0.007264112588018179,
-0.07664340734481812,
0.007947796024382114,
-0.0505010262131691,
-0.0031156602781265974,
0.058550622314214706,
0.008204806596040726,
-0.08858436346054077,
-0.041864797472953796,
0.026488879695534706,
0.040364816784858704,
-0.04277779534459114,
-0.0026699311565607786,
0.04029996693134308,
-0.05771501362323761,
0.014436201192438602,
0.01899760589003563,
0.04520862177014351,
-0.04584339261054993,
-0.01800619624555111,
-0.0630941167473793,
0.009428963996469975,
0.013849176466464996,
-0.16601422429084778,
0.057544734328985214,
-0.018458150327205658,
-0.03328719735145569,
-0.010564656928181648,
0.011397946625947952,
-0.01986268162727356,
-0.04217775538563728,
0.06253926455974579,
-0.01176108606159687,
0.11645058542490005,
0.08084803819656372,
0.10916201770305634,
-0.029065394774079323,
0.0646403580904007,
0.012991888448596,
0.04381106048822403,
0.04534103721380234,
0.03134481981396675,
0.056422747671604156,
0.014251282438635826,
0.04106420278549194,
0.037284962832927704,
0.05331982672214508,
-0.05837578698992729,
0.03880004212260246,
-0.01212142314761877,
-0.03924350440502167,
0.0012665829854086041,
0.006345470901578665,
0.025409448891878128,
-0.00490968581289053,
0.037321556359529495,
-0.027806617319583893,
0.01730192080140114,
0.06159772351384163,
0.02412399835884571,
0.08365844190120697,
-0.046269726008176804,
0.028804533183574677,
-0.00434093689545989,
0.10684359818696976,
0.06421931087970734,
0.03193911910057068,
0.007459735497832298,
-0.031355660408735275,
0.007931379601359367,
0.02227804809808731,
0.011194386519491673,
-0.003732027718797326,
0.0443294495344162,
0.04870108142495155,
-0.0519879125058651,
0.08865261822938919,
-0.004534139763563871,
0.013590812683105469,
-0.010203411802649498,
0.0566742867231369,
-0.0743793472647667,
-0.020926281809806824,
0.04317392036318779,
-0.014571083709597588,
0.0038824165239930153,
-0.05178790166974068,
-0.07710147649049759,
0.05194437876343727,
-0.030329888686537743,
0.0007347528007812798,
0.04501067474484444,
0.0821416825056076,
0.01389304269105196,
-0.023632511496543884,
0.03948966786265373,
-0.01808268204331398,
0.02919064089655876,
-0.06402875483036041,
0.015485201962292194,
-0.04272337630391121,
0.0496448278427124,
0.010148324072360992,
0.028956778347492218,
-0.01950920931994915,
0.003499696496874094,
-0.0035483436658978462,
-0.022738385945558548,
0.01969377137720585,
-0.01659969799220562,
-0.04427091032266617,
-0.016347553580999374,
0.07858800888061523,
-0.015248266980051994,
0.006856831721961498,
-0.08714953809976578,
3.478107786510679e-33,
-0.0009659555507823825,
0.061716701835393906,
0.015040933154523373,
-0.01791566237807274,
-0.016125144436955452,
-0.07974832504987717,
0.012524668127298355,
0.009428909048438072,
-0.09326149523258209,
0.06533349305391312,
-0.06487062573432922,
0.02132323943078518,
-0.0330747626721859,
0.047542545944452286,
-0.00036141512100584805,
-0.025846285745501518,
0.05577310919761658,
0.00661808717995882,
-0.053259916603565216,
-0.0768536925315857,
0.05463182181119919,
-0.053406234830617905,
0.003818125929683447,
-0.07873120903968811,
-0.056934405118227005,
0.03190430998802185,
0.020650245249271393,
0.004297588486224413,
-0.14962461590766907,
-0.005159804597496986,
-0.06252231448888779,
-0.023048685863614082,
0.022473972290754318,
-0.06438864022493362,
0.015441365540027618,
-0.025374364107847214,
0.03504263609647751,
0.038498181849718094,
0.056122541427612305,
-0.006483322940766811,
0.013780671171844006,
0.08163678646087646,
0.03690216690301895,
-0.0030178201850503683,
-0.08950194716453552,
-0.029214009642601013,
-0.05122460052371025,
0.015975981950759888,
0.04943738132715225,
-0.0014016492059454322,
-0.048517387360334396,
-0.02970890887081623,
-0.03564324975013733,
-0.018021728843450546,
0.019975023344159126,
-0.02519010193645954,
0.051543399691581726,
0.036527469754219055,
-0.047371890395879745,
0.06586788594722748,
-0.0038199725095182657,
-0.0223899707198143,
-0.06670331209897995,
0.020941592752933502,
-0.06449934840202332,
0.02847140282392502,
-0.02187599055469036,
-0.02501482143998146,
0.034612987190485,
0.029244786128401756,
0.02293296903371811,
0.04202616214752197,
-0.011602750048041344,
-0.04381774738430977,
0.06955312937498093,
0.016606681048870087,
-0.0232037715613842,
-0.04662913829088211,
-0.016043750569224358,
0.01925896666944027,
-0.05877619981765747,
0.09807848930358887,
0.0032285789493471384,
0.0012206174433231354,
-0.11636041849851608,
0.02037506364285946,
0.031420379877090454,
0.03091934509575367,
-0.03840101510286331,
-0.04034793749451637,
-0.056467823684215546,
0.04183291643857956,
0.03443990647792816,
-0.08766967803239822,
-0.027362532913684845,
-3.35316779803964e-33,
-0.04497907683253288,
0.07935250550508499,
-0.14546790719032288,
0.04399840161204338,
-0.025775432586669922,
0.04687587916851044,
-0.04317954555153847,
0.04370718076825142,
0.00665343226864934,
-0.020051956176757812,
0.07329468429088593,
0.021613050252199173,
-0.03982604667544365,
-0.05271201953291893,
0.09529034793376923,
0.0013346803607419133,
-0.10772635042667389,
-0.031436506658792496,
0.05075402557849884,
-0.03538922965526581,
0.06682759523391724,
0.0007688578334636986,
-0.10818619281053543,
0.01289043016731739,
0.03594057261943817,
-0.01783864200115204,
-0.06431109458208084,
0.08556251227855682,
0.04518305882811546,
-0.0038324189372360706,
-0.07220983505249023,
0.026158928871154785,
-0.0010756347328424454,
0.052794091403484344,
0.011945397593080997,
0.09174223989248276,
0.034586161375045776,
0.01837444119155407,
-0.007073680870234966,
0.04153452441096306,
0.10118630528450012,
0.012393015436828136,
-0.07683724910020828,
0.002827028278261423,
-0.00005672892075381242,
-0.044957660138607025,
-0.005983573850244284,
0.05444851890206337,
-0.05949978157877922,
0.061028771102428436,
-0.012486966326832771,
-0.13409686088562012,
0.008165537379682064,
0.009855501353740692,
-0.04840980842709541,
0.04244418069720268,
0.02425365336239338,
0.05156022310256958,
0.05248391255736351,
-0.029670046642422676,
-0.03126521036028862,
-0.04524005576968193,
-0.010280401445925236,
-0.012218219228088856,
-0.11509259790182114,
-0.03146075829863548,
0.013398614712059498,
0.0894111916422844,
0.029269106686115265,
-0.011258253827691078,
-0.030426369979977608,
0.04341241344809532,
0.12680955231189728,
-0.00734490342438221,
-0.11161448806524277,
-0.09032966196537018,
0.02467060461640358,
0.0026838271878659725,
0.0035674571990966797,
-0.018249113112688065,
-0.010996270924806595,
0.07003701478242874,
0.03570371866226196,
0.02996021881699562,
0.13705411553382874,
0.14702439308166504,
0.0164963286370039,
0.07455845177173615,
0.023172039538621902,
-0.006678270176053047,
-0.045810800045728683,
0.05122138559818268,
-0.01854007877409458,
0.0070252325385808945,
-0.07121878117322922,
-4.404233777677291e-8,
-0.08446255326271057,
-0.010148001834750175,
-0.0687723159790039,
0.08544623851776123,
0.023912953212857246,
-0.008296876214444637,
0.0557849258184433,
0.05014079436659813,
-0.031415361911058426,
-0.07688549160957336,
0.02880351059138775,
0.003445503069087863,
0.006082639098167419,
-0.09964890033006668,
-0.0621015802025795,
0.12715797126293182,
0.026679333299398422,
-0.0012927236966788769,
-0.006551219150424004,
0.008244424127042294,
-0.027448095381259918,
0.034638941287994385,
0.022873565554618835,
-0.07721927762031555,
0.017773926258087158,
-0.08431795984506607,
-0.11042284220457077,
0.02040453441441059,
0.06469311565160751,
0.04223436862230301,
-0.050483133643865585,
0.014179392717778683,
0.02537020482122898,
0.02706526592373848,
0.06557977944612503,
0.11762500554323196,
-0.03169800713658333,
0.06607086211442947,
-0.030271917581558228,
-0.036416199058294296,
-0.061334189027547836,
0.06716863811016083,
0.01751112937927246,
-0.039013076573610306,
0.01638965681195259,
-0.006592292804270983,
-0.015675082802772522,
-0.08058317005634308,
0.059822726994752884,
-0.06493184715509415,
0.03964774310588837,
0.12873877584934235,
-0.021239381283521652,
0.03728132322430611,
0.00012658217747230083,
0.023693116381764412,
-0.013332477770745754,
-0.09068312495946884,
-0.07138092070817947,
0.057955555617809296,
-0.03636611998081207,
0.006432780530303717,
-0.04511668160557747,
-0.019681638106703758
] | 0 | 0 | 0 | 0 |
|
Stiefel Flow Matching for Moment-Constrained Structure Elucidation | https://openreview.net/forum?id=84WmbzikPP | [
"Austin Henry Cheng",
"Alston Lo",
"Kin Long Kelvin Lee",
"Santiago Miret",
"Alan Aspuru-Guzik"
] | Poster | Molecular structure elucidation is a fundamental step in understanding chemical phenomena, with applications in identifying molecules in natural products, lab syntheses, forensic samples, and the interstellar medium.
We consider the task of predicting a molecule's all-atom 3D structure given only its molecular formula and moments of inertia, motivated by the ability of rotational spectroscopy to measure these moments.
While existing generative models can conditionally sample 3D structures with approximately correct moments, this soft conditioning fails to leverage the many digits of precision afforded by experimental rotational spectroscopy.
To address this, we first show that the space of $n$-atom point clouds with a fixed set of moments of inertia is embedded in the Stiefel manifold $\mathrm{St}(n, 4)$.
We then propose Stiefel Flow Matching as a generative model for elucidating 3D structure under exact moment constraints.
Additionally, we learn simpler and shorter flows by finding approximate solutions for equivariant optimal transport on the Stiefel manifold.
Empirically, enforcing exact moment constraints allows Stiefel Flow Matching to achieve higher success rates and faster sampling than Euclidean diffusion models, even on high-dimensional manifolds corresponding to large molecules in the GEOM dataset. | 3D molecular generative models, flow matching, Stiefel manifold, structure elucidation | We propose Stiefel flow matching as a generative model on the Stiefel manifold and apply it for molecular structure elucidation by rotational spectroscopy. | 13,419 | 2412.12540 | [
-0.060837581753730774,
-0.03518743813037872,
0.005348757840692997,
-0.017349494621157646,
0.053510632365942,
0.04106036573648453,
-0.016849033534526825,
0.004930456634610891,
0.13708536326885223,
-0.08014190196990967,
0.012224230915307999,
-0.06414507329463959,
-0.03975025191903114,
0.0005574444076046348,
0.01914704218506813,
-0.035485029220581055,
0.002956582233309746,
0.1006757915019989,
-0.05761220306158066,
0.057215698063373566,
0.03615548461675644,
0.017228467389941216,
0.02925794944167137,
0.05800868570804596,
0.05053888261318207,
0.07302485406398773,
0.03575584664940834,
-0.04826119914650917,
0.01760486699640751,
-0.07520610839128494,
0.13990125060081482,
0.054239097982645035,
-0.050320740789175034,
-0.04212099313735962,
0.09478582441806793,
-0.042164720594882965,
-0.06465309113264084,
-0.019009539857506752,
-0.020838268101215363,
-0.0233590230345726,
0.04234524071216583,
0.014582677744328976,
0.007009111810475588,
0.06227730214595795,
-0.03260849043726921,
-0.017167815938591957,
0.047794245183467865,
-0.022862227633595467,
0.026479583233594894,
-0.03799125552177429,
-0.008353055454790592,
-0.006115289870649576,
-0.08557481318712234,
0.026731878519058228,
0.012214639224112034,
0.031043576076626778,
-0.06405331194400787,
-0.08352052420377731,
0.08480683714151382,
-0.06086808815598488,
0.11160600930452347,
-0.08629640936851501,
0.001182154635898769,
-0.017789222300052643,
-0.026204414665699005,
0.021490097045898438,
0.003886821446940303,
0.0017196066910400987,
0.020255617797374725,
0.06335419416427612,
-0.030979176983237267,
0.07464233040809631,
-0.0750996395945549,
0.08553116023540497,
0.05459509789943695,
0.028305262327194214,
0.023713605478405952,
0.005243364255875349,
-0.016689706593751907,
-0.011256505735218525,
-0.0001821885525714606,
0.08643262833356857,
0.0441271997988224,
-0.0005925831501372159,
0.008502352982759476,
0.04837580397725105,
-0.07568302005529404,
0.053090453147888184,
0.05346798524260521,
-0.05651846155524254,
-0.009080064482986927,
-0.04469709470868111,
-0.14888323843479156,
-0.04824533686041832,
0.04529808461666107,
0.04039236530661583,
0.05677999556064606,
0.021660570055246353,
0.09251667559146881,
0.03077285923063755,
0.10032002627849579,
0.0120287099853158,
-0.03429163619875908,
0.030133159831166267,
-0.0050682183355093,
-0.012234563007950783,
0.03279193490743637,
0.03082604520022869,
-0.019567491486668587,
-0.02648267149925232,
0.0322704054415226,
-0.03948226571083069,
-0.00832100585103035,
-0.034097470343112946,
0.0022941369097679853,
-0.025740541517734528,
-0.00017102283891290426,
0.03578309714794159,
0.0004052500007674098,
0.029796333983540535,
-0.06415852159261703,
0.03535031899809837,
0.01251492090523243,
-0.002700014505535364,
0.02767500840127468,
0.020466243848204613,
-0.1194973737001419,
1.2226487556925466e-33,
0.02557390369474888,
0.07644080370664597,
0.09097949415445328,
0.054409708827733994,
0.060572266578674316,
-0.04669542238116264,
-0.028712507337331772,
0.001725327572785318,
-0.027434969320893288,
0.02643437683582306,
-0.09887709468603134,
-0.054867085069417953,
-0.012770013883709908,
-0.08096815645694733,
-0.04615888372063637,
-0.03054393269121647,
-0.03805308789014816,
0.038921426981687546,
0.03392636030912399,
0.030096711590886116,
0.019101813435554504,
-0.037056196480989456,
0.004084375221282244,
0.029008060693740845,
-0.08868561685085297,
0.07444177567958832,
-0.016304468736052513,
0.00967314001172781,
0.029966864734888077,
-0.009178672917187214,
0.004358435049653053,
-0.022706741467118263,
-0.05399002134799957,
0.001642319024540484,
0.05525650456547737,
-0.05263137072324753,
-0.0465836375951767,
-0.015141870826482773,
-0.04596530646085739,
-0.05599668622016907,
-0.018178727477788925,
0.07331318408250809,
-0.02647036872804165,
0.025696570053696632,
-0.11377191543579102,
0.016642309725284576,
0.04567123204469681,
-0.008781778626143932,
0.08155667036771774,
0.017068631947040558,
0.02980123646557331,
0.028465893119573593,
0.00038823828799650073,
-0.024515029042959213,
-0.032800592482089996,
0.029692981392145157,
0.00018934853142127395,
0.033020518720149994,
0.050145987421274185,
0.06326019018888474,
-0.04104124754667282,
0.06998620182275772,
0.03775167465209961,
-0.007509297225624323,
0.06863219290971756,
0.07383081316947937,
-0.12275511026382446,
0.034322340041399,
0.06736978143453598,
0.037237536162137985,
-0.042636144906282425,
0.02723875641822815,
0.004735270980745554,
-0.04362925514578819,
0.08320579677820206,
0.034471381455659866,
-0.0052353255450725555,
-0.06876213103532791,
0.1080077737569809,
0.032552238553762436,
-0.08592715859413147,
-0.0008688387461006641,
-0.02919909916818142,
-0.09516256302595139,
-0.08834563940763474,
-0.0588519461452961,
-0.0632963478565216,
-0.11391057819128036,
-0.019153272733092308,
0.013166310265660286,
0.0016707135364413261,
-0.132404625415802,
0.028272375464439392,
-0.04311543330550194,
-0.07433607429265976,
-2.431893937049842e-33,
0.02564273215830326,
-0.04589663818478584,
0.08191415667533875,
0.021991191431879997,
0.003522801212966442,
-0.003059565555304289,
-0.05175510793924332,
0.024447500705718994,
0.0010398312006145716,
-0.030821900814771652,
0.06779132038354874,
0.06229807063937187,
-0.042991798371076584,
-0.011509605683386326,
0.07070856541395187,
0.009002760984003544,
0.008221520110964775,
-0.04928148537874222,
0.05774867907166481,
0.01023826003074646,
0.01985834166407585,
0.0044688452035188675,
-0.04662979394197464,
-0.01191656943410635,
-0.05243953689932823,
0.05850415304303169,
0.011032326146960258,
-0.057236917316913605,
0.028182923793792725,
0.01986154355108738,
-0.015015071257948875,
-0.03793435916304588,
-0.0845872089266777,
0.045694828033447266,
-0.14170590043067932,
0.05437253788113594,
-0.011539855971932411,
0.12722744047641754,
-0.06134134903550148,
0.04116615653038025,
-0.0028470163233578205,
0.03466951474547386,
-0.11581646651029587,
0.12850287556648254,
0.06759894639253616,
0.0008086437592282891,
-0.05358247458934784,
-0.01805834099650383,
0.040807828307151794,
-0.03718724101781845,
-0.03273722529411316,
0.04564964398741722,
0.006706420797854662,
-0.0008675439166836441,
0.02918202430009842,
-0.0009812329662963748,
-0.042611777782440186,
-0.04375461861491203,
0.06272275000810623,
-0.016288818791508675,
-0.063860222697258,
-0.053228989243507385,
-0.006445946637541056,
0.011003129184246063,
0.023871196433901787,
-0.09670525789260864,
-0.010919013060629368,
-0.00814068503677845,
-0.03901209682226181,
0.005923435091972351,
-0.012896192260086536,
-0.06848777830600739,
-0.01585415005683899,
0.0640440508723259,
0.10899084061384201,
-0.05788377672433853,
0.08095766603946686,
-0.06681407988071442,
-0.022917702794075012,
-0.014401499181985855,
-0.053126588463783264,
-0.00026357537717558444,
0.010959723964333534,
-0.008574736304581165,
0.09174881875514984,
0.020517142489552498,
0.04688749834895134,
-0.05328755080699921,
0.032114528119564056,
0.018528806045651436,
0.018933730199933052,
0.045786771923303604,
0.05999108776450157,
0.020108433440327644,
-0.010481931269168854,
-4.540028442079347e-8,
-0.005934716667979956,
-0.04800589755177498,
0.041737232357263565,
-0.027543559670448303,
0.04738624766469002,
0.07605563849210739,
-0.02619401179254055,
0.057397034019231796,
0.017837755382061005,
-0.014799283817410469,
0.025861265137791634,
-0.09051705151796341,
-0.04985593259334564,
-0.04743245616555214,
0.017070360481739044,
0.05250503122806549,
-0.05057115480303764,
-0.0041807848028838634,
-0.002489935839548707,
0.023731784895062447,
-0.062212031334638596,
0.11345374584197998,
0.023903818801045418,
-0.010860075242817402,
0.09770101308822632,
-0.039511602371931076,
0.04200597479939461,
-0.01612103171646595,
-0.00440028915181756,
-0.004652794450521469,
-0.05028265714645386,
-0.024450311437249184,
0.0542152114212513,
0.05243407189846039,
-0.03395628184080124,
-0.05038183555006981,
-0.005490151233971119,
0.00940081849694252,
-0.10092198848724365,
-0.001325375633314252,
-0.07141505181789398,
0.10088568180799484,
-0.033308785408735275,
-0.029850929975509644,
0.004546085838228464,
0.13428258895874023,
-0.005768097471445799,
0.02893928438425064,
-0.031802620738744736,
0.05116618424654007,
0.003873024135828018,
-0.03718351945281029,
-0.054667212069034576,
0.01726951077580452,
-0.06765726953744888,
0.04482703283429146,
-0.08356045186519623,
-0.030200203880667686,
0.004384591709822416,
-0.031347110867500305,
-0.02437327802181244,
-0.01021561399102211,
-0.03915496915578842,
-0.06863995641469955
] | 0 | 0 | 0 | 0 |
|
Lawma: The Power of Specialization for Legal Annotation | https://openreview.net/forum?id=7El7K1DoyX | [
"Ricardo Dominguez-Olmedo",
"Vedant Nanda",
"Rediet Abebe",
"Stefan Bechtold",
"Christoph Engel",
"Jens Frankenreiter",
"Krishna P. Gummadi",
"Moritz Hardt",
"Michael Livermore"
] | Poster | Annotation and classification of legal text are central components of empirical legal research. Traditionally, these tasks are often delegated to trained research assistants. Motivated by the advances in language modeling, empirical legal scholars are increasingly turning to commercial models, hoping that it will alleviate the significant cost of human annotation. In this work, we present a comprehensive analysis of large language models' current abilities to perform legal annotation tasks. To do so, we construct CaselawQA, a benchmark comprising 260 legal text classification tasks, nearly all new to the machine learning community. Starting from GPT-4 as a baseline, we show that it has non-trivial but highly varied accuracy, often exhibiting performance that may be insufficient for legal work. We then demonstrate that a lightly fine-tuned Llama 3 8B model vastly outperforms GPT-4 on almost all tasks, typically by double-digit percentage points. A few tens to hundreds of examples suffice to achieve high classification accuracy. Our work points to a viable alternative to the predominant practice of prompting commercial models. For concrete legal tasks with some available labeled data, researchers are better off using a specialized open-source model. | large language models, legal classification tasks, benchmarks | A lightly finetuned Llama-3 model outperforms GPT-4 on 260 new legal classification tasks | 13,416 | null | [
-0.0682501271367073,
0.019885318353772163,
-0.044418394565582275,
-0.07455131411552429,
0.06319302320480347,
0.05029095336794853,
0.009544990956783295,
0.027336444705724716,
0.029561897739768028,
0.024942200630903244,
-0.0932556763291359,
-0.008267698809504509,
-0.05222570151090622,
0.02997424639761448,
0.03067071922123432,
0.028979450464248657,
0.08852044492959976,
0.05382721126079559,
-0.07430573552846909,
-0.009130665101110935,
0.010501042939722538,
0.10931209474802017,
0.08038270473480225,
0.006899476516991854,
0.012319719418883324,
-0.04413513466715813,
0.005658538080751896,
-0.05520271509885788,
0.006185214500874281,
-0.04197854548692703,
-0.012892448343336582,
0.0315093956887722,
0.002233446342870593,
0.06379915773868561,
-0.018475105985999107,
-0.04102247580885887,
-0.024439549073576927,
0.023785708472132683,
0.0015791381010785699,
-0.002843014895915985,
0.059344008564949036,
-0.024427063763141632,
0.0039826142601668835,
0.02772039920091629,
0.10059984773397446,
-0.03867311775684357,
0.010334442369639874,
0.025621939450502396,
0.004320161882787943,
0.07535979896783829,
-0.1238127127289772,
-0.04777024686336517,
0.03029944747686386,
0.049084004014730453,
-0.06700492650270462,
-0.06917200237512589,
0.07389340549707413,
0.02005660906434059,
-0.04602062702178955,
-0.07521114498376846,
0.036747489124536514,
-0.0467844232916832,
-0.061490174382925034,
-0.045842621475458145,
-0.04369739443063736,
-0.04625120386481285,
0.02877986989915371,
0.07291148602962494,
-0.058985572308301926,
-0.02982894890010357,
0.03949592635035515,
0.11438320577144623,
-0.02979334071278572,
0.09631674736738205,
-0.06630638241767883,
0.01838170737028122,
0.0143214650452137,
0.0629981979727745,
0.07968395948410034,
-0.1278756558895111,
-0.005411549471318722,
0.013847080990672112,
0.00253975298255682,
0.005238572135567665,
-0.016987204551696777,
-0.04693574458360672,
-0.0008315208251588047,
0.08463913202285767,
0.032630469650030136,
0.051230475306510925,
-0.001320625189691782,
-0.025304382666945457,
0.12128793448209763,
-0.053763650357723236,
-0.026148132979869843,
0.025011716410517693,
-0.002342891413718462,
-0.1077442467212677,
0.02512999437749386,
0.11339160799980164,
-0.01639459654688835,
0.0841674953699112,
-0.011838552542030811,
-0.08561182767152786,
0.003272128291428089,
0.022221803665161133,
0.02561265043914318,
-0.008699314668774605,
-0.0023696960415691137,
-0.0691501796245575,
0.086033396422863,
0.047587547451257706,
-0.03806281462311745,
-0.015903793275356293,
0.0410022959113121,
0.016398102045059204,
-0.020572056993842125,
0.028780102729797363,
0.007073408458381891,
0.05846351385116577,
-0.10157038271427155,
0.07295822352170944,
-0.019698217511177063,
-0.028874067589640617,
-0.023840177804231644,
-0.01449272409081459,
-0.13034728169441223,
1.0501052867750357e-33,
0.05529516935348511,
0.012118306010961533,
-0.05879371613264084,
0.04485281556844711,
0.014297904446721077,
-0.03437274321913719,
-0.07176564633846283,
0.008387457579374313,
-0.04002819582819939,
-0.008223273791372776,
-0.025259921327233315,
-0.03025795705616474,
-0.015913119539618492,
-0.015162755735218525,
-0.017136067152023315,
0.0001287755003431812,
-0.06333749741315842,
-0.018903454765677452,
-0.07119955122470856,
-0.012274292297661304,
0.09354619681835175,
-0.02050846628844738,
0.05146109685301781,
0.01392270252108574,
-0.04412849247455597,
0.040254343301057816,
0.02470645122230053,
-0.09822336584329605,
0.01274527981877327,
0.023869024589657784,
-0.006415945012122393,
-0.03136136010289192,
0.10956104099750519,
0.031057674437761307,
0.09181970357894897,
0.03820556774735451,
-0.026310233399271965,
-0.017523260787129402,
0.11707665026187897,
0.03486348316073418,
-0.09265873581171036,
-0.020164569839835167,
0.09425321966409683,
-0.022882850840687752,
-0.026179254055023193,
-0.021077480167150497,
-0.07872968912124634,
-0.04089732468128204,
0.05127844959497452,
0.0027796111535280943,
0.06510334461927414,
-0.01621777005493641,
-0.013814561069011688,
0.0020048595033586025,
-0.04576706141233444,
0.06540583819150925,
-0.07030069828033447,
0.0388706736266613,
-0.03069724701344967,
-0.007325975224375725,
-0.0023575443774461746,
0.09253186732530594,
0.03289628401398659,
0.07767585664987564,
-0.09757217764854431,
0.01768595166504383,
-0.029908519238233566,
0.06834723800420761,
0.053197264671325684,
-0.019027313217520714,
0.009440025314688683,
0.015236764214932919,
0.02610171213746071,
-0.020482316613197327,
0.005425817798823118,
0.00858025997877121,
0.09282810240983963,
-0.10070610046386719,
-0.012259121052920818,
0.01370211597532034,
-0.0051564848981797695,
0.0036348069552332163,
0.07442282140254974,
0.0093076191842556,
-0.02874775603413582,
-0.08500263094902039,
0.042822789400815964,
-0.047497574239969254,
0.019375227391719818,
0.022916976362466812,
-0.04521562159061432,
-0.018912479281425476,
-0.09466012567281723,
0.0241326242685318,
0.035698581486940384,
-9.615837663314017e-34,
-0.05237292870879173,
-0.06160811334848404,
-0.007607669569551945,
0.043911222368478775,
-0.053483497351408005,
-0.02718406729400158,
-0.05350913852453232,
-0.006890990771353245,
-0.027471959590911865,
-0.003825392806902528,
-0.01706693507730961,
-0.0644608736038208,
0.020578335970640182,
0.014280717819929123,
0.05019238963723183,
-0.052760250866413116,
-0.050863251090049744,
0.018688132986426353,
0.019664090126752853,
0.09048662334680557,
0.025148905813694,
0.06570778042078018,
-0.09342285245656967,
0.08577890694141388,
0.00878229271620512,
-0.0011169627541676164,
-0.037095870822668076,
-0.0029374659061431885,
0.018518447875976562,
-0.020334286615252495,
-0.06881845742464066,
0.006025087088346481,
-0.08462042361497879,
-0.03019782528281212,
-0.07237707078456879,
-0.00304407044313848,
0.15044766664505005,
-0.06222851201891899,
-0.030626598745584488,
0.054927412420511246,
0.039238911122083664,
0.025969507172703743,
-0.06954119354486465,
-0.008261483162641525,
-0.06572265923023224,
-0.03191400319337845,
-0.08489684015512466,
0.003684116993099451,
0.07869952917098999,
-0.01780039444565773,
0.006817531306296587,
-0.026729006320238113,
-0.009309148415923119,
0.04960592836141586,
-0.01897943578660488,
-0.11056870967149734,
-0.035889409482479095,
-0.0570925697684288,
-0.04011702165007591,
0.05606602132320404,
-0.08891437947750092,
0.0416136234998703,
0.041064418852329254,
-0.00005965074524283409,
0.016921810805797577,
-0.08933137357234955,
-0.044825971126556396,
0.0084464056417346,
-0.019177883863449097,
-0.028603840619325638,
0.03504125773906708,
-0.05011383816599846,
-0.047625962644815445,
-0.012152167968451977,
-0.011799115687608719,
0.0700521394610405,
0.027776015922427177,
-0.06997577846050262,
-0.02124476432800293,
-0.04899248480796814,
0.07898716628551483,
-0.04378635808825493,
0.023614143952727318,
0.05863465741276741,
0.07296308130025864,
0.061237361282110214,
0.0330660343170166,
0.09703335911035538,
0.015056396834552288,
0.09377236664295197,
-0.03244103863835335,
0.030598893761634827,
-0.009489921852946281,
0.09416154026985168,
-0.05058695003390312,
-3.850504270985766e-8,
-0.03625587373971939,
0.07067414373159409,
-0.04440005123615265,
0.09622608870267868,
-0.04688715562224388,
0.011266672052443027,
-0.03252791240811348,
0.028057599440217018,
-0.026931792497634888,
-0.025005385279655457,
0.036900535225868225,
-0.023028753697872162,
-0.08111852407455444,
-0.08823656290769577,
-0.0352853387594223,
0.048542048782110214,
0.04048445075750351,
0.007718828972429037,
-0.011068294756114483,
0.06953004002571106,
-0.01506802998483181,
0.05652957037091255,
0.06030619144439697,
-0.05432543531060219,
0.03148934990167618,
-0.07086879760026932,
-0.028589829802513123,
0.08624231070280075,
0.0617867149412632,
0.04216768220067024,
-0.011009164154529572,
0.0428045429289341,
-0.09414371103048325,
-0.04171709343791008,
0.04226351156830788,
0.017197635024785995,
0.035408515483140945,
0.009996555745601654,
0.017707400023937225,
0.015400074422359467,
0.04012693837285042,
0.044672805815935135,
-0.04240867868065834,
-0.010496859438717365,
0.027877237647771835,
-0.061119444668293,
-0.047311920672655106,
-0.051674965769052505,
0.06933247298002243,
-0.03878571838140488,
-0.010369781404733658,
-0.07588671147823334,
0.010749931447207928,
0.011707458645105362,
0.027679480612277985,
0.090250663459301,
0.025980297476053238,
0.02199995517730713,
-0.04649381712079048,
-0.00223582168109715,
0.093396857380867,
-0.0026833920273929834,
0.033569250255823135,
0.03560321033000946
] | 0 | 0 | 0 | 0 |
|
OpenRCA: Can Large Language Models Locate the Root Cause of Software Failures? | https://openreview.net/forum?id=M4qNIzQYpd | [
"Junjielong Xu",
"Qinan Zhang",
"Zhiqing Zhong",
"Shilin He",
"Chaoyun Zhang",
"Qingwei Lin",
"Dan Pei",
"Pinjia He",
"Dongmei Zhang",
"Qi Zhang"
] | Poster | Large language models (LLMs) are driving substantial advancements in software engineering, with successful applications like Copilot and Cursor transforming real-world development practices. However, current research predominantly focuses on the early stages of development, such as code generation, while overlooking the post-development phases that are crucial to user experience. To explore the potential of LLMs in this direction, we propose OpenRCA, a benchmark dataset and evaluation framework for assessing LLMs’ ability to identify the root cause of software failures. OpenRCA includes 335 failures from three enterprise software systems, along with over 68 GB of telemetry data (logs, metrics, and traces). Given a failure case and its associated telemetry, the LLM is tasked to identify the root cause that triggered the failure, requiring comprehension of software dependencies and reasoning over heterogeneous, long-context telemetry data. Our results show substantial room for improvement, as current models can only handle the simplest cases. Even with the specially designed RCA-agent, the best-performing model, Claude 3.5, solved only 11.34% failure cases. Our work paves the way for future research in this direction. | Language models, Natural language processing, Software engineering | A novel benchmark for evaluating language models that introduces real-world root cause analysis as a task. | 13,411 | null | [
-0.010430090129375458,
-0.05994168668985367,
0.034539490938186646,
-0.03357161208987236,
0.07449787110090256,
-0.10426080226898193,
-0.029852507635951042,
0.09744880348443985,
0.006298490334302187,
-0.01931552216410637,
-0.029980989173054695,
-0.03403404727578163,
0.03676546737551689,
-0.03822821378707886,
0.01744384877383709,
0.022673342376947403,
-0.006920262239873409,
-0.0537123829126358,
-0.037195295095443726,
-0.0416463240981102,
-0.044930219650268555,
0.08150026947259903,
-0.02287016063928604,
0.013767286203801632,
0.006544604431837797,
0.0015116502763703465,
-0.006892390549182892,
-0.05745994299650192,
-0.0058786808513104916,
-0.020588964223861694,
0.011822198517620564,
0.08390196412801743,
-0.01736288331449032,
0.05626070871949196,
0.05033130198717117,
0.049074266105890274,
0.055461857467889786,
-0.015289046801626682,
-0.0885608047246933,
-0.068699412047863,
-0.04717779532074928,
0.011773782782256603,
0.04080583155155182,
0.07323318719863892,
0.019199565052986145,
-0.15874789655208588,
0.05741600692272186,
-0.05759962275624275,
-0.019854355603456497,
-0.01901879906654358,
-0.10434505343437195,
-0.04382823407649994,
0.03762386739253998,
0.031467318534851074,
0.010062342509627342,
0.009003260172903538,
0.04301099479198456,
-0.017829326912760735,
-0.025245102122426033,
-0.02711573801934719,
-0.02528572641313076,
-0.025083208456635475,
-0.07815205305814743,
-0.026559220626950264,
0.0014986877795308828,
0.048245832324028015,
0.05864116922020912,
0.011061441153287888,
-0.004041008185595274,
0.04012894630432129,
-0.11482047289609909,
0.055048324167728424,
-0.006780002266168594,
0.08437959104776382,
-0.0587722510099411,
0.06414899975061417,
0.019593175500631332,
0.016581716015934944,
0.06798195838928223,
-0.07544097304344177,
-0.005140101537108421,
-0.009037679992616177,
-0.05257983133196831,
0.06892195343971252,
0.06740891933441162,
0.05140835419297218,
0.08269894123077393,
0.046295564621686935,
0.041384316980838776,
0.03876733034849167,
0.015795346349477768,
-0.007532594259828329,
0.09229175001382828,
0.005883702076971531,
0.013562073931097984,
0.03247958421707153,
0.03658875450491905,
-0.07228302210569382,
0.010507067665457726,
0.07082231342792511,
-0.02411101944744587,
0.05847664177417755,
0.049944017082452774,
-0.05559825897216797,
-0.0648183822631836,
0.010172335430979729,
0.06832192838191986,
-0.022700559347867966,
0.02945120818912983,
-0.030022041872143745,
0.017874184995889664,
-0.02198885940015316,
0.047071874141693115,
-0.07447895407676697,
0.045582883059978485,
0.003152229357510805,
-0.06248176842927933,
-0.04696786031126976,
-0.01959249936044216,
0.06489069759845734,
-0.003432840807363391,
0.05972748622298241,
0.019070588052272797,
-0.003939428832381964,
0.03475692495703697,
0.058804918080568314,
0.015491553582251072,
1.5931405141717383e-33,
0.09971213340759277,
0.08439508825540543,
-0.04395865276455879,
0.06562381237745285,
0.06735219061374664,
-0.04407801479101181,
0.004572050180286169,
0.06288452446460724,
-0.08645879477262497,
-0.005210178904235363,
-0.021868787705898285,
0.010926702059805393,
-0.07399825751781464,
-0.06376011669635773,
0.1455891728401184,
0.02069101668894291,
-0.05645104497671127,
0.08301972597837448,
-0.0801154151558876,
-0.008284068666398525,
-0.011783525347709656,
-0.08097051084041595,
0.003182209562510252,
-0.04758033528923988,
0.09508751332759857,
0.06091322749853134,
0.02842721901834011,
0.003601674223318696,
0.017240582033991814,
0.01731247641146183,
0.011405096389353275,
-0.0037916928995400667,
0.018797054886817932,
0.048568665981292725,
0.0639742761850357,
-0.044820286333560944,
-0.0843706950545311,
-0.054744988679885864,
0.03448617085814476,
0.0037389560602605343,
-0.09218890964984894,
0.02700979635119438,
-0.003083993913605809,
-0.003334064269438386,
0.02380804345011711,
-0.09901916235685349,
0.014414663426578045,
-0.06427380442619324,
0.017963776364922523,
-0.060929521918296814,
0.044351864606142044,
0.033684924244880676,
0.020631704479455948,
0.053397949784994125,
0.016296308487653732,
-0.010423682630062103,
-0.004053538199514151,
-0.07322152704000473,
0.05454777553677559,
0.12868735194206238,
-0.01702197641134262,
0.026650171726942062,
0.022486260160803795,
0.014322339557111263,
0.04305111989378929,
-0.011425516568124294,
-0.011353355832397938,
0.04130328446626663,
0.0544474720954895,
0.04233890771865845,
-0.0499836802482605,
-0.05020865797996521,
0.012130249291658401,
-0.0035794179420918226,
0.024684112519025803,
0.010940935462713242,
0.013843022286891937,
-0.05702998489141464,
0.006732896436005831,
-0.05331889167428017,
-0.010864370502531528,
-0.04955887421965599,
0.01532685849815607,
-0.027384379878640175,
0.017361685633659363,
0.01093804556876421,
0.01958972029387951,
-0.041171714663505554,
-0.09970196336507797,
0.06736559420824051,
-0.05417365953326225,
-0.019604980945587158,
-0.08648326992988586,
-0.0006242815870791674,
-0.005567962769418955,
-1.2201172184054554e-33,
-0.08729303628206253,
-0.053248170763254166,
0.004024789202958345,
0.04619920626282692,
-0.08939153701066971,
-0.11894027888774872,
-0.041360046714544296,
0.004340596031397581,
-0.043136343359947205,
-0.03599758818745613,
-0.047766607254743576,
-0.09942732006311417,
0.04384341463446617,
0.037397466599941254,
0.04497488588094711,
-0.056418176740407944,
0.006302915047854185,
-0.08749718964099884,
0.005532213486731052,
0.051802877336740494,
0.01253572292625904,
0.05047839879989624,
-0.07335080206394196,
-0.018362611532211304,
-0.012076864950358868,
0.03599228337407112,
-0.0056311157532036304,
-0.037473004311323166,
0.03576700761914253,
-0.06184288486838341,
0.04665530472993851,
0.03760537505149841,
-0.042679786682128906,
0.0746273621916771,
-0.023106450214982033,
0.06872887909412384,
0.06997165083885193,
0.00274693313986063,
0.030106419697403908,
0.022411907091736794,
0.13060057163238525,
-0.011382504366338253,
-0.04655345901846886,
-0.06289664655923843,
-0.07614270597696304,
0.012834086082875729,
-0.09131038933992386,
0.060483116656541824,
-0.03461863473057747,
-0.10563240200281143,
0.018436817452311516,
-0.007429289165884256,
0.025836963206529617,
0.021060854196548462,
0.024107689037919044,
-0.010166758671402931,
0.04481185972690582,
-0.08418235182762146,
-0.023807020857930183,
-0.005673752631992102,
-0.05522218719124794,
-0.03169522061944008,
0.05048275738954544,
-0.01878729648888111,
0.04307008162140846,
-0.04622051492333412,
0.0009711026214063168,
0.06430908292531967,
-0.08806699514389038,
0.004320991691201925,
0.04639224335551262,
-0.0410618931055069,
-0.05863260477781296,
-0.00484539894387126,
0.0036292856093496084,
0.10594162344932556,
-0.06182549521327019,
-0.07604100555181503,
-0.04150860011577606,
-0.010408781468868256,
0.040927816182374954,
-0.03135596588253975,
0.059086740016937256,
0.027717500925064087,
0.0037257366348057985,
0.09604291617870331,
0.07247557491064072,
0.04074500501155853,
0.033763956278562546,
0.02195340394973755,
-0.1346149891614914,
0.0638335794210434,
-0.09150484204292297,
0.06643477082252502,
0.015463924035429955,
-4.6179341239849236e-8,
-0.08502069860696793,
0.004954359494149685,
-0.016353297978639603,
0.030601685866713524,
0.02866607904434204,
-0.013098331168293953,
-0.0322100855410099,
0.09574755281209946,
0.00248472997918725,
0.05458111688494682,
-0.05894976854324341,
-0.004821212030947208,
-0.10587839037179947,
0.03825872018933296,
-0.002248057397082448,
0.09166725724935532,
-0.029059890657663345,
0.020825151354074478,
-0.04864617809653282,
-0.014636626467108727,
0.03685387596487999,
0.069793701171875,
0.05644621700048447,
0.017095886170864105,
-0.018124567344784737,
-0.0869310051202774,
0.015600902959704399,
0.13618041574954987,
0.0428142286837101,
-0.053656838834285736,
0.07019049674272537,
-0.04590463638305664,
0.0271298848092556,
-0.056403420865535736,
0.005116785876452923,
0.04723811149597168,
0.0036161611787974834,
-0.042246341705322266,
0.034710466861724854,
-0.0043521784245967865,
0.09850477427244186,
0.053692709654569626,
-0.02634330280125141,
0.021949809044599533,
-0.0024436309468001127,
-0.04744293913245201,
-0.03184711933135986,
-0.05191764608025551,
0.011529439128935337,
-0.09553030133247375,
-0.018360311165452003,
-0.030463235452771187,
-0.03338389843702316,
0.08823149651288986,
0.04414775222539902,
0.011771241202950478,
-0.01632038503885269,
0.004951197654008865,
0.020730331540107727,
0.0768430307507515,
-0.005448071751743555,
0.0020020913798362017,
0.013505169190466404,
0.02635105699300766
] | 0 | 0 | 0 | 0 |
|
Wavelet-based Positional Representation for Long Context | https://openreview.net/forum?id=OhauMUNW8T | [
"Yui Oka",
"Taku Hasegawa",
"Kyosuke Nishida",
"Kuniko Saito"
] | Poster | In the realm of large-scale language models, a significant challenge arises when extrapolating sequences beyond the maximum allowable length.
This is because the model's position embedding mechanisms are limited to positions encountered during training, thus preventing effective representation of positions in longer sequences.
We analyzed conventional position encoding methods for long contexts and found the following characteristics.
(1) When the representation dimension is regarded as the time axis, Rotary Position Embedding (RoPE) can be interpreted as a restricted wavelet transform using Haar-like wavelets.
However, because it uses only a fixed scale parameter, it does not fully exploit the advantages of wavelet transforms, which capture the fine movements of non-stationary signals using multiple scales (window sizes).
This limitation could explain why RoPE performs poorly in extrapolation.
(2)
Previous research as well as our own analysis indicates that Attention with Linear Biases (ALiBi) functions similarly to windowed attention, using windows of varying sizes.
However, it has limitations in capturing deep dependencies because it restricts the receptive field of the model.
From these insights, we propose a new position representation method that captures multiple scales (i.e., window sizes) by leveraging wavelet transforms without limiting the model's attention field.
Experimental results show that this new method improves the performance of the model in both short and long contexts.
In particular, our method allows extrapolation of position information without limiting the model's attention field. | Positional Encoding, Extrapolation, Wavelet Transform, Transformers, RoPE, ALiBi, NLP | We found that RoPE can be interpreted as a restricted wavelet transform. And we propose a new position representation method that captures window sizes by leveraging wavelet transforms without limiting the model's attention field. | 13,404 | 2502.02004 | [
-0.03938936069607735,
-0.09285444021224976,
-0.028886817395687103,
-0.06923221796751022,
0.052869390696287155,
0.07860143482685089,
-0.0198979452252388,
-0.02997460775077343,
0.09868934750556946,
-0.11686984449625015,
-0.004937058314681053,
0.025092855095863342,
-0.04951099306344986,
0.005193531978875399,
-0.005713983904570341,
-0.03560804948210716,
0.02536587417125702,
0.10477888584136963,
-0.08396556228399277,
-0.078093521296978,
0.08723223954439163,
-0.017030734568834305,
0.06758478283882141,
-0.008009515702724457,
0.05044182017445564,
0.002110328059643507,
-0.057994406670331955,
-0.001879338058643043,
0.08090487122535706,
0.0008342144428752363,
0.07454223185777664,
0.04182688519358635,
0.04885218292474747,
0.03684953227639198,
-0.12352573126554489,
0.06504706293344498,
-0.09544670581817627,
0.014037185348570347,
-0.026073353365063667,
-0.06247284263372421,
0.05580207332968712,
0.05959257110953331,
0.011841148138046265,
0.08694176375865936,
0.055555135011672974,
-0.011796771548688412,
0.06520912051200867,
-0.0239239651709795,
-0.05207248404622078,
-0.004546238109469414,
-0.05378687381744385,
-0.04134157672524452,
-0.02658497728407383,
0.053406115621328354,
-0.08491232246160507,
-0.006198615301400423,
-0.005531882867217064,
0.038628268986940384,
0.08879884332418442,
0.009656190872192383,
-0.014236441813409328,
-0.023533884435892105,
0.002608712064102292,
-0.02021227777004242,
-0.06463183462619781,
-0.034551192075014114,
0.016827180981636047,
0.006643627304583788,
-0.060389380902051926,
0.05151784047484398,
-0.03373175859451294,
-0.04687971994280815,
-0.007521048188209534,
0.044291745871305466,
0.04647742211818695,
0.005743367597460747,
0.024709083139896393,
-0.03296028822660446,
0.07609695941209793,
-0.0344359315931797,
0.0015903369057923555,
-0.06764651089906693,
0.13588517904281616,
-0.008520197123289108,
0.15607760846614838,
0.04777842015028,
0.005607135593891144,
0.049476027488708496,
-0.037794265896081924,
0.022453831508755684,
-0.025706041604280472,
-0.0792776346206665,
0.021126914769411087,
0.009591076523065567,
0.06985727697610855,
-0.017128659412264824,
-0.036544568836688995,
0.029398247599601746,
-0.021919751539826393,
0.05448543652892113,
0.047040607780218124,
0.00540581438690424,
-0.016765201464295387,
0.04435635730624199,
0.06031271815299988,
-0.02680303156375885,
-0.021451598033308983,
0.05433604493737221,
0.011070764623582363,
-0.1006338894367218,
-0.03347489982843399,
0.001983825583010912,
-0.009161506779491901,
0.030284332111477852,
0.0699470043182373,
-0.10071554034948349,
0.010268007405102253,
-0.033093877136707306,
-0.02417750470340252,
0.02490837872028351,
0.009064668789505959,
-0.00012316819629631937,
-0.06333029270172119,
0.005429118871688843,
-0.01442714687436819,
0.012859645299613476,
0.02645321562886238,
2.2541152938383487e-33,
0.003227420849725604,
0.07822711020708084,
0.030846448615193367,
-0.006728133652359247,
0.04095398262143135,
0.03416149690747261,
-0.058454569429159164,
0.03982236608862877,
0.02780003845691681,
0.04689602181315422,
-0.02031688205897808,
0.034435197710990906,
-0.029489822685718536,
0.07971906661987305,
0.009460847824811935,
0.007429660297930241,
0.007137124426662922,
0.02719043381512165,
-0.05949186161160469,
0.014566940255463123,
-0.043036434799432755,
-0.023820871487259865,
0.019226159900426865,
-0.07744580507278442,
-0.02367340587079525,
0.009193053469061852,
0.02314264141023159,
-0.0657246932387352,
-0.07986212521791458,
0.0012279185466468334,
-0.1439216136932373,
-0.03338557481765747,
-0.06508912891149521,
-0.0063655078411102295,
0.018283197656273842,
-0.08051915466785431,
0.011735622771084309,
-0.042680997401475906,
0.017899775877594948,
0.030013924464583397,
-0.02303873375058174,
0.028540022671222687,
0.002618910279124975,
0.05394066125154495,
-0.009781665168702602,
-0.07071888446807861,
0.05764099210500717,
0.08325419574975967,
-0.07417769730091095,
0.03196440264582634,
-0.00838540494441986,
0.03879562392830849,
0.015794310718774796,
-0.08856397867202759,
0.03832114860415459,
0.006739445962011814,
0.07867163419723511,
0.10461460053920746,
0.002589848358184099,
0.026562979444861412,
-0.06640487909317017,
0.010913027450442314,
0.035315077751874924,
0.0725792646408081,
0.07140747457742691,
-0.0035874061286449432,
-0.047655168920755386,
0.03473805636167526,
-0.006554008927196264,
0.01880234107375145,
0.017029745504260063,
0.024931859225034714,
-0.01415286399424076,
-0.03484385088086128,
-0.04558528587222099,
-0.005967429373413324,
0.026753798127174377,
-0.07874352484941483,
-0.07882896810770035,
0.05968992039561272,
0.020694833248853683,
-0.10522917658090591,
0.08154762536287308,
-0.061754219233989716,
-0.047491393983364105,
-0.05975286290049553,
0.015620600432157516,
-0.17314328253269196,
0.02370350994169712,
-0.06073712557554245,
-0.08951250463724136,
-0.024110501632094383,
-0.04470730572938919,
-0.003745070192962885,
0.022823845967650414,
-1.8338834102590736e-33,
0.019686127081513405,
0.05488240718841553,
-0.020840661600232124,
-0.03545648232102394,
-0.033978719264268875,
0.04518964886665344,
0.032990504056215286,
0.05329650267958641,
-0.06264156103134155,
-0.02516527846455574,
-0.037044458091259,
-0.10245558619499207,
0.05096888169646263,
-0.029696812853217125,
0.02831839956343174,
0.018982650712132454,
0.0003852456866297871,
0.026046426966786385,
-0.0282977856695652,
-0.022049183025956154,
0.05940517410635948,
0.038683921098709106,
-0.05623820796608925,
0.08222266286611557,
-0.0006039715954102576,
0.01346665620803833,
-0.014816265553236008,
-0.011529031209647655,
-0.04639545455574989,
-0.09635770320892334,
-0.060915276408195496,
-0.023957479745149612,
0.059709332883358,
-0.056463856250047684,
-0.053102828562259674,
0.039379723370075226,
0.051138244569301605,
-0.011226953938603401,
-0.053291767835617065,
0.05516654998064041,
0.10106652230024338,
-0.02043415792286396,
0.05450475588440895,
-0.024126529693603516,
-0.03797099366784096,
0.037975311279296875,
-0.1082298681139946,
0.018423397094011307,
-0.05888919532299042,
0.07297680526971817,
-0.035135235637426376,
0.11481697112321854,
-0.05744171887636185,
0.03837057203054428,
-0.030223732814192772,
-0.032526880502700806,
-0.09625855833292007,
-0.0428643524646759,
0.051482003182172775,
-0.08176924288272858,
0.00008167113264789805,
-0.032232508063316345,
-0.001564306323416531,
-0.07420019805431366,
0.04945630952715874,
-0.042486026883125305,
-0.0007805372006259859,
-0.04685676842927933,
0.02906337007880211,
-0.020633134990930557,
0.04826369136571884,
-0.033640000969171524,
-0.029797546565532684,
0.030922725796699524,
-0.0804070308804512,
0.035793449729681015,
0.049572646617889404,
-0.07723284512758255,
-0.046591464430093765,
-0.04205996170639992,
-0.007549870293587446,
-0.022030925378203392,
0.05774157866835594,
0.05354929715394974,
0.01427539810538292,
0.1111629381775856,
0.090171217918396,
0.030761394649744034,
0.04325256496667862,
0.012764737010002136,
0.04518688842654228,
0.08404857665300369,
-0.033589497208595276,
0.07224954664707184,
0.024353640154004097,
-4.464331837539248e-8,
-0.1111895814538002,
0.07490836828947067,
-0.038900963962078094,
0.0024078465066850185,
-0.006301175337284803,
-0.05196231231093407,
0.0363260880112648,
0.07334718853235245,
-0.04228406772017479,
-0.00825634878128767,
0.034852661192417145,
-0.040248073637485504,
-0.013929633423686028,
-0.042947955429553986,
0.038877423852682114,
0.03523483872413635,
0.02977694384753704,
-0.04750936105847359,
0.015476357191801071,
0.008862660266458988,
0.06375402212142944,
0.07572218775749207,
0.004709811415523291,
0.029748134315013885,
-0.024918386712670326,
-0.03005743958055973,
-0.0463838130235672,
0.11074194312095642,
0.06424472481012344,
-0.036591462790966034,
0.0012471976224333048,
0.056831106543540955,
0.007140791974961758,
-0.037325311452150345,
-0.03892667591571808,
0.03240783512592316,
0.04757505655288696,
-0.015274742618203163,
-0.014711295254528522,
0.015807176008820534,
0.027962341904640198,
0.004586058668792248,
-0.04485752061009407,
0.04822952300310135,
0.027638476341962814,
0.039991818368434906,
0.05505838617682457,
-0.11040575057268143,
0.03339211270213127,
-0.03922295570373535,
0.07600095123052597,
-0.01989818923175335,
0.01741214282810688,
-0.0076177106238901615,
0.02281610295176506,
0.10681114345788956,
-0.004477992653846741,
-0.012914400547742844,
0.007999125868082047,
0.059071194380521774,
-0.058524973690509796,
0.051851581782102585,
-0.07049039006233215,
-0.006694103591144085
] | 0 | 0 | 0 | 0 |
|
LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code | https://openreview.net/forum?id=chfJJYC3iL | [
"Naman Jain",
"King Han",
"Alex Gu",
"Wen-Ding Li",
"Fanjia Yan",
"Tianjun Zhang",
"Sida Wang",
"Armando Solar-Lezama",
"Koushik Sen",
"Ion Stoica"
] | Poster | Large Language Models (LLMs) applied to code-related applications have emerged as a prominent field, attracting significant interest from academia and industry. However, as new and improved LLMs are developed, existing evaluation benchmarks (e.g., HumanEvla, MBPP) are no longer sufficient for assessing their capabilities suffering from data contamination, overfitting, saturation, and focus on merely code generation. In this work, we propose LiveCodeBench, a comprehensive and contamination-free evaluation of LLMs for code, which collects new problems over time from contests across three competition platforms, Leetcode, Atcoder, and Codeforces. Notably, our benchmark also focuses on a broader range of code-related capabilities, such as self-repair, code execution, and test output prediction, beyond just code generation. Currently, LiveCodeBench hosts over six hundred coding problems that were published between May 2023 and Aug 2024. We evaluate over 50 LLMs on LiveCodeBench (LCB for brevity) presenting the largest evaluation study of code LLMs on competition problems. Based on the study, we present novel empirical findings on contamination, overfitting, and holistic evaluations. We demonstrate that time-segmented evaluations serve as a robust approach to evade contamination; they are successful at detecting contamination across a wide range of open and closed models including GPT-4O, Claude, Deepseek, and Codestral. Next, we highlight overfitting and saturation of traditional coding benchmarks like HumanEvla and demonstrate LCB allows more reliable evaluations. Finally, our holistic evaluation scenarios allow for measuring the different capabilities of programming agents in isolation. | Code LLMs; Evaluation; Contaminationl; Overfitting | We build evaluation of code LLMs on new problems highlighting challenges like contamination and overfitting | 13,402 | 2403.07974 | [
-0.10955946892499924,
-0.046722013503313065,
-0.035272207111120224,
0.0020820279605686665,
0.07296475768089294,
0.014889377169311047,
-0.06745096296072006,
-0.04150700569152832,
-0.07455950975418091,
-0.00035395153099671006,
-0.03776209428906441,
-0.0863131731748581,
0.023617850616574287,
-0.05289267376065254,
-0.0100424624979496,
-0.027493562549352646,
0.015872543677687645,
-0.02336556650698185,
-0.06764934957027435,
-0.048198357224464417,
-0.0054941847920417786,
-0.00210717529989779,
-0.031064338982105255,
0.04445934668183327,
0.08181656897068024,
-0.0031418930739164352,
-0.06771670281887054,
-0.010624461807310581,
0.03955477848649025,
0.007684940937906504,
0.024213798344135284,
0.07248778641223907,
0.02775559201836586,
0.08016137778759003,
0.058561936020851135,
0.04485077038407326,
-0.05260280892252922,
-0.08146333694458008,
-0.08340628445148468,
-0.007466427981853485,
-0.09878634661436081,
0.005727364681661129,
-0.0013468218967318535,
0.02351459674537182,
0.06026603654026985,
-0.07372552901506424,
-0.01737779565155506,
-0.022673122584819794,
-0.08330383896827698,
-0.037085458636283875,
-0.08191291242837906,
-0.0479276105761528,
0.04985680431127548,
-0.02378990314900875,
-0.02870657667517662,
-0.07286461442708969,
0.060685910284519196,
-0.01609279029071331,
-0.019191311672329903,
0.0022362684831023216,
-0.03853654861450195,
-0.02558770589530468,
-0.01600104570388794,
0.06612206250429153,
-0.00827816128730774,
0.0025130310095846653,
0.009683222509920597,
0.0893922671675682,
-0.0023415435571223497,
-0.014921177178621292,
-0.052847083657979965,
-0.0005016529466956854,
-0.0573081448674202,
0.12757381796836853,
0.0724029392004013,
0.035638418048620224,
-0.02740742824971676,
0.003013099543750286,
0.06936857849359512,
-0.053484298288822174,
-0.019508523866534233,
-0.08588463068008423,
-0.040491726249456406,
0.06769100576639175,
0.09385477006435394,
0.023692457005381584,
0.08284609764814377,
0.045906912535429,
0.04017644748091698,
0.07770238816738129,
-0.014319563284516335,
-0.012199245393276215,
0.11402548104524612,
0.058803100138902664,
-0.016485899686813354,
0.038310013711452484,
0.04939085245132446,
-0.02874344401061535,
0.04472145065665245,
0.04937415570020676,
-0.035595666617155075,
0.04301384463906288,
-0.015807928517460823,
-0.0968230664730072,
-0.005512004718184471,
0.011021348647773266,
0.09844078123569489,
0.02856076881289482,
0.053447216749191284,
-0.038569070398807526,
0.03011535294353962,
0.06005224585533142,
0.026296865195035934,
-0.02432367019355297,
0.07306450605392456,
0.021316058933734894,
-0.032902274280786514,
-0.006801444571465254,
0.01152067631483078,
0.06542632728815079,
0.023999348282814026,
0.016307925805449486,
-0.024482108652591705,
-0.060731541365385056,
0.027107195928692818,
0.03461518883705139,
-0.017645692452788353,
1.8044664804587535e-33,
0.034475650638341904,
-0.036276012659072876,
-0.06592383235692978,
0.03849821165204048,
-0.017415154725313187,
-0.0741993710398674,
0.05119739845395088,
0.09063026309013367,
-0.09350643306970596,
-0.008711608126759529,
-0.0072604757733643055,
-0.005370087455958128,
-0.06411439925432205,
0.03923148289322853,
0.07898511737585068,
0.03281799703836441,
-0.046699266880750656,
0.00956935714930296,
-0.061664264649152756,
0.017098864540457726,
0.05818565934896469,
-0.05827587842941284,
-0.017144525423645973,
-0.05783693492412567,
0.040462419390678406,
0.059721991419792175,
0.025317762047052383,
-0.09556500613689423,
-0.03346443548798561,
0.05212859809398651,
-0.08880306780338287,
-0.0645187571644783,
0.05889679491519928,
0.06153655797243118,
-0.006846263539046049,
-0.0068465727381408215,
-0.05353385582566261,
-0.04661945998668671,
-0.026385946199297905,
0.055752232670784,
-0.045130763202905655,
0.009236975573003292,
-0.042891573160886765,
-0.039621494710445404,
0.06701380014419556,
-0.061204683035612106,
-0.014220108278095722,
-0.036536525934934616,
-0.018989769741892815,
0.01348094455897808,
0.05323580652475357,
0.07629379630088806,
0.00588952424004674,
0.05497487634420395,
-0.07982730865478516,
0.05199144035577774,
0.030206847935914993,
-0.05733150616288185,
0.0429375134408474,
0.10830439627170563,
-0.05182589590549469,
0.053410694003105164,
0.006722640246152878,
0.02280719392001629,
0.0901879370212555,
-0.010637667961418629,
0.02853245474398136,
-0.03778039664030075,
0.013308861292898655,
-0.02174205891788006,
-0.028946470469236374,
-0.06046951189637184,
-0.006096614059060812,
-0.0020767953246831894,
0.007058191578835249,
-0.05405254662036896,
0.031906891614198685,
-0.04748108610510826,
0.019359547644853592,
-0.072840116918087,
0.04611140489578247,
-0.0030672976281493902,
0.0255923792719841,
-0.048005979508161545,
-0.034058574587106705,
-0.04584084823727608,
0.001828908221796155,
-0.08145312964916229,
-0.03648882731795311,
0.0467686802148819,
-0.000906755740288645,
-0.07377856224775314,
0.01952766440808773,
-0.09102203696966171,
0.026467276737093925,
-2.7910732154413e-33,
-0.03507941961288452,
-0.036427050828933716,
-0.07233747094869614,
0.10849877446889877,
-0.06658309698104858,
-0.07923637330532074,
-0.021076759323477745,
0.023266037926077843,
-0.00046278038644231856,
-0.023916037753224373,
-0.04560612142086029,
-0.06315456330776215,
0.0073732296004891396,
0.02767559513449669,
0.06568500399589539,
-0.06540955603122711,
-0.03970547392964363,
-0.0004092584422323853,
0.04128881171345711,
0.09220284968614578,
0.049472931772470474,
0.11056940257549286,
-0.010038288310170174,
-0.03466866910457611,
-0.05982758104801178,
0.022799890488386154,
-0.04752826690673828,
0.02943938784301281,
0.09164062887430191,
-0.03320327773690224,
0.04299173504114151,
0.03189969062805176,
-0.05541566014289856,
0.015479913912713528,
0.019307458773255348,
0.014515422284603119,
0.11991728097200394,
-0.04491874948143959,
0.009807691909372807,
0.04891423508524895,
0.1020875796675682,
-0.019237343221902847,
-0.09483885020017624,
0.028494739904999733,
0.007069455459713936,
0.027641458436846733,
-0.13625214993953705,
-0.03793780878186226,
0.10324996709823608,
-0.050362396985292435,
-0.0006926655187271535,
0.027189936488866806,
-0.0069676595740020275,
-0.007624841295182705,
-0.05341691896319389,
-0.07511522620916367,
-0.05288473516702652,
-0.05529717728495598,
-0.03055150806903839,
0.055554211139678955,
-0.039102017879486084,
0.02146788313984871,
0.04390673339366913,
-0.005018414929509163,
0.034092411398887634,
-0.043616387993097305,
-0.03888349607586861,
0.07151474803686142,
-0.06363823264837265,
-0.05780508741736412,
0.040197573602199554,
-0.03808753564953804,
-0.05811682343482971,
0.007100064307451248,
-0.03436482325196266,
0.029388919472694397,
-0.028633292764425278,
-0.00967935100197792,
0.013081367127597332,
-0.003240045392885804,
0.015572728589177132,
0.029484162107110023,
0.06404291838407516,
0.02454008162021637,
0.04244428873062134,
0.062047794461250305,
0.03240358829498291,
0.06217533349990845,
-0.015308945439755917,
0.018259892240166664,
-0.04030245169997215,
0.03259208798408508,
0.010291033424437046,
-0.038332123309373856,
-0.0035637698601931334,
-4.2446547610097696e-8,
-0.01808592677116394,
0.016856957226991653,
-0.02651120349764824,
0.08122358471155167,
0.006667028181254864,
-0.0029316346626728773,
-0.06192638352513313,
0.056128811091184616,
0.009065025486052036,
0.0434839129447937,
0.07816743850708008,
-0.0007924022502265871,
-0.1136246994137764,
0.0027875483501702547,
0.008212106302380562,
0.058440010994672775,
-0.12633170187473297,
0.02294275164604187,
-0.04620180279016495,
-0.007278730161488056,
0.030044661834836006,
0.11506250500679016,
0.046807967126369476,
-0.0527498796582222,
-0.05412088707089424,
-0.04005209729075432,
0.0374104380607605,
0.13951712846755981,
0.05416697636246681,
-0.004118818789720535,
0.001983295427635312,
0.011677440255880356,
-0.011287126690149307,
-0.024026203900575638,
0.06064751371741295,
-0.011572661809623241,
-0.04174300283193588,
-0.03750698268413544,
0.038840536028146744,
0.019401073455810547,
0.00633242167532444,
0.03823210671544075,
-0.06969400495290756,
-0.006363390013575554,
0.03607785329222679,
-0.06411072611808777,
-0.028619511052966118,
-0.08840613067150116,
0.025989193469285965,
-0.1006188914179802,
-0.0021815458312630653,
-0.054913297295570374,
-0.11041413247585297,
0.0512472428381443,
0.10144267976284027,
0.0924396812915802,
-0.04470941796898842,
-0.017114147543907166,
0.06324122101068497,
0.06770244985818863,
0.04445822164416313,
0.005418638698756695,
0.016553353518247604,
0.04366939887404442
] | 0 | 0 | 0 | 0 |
|
ActionReasoningBench: Reasoning about Actions with and without Ramification Constraints | https://openreview.net/forum?id=NUD03NBDOE | [
"Divij Handa",
"Pavel Dolin",
"Shrinidhi Kumbhar",
"Tran Cao Son",
"Chitta Baral"
] | Poster | Reasoning about Actions and Change (RAC) has historically played a pivotal role in solving foundational AI problems, such as the frame problem. It has driven advancements in AI fields, such as non-monotonic and commonsense reasoning. RAC remains crucial for AI systems that operate in dynamic environments, engage in interactive scenarios, or rely on commonsense reasoning. Despite substantial advances made by Large Language Models (LLMs) in various AI domains, their performance in RAC remains underexplored. To address this gap, we introduce a new diagnostic benchmark, $\textbf{ActionReasoningBench}$, which encompasses 8 domains and includes questions for up to 19 action sequences. This benchmark rigorously evaluates LLMs across six key RAC dimensions: $\textit{Fluent Tracking}$, $\textit{State Tracking}$, $\textit{Action Executability}$, $\textit{Effects of Actions}$, $\textit{Numerical RAC}$, and $\textit{Composite Questions}$. LLMs demonstrate average accuracy rates of 73.55%, 65.63%, 58.73%, and 62.38% on the former four dimensions, which are frequently discussed in RAC literature. However, the performance on the latter two dimensions, which introduce complex and novel reasoning questions, the average performance of LLMs is lowered to 33.16% and 51.19%, respectively, reflecting a 17.9% performance decline. We also introduce new ramification constraints to capture the indirect effects of actions, providing deeper insights into RAC challenges. Our evaluation of state-of-the-art LLMs, including both open-source and commercial models, reveals challenges across all RAC dimensions, particularly in handling ramifications, with GPT-4o failing to solve any question and o1-preview achieving a score of only 18.4%. | Reasoning about Actions and Change (RAC), Benchmark, Large Language Models (LLMs) | The paper introduces ActionReasoningBench, a diagnostic benchmark for Reasoning about Actions and Change (RAC) covering novel complex questions and indirect effects of actions, known as ramifications. | 13,401 | 2406.04046 | [
-0.052695393562316895,
-0.12652459740638733,
-0.03459366038441658,
0.018704241141676903,
0.1115034818649292,
-0.002508159028366208,
0.04014115408062935,
0.03894798085093498,
0.00921118538826704,
0.006573765072971582,
-0.0722322016954422,
-0.10622409731149673,
-0.006312113720923662,
0.053695954382419586,
-0.023504799231886864,
0.012749869376420975,
0.06623321771621704,
-0.028951741755008698,
-0.06850895285606384,
-0.02002711035311222,
-0.06421657651662827,
-0.0063531589694321156,
0.04797481745481491,
-0.009811549447476864,
-0.010954396799206734,
-0.02208954282104969,
-0.02851383201777935,
-0.0985240489244461,
-0.04000885784626007,
-0.0042562829330563545,
0.07752404361963272,
-0.0008922481210902333,
0.02083025872707367,
0.007223981898277998,
0.02365919016301632,
0.0738799199461937,
-0.012922010384500027,
-0.061007168143987656,
-0.05079644173383713,
-0.0756048932671547,
-0.027134791016578674,
-0.022913504391908646,
0.039185505360364914,
-0.02918507158756256,
0.09770352393388748,
0.008367392234504223,
-0.005977322813123465,
-0.0038044629618525505,
0.008287545293569565,
0.0019110724097117782,
-0.08809906989336014,
-0.034314852207899094,
0.012535002082586288,
-0.025347476825118065,
-0.06757117062807083,
0.03380001336336136,
0.03164324164390564,
0.026252713054418564,
0.03545527905225754,
-0.010326001793146133,
-0.034447044134140015,
-0.11265198886394501,
-0.021803203970193863,
0.025644464418292046,
-0.021328790113329887,
0.0323372557759285,
0.016042476519942284,
-0.005552586168050766,
0.007150936871767044,
0.04289822280406952,
-0.04193156212568283,
0.08611451089382172,
-0.03364386409521103,
0.020991185680031776,
-0.029652219265699387,
0.021842261776328087,
-0.014786981977522373,
-0.04019412770867348,
0.03899666666984558,
-0.0985901728272438,
-0.011171001940965652,
-0.05981695279479027,
0.017551453784108162,
0.02780485339462757,
0.14102044701576233,
0.013967505656182766,
0.007887190207839012,
0.010663555935025215,
0.035019826143980026,
0.09658824652433395,
-0.027809998020529747,
-0.08511188626289368,
0.09711568802595139,
0.07049590349197388,
0.04674079269170761,
0.07734024524688721,
0.037231843918561935,
-0.007724303752183914,
-0.029396364465355873,
0.09626137465238571,
0.018130214884877205,
0.10172818601131439,
-0.022060716524720192,
-0.044219791889190674,
0.021522805094718933,
-0.015195317566394806,
0.019856130704283714,
0.024052297696471214,
0.004804679192602634,
-0.0883561372756958,
0.03047579899430275,
0.03218650072813034,
0.10133620351552963,
-0.030568473041057587,
0.06012775003910065,
-0.04673197120428085,
0.07861091941595078,
-0.026371555402874947,
-0.01656169630587101,
0.009278355166316032,
0.0668625459074974,
0.026246191933751106,
0.02178596705198288,
0.022802911698818207,
0.08889742940664291,
-0.007804433349519968,
-0.014153644442558289,
2.582571190661143e-33,
0.06055647134780884,
0.0005111229256726801,
0.06374015659093857,
0.0269333329051733,
0.04671662300825119,
0.03247891366481781,
0.02314394898712635,
0.05872800201177597,
-0.022642845287919044,
0.023563968017697334,
0.00319392466917634,
0.008051023818552494,
-0.033914849162101746,
0.0013117474736645818,
0.0713857039809227,
-0.023161232471466064,
-0.048707060515880585,
0.04601243510842323,
-0.04746592417359352,
-0.00047252385411411524,
0.029752984642982483,
-0.05495010316371918,
-0.00982377864420414,
-0.035568345338106155,
0.0643855407834053,
-0.00017440046940464526,
0.06281732767820358,
-0.02280130796134472,
-0.00006488020881079137,
0.0008780139614827931,
-0.11122649163007736,
-0.06566418707370758,
-0.08466468006372452,
0.11568998545408249,
0.037407007068395615,
-0.028696328401565552,
-0.0569496788084507,
-0.09889574348926544,
0.023931140080094337,
-0.030452465638518333,
-0.06300569325685501,
0.05035768821835518,
-0.03864765539765358,
0.03092583827674389,
-0.04585947096347809,
-0.05082235112786293,
0.01888539083302021,
0.05131279304623604,
-0.04317041113972664,
-0.005099922884255648,
0.034356702119112015,
0.04215383902192116,
0.0055595943704247475,
-0.053445156663656235,
0.036460068076848984,
-0.028795238584280014,
0.007768307812511921,
0.036195095628499985,
0.02568214014172554,
0.07629670947790146,
0.05845018848776817,
0.014876195229589939,
-0.04526112973690033,
0.11094696074724197,
0.01669672690331936,
0.0877857431769371,
-0.0747847929596901,
0.07670574635267258,
0.06594972312450409,
-0.03476928547024727,
-0.02126796543598175,
0.022653132677078247,
0.036310210824012756,
-0.011335078626871109,
0.06473143398761749,
-0.00313769094645977,
0.028809957206249237,
-0.14719441533088684,
0.02422546036541462,
-0.01962750218808651,
-0.0475674569606781,
0.0024808221496641636,
-0.054899267852306366,
0.02786390110850334,
-0.04161481186747551,
-0.07515692710876465,
0.02710721828043461,
-0.04550931230187416,
-0.0445345975458622,
-0.05830758437514305,
0.002074297284707427,
-0.07410125434398651,
-0.01868933066725731,
0.0072594271041452885,
0.03520620986819267,
-1.894063027220413e-33,
-0.02285483479499817,
-0.06901512295007706,
-0.09295183420181274,
0.07898525148630142,
-0.054683271795511246,
-0.0649186447262764,
-0.035511843860149384,
0.005997251719236374,
0.04063720628619194,
-0.06642482429742813,
-0.03459802269935608,
-0.021138709038496017,
-0.03194937855005264,
0.04601069912314415,
0.022984594106674194,
-0.05222940444946289,
-0.011668183840811253,
-0.03856922313570976,
0.015859978273510933,
0.03679754212498665,
0.023985307663679123,
0.09695612639188766,
-0.10825896263122559,
-0.015013677068054676,
-0.08850012719631195,
0.03153514862060547,
-0.02427215874195099,
0.0037386633921414614,
-0.012826701626181602,
0.0014956197701394558,
-0.041546329855918884,
-0.014181801117956638,
-0.056029267609119415,
0.05580557510256767,
-0.056876808404922485,
0.020472794771194458,
0.05608528479933739,
-0.017810655757784843,
-0.039497584104537964,
0.03699551522731781,
0.09095145761966705,
-0.0504903681576252,
-0.07228146493434906,
0.0047449227422475815,
-0.006813678424805403,
0.06882096081972122,
-0.12213850021362305,
0.06274382770061493,
-0.012584337964653969,
-0.022962527349591255,
-0.03843865916132927,
-0.05184154212474823,
-0.0021257612388581038,
-0.02225983329117298,
-0.027253033593297005,
-0.013416429981589317,
-0.026102975010871887,
-0.040164295583963394,
-0.032569967210292816,
-0.038026873022317886,
-0.04985098913311958,
0.1256849616765976,
0.02498205006122589,
-0.009459967724978924,
0.04849258065223694,
0.04644007980823517,
-0.03549593687057495,
0.007584625389426947,
-0.035189904272556305,
-0.03562687337398529,
0.03767324611544609,
-0.06659315526485443,
-0.049065180122852325,
0.04185102880001068,
0.08372797071933746,
-0.02188638411462307,
0.028113041073083878,
-0.04093468561768532,
-0.0007612588233314455,
-0.04869905859231949,
0.022740131244063377,
-0.028760189190506935,
0.07292809337377548,
0.023874863982200623,
-0.025375250726938248,
0.1107739731669426,
0.022562218829989433,
0.09489891678094864,
0.02459125407040119,
0.0656006932258606,
-0.009510126896202564,
0.05549080669879913,
0.008023717440664768,
0.04453350603580475,
-0.06804122030735016,
-4.580533286002719e-8,
-0.08437948673963547,
0.12277291715145111,
0.00384452729485929,
-0.009623457677662373,
0.04709324985742569,
-0.015753915533423424,
-0.035262320190668106,
0.056337688118219376,
-0.014009143225848675,
0.007960244081914425,
0.07425897568464279,
-0.018425410613417625,
-0.009883282706141472,
-0.02315129153430462,
0.0028064155485481024,
0.03254859149456024,
0.02625386230647564,
-0.022114457562565804,
-0.06708553433418274,
-0.02259148098528385,
0.09133550524711609,
0.06563684344291687,
-0.0076995668932795525,
-0.020583875477313995,
-0.05742493271827698,
-0.07188867777585983,
-0.0032604325097054243,
0.15880195796489716,
0.044903554022312164,
0.02484113536775112,
0.03924708068370819,
0.019770460203289986,
0.02802569791674614,
-0.014036142267286777,
0.048646487295627594,
0.045852068811655045,
0.004438657313585281,
-0.0576227642595768,
0.05570835620164871,
-0.005920977331697941,
-0.02326119877398014,
0.047331590205430984,
-0.03242567181587219,
0.026080800220370293,
0.01608896441757679,
-0.001610119710676372,
-0.12241993844509125,
-0.10509710758924484,
0.0835551917552948,
-0.07635308057069778,
-0.016091128811240196,
-0.017450081184506416,
-0.054688457399606705,
0.1204143837094307,
0.050673503428697586,
0.09613177925348282,
0.007491270545870066,
-0.05284508690237999,
-0.024863459169864655,
0.08406175673007965,
0.03514176234602928,
-0.011188989505171776,
-0.06984549015760422,
0.032212816178798676
] | 0 | 0 | 0 | 0 |
|
No Preference Left Behind: Group Distributional Preference Optimization | https://openreview.net/forum?id=bgpNJBD6Va | [
"Binwei Yao",
"Zefan Cai",
"Yun-Shiuan Chuang",
"Shanglin Yang",
"Ming Jiang",
"Diyi Yang",
"Junjie Hu"
] | Poster | Preferences within a group of people are not uniform but follow a distribution. While existing alignment methods like Direct Preference Optimization (DPO) attempt to steer models to reflect human preferences, they struggle to capture the distributional pluralistic preferences within a group. These methods often skew toward dominant preferences, overlooking the diversity of opinions, especially when conflicting preferences arise. To address this issue, we propose Group Distributional Preference Optimization (GDPO), a novel framework that aligns language models with the distribution of preferences within a group by incorporating the concept of beliefs that shape individual preferences. GDPO calibrates a language model using statistical estimation of the group's belief distribution and aligns the model with belief-conditioned preferences, offering a more inclusive alignment framework than traditional methods. In experiments using both synthetic controllable opinion generation and real-world movie review datasets, we show that DPO fails to align with the targeted belief distributions, while GDPO consistently reduces this alignment gap during training. Additionally, our evaluation metrics demonstrate that GDPO outperforms existing approaches in aligning with group distributional preferences, marking a significant advance in pluralistic alignment. | preference alignment; large language model; fairness; group preferences | null | 13,392 | 2412.20299 | [
-0.05997229740023613,
-0.13249430060386658,
-0.05103893205523491,
-0.016567280516028404,
-0.007869802415370941,
0.025455204769968987,
-0.00029144861036911607,
0.0024737485218793154,
0.10476823151111603,
-0.013134298846125603,
0.012074907310307026,
-0.09097925573587418,
0.052018146961927414,
0.027242228388786316,
0.021118400618433952,
-0.05157472565770149,
0.12502555549144745,
0.0386420376598835,
-0.1010451391339302,
-0.02790551260113716,
-0.04793105646967888,
0.01831229217350483,
0.020695244893431664,
-0.02936931885778904,
-0.02657265029847622,
-0.017126210033893585,
0.01055513322353363,
-0.008320428431034088,
-0.006426980718970299,
-0.0034317991230636835,
0.047613777220249176,
0.014669831842184067,
0.002358326455578208,
0.005860129836946726,
-0.05461978539824486,
0.08035989850759506,
-0.10338451713323593,
0.0001763263571774587,
-0.0147938784211874,
-0.0440191775560379,
-0.07884234935045242,
0.008984982036054134,
0.040901944041252136,
-0.007732204627245665,
0.02193315699696541,
0.02486485056579113,
-0.013846575282514095,
0.04259578138589859,
0.005063401535153389,
0.06634557247161865,
-0.04976712167263031,
-0.01650579273700714,
0.06590627878904343,
0.03758963197469711,
-0.013259563595056534,
-0.025754818692803383,
0.021560121327638626,
0.05170105770230293,
0.051000624895095825,
-0.0405779629945755,
-0.08905838429927826,
-0.15702706575393677,
-0.017154140397906303,
0.05699722096323967,
-0.05061966925859451,
-0.061981603503227234,
0.03165833652019501,
0.0506192147731781,
-0.07722178846597672,
0.029389595612883568,
-0.05487433820962906,
0.13753342628479004,
0.0208932813256979,
0.06890571117401123,
0.025600649416446686,
0.06794514507055283,
0.048252105712890625,
-0.09835444390773773,
0.03454006090760231,
-0.058099642395973206,
0.03492496907711029,
0.031128723174333572,
0.043833933770656586,
0.044871117919683456,
0.12173853069543839,
0.01814883016049862,
-0.023173341527581215,
0.07515376806259155,
-0.0396142415702343,
0.026000598445534706,
-0.13782303035259247,
-0.017824945971369743,
0.02821161039173603,
-0.022264013066887856,
-0.025262361392378807,
0.12724819779396057,
-0.021114615723490715,
-0.0958208218216896,
-0.027583081275224686,
0.06562293320894241,
0.020366737619042397,
0.05545606464147568,
-0.03761303424835205,
-0.0679420530796051,
0.002550592180341482,
-0.10142187029123306,
-0.023605724796652794,
0.019659103825688362,
0.007674662861973047,
-0.05749695748090744,
-0.006730223074555397,
0.0649624839425087,
-0.00037014420377090573,
0.01771128922700882,
-0.00599835067987442,
-0.06123945862054825,
0.054389454424381256,
0.03486327826976776,
-0.018112173303961754,
-0.007113137748092413,
-0.03329095244407654,
-0.008545690216124058,
0.07040973752737045,
0.0024153601843863726,
0.05798172578215599,
0.016034208238124847,
-0.0051124426536262035,
1.1655529746921926e-33,
-0.04908168315887451,
0.05381380021572113,
0.07152825593948364,
-0.05276181921362877,
0.037005715072155,
-0.04877709224820137,
-0.08221018314361572,
0.07709794491529465,
-0.07291518151760101,
-0.012908685952425003,
-0.024489281699061394,
0.030599482357501984,
-0.06474360823631287,
0.07794796675443649,
0.019667543470859528,
-0.01223081722855568,
-0.0506935641169548,
0.09208608418703079,
-0.08975948393344879,
0.00001561894350743387,
0.030114561319351196,
0.11160870641469955,
0.03678877279162407,
-0.010098643600940704,
-0.016640853136777878,
-0.03207365795969963,
0.012627319432795048,
-0.027322517707943916,
-0.06284312158823013,
0.006827718112617731,
-0.08097109198570251,
-0.01865510083734989,
0.021519701927900314,
0.01457217801362276,
0.04486573114991188,
-0.03524229675531387,
-0.04211891442537308,
-0.045022882521152496,
0.0027722010854631662,
0.012246562168002129,
0.0012714946642518044,
0.0656731054186821,
-0.022492770105600357,
0.05698198080062866,
-0.04576607421040535,
-0.005221965257078409,
0.04615223780274391,
-0.05756020545959473,
0.031251780688762665,
-0.0801885649561882,
0.02190243825316429,
0.018495719879865646,
0.06287616491317749,
-0.057646963745355606,
-0.04741916060447693,
0.006453476846218109,
0.010939504019916058,
-0.003464335575699806,
0.04345330223441124,
-0.09100817143917084,
-0.08178950101137161,
0.08080105483531952,
0.017117813229560852,
-0.03058609925210476,
0.03994730859994888,
0.0184645913541317,
-0.0009427507757209241,
0.011823990382254124,
0.04778726026415825,
0.013961694203317165,
0.07815290242433548,
0.02309134602546692,
-0.0672774687409401,
-0.00885916780680418,
0.004981465637683868,
0.06902609765529633,
0.06127474084496498,
0.004065227694809437,
0.052472133189439774,
-0.03818614408373833,
0.02896619588136673,
-0.06335730850696564,
-0.01843719743192196,
-0.09395827353000641,
-0.01160603016614914,
-0.05634725093841553,
0.06334011256694794,
-0.029761604964733124,
-0.0027409910690039396,
-0.027215437963604927,
-0.021723249927163124,
0.018174683675169945,
-0.010628985241055489,
0.004360502120107412,
0.0306497011333704,
-2.645758052779672e-33,
-0.08793958276510239,
-0.0021746426355093718,
0.05662643536925316,
0.04300638288259506,
-0.007854992523789406,
-0.03894937038421631,
-0.005907122511416674,
0.06000985950231552,
0.017409833148121834,
-0.07194893062114716,
-0.09333300590515137,
-0.06270159780979156,
0.04210905730724335,
-0.022161994129419327,
-0.04572947323322296,
-0.0025615901686251163,
-0.009771336801350117,
0.0639200508594513,
-0.0005737415631301701,
0.014993133954703808,
0.04101071506738663,
0.06518259644508362,
-0.12818840146064758,
0.0932236835360527,
0.02954675257205963,
0.008952758274972439,
0.002269303658977151,
0.04374169185757637,
-0.020699655637145042,
-0.002393813803792,
-0.02626972831785679,
0.005329510197043419,
-0.08206697553396225,
-0.010347598232328892,
-0.04055080562829971,
-0.010433128103613853,
-0.0625915676355362,
0.0038017660845071077,
-0.016627319157123566,
0.13544531166553497,
-0.06765766441822052,
-0.03384314477443695,
-0.09444532543420792,
-0.01744716614484787,
-0.03148152679204941,
0.012675309553742409,
0.01498129777610302,
-0.04017059504985809,
-0.026905249804258347,
-0.03835119307041168,
-0.0679832473397255,
0.0952976793050766,
-0.018850309774279594,
0.05484015867114067,
0.01040652859956026,
-0.11406654864549637,
-0.05816807225346565,
-0.07111653685569763,
0.05614275485277176,
0.026418965309858322,
-0.00848858617246151,
0.02678937278687954,
0.09750138968229294,
-0.0697966143488884,
-0.014146062545478344,
-0.040222473442554474,
0.03894703835248947,
0.025106143206357956,
-0.0031431149691343307,
0.00646212138235569,
0.060198329389095306,
-0.03966263681650162,
-0.030979465693235397,
-0.016412852331995964,
-0.003632462816312909,
0.03044969215989113,
0.04465902969241142,
-0.01642710715532303,
0.015170572325587273,
0.01530328206717968,
0.06584367156028748,
0.048290759325027466,
0.027069000527262688,
0.031890109181404114,
0.07500337064266205,
0.04044153913855553,
0.02889026328921318,
0.11832766234874725,
0.021466737613081932,
0.03737610578536987,
0.021679775789380074,
0.009956451132893562,
0.06033346429467201,
0.05034426599740982,
-0.1051965206861496,
-4.4051475356354786e-8,
-0.07054399698972702,
-0.04104907438158989,
0.06242058053612709,
0.09712421894073486,
-0.03545588627457619,
-0.03542875498533249,
-0.005498353857547045,
-0.001512190792709589,
-0.052002888172864914,
-0.05767591670155525,
0.004828812088817358,
0.12611345946788788,
-0.038236163556575775,
-0.053650353103876114,
-0.0492616668343544,
0.02859317511320114,
-0.013326190412044525,
0.04649519547820091,
0.0164178516715765,
0.04583055153489113,
0.04829841107130051,
0.022128326818346977,
0.013021046295762062,
-0.0779181718826294,
0.06216650828719139,
-0.07260860502719879,
-0.07901713997125626,
-0.021122945472598076,
0.017513921484351158,
-0.00475172558799386,
-0.029688728973269463,
-0.035604093223810196,
-0.10853797197341919,
-0.00286508584395051,
-0.017464589327573776,
0.10414514690637589,
-0.01440238393843174,
0.019809691235423088,
0.04395707696676254,
-0.025610029697418213,
0.09733860194683075,
0.036465972661972046,
-0.023020775988698006,
0.04263125732541084,
-0.04051581025123596,
0.04786277562379837,
-0.004683255217969418,
-0.06867944449186325,
0.0009186980896629393,
-0.030074873939156532,
0.031476929783821106,
-0.027657100930809975,
-0.031155534088611603,
0.07966212183237076,
0.056459639221429825,
0.01701788604259491,
-0.053126417100429535,
-0.02023054100573063,
0.0334254689514637,
0.0242341086268425,
0.08273167908191681,
0.03454839438199997,
-0.04884399101138115,
-0.009183503687381744
] | https://github.com/BigBinnie/GDPO | 9 | 0 | 0 | 0 |
MagicDec: Breaking the Latency-Throughput Tradeoff for Long Context Generation with Speculative Decoding | https://openreview.net/forum?id=CS2JWaziYr | [
"Ranajoy Sadhukhan",
"Jian Chen",
"Zhuoming Chen",
"Vashisth Tiwari",
"Ruihang Lai",
"Jinyuan Shi",
"Ian En-Hsu Yen",
"Avner May",
"Tianqi Chen",
"Beidi Chen"
] | Poster | Large Language Models (LLMs) have become more prevalent in long-context applications such as interactive chatbots, document analysis, and agent workflows, but it is challenging to serve long-context requests with low latency and high throughput. Speculative decoding (SD) is a widely used technique to reduce latency losslessly, but the conventional wisdom suggests that its efficacy is limited to small batch sizes. In MagicDec, we show that surprisingly SD can achieve speedup even for a high throughput inference regime for moderate to long sequences. More interestingly, an intelligent drafting strategy can achieve better speedup with increasing batch size based on our rigorous analysis. MagicDec first identifies the bottleneck shifts with increasing batch size and sequence length, and uses these insights to deploy SD more effectively for high throughput inference. We leverage draft model with sparse KV cache to address the KV bottleneck, which scales with both sequence length and batch size. Additionally, we propose a theoretical model to select the optimal drafting strategy for maximum speedup. Our work highlights the broad applicability of speculative decoding in long-context serving, as it can enhance throughput and reduce latency without compromising accuracy. For moderate to long sequences, we demonstrate up to 2.51x speedup for LLaMA-3.1-8B when serving batch sizes ranging from 32 to 256 on various types of hardware and tasks. | LLM Inference, Speculative Decoding, Performance Analysis | Long-context serving, Speculative Decoding, Sparse KV Cache | 13,385 | 2408.11049 | [
-0.03150095418095589,
-0.026245705783367157,
-0.025386646389961243,
0.03860035538673401,
0.00879589468240738,
-0.041924383491277695,
-0.0341031551361084,
-0.036584772169589996,
0.056433893740177155,
-0.026987923309206963,
-0.06562279909849167,
-0.024621732532978058,
-0.031462594866752625,
-0.03508227691054344,
0.04243914783000946,
-0.05089936405420303,
0.13783793151378632,
0.018354300409555435,
-0.0482511892914772,
0.0008507866295985878,
0.016696171835064888,
-0.016469670459628105,
0.07786229252815247,
-0.02942449226975441,
0.04164564236998558,
-0.021940769627690315,
-0.04846964776515961,
-0.03918895870447159,
0.016630804166197777,
0.026054054498672485,
0.07079485803842545,
0.14991191029548645,
-0.04850410297513008,
0.01732349768280983,
0.018550625070929527,
0.055060192942619324,
-0.03708286210894585,
0.021727804094552994,
0.0666164979338646,
-0.043241892009973526,
-0.02173241600394249,
0.01704241707921028,
-0.04341238737106323,
0.08144766837358475,
0.02689080685377121,
-0.04640977457165718,
-0.01296233106404543,
-0.038854967802762985,
0.008718331344425678,
0.022653955966234207,
-0.12365055084228516,
-0.002449901308864355,
0.016006387770175934,
0.05348597466945648,
-0.07080858200788498,
0.027530338615179062,
0.0481128990650177,
0.03716444969177246,
-0.012722313404083252,
0.02786089852452278,
-0.08310020714998245,
-0.11915742605924606,
-0.0010394877754151821,
-0.036498911678791046,
-0.05891498550772667,
-0.02863629162311554,
0.0449163019657135,
0.02487751469016075,
0.020729465410113335,
-0.023028306663036346,
-0.011681023053824902,
0.044963303953409195,
-0.03757191076874733,
0.05212417617440224,
-0.021339572966098785,
0.038606494665145874,
0.02712547779083252,
-0.06297175586223602,
0.05851859226822853,
-0.11756007373332977,
-0.0038927551358938217,
-0.0007004653452895582,
0.053115684539079666,
-0.027737382799386978,
0.015290949493646622,
-0.031094204634428024,
0.06784534454345703,
0.04087315872311592,
0.01732059009373188,
0.038620442152023315,
-0.03180564567446709,
-0.0988907516002655,
0.07336518168449402,
0.04512859135866165,
0.017807679250836372,
0.08926008641719818,
0.0019362317398190498,
-0.019711719825863838,
-0.0274795051664114,
0.0552707239985466,
0.013081102631986141,
0.03536306694149971,
-0.013907195068895817,
-0.020182309672236443,
-0.027261076495051384,
-0.040816646069288254,
0.0672784373164177,
0.05764055252075195,
-0.049506112933158875,
-0.041496261954307556,
0.022358806803822517,
0.025427494198083878,
0.05969565361738205,
-0.03403234854340553,
0.03446035832166672,
0.009951543994247913,
0.04199331998825073,
-0.02184535562992096,
-0.007634049281477928,
0.10447090119123459,
0.05019683763384819,
0.018039237707853317,
-0.018493754789233208,
0.01556987501680851,
-0.00008877823711372912,
-0.026259303092956543,
0.04718564823269844,
2.0002233914111122e-33,
0.0539383701980114,
-0.008533980697393417,
-0.004121540114283562,
0.09449651092290878,
0.0002970737114083022,
-0.01606122776865959,
0.07872002571821213,
0.007610874250531197,
-0.03675103932619095,
-0.0821617841720581,
-0.014445887878537178,
-0.013877510093152523,
-0.028264658525586128,
0.06957825273275375,
-0.032213203608989716,
-0.020559705793857574,
-0.040184326469898224,
0.07174202799797058,
-0.03828732296824455,
-0.006678510922938585,
0.05588617920875549,
-0.07300753146409988,
-0.002822721842676401,
-0.04435230791568756,
0.06758665293455124,
0.040790215134620667,
0.01868271641433239,
-0.10588997602462769,
0.001969133270904422,
0.024831363931298256,
-0.13296370208263397,
-0.058078717440366745,
-0.0007051865686662495,
0.03934758901596069,
0.04602722451090813,
0.0006792545318603516,
-0.031019601970911026,
-0.07381869107484818,
0.011723365634679794,
0.009050756692886353,
-0.051693812012672424,
0.013044136576354504,
0.012477590702474117,
-0.025215310975909233,
-0.14844673871994019,
-0.058697402477264404,
-0.08459398150444031,
0.015673262998461723,
-0.08191993832588196,
0.06681878119707108,
0.0756336972117424,
0.07315034419298172,
-0.01371679361909628,
0.030689191073179245,
-0.006872977130115032,
-0.08514080941677094,
0.06287698447704315,
-0.04140618070960045,
0.07072891294956207,
0.10284437984228134,
-0.003010238753631711,
-0.0003340180264785886,
-0.05366833508014679,
0.12279097735881805,
0.05621499568223953,
-0.02155170775949955,
-0.0531303733587265,
0.12131565809249878,
0.015939002856612206,
-0.030040526762604713,
-0.05491458997130394,
-0.0025952954310923815,
0.06762775778770447,
-0.02506496198475361,
0.008537744171917439,
0.010094701312482357,
0.03250037506222725,
-0.04668809100985527,
-0.048044029623270035,
-0.022227292880415916,
0.004287506453692913,
0.052060727030038834,
-0.060206376016139984,
0.0008984726737253368,
-0.02976907230913639,
-0.06437928974628448,
-0.05692168325185776,
-0.05439767614006996,
-0.030703069642186165,
-0.027669256553053856,
-0.09141916781663895,
0.004154828377068043,
-0.013503020629286766,
-0.05139262229204178,
-0.049409788101911545,
-1.8169869646629494e-33,
0.011789210140705109,
-0.040729232132434845,
-0.08335154503583908,
0.10278426855802536,
0.035091500729322433,
-0.10969559848308563,
-0.006987793371081352,
-0.0013811729149892926,
-0.007084061857312918,
-0.07413743436336517,
-0.08233042806386948,
-0.02739027515053749,
0.016788607463240623,
-0.0027306322008371353,
0.06842585653066635,
-0.03260083124041557,
0.0315282829105854,
-0.007109550759196281,
0.03056689165532589,
0.07631608843803406,
0.02952003665268421,
0.007682253606617451,
-0.10694985091686249,
0.017216317355632782,
0.02479041926562786,
0.02353917434811592,
-0.07588256895542145,
0.04258706793189049,
0.01572415977716446,
-0.05089423432946205,
-0.0229953546077013,
0.04070613533258438,
0.0006684114341624081,
-0.02367096021771431,
-0.017621103674173355,
0.09939072281122208,
0.04469786956906319,
0.03257908299565315,
-0.04055386036634445,
0.08343958854675293,
0.1347496509552002,
-0.04876802861690521,
-0.029025675728917122,
0.0005995657411403954,
-0.030818458646535873,
0.03535028547048569,
-0.1106414720416069,
-0.02899038791656494,
0.03570619598031044,
0.06267747282981873,
0.03729109838604927,
0.04111894220113754,
-0.04259590059518814,
-0.006389234680682421,
-0.013716282323002815,
-0.13003551959991455,
-0.07409823685884476,
-0.06871189922094345,
0.024723779410123825,
-0.022651925683021545,
0.0077746654860675335,
0.017175935208797455,
0.052475303411483765,
-0.07635969668626785,
0.0530119389295578,
-0.059221118688583374,
-0.007250544615089893,
-0.04943481832742691,
-0.0015311207389459014,
-0.010216081514954567,
0.06939810514450073,
-0.04418684169650078,
-0.02527717500925064,
0.11662151664495468,
0.03910963609814644,
0.01228071004152298,
0.0075401305221021175,
-0.044427115470170975,
0.05725177749991417,
0.06618986278772354,
-0.009664743207395077,
0.11014588922262192,
0.004136653151363134,
0.013916993513703346,
0.07351068407297134,
0.11067651212215424,
-0.04724346846342087,
0.0863153412938118,
-0.01951824687421322,
0.003667454468086362,
-0.009448372758924961,
0.005344586446881294,
0.03992550075054169,
0.03566761314868927,
-0.03433523327112198,
-4.047475954394031e-8,
-0.003969904035329819,
0.045904334634542465,
-0.0019850186072289944,
0.0499226413667202,
0.07134458422660828,
0.003032459644600749,
-0.011324804276227951,
0.05003705248236656,
0.08583397418260574,
-0.021330956369638443,
0.17271395027637482,
-0.053953107446432114,
0.005814273376017809,
-0.03393257409334183,
0.0028818915598094463,
0.03047838620841503,
-0.01770883798599243,
-0.07997956871986389,
-0.03234142065048218,
-0.037998348474502563,
0.027593472972512245,
0.06638317555189133,
-0.036246802657842636,
-0.016718944534659386,
0.04660153016448021,
-0.04535650089383125,
0.0319434255361557,
0.1309320032596588,
0.035552047193050385,
-0.05493856593966484,
-0.03780407831072807,
-0.00853513553738594,
-0.02770194225013256,
-0.04609411954879761,
-0.0000029165801151975757,
0.020266423001885414,
0.020571067929267883,
-0.007663109805434942,
0.02346225082874298,
-0.02280307374894619,
0.025999007746577263,
-0.009143375791609287,
-0.04955804720520973,
-0.03969040513038635,
0.006841226946562529,
-0.06997616589069366,
-0.06685459613800049,
-0.06519652158021927,
-0.012331606820225716,
-0.02505761943757534,
-0.020916610956192017,
-0.06249421834945679,
0.02186974510550499,
0.10684660822153091,
0.031637564301490784,
0.04222601652145386,
-0.030070926994085312,
-0.0023937092628329992,
0.03861966356635094,
0.012825229205191135,
-0.0016404688358306885,
0.01511321496218443,
-0.0751790702342987,
0.02311882935464382
] | https://github.com/infini-ai-lab/magicdec | 115 | 0 | 0 | 0 |
Parameter Expanded Stochastic Gradient Markov Chain Monte Carlo | https://openreview.net/forum?id=exgLs4snap | [
"Hyunsu Kim",
"Giung Nam",
"Chulhee Yun",
"Hongseok Yang",
"Juho Lee"
] | Poster | Bayesian Neural Networks (BNNs) provide a promising framework for modeling predictive uncertainty and enhancing out-of-distribution robustness (OOD) by estimating the posterior distribution of network parameters. Stochastic Gradient Markov Chain Monte Carlo (SGMCMC) is one of the most powerful methods for scalable posterior sampling in BNNs, achieving efficiency by combining stochastic gradient descent with second-order Langevin dynamics. However, SGMCMC often suffers from limited sample diversity in practice, which affects uncertainty estimation and model performance. We propose a simple yet effective approach to enhance sample diversity in SGMCMC without the need for tempering or running multiple chains. Our approach reparameterizes the neural network by decomposing each of its weight matrices into a product of matrices, resulting in a sampling trajectory that better explores the target parameter space. This approach produces a more diverse set of samples, allowing faster mixing within the same computational budget. Notably, our sampler achieves these improvements without increasing the inference cost compared to the standard SGMCMC. Extensive experiments on image classification tasks, including OOD robustness, diversity, loss surface analyses, and a comparative study with Hamiltonian Monte Carlo, demonstrate the superiority of the proposed approach. | SGMCMC, Bayesian Neural Network, Parameter Expansion | null | 13,380 | 2503.00699 | [
-0.07653748244047165,
-0.10622162371873856,
0.020665334537625313,
0.0499095655977726,
0.057532645761966705,
0.020906973630189896,
-0.02980159781873226,
-0.05554110184311867,
0.013628702610731125,
-0.07602860778570175,
-0.06913872063159943,
-0.041332751512527466,
0.07765036821365356,
-0.007410241756588221,
-0.037261977791786194,
0.007438196334987879,
0.05414539948105812,
0.16183625161647797,
-0.06747108697891235,
-0.013637630268931389,
0.0025110822170972824,
0.05852841958403587,
0.00998625997453928,
-0.006738951895385981,
0.007652070838958025,
-0.03675369918346405,
0.06115547567605972,
-0.055328480899333954,
-0.0688510611653328,
-0.008212552405893803,
0.07604201883077621,
-0.0852551981806755,
-0.09219980984926224,
-0.10370158404111862,
0.04984097182750702,
0.05325816944241524,
-0.0755174309015274,
0.04982411488890648,
0.057854462414979935,
-0.0005527829052880406,
0.02847208082675934,
0.07806556671857834,
0.06983207166194916,
0.003902147524058819,
0.0755571648478508,
-0.021023254841566086,
0.03677637502551079,
-0.10938844084739685,
0.05906294286251068,
0.03464841842651367,
-0.0006669111899100244,
-0.031106971204280853,
-0.06959963589906693,
-0.011202126741409302,
-0.04173032566905022,
0.00506175821647048,
-0.013539261184632778,
-0.057055942714214325,
0.00881657749414444,
0.009154490195214748,
-0.03608141839504242,
-0.09988836944103241,
0.04392985999584198,
0.0016527597326785326,
0.006007150746881962,
0.020320700481534004,
0.03012048825621605,
0.05284977704286575,
0.06371799856424332,
0.03249281644821167,
0.016452722251415253,
0.10573631525039673,
-0.00947581883519888,
0.05290713906288147,
-0.02367795631289482,
0.05121815577149391,
0.050748467445373535,
-0.05042048916220665,
0.03509445860981941,
-0.05402252823114395,
0.029213503003120422,
0.03814927488565445,
0.0305247213691473,
-0.029075181111693382,
0.04991314187645912,
0.044070545583963394,
-0.05915967747569084,
0.0626097023487091,
-0.002713781548663974,
-0.0019122478552162647,
-0.05636531859636307,
0.055307451635599136,
-0.029520530253648758,
-0.07068350911140442,
-0.006085776258260012,
-0.02708016335964203,
0.025242576375603676,
-0.07273923605680466,
-0.006434980314224958,
0.06473306566476822,
0.04302477091550827,
0.04411591961979866,
0.12330859899520874,
-0.011611557565629482,
0.024076523259282112,
-0.04134652018547058,
0.051800526678562164,
-0.014168896712362766,
0.03979773074388504,
-0.0409964919090271,
0.019720885902643204,
0.03171239420771599,
-0.008701442740857601,
-0.044623181223869324,
0.017286958172917366,
0.004377226810902357,
0.028973927721381187,
-0.01911325566470623,
-0.03455511853098869,
0.15053528547286987,
-0.10365273803472519,
-0.005857162177562714,
-0.059633828699588776,
-0.01562273595482111,
0.032396506518125534,
0.014536353759467602,
-0.04888436570763588,
5.8394614095936296e-33,
0.006479789037257433,
-0.021439921110868454,
0.13815072178840637,
0.07042530179023743,
0.07934684306383133,
0.0017735441215336323,
-0.009709246456623077,
-0.07108943164348602,
-0.07427366822957993,
0.01129190530627966,
-0.014768083579838276,
-0.008409533649682999,
-0.10652823746204376,
0.05301135778427124,
0.044855549931526184,
0.007207425776869059,
-0.0664854571223259,
0.05393684282898903,
0.0357150174677372,
-0.02230115421116352,
0.020036857575178146,
-0.028788458555936813,
-0.012109325267374516,
0.02472074329853058,
-0.030926251783967018,
0.012740246951580048,
0.10847143083810806,
0.0548609234392643,
-0.007096669636666775,
0.020303241908550262,
-0.06417539715766907,
-0.0065511902794241905,
-0.021641433238983154,
0.016810592263936996,
0.10928536206483841,
-0.010910941287875175,
-0.058708589524030685,
-0.011356349103152752,
0.012931408360600471,
-0.017293378710746765,
-0.007485503330826759,
-0.0026867275591939688,
0.05307392403483391,
0.03662848100066185,
-0.09724995493888855,
-0.06840458512306213,
0.03272790461778641,
-0.037604980170726776,
-0.005992901045829058,
-0.0202963724732399,
0.020707247778773308,
-0.035499002784490585,
0.00746541703119874,
0.003322154749184847,
-0.029843665659427643,
-0.03880959376692772,
0.004826581105589867,
-0.01898789592087269,
0.02297656238079071,
0.0024937507696449757,
0.013579504564404488,
0.040230125188827515,
-0.07810887694358826,
0.07627058029174805,
0.1041700690984726,
0.08510781079530716,
-0.05039897933602333,
0.09023003280162811,
-0.020078031346201897,
0.018645480275154114,
0.036822348833084106,
-0.00628996267914772,
-0.01709420047700405,
-0.06340885907411575,
0.04246458783745766,
0.056503284722566605,
0.02008407562971115,
-0.0277620330452919,
0.010538499802350998,
0.0374312698841095,
-0.07689168304204941,
-0.036760564893484116,
-0.06336521357297897,
0.04460158571600914,
-0.04453928396105766,
0.03030669130384922,
0.07646434754133224,
-0.03208429366350174,
0.005529256071895361,
0.06144876778125763,
-0.06538853794336319,
-0.05685080960392952,
-0.02525351382791996,
0.0016804346814751625,
-0.07515381276607513,
-5.037635774404623e-33,
-0.050580285489559174,
0.05395813286304474,
0.02243945375084877,
0.027191972360014915,
-0.04202544316649437,
-0.057686761021614075,
0.005030504893511534,
0.03657147288322449,
-0.0015571343246847391,
-0.08896146714687347,
-0.07591512799263,
-0.01781443879008293,
0.029350483790040016,
-0.0399220734834671,
0.03537273034453392,
0.008953840471804142,
-0.048773400485515594,
0.07251696288585663,
-0.010378791019320488,
-0.030973851680755615,
0.03705831989645958,
0.023286888375878334,
0.03019574098289013,
-0.013173455372452736,
-0.05492302402853966,
0.05622216314077377,
-0.01876467652618885,
0.08037466555833817,
0.03105158917605877,
-0.029727216809988022,
-0.06702456623315811,
-0.01001669466495514,
0.017334943637251854,
0.0513124018907547,
-0.05759616568684578,
0.05087656155228615,
0.056819673627614975,
0.00783543474972248,
0.0221652053296566,
0.05972617119550705,
-0.029237128794193268,
-0.06539860367774963,
-0.06251824647188187,
-0.03896353021264076,
0.010331913828849792,
-0.02763538807630539,
0.017934061586856842,
-0.004793407395482063,
0.03456699475646019,
-0.010662213899195194,
-0.05392526835203171,
0.04402240738272667,
-0.14273327589035034,
0.08365922421216965,
-0.07032664120197296,
-0.0725007876753807,
-0.02789568528532982,
0.013267033733427525,
0.11991454660892487,
0.020608386024832726,
-0.08587164431810379,
-0.016418516635894775,
-0.055453039705753326,
-0.027644312009215355,
-0.01078325230628252,
-0.015835609287023544,
-0.03451387211680412,
-0.05007389560341835,
0.04682818427681923,
0.05595381557941437,
-0.031844787299633026,
-0.012823120690882206,
0.08196015655994415,
0.003994252532720566,
0.014318709261715412,
-0.07564616948366165,
0.029224423691630363,
-0.03560250625014305,
0.014476591721177101,
-0.03398602455854416,
0.0602182112634182,
-0.04877964034676552,
0.027037356048822403,
0.05364195629954338,
0.09290847182273865,
0.10151007771492004,
0.0069748167879879475,
-0.02607067860662937,
0.07886743545532227,
-0.019980402663350105,
0.010401255451142788,
0.024507714435458183,
0.04101834073662758,
0.007072648964822292,
-0.11134015023708344,
-4.6020499411270066e-8,
-0.06464622914791107,
0.1529345065355301,
0.08371269702911377,
0.002011471427977085,
0.06943471729755402,
-0.01899711787700653,
0.02298334427177906,
0.027911720797419548,
0.06153178587555885,
0.026286741718649864,
0.11857441812753677,
-0.03453445807099342,
0.01383891236037016,
-0.07242652773857117,
-0.018999893218278885,
0.061018168926239014,
-0.020561546087265015,
0.04072634503245354,
0.04823989421129227,
-0.022381296381354332,
-0.061407364904880524,
-0.013929690234363079,
0.050877850502729416,
0.009779401123523712,
0.024142544716596603,
-0.07244464755058289,
-0.018353870138525963,
0.05383339896798134,
0.0932890921831131,
-0.009260296821594238,
-0.0013074098387733102,
-0.032845254987478256,
-0.04588652402162552,
0.0216037780046463,
-0.03414199501276016,
0.06222464144229889,
-0.07401144504547119,
0.008819513022899628,
-0.001306756865233183,
-0.04011107236146927,
-0.07974433153867722,
0.035822466015815735,
-0.0663633942604065,
0.04027334228157997,
-0.004191678483039141,
-0.011598199605941772,
-0.018273785710334778,
-0.0066666738130152225,
0.010551934130489826,
0.02066698856651783,
-0.031272683292627335,
-0.06946870684623718,
0.0018294164910912514,
0.05658412352204323,
0.03985831141471863,
-0.039115749299526215,
0.03564835339784622,
-0.06403332948684692,
-0.03451309725642204,
0.09335912764072418,
0.027385788038372993,
-0.042729511857032776,
-0.10734518617391586,
-0.043594036251306534
] | 0 | 0 | 0 | 0 |
|
Hyperbolic Genome Embeddings | https://openreview.net/forum?id=NkGDNM8LB0 | [
"Raiyan R. Khan",
"Philippe Chlenski",
"Itsik Pe'er"
] | Poster | Current approaches to genomic sequence modeling often struggle to align the inductive biases of machine learning models with the evolutionarily-informed structure of biological systems. To this end, we formulate a novel application of hyperbolic CNNs that exploits this structure, enabling more expressive DNA sequence representations. Our strategy circumvents the need for explicit phylogenetic mapping while discerning key properties of sequences pertaining to core functional and regulatory behavior. Across 37 out of 43 genome interpretation benchmark datasets, our hyperbolic models outperform their Euclidean equivalents. Notably, our approach even surpasses state-of-the-art performance on seven GUE benchmark datasets, consistently outperforming many DNA language models while using orders of magnitude fewer parameters and avoiding pretraining. Our results include a novel set of benchmark datasets - the Transposable Elements Benchmark - which explores a major but understudied component of the genome with deep evolutionary significance. We further motivate our work by exploring how our hyperbolic models recognize genomic signal under various data-generating conditions and by constructing an empirical method for interpreting the hyperbolicity of dataset embeddings. Throughout these assessments, we find persistent evidence highlighting the potential of our hyperbolic framework as a robust paradigm for genome representation learning. Our code and benchmark datasets are available at https://github.com/rrkhan/HGE. | genomics, representation learning, hyperbolic geometry | A hyperbolic geometry-based approach for genomic sequence representation learning | 13,373 | null | [
-0.10588569939136505,
-0.08995363116264343,
0.006980631500482559,
-0.00972041767090559,
0.04904112592339516,
0.025462010875344276,
-0.07711181044578552,
-0.06493975222110748,
-0.00346895819529891,
-0.10034756362438202,
0.05725196748971939,
-0.10803990811109543,
-0.028306564316153526,
0.018419990316033363,
-0.08221317082643509,
-0.013874118216335773,
0.0420086644589901,
-0.012562237679958344,
-0.04905954375863075,
-0.07537955790758133,
0.05765770748257637,
0.09906726330518723,
0.09561249613761902,
0.0097047733142972,
0.06037084013223648,
-0.0012042280286550522,
-0.017899082973599434,
0.039753127843141556,
0.01826160028576851,
-0.019945187494158745,
0.035830844193696976,
-0.017563972622156143,
-0.06665422022342682,
0.019538698717951775,
-0.002517277142032981,
0.08222360163927078,
-0.01292690634727478,
0.016887854784727097,
0.028310690075159073,
0.0062337107956409454,
0.0718182846903801,
-0.056363385170698166,
0.03407780081033707,
0.0777905210852623,
0.01800372079014778,
-0.022353312000632286,
0.023551302030682564,
-0.0453861728310585,
-0.02406095527112484,
0.02436991222202778,
-0.07004006206989288,
-0.016882671043276787,
0.015576439909636974,
0.08946049958467484,
-0.023686612024903297,
-0.02517956867814064,
0.021332595497369766,
0.007116528227925301,
-0.025307713076472282,
-0.06899809092283249,
0.019626423716545105,
-0.10589791089296341,
-0.044706691056489944,
-0.02396020106971264,
-0.05864608660340309,
0.005651799961924553,
0.04014378786087036,
0.01141513790935278,
-0.04253234714269638,
0.0168449804186821,
-0.006062529049813747,
0.04191012308001518,
-0.0524514801800251,
0.041946981102228165,
0.03448188677430153,
0.10729287564754486,
0.0351562462747097,
0.029946984723210335,
0.10261848568916321,
-0.15720269083976746,
0.051260218024253845,
0.020032303407788277,
0.07550103962421417,
0.004670904017984867,
0.11142060905694962,
-0.0023656112607568502,
0.03904811292886734,
0.030820926651358604,
-0.010235975496470928,
0.010319390334188938,
-0.035384103655815125,
-0.08864552527666092,
0.1054159626364708,
-0.030335891991853714,
0.03733314201235771,
0.10425496101379395,
-0.0080956956371665,
-0.004030621610581875,
-0.043695997446775436,
0.04287710785865784,
-0.04029335454106331,
0.03657640516757965,
0.04962676391005516,
0.04799734801054001,
-0.006337845232337713,
-0.08128403127193451,
0.028446877375245094,
0.02672821469604969,
-0.020256340503692627,
-0.04794579744338989,
0.003848959458991885,
0.06868728250265121,
0.032101694494485855,
0.06309177726507187,
0.04538847878575325,
-0.017933018505573273,
-0.008915816433727741,
-0.015031936578452587,
-0.0039896974340081215,
0.030499225482344627,
-0.04926253855228424,
0.07578234374523163,
-0.046820059418678284,
0.0710141509771347,
0.007358441594988108,
-0.0012771021574735641,
-0.07128027826547623,
2.716545036889244e-33,
0.05035717412829399,
0.06288453936576843,
0.05204858258366585,
0.0073715560138225555,
0.022935857996344566,
-0.04992144927382469,
-0.049651604145765305,
0.007400063332170248,
-0.09461180865764618,
0.05249267444014549,
-0.08053046464920044,
0.030039269477128983,
-0.06038827449083328,
0.10216288268566132,
-0.05565393716096878,
0.0264018215239048,
-0.03519734740257263,
0.03960814327001572,
-0.060008302330970764,
-0.001629435340873897,
-0.03632636368274689,
-0.025131115689873695,
-0.006907855626195669,
-0.12578071653842926,
-0.01861215941607952,
0.03944019228219986,
-0.032135386019945145,
-0.04585212469100952,
-0.09272309392690659,
0.026751549914479256,
-0.1298634260892868,
-0.017480265349149704,
0.05721740424633026,
0.054633695632219315,
0.03873821720480919,
-0.06482700258493423,
0.024206271395087242,
-0.054843444377183914,
0.0325445756316185,
0.0011858503567054868,
-0.011625101789832115,
0.05951354280114174,
-0.016836317256093025,
0.008428320288658142,
-0.07440279424190521,
-0.03525485843420029,
0.04204639419913292,
-0.053567513823509216,
-0.04532027989625931,
-0.011313569732010365,
0.05013788118958473,
0.01892975904047489,
0.025335924699902534,
0.007149933837354183,
-0.05495085194706917,
-0.01633206009864807,
-0.01224670372903347,
0.0033174369018524885,
-0.01865214668214321,
0.037429045885801315,
0.019584398716688156,
0.03587222471833229,
0.018991215154528618,
0.06552527099847794,
-0.00328297121450305,
-0.03166000917553902,
-0.08503305912017822,
0.0765363872051239,
0.004576967563480139,
0.06952521950006485,
0.010875798761844635,
-0.007664710283279419,
-0.017508385702967644,
-0.031032640486955643,
0.054616957902908325,
0.004198733251541853,
0.06313124299049377,
-0.1177486777305603,
-0.06803180277347565,
0.0754140093922615,
-0.026514600962400436,
0.00792127475142479,
-0.029687173664569855,
-0.048529140651226044,
-0.12519115209579468,
-0.0628541111946106,
0.07373813539743423,
-0.07163261622190475,
0.009142320603132248,
-0.07536637037992477,
-0.03187376633286476,
-0.027027608826756477,
-0.01830686815083027,
-0.04374239966273308,
0.03389064222574234,
-4.463912613940774e-33,
-0.0062267836183309555,
0.01303828228265047,
-0.03698458895087242,
-0.047464519739151,
-0.0724797397851944,
-0.04852481186389923,
-0.0006192304426804185,
0.0357716903090477,
-0.06347796320915222,
-0.08191292732954025,
0.08315952867269516,
-0.035006601363420486,
0.017543232068419456,
-0.06599105149507523,
0.09670516848564148,
-0.02496339939534664,
-0.022269446402788162,
0.026956787332892418,
0.07591305673122406,
0.005385877098888159,
0.061855461448431015,
0.09276128560304642,
-0.10424039512872696,
0.061281416565179825,
-0.04161321371793747,
0.01617565006017685,
-0.003022100543603301,
0.09463390707969666,
-0.008064433932304382,
-0.041436195373535156,
-0.13990913331508636,
0.01802157610654831,
-0.03891081362962723,
-0.02111496776342392,
-0.07370173186063766,
0.08909555524587631,
0.06421440094709396,
0.06326087564229965,
-0.02491792105138302,
0.015114889480173588,
0.04372483864426613,
0.01700124889612198,
-0.11358775198459625,
0.07663045078516006,
0.0183408223092556,
0.07147547602653503,
-0.01784541644155979,
0.0365879088640213,
0.026242269203066826,
0.018452798947691917,
0.0412851944565773,
0.05307526886463165,
-0.029497113078832626,
0.02172749489545822,
-0.06541337072849274,
-0.0829719752073288,
-0.0780913308262825,
-0.0038633355870842934,
0.000764146214351058,
-0.023913761600852013,
-0.09313531219959259,
-0.004991406574845314,
-0.04457499459385872,
-0.021114828065037727,
0.01153161097317934,
-0.07774334400892258,
-0.028601277619600296,
-0.06292906403541565,
0.04737785831093788,
0.0038513443432748318,
-0.05204113945364952,
-0.06329900026321411,
0.008246354758739471,
0.018125973641872406,
0.006303589791059494,
-0.026779036968946457,
-0.02584114670753479,
-0.06440651416778564,
-0.025723202154040337,
-0.004887291695922613,
0.03307938203215599,
0.021999942138791084,
0.0007393668056465685,
0.00009719199442770332,
0.09707264602184296,
0.0799049586057663,
0.002150941640138626,
0.050877321511507034,
0.044831134378910065,
0.020698966458439827,
-0.008242279291152954,
-0.013258355669677258,
-0.10726834833621979,
0.0968082845211029,
-0.04022028297185898,
-4.628543237572558e-8,
0.045037951320409775,
0.036077409982681274,
0.023593364283442497,
-0.019619878381490707,
-0.03921206295490265,
0.0013897663448005915,
-0.047124434262514114,
0.1313360631465912,
0.01473052054643631,
-0.03451886773109436,
0.08720337599515915,
0.028953012079000473,
-0.0025342691224068403,
0.004851687699556351,
-0.016919316723942757,
0.1080072745680809,
0.010540534742176533,
0.05428166687488556,
0.008016154170036316,
0.02196010947227478,
-0.018524963408708572,
0.058591749519109726,
-0.01319847535341978,
-0.016074148938059807,
0.042851366102695465,
-0.04458971321582794,
0.0016056002350524068,
0.03409339487552643,
0.036631882190704346,
-0.018188070505857468,
0.03475512936711311,
0.015074584633111954,
0.007257178891450167,
-0.06552908569574356,
-0.0450570322573185,
0.05672817677259445,
-0.04598859325051308,
-0.02915741689503193,
0.04726275056600571,
-0.025027954950928688,
0.04800146073102951,
0.06995869427919388,
-0.08475617319345474,
-0.006430990062654018,
0.02640877477824688,
0.0351388156414032,
-0.005546422675251961,
0.0018345286371186376,
0.07423736155033112,
0.007904269732534885,
0.0026634354144334793,
-0.01609291508793831,
0.05084621533751488,
0.0030112392269074917,
0.04093397408723831,
-0.03960737586021423,
-0.007117270492017269,
-0.059927232563495636,
0.0038869346026331186,
0.039548423141241074,
0.0669446736574173,
0.008857435546815395,
0.004538464825600386,
-0.07825926691293716
] | 0 | 0 | 0 | 0 |
|
Deep Distributed Optimization for Large-Scale Quadratic Programming | https://openreview.net/forum?id=hzuumhfYSO | [
"Augustinos D Saravanos",
"Hunter Kuperman",
"Alex Oshin",
"Arshiya Taj Abdul",
"Vincent Pacelli",
"Evangelos Theodorou"
] | Poster | Quadratic programming (QP) forms a crucial foundation in optimization, appearing in a broad spectrum of domains and serving as the basis for more advanced algorithms. Consequently, as the scale and complexity of modern applications continue to grow, the development of efficient and reliable QP algorithms becomes increasingly vital. In this context, this paper introduces a novel deep learning-aided distributed optimization architecture designed for tackling large-scale QP problems. First, we combine the state-of-the-art Operator Splitting QP (OSQP) method with a consensus approach to derive **DistributedQP**, a new method tailored for network-structured problems, with convergence guarantees to optimality. Subsequently, we unfold this optimizer into a deep learning framework, leading to **DeepDistributedQP**, which leverages learned policies to accelerate reaching to desired accuracy within a restricted amount of iterations. Our approach is also theoretically grounded through Probably Approximately Correct (PAC)-Bayes theory, providing generalization bounds on the expected optimality gap for unseen problems. The proposed framework, as well as its centralized version **DeepQP**, significantly outperform their standard optimization counterparts on a variety of tasks such as randomly generated problems, optimal control, linear regression, transportation networks and others. Notably, DeepDistributedQP demonstrates strong generalization by training on small problems and scaling to solve much larger ones (up to 50K variables and 150K constraints) using the same policy. Moreover, it achieves orders-of-magnitude improvements in wall-clock time compared to OSQP. The certifiable performance guarantees of our approach are also demonstrated, ensuring higher-quality solutions over traditional optimizers. | Learning-to-Optimize, Distributed Optimization, Large-Scale Quadratic Programming | We propose DeepDistributedQP, a new deep learning-aided distributed optimization architecture for large-scale quadratic programming. | 13,359 | 2412.12156 | [
-0.08510301262140274,
-0.01924038492143154,
0.04039089381694794,
0.024212006479501724,
-0.06019148975610733,
0.02425931952893734,
-0.06228505074977875,
-0.03259465470910072,
-0.04950200021266937,
-0.018391795456409454,
-0.07792259007692337,
-0.012621375732123852,
0.00033867001184262335,
0.04102681204676628,
-0.04231990873813629,
0.08758271485567093,
0.11235883086919785,
0.05696408450603485,
-0.10994299501180649,
-0.05767558515071869,
-0.041762564331293106,
0.009824353270232677,
-0.016798129305243492,
-0.04460257664322853,
0.0765327736735344,
-0.01888909749686718,
0.025668932124972343,
-0.059002745896577835,
0.018414951860904694,
0.03729772940278053,
0.09354833513498306,
-0.0543847493827343,
-0.024881714954972267,
-0.06184489279985428,
0.021487019956111908,
0.10248903930187225,
-0.09289876371622086,
-0.030365221202373505,
0.047332923859357834,
-0.017632361501455307,
0.04533790051937103,
0.05653462931513786,
-0.02743469923734665,
0.012335852719843388,
0.04631292074918747,
0.006881566718220711,
0.02453019469976425,
0.03527321293950081,
0.07089375704526901,
-0.05762716010212898,
-0.03099512867629528,
-0.056633949279785156,
-0.0012188772670924664,
0.09517351537942886,
0.03417781740427017,
-0.012809514068067074,
0.016520509496331215,
-0.04291967302560806,
0.009021342732012272,
0.0003792368806898594,
0.017782265320420265,
-0.10631373524665833,
0.02257775515317917,
0.014213684014976025,
-0.023602887988090515,
0.025701330974698067,
0.03198199346661568,
0.02641887404024601,
-0.010523215867578983,
0.007730023004114628,
-0.03236440196633339,
0.0905173048377037,
-0.02433621883392334,
0.030054539442062378,
0.01829490438103676,
-0.015282579697668552,
0.08353891223669052,
0.001194590120576322,
0.0448177345097065,
-0.035454850643873215,
0.07854720205068588,
0.0684090256690979,
-0.03469304367899895,
-0.032534822821617126,
0.15126386284828186,
-0.0658552497625351,
-0.01789202354848385,
0.025851156562566757,
0.006961862090975046,
-0.028698738664388657,
-0.03557441383600235,
-0.007956355810165405,
-0.060591213405132294,
-0.04803479462862015,
0.029829835519194603,
-0.05652693659067154,
0.022959480062127113,
-0.07349712401628494,
-0.06522160023450851,
0.09791531413793564,
0.036614347249269485,
0.03771410882472992,
0.04653339460492134,
-0.08444491028785706,
-0.010654859244823456,
-0.013391256332397461,
0.014778543263673782,
0.07539555430412292,
0.013769340701401234,
-0.08438699692487717,
0.053897034376859665,
0.07253700494766235,
0.04516034200787544,
0.05350199714303017,
-0.0009492353419773281,
0.026118427515029907,
0.039229169487953186,
0.013388989493250847,
-0.024399492889642715,
0.03164738416671753,
-0.10126134008169174,
-0.05512787774205208,
0.02418760396540165,
0.035525836050510406,
0.017217978835105896,
0.04100632667541504,
-0.09587113559246063,
8.337363578874966e-34,
0.0043927691876888275,
-0.018296172842383385,
0.07791046053171158,
-0.08594082295894623,
0.1168724074959755,
-0.019219551235437393,
0.06283263862133026,
0.05735531449317932,
-0.11886764317750931,
0.030864350497722626,
0.031943704932928085,
-0.05556764826178551,
-0.08713123947381973,
0.05374857783317566,
0.05716739967465401,
-0.0307980515062809,
-0.060917675495147705,
0.03263929858803749,
0.021870208904147148,
-0.1017511710524559,
0.07700134813785553,
-0.05712738633155823,
-0.07181259989738464,
0.029656806960701942,
0.029702510684728622,
-0.0681113749742508,
0.09462721645832062,
-0.015054328367114067,
0.03175778687000275,
0.03477982059121132,
-0.14438867568969727,
0.015041658654808998,
-0.009016869589686394,
0.06911426037549973,
0.03616413474082947,
0.05256874859333038,
-0.051469601690769196,
-0.008654524572193623,
-0.00835186056792736,
-0.00500805489718914,
-0.016706664115190506,
-0.017104333266615868,
0.03602415695786476,
-0.07071442902088165,
-0.08319239318370819,
-0.08981809765100479,
-0.013232989236712456,
-0.01771673560142517,
-0.01430879533290863,
-0.126052126288414,
0.020959103479981422,
0.012836244888603687,
-0.021230828016996384,
-0.016496319323778152,
0.035977452993392944,
-0.03332764282822609,
0.06398448348045349,
0.0396374836564064,
-0.012401794083416462,
0.045125603675842285,
0.06910424679517746,
-0.012018373236060143,
-0.06592127680778503,
0.06495800614356995,
-0.034580402076244354,
0.05438229814171791,
-0.0030414932407438755,
0.05713605135679245,
0.053305722773075104,
-0.007310870103538036,
0.018450016155838966,
0.03249172121286392,
0.04147585853934288,
-0.061285730451345444,
0.08285582810640335,
-0.0004934699973091483,
0.06608888506889343,
-0.02431941218674183,
-0.02118965983390808,
-0.021310677751898766,
0.005484397057443857,
0.10336616635322571,
-0.06364057958126068,
0.0071450937539339066,
0.021401166915893555,
-0.07991392910480499,
0.046017445623874664,
0.011063000187277794,
-0.0627763420343399,
0.009714058600366116,
-0.12664134800434113,
-0.004330079536885023,
0.013034111820161343,
0.015439414419233799,
0.006453086622059345,
-1.4964617976173669e-33,
-0.04440386965870857,
0.010502435266971588,
-0.021577829495072365,
0.022569870576262474,
0.010629608295857906,
-0.09312336891889572,
-0.09638138115406036,
-0.08482930064201355,
-0.028960613533854485,
-0.07635942846536636,
-0.03924253210425377,
-0.0062761493027210236,
-0.006544174626469612,
-0.019205331802368164,
0.00048234444693662226,
-0.04007713124155998,
0.004199740011245012,
0.03202587366104126,
0.021133583039045334,
-0.019514424726366997,
-0.013688139617443085,
0.03741026297211647,
-0.004087154287844896,
0.07476001232862473,
0.04901961609721184,
0.0006850989884696901,
0.027811406180262566,
0.03851927071809769,
0.028023308143019676,
0.013302711769938469,
-0.11885351687669754,
-0.03922522813081741,
-0.12442824244499207,
-0.05977427214384079,
-0.08643341064453125,
0.04485996440052986,
0.002334196586161852,
-0.019346017390489578,
-0.026546118780970573,
0.08575693517923355,
0.08314862847328186,
-0.09570643305778503,
-0.10795588046312332,
0.0682748407125473,
0.018496684730052948,
-0.020410967990756035,
0.005390552803874016,
-0.038838744163513184,
0.011898962780833244,
0.07410915195941925,
0.029502974823117256,
0.04163411259651184,
-0.04285416007041931,
0.07559797167778015,
-0.06336690485477448,
0.004886290989816189,
-0.01409329567104578,
0.04337625205516815,
0.057237621396780014,
0.011240053921937943,
-0.07312938570976257,
-0.041136499494314194,
-0.028863780200481415,
0.03089287504553795,
0.012981952168047428,
-0.00625309394672513,
0.0011509787291288376,
0.0796871930360794,
0.013021459802985191,
0.04286007955670357,
-0.013590237125754356,
0.018183372914791107,
0.01602792739868164,
0.08993469923734665,
0.011653772555291653,
0.03441004827618599,
0.005634253844618797,
-0.005673923995345831,
-0.021428657695651054,
0.03228555992245674,
0.07418198138475418,
-0.0001113066027755849,
0.023501383140683174,
0.054941046983003616,
0.055718328803777695,
-0.013989533297717571,
0.08611354231834412,
0.07518106698989868,
0.015110956504940987,
-0.022861871868371964,
-0.02270331420004368,
0.008087974973022938,
-0.017620109021663666,
-0.03760494664311409,
-0.08650609850883484,
-4.204011005981556e-8,
-0.05804644897580147,
0.07140350341796875,
-0.013637247495353222,
-0.02640604041516781,
0.06640937179327011,
-0.018612176179885864,
0.03305691480636597,
0.02935393527150154,
-0.015154759399592876,
-0.02174622006714344,
0.01299214456230402,
0.06896475702524185,
0.002188665559515357,
-0.029935767874121666,
-0.10875757783651352,
0.045902643352746964,
-0.04618711397051811,
-0.07322988659143448,
0.007167413365095854,
-0.03651959449052811,
-0.015077916905283928,
0.029429925605654716,
0.020954487845301628,
0.0020057677756994963,
-0.04337134584784508,
-0.05436838045716286,
0.010224924422800541,
0.050848204642534256,
0.09047112613916397,
0.04450013116002083,
-0.08762103319168091,
-0.021130869165062904,
-0.031789060682058334,
-0.0457552894949913,
-0.02461284212768078,
0.09206776320934296,
0.0202657338231802,
-0.014178217388689518,
0.04854876548051834,
-0.08049428462982178,
0.017713207751512527,
0.05391314998269081,
0.0098050432279706,
-0.002750647021457553,
-0.005986475385725498,
-0.0023640459403395653,
-0.02699659764766693,
0.08114994317293167,
0.0695827454328537,
0.013529792428016663,
0.026181451976299286,
0.020094487816095352,
0.019896386191248894,
0.07762335240840912,
0.043653856962919235,
0.0004165750870015472,
-0.00212213140912354,
-0.08503472059965134,
0.051551803946495056,
0.05079406127333641,
-0.01217611227184534,
0.0015537934377789497,
-0.07926320284605026,
0.05181674286723137
] | 0 | 0 | 0 | 0 |
|
Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining | https://openreview.net/forum?id=gU4ZgQNsOC | [
"Daouda Sow",
"Herbert Woisetschläger",
"Saikiran Bulusu",
"Shiqiang Wang",
"Hans Arno Jacobsen",
"Yingbin Liang"
] | Poster | Pretraining large language models (LLMs) on vast and heterogeneous datasets is crucial for achieving state-of-the-art performance across diverse downstream tasks. However, current training paradigms treat all samples equally, overlooking the importance or relevance of individual samples throughout the training process. Existing reweighting strategies, which primarily focus on group-level data importance, fail to leverage fine-grained instance-level information and do not adapt dynamically to individual sample importance as training progresses. In this paper, we introduce novel algorithms for dynamic, instance-level data reweighting aimed at improving both the efficiency and effectiveness of LLM pretraining. Our methods adjust the weight of each training sample based on its loss value in an online fashion, allowing the model to dynamically focus on more informative or important samples at the current training stage. In particular, our framework allows us to systematically devise reweighting strategies deprioritizing redundant or uninformative data, which we find tend to work best.
Furthermore, we develop a new theoretical framework for analyzing the impact of loss-based reweighting on the convergence of gradient-based optimization, providing the first formal characterization of how these strategies affect convergence bounds. We empirically validate our approach across a spectrum of tasks, from pretraining 7B and 1.4B parameter LLMs to smaller-scale language models and linear regression problems, demonstrating that our loss-based reweighting approach can lead to faster convergence and significantly improved performance. | sample reweighing, large language models, pretraining | We propose new loss-based sample reweighing techniques for improved efficiency and effectiveness of LLMs pretraining. | 13,352 | 2502.06733 | [
-0.015575848519802094,
-0.04583363980054855,
0.06624405831098557,
0.08347561210393906,
0.08033335953950882,
0.04562985152006149,
-0.019895052537322044,
0.004288932308554649,
-0.030170800164341927,
-0.08852244913578033,
-0.10314597934484482,
-0.04732238128781319,
0.020437274128198624,
0.026027265936136246,
0.0008436777861788869,
-0.02169688045978546,
0.11525071412324905,
0.11915530264377594,
-0.13462500274181366,
-0.10506340861320496,
0.05257303640246391,
0.036644626408815384,
0.055394090712070465,
0.07812008261680603,
0.055196456611156464,
-0.03372463583946228,
0.00952273141592741,
0.003991244360804558,
0.05435405299067497,
-0.019441870972514153,
-0.0003911419480573386,
0.01747754216194153,
-0.020398929715156555,
0.01263261679559946,
-0.09189597517251968,
0.03723585605621338,
-0.04626478999853134,
0.00423314468935132,
0.0026330603286623955,
0.028746744617819786,
-0.012159980833530426,
0.00577121414244175,
0.009980601258575916,
0.03007037192583084,
0.025254348292946815,
-0.03612677752971649,
0.04233020171523094,
-0.024331163614988327,
-0.006875989027321339,
0.08941803127527237,
-0.04622303321957588,
-0.01601225882768631,
-0.05128124728798866,
0.08684764802455902,
-0.039413273334503174,
0.02030273899435997,
0.08095794171094894,
0.028866806998848915,
-0.05020773410797119,
0.03626007214188576,
-0.07107456773519516,
-0.10445527732372284,
0.054428309202194214,
-0.032656412571668625,
-0.05021078139543533,
-0.047211144119501114,
0.017035359516739845,
0.0786258727312088,
0.0014397939667105675,
0.08609866350889206,
-0.00896730087697506,
0.07387784123420715,
-0.014543276280164719,
0.08530142903327942,
-0.031949229538440704,
0.08043580502271652,
-0.0098948422819376,
0.023807527497410774,
0.015408444218337536,
-0.011076469905674458,
-0.005251288414001465,
0.05348430573940277,
-0.0006896016420796514,
-0.03401434049010277,
0.09407737106084824,
0.005784810986369848,
0.020659862086176872,
-0.06850624829530716,
0.07378656417131424,
0.013225119560956955,
-0.08368116617202759,
0.00986375194042921,
0.0476696714758873,
0.02210623398423195,
0.04801827669143677,
0.0113359484821558,
-0.01763796992599964,
-0.0896158367395401,
0.004781398456543684,
0.05155942589044571,
-0.004726184997707605,
0.08988823741674423,
0.000037359761336119846,
0.01795203424990177,
-0.043211571872234344,
-0.06653376668691635,
0.04329642653465271,
0.05430098995566368,
0.008589987643063068,
-0.05677129700779915,
0.09472767263650894,
0.04716828465461731,
0.005518556106835604,
0.007064358331263065,
0.036531589925289154,
0.021976009011268616,
0.036862563341856,
-0.058724526315927505,
-0.08288867026567459,
0.12371287494897842,
-0.04155484959483147,
0.012720507569611073,
0.04634847491979599,
-0.058678045868873596,
0.02708069607615471,
-0.009965541772544384,
-0.07677322626113892,
3.462784850305726e-33,
0.04087641462683678,
0.019857296720147133,
0.04463251307606697,
0.047193337231874466,
-0.002539814915508032,
-0.10191097855567932,
0.00309277162887156,
0.05458442121744156,
-0.036435581743717194,
0.006032339297235012,
0.0021585423965007067,
0.031601984053850174,
-0.037179313600063324,
0.03264414891600609,
0.026640236377716064,
-0.054352059960365295,
-0.04655062034726143,
0.10897966474294662,
-0.007852121256291866,
-0.02893275022506714,
-0.0009433826780878007,
0.02041543275117874,
-0.029969779774546623,
-0.10374345630407333,
0.03078484907746315,
0.06257440149784088,
0.055430538952350616,
-0.04060421884059906,
-0.056457601487636566,
0.01582309976220131,
-0.0507318489253521,
-0.07068829983472824,
-0.030476955696940422,
-0.04549768567085266,
-0.021493077278137207,
-0.05494166910648346,
-0.035479478538036346,
-0.0074193235486745834,
0.08073106408119202,
0.009634224697947502,
0.04183158278465271,
0.027481863275170326,
0.060870300978422165,
-0.05067041143774986,
-0.058482684195041656,
-0.09327389299869537,
-0.006525631994009018,
0.008000687696039677,
-0.00536262197420001,
-0.05823850259184837,
0.06696602702140808,
-0.03978902846574783,
-0.024747047573328018,
-0.0019873245619237423,
-0.07260911911725998,
0.0020661987364292145,
-0.059285134077072144,
0.03926241770386696,
0.046913232654333115,
-0.002056432655081153,
-0.01359097845852375,
0.05106775835156441,
-0.06100449338555336,
0.12368683516979218,
0.016743574291467667,
-0.03161158785223961,
0.026304353028535843,
0.01868456043303013,
-0.032496120780706406,
-0.016237713396549225,
0.02423868328332901,
-0.028251616284251213,
-0.02034594677388668,
-0.07588637620210648,
0.10056361556053162,
-0.024026120081543922,
0.056727588176727295,
-0.06610625237226486,
0.009713136591017246,
0.005928698927164078,
0.025105444714426994,
0.006410505622625351,
0.020427878946065903,
-0.10662108659744263,
-0.045813798904418945,
0.004786190576851368,
0.09524219483137131,
-0.0973089337348938,
0.025430750101804733,
0.02169855311512947,
-0.05288965627551079,
-0.0030324917752295732,
-0.052527107298374176,
-0.03321409225463867,
-0.0030637122690677643,
-1.9418775456370557e-33,
-0.02923797257244587,
0.002643843414261937,
-0.004870252683758736,
0.06637637317180634,
0.03555890917778015,
-0.0309139396995306,
-0.00430881604552269,
0.01658872701227665,
-0.042276252061128616,
-0.02725302241742611,
-0.027880877256393433,
-0.11081419140100479,
-0.0034046457149088383,
-0.011167958378791809,
0.0392918735742569,
-0.00780254602432251,
-0.07467760890722275,
0.07409297674894333,
0.041498951613903046,
0.05988346040248871,
0.029450509697198868,
0.07460857927799225,
-0.08152497559785843,
0.01640821248292923,
0.021292397752404213,
0.020243825390934944,
-0.030201641842722893,
0.11840660125017166,
-0.002926378743723035,
-0.04470869526267052,
-0.07126365602016449,
-0.053406961262226105,
-0.060740478336811066,
-0.06339959055185318,
-0.11060180515050888,
0.04549103602766991,
0.05155700817704201,
-0.02298162505030632,
-0.024602025747299194,
0.1473221331834793,
0.07095006853342056,
-0.023813316598534584,
-0.09376032650470734,
-0.03377745300531387,
-0.05654921010136604,
-0.041731592267751694,
-0.06858466565608978,
0.006822541821748018,
0.061635278165340424,
0.043816544115543365,
-0.05930078402161598,
0.02125663124024868,
-0.09653680771589279,
0.06324968487024307,
0.007678487803786993,
-0.049759477376937866,
-0.006199476774781942,
-0.06367382407188416,
-0.035711001604795456,
0.018825476989150047,
-0.05717339366674423,
0.01462949812412262,
0.024014104157686234,
-0.033869434148073196,
0.017325572669506073,
-0.07251827418804169,
0.009943854063749313,
-0.037111468613147736,
0.0039732493460178375,
0.027870774269104004,
0.019847484305500984,
-0.008645647205412388,
0.0003061034076381475,
-0.011723696254193783,
-0.04016018658876419,
0.014458917081356049,
0.018115317448973656,
-0.0438477098941803,
-0.01720910705626011,
-0.058162666857242584,
0.04614487662911415,
0.0017067758599296212,
-0.031379882246255875,
0.03506360203027725,
0.06987198442220688,
0.1584937572479248,
0.01467067189514637,
0.119583360850811,
0.02189456857740879,
0.04817107692360878,
-0.02462857775390148,
-0.054482609033584595,
0.004953195806592703,
0.006611906923353672,
-0.06511739641427994,
-4.3192461163243934e-8,
-0.06967924535274506,
0.0516471192240715,
0.019383331760764122,
0.1323988437652588,
0.02598850429058075,
-0.07065598666667938,
-0.023569459095597267,
0.12577812373638153,
0.023150043562054634,
-0.005236055701971054,
0.04045424982905388,
-0.04482046514749527,
0.0006579336477443576,
-0.054746221750974655,
-0.030766000971198082,
0.07992448657751083,
-0.04579021409153938,
0.022227447479963303,
-0.032708778977394104,
0.005942745599895716,
0.018654905259609222,
0.03192722052335739,
0.022090014070272446,
0.02391786500811577,
0.0532560758292675,
-0.05883793532848358,
0.042532067745923996,
0.09746773540973663,
0.0734327882528305,
-0.04371387138962746,
-0.051518335938453674,
0.006372263189405203,
-0.03747846186161041,
0.01744239777326584,
0.021797072142362595,
0.020648254081606865,
0.0018408283358439803,
-0.0017609219066798687,
-0.022601651027798653,
-0.010360163636505604,
-0.002542126690968871,
0.039147596806287766,
-0.01464463584125042,
0.033489469438791275,
0.0841200202703476,
-0.03477928787469864,
-0.0830797478556633,
-0.017378924414515495,
0.014752156101167202,
-0.05647847428917885,
-0.035918064415454865,
-0.0433364100754261,
0.0016133434837684035,
0.05983391031622887,
0.09207156300544739,
0.024881262332201004,
-0.0435560904443264,
0.019946368411183357,
0.04696274548768997,
0.030842328444123268,
0.051752544939517975,
-0.02580006793141365,
-0.03837825730443001,
-0.028098467737436295
] | 0 | 0 | 0 | 0 |
|
Frame-Voyager: Learning to Query Frames for Video Large Language Models | https://openreview.net/forum?id=LNL7zKvm7e | [
"Sicheng Yu",
"CHENGKAI JIN",
"Huanyu Wang",
"Zhenghao Chen",
"Sheng Jin",
"ZHONGRONG ZUO",
"XU XIAOLEI",
"Zhenbang Sun",
"Bingni Zhang",
"Jiawei Wu",
"Hao Zhang",
"Qianru Sun"
] | Poster | Video Large Language Models (Video-LLMs) have made remarkable progress in video understanding tasks. However, they are constrained by the maximum length of input tokens, making it impractical to input entire videos. Existing frame selection approaches, such as uniform frame sampling and text-frame retrieval, fail to account for the information density variations in the videos or the complex instructions in the tasks, leading to sub-optimal performance. In this paper, we propose Frame-Voyager that learns to query informative frame combinations, based on the given textual queries in the task. To train Frame-Voyager, we introduce a new data collection and labeling pipeline, by ranking frame combinations using a pre-trained Video-LLM. Given a video of M frames, we traverse its T-frame combinations, feed them into a Video-LLM, and rank them based on Video-LLM's prediction losses. Using this ranking as supervision, we train Frame-Voyager to query the frame combinations with lower losses. In experiments, we evaluate Frame-Voyager on four Video Question Answering benchmarks by plugging it into two different Video-LLMs. The experimental results demonstrate that Frame-Voyager achieves impressive results in all settings, highlighting its potential as a plug-and-play solution for Video-LLMs. | Video-LLM, Adaptive Frame Sampling | Frame-Voyager queries optimal frame combinations for Video-LLMs based on task-specific queries, outperforming existing SOTA methods and achieving best results in Video Question Answering benchmarks as a plug-and-play solution. | 13,344 | null | [
0.009470322169363499,
-0.10317957401275635,
0.006996359676122665,
-0.021154798567295074,
0.1014433205127716,
0.08812453597784042,
0.002526845782995224,
0.057847436517477036,
0.05332096293568611,
-0.08044897019863129,
-0.050997354090213776,
-0.02596357837319374,
-0.028906263411045074,
0.10190290212631226,
-0.026378702372312546,
-0.004602418746799231,
0.0900486633181572,
0.03198505565524101,
-0.06619707494974136,
-0.06659252941608429,
-0.018151750788092613,
-0.015566256828606129,
0.12462425231933594,
-0.005667960271239281,
0.07115179300308228,
-0.018013661727309227,
-0.030269721522927284,
-0.01493688952177763,
0.038187168538570404,
-0.012303486466407776,
0.021694205701351166,
0.031442634761333466,
-0.007006336934864521,
0.016823414713144302,
-0.061787351965904236,
-0.019583161920309067,
-0.08244562894105911,
-0.034770384430885315,
-0.031568337231874466,
-0.054537586867809296,
0.009487608447670937,
0.032700490206480026,
0.04719449207186699,
0.021277351304888725,
0.0346209779381752,
0.012718805111944675,
0.017574112862348557,
-0.07925046235322952,
0.03046312741935253,
0.04968874156475067,
-0.09628870338201523,
0.0332946702837944,
-0.01737368293106556,
0.06980326026678085,
-0.09932003915309906,
0.0056604184210300446,
-0.018902413547039032,
-0.02297801338136196,
0.02771512046456337,
-0.034319985657930374,
-0.03951245918869972,
-0.14644432067871094,
0.0025791446678340435,
0.03378654271364212,
-0.009227980859577656,
-0.050414469093084335,
0.05905024707317352,
0.027551697567105293,
-0.03535423427820206,
0.016920428723096848,
-0.0627913549542427,
0.07247766107320786,
0.008375412784516811,
0.11388937383890152,
-0.02711201272904873,
0.04882277920842171,
0.03406510502099991,
-0.04985782876610756,
0.021820034831762314,
-0.0531264990568161,
0.018495500087738037,
-0.009127902798354626,
0.06269998848438263,
0.03358093276619911,
0.14716599881649017,
0.004528965335339308,
0.05372585356235504,
0.028019053861498833,
-0.004782387055456638,
0.02002030797302723,
-0.08750716596841812,
-0.031135626137256622,
0.0935288816690445,
-0.01749141700565815,
0.011388054117560387,
0.037903185933828354,
-0.04797874391078949,
-0.04619557410478592,
-0.03978605195879936,
0.061341483145952225,
0.05018308758735657,
0.021857742220163345,
0.005462469067424536,
-0.037676285952329636,
0.004637415055185556,
0.031199423596262932,
-0.01671508513391018,
0.054208673536777496,
0.04512747377157211,
-0.10019690543413162,
0.05003851279616356,
0.029635107144713402,
0.027770143002271652,
0.003368993056938052,
0.02124827913939953,
0.03097333386540413,
0.03440738096833229,
-0.0011987722245976329,
-0.016331758350133896,
0.010227913036942482,
-0.001004807185381651,
0.018605582416057587,
-0.03640880063176155,
-0.05384596437215805,
-0.052552007138729095,
-0.04900302365422249,
0.040167782455682755,
-4.993679999865547e-34,
0.09964237362146378,
0.006640351377427578,
0.11741653829813004,
0.02085723914206028,
0.02408250793814659,
0.02950880490243435,
0.0038976622745394707,
0.11670804768800735,
-0.045827850699424744,
0.009808090515434742,
-0.030194664373993874,
-0.07489040493965149,
-0.0028348476625978947,
0.01228028628975153,
-0.014336539432406425,
-0.04350637272000313,
-0.059431444853544235,
0.009996850043535233,
-0.022727813571691513,
0.04438568651676178,
0.04244980588555336,
0.031125608831644058,
0.055314913392066956,
-0.04013986140489578,
0.051602210849523544,
-0.054135724902153015,
0.04420118033885956,
-0.09205814450979233,
-0.060790445655584335,
-0.01777748577296734,
-0.16287028789520264,
-0.1108710840344429,
0.03859075903892517,
0.0491078719496727,
0.07151008397340775,
-0.027079785242676735,
-0.05488439276814461,
-0.018349986523389816,
0.00864656176418066,
-0.05093177407979965,
0.001393654616549611,
-0.006159622687846422,
0.012601513415575027,
-0.017688298597931862,
-0.12440960109233856,
-0.04925434663891792,
0.01500457338988781,
0.04154781624674797,
-0.0582127645611763,
0.051809269934892654,
0.05710449442267418,
-0.0016393130645155907,
-0.04433440789580345,
-0.034890804439783096,
0.0655275508761406,
0.014219805598258972,
0.07263071089982986,
0.05758321285247803,
-0.01668653078377247,
-0.014327160082757473,
-0.03291967138648033,
0.02730139158666134,
-0.004837421700358391,
0.07717055827379227,
0.019478285685181618,
0.054858751595020294,
0.03126382455229759,
0.07923374325037003,
0.07063134759664536,
-0.006986996158957481,
-0.03075375221669674,
0.0017334169242531061,
0.03646177798509598,
-0.05210844799876213,
0.08638438582420349,
0.07629793137311935,
0.0026514220517128706,
-0.1143048033118248,
-0.01292736642062664,
0.06423350423574448,
-0.0165287796407938,
-0.08642448484897614,
0.03794795647263527,
-0.053653642535209656,
0.003857838222756982,
-0.035054583102464676,
0.06404674053192139,
-0.05856161564588547,
0.08001973479986191,
-0.03990717977285385,
-0.01605626754462719,
-0.05099620670080185,
-0.027626486495137215,
0.004169976804405451,
0.026694821193814278,
2.5272797934893013e-33,
0.03508884832262993,
0.018552551046013832,
0.004553659353405237,
-0.01085086353123188,
-0.009297164157032967,
-0.06904146820306778,
-0.0045504458248615265,
0.058253154158592224,
-0.03456082195043564,
-0.10274583101272583,
-0.03746238350868225,
-0.08228964358568192,
0.00970913004130125,
-0.04851312190294266,
0.0008594362298026681,
0.03899988904595375,
-0.000016791585949249566,
-0.026114903390407562,
-0.05567612126469612,
0.014081399887800217,
-0.032154109328985214,
0.07478722929954529,
-0.10262908041477203,
0.026239363476634026,
-0.06975633651018143,
0.05057638883590698,
0.06075180694460869,
0.018458973616361618,
0.0028494263533502817,
-0.04372285678982735,
-0.002051080111414194,
-0.07555348426103592,
-0.038728002458810806,
-0.034551188349723816,
-0.09253936260938644,
0.06654224544763565,
0.07913043349981308,
-0.0085551543161273,
-0.08388595283031464,
0.07624556869268417,
0.06876402348279953,
-0.005977720953524113,
-0.014805778861045837,
0.04140997305512428,
-0.06325392425060272,
-0.01565556600689888,
-0.1029699519276619,
-0.01841166242957115,
-0.03269501030445099,
0.020105190575122833,
-0.06853944808244705,
0.03880202770233154,
-0.03781222924590111,
0.01841071806848049,
0.005687015131115913,
-0.10697773098945618,
-0.06650496274232864,
0.046076737344264984,
0.050964392721652985,
-0.0028461641632020473,
0.01229135226458311,
0.010845381766557693,
0.0029873985331505537,
-0.07610780745744705,
0.04259681701660156,
-0.0545036680996418,
-0.043752823024988174,
0.02178850583732128,
-0.037579987198114395,
-0.03548474237322807,
0.050922930240631104,
-0.05315983667969704,
0.008389837108552456,
0.0357942096889019,
-0.00012234704627189785,
0.034000273793935776,
0.039122458547353745,
-0.06691576540470123,
0.01862230896949768,
-0.02580704540014267,
-0.014468180947005749,
-0.024746933951973915,
0.05436338111758232,
0.07011537998914719,
0.07500993460416794,
0.10349173098802567,
0.0828404426574707,
0.05712459236383438,
0.06542273610830307,
0.03919016569852829,
-0.02212471142411232,
0.03440888971090317,
0.027571195736527443,
0.04214613139629364,
0.04327137768268585,
-3.121640190784092e-8,
-0.12593868374824524,
0.03787139430642128,
-0.09728267043828964,
0.0509626641869545,
-0.022763915359973907,
0.006059759762138128,
-0.003984255250543356,
0.12283574789762497,
0.0023811818100512028,
-0.07083173096179962,
0.05857694521546364,
0.04847357049584389,
-0.02279338240623474,
-0.014477518387138844,
-0.056241199374198914,
0.022358570247888565,
0.009808117523789406,
0.01890263520181179,
0.01643236167728901,
0.009897354990243912,
0.06160130351781845,
0.019871221855282784,
0.08400756120681763,
-0.023393575102090836,
-0.053101155906915665,
0.007408632896840572,
0.03077101893723011,
0.10191874951124191,
0.11617271602153778,
0.0020393794402480125,
-0.03859966993331909,
-0.017259815707802773,
-0.018774794414639473,
-0.08423037827014923,
0.029437905177474022,
-0.023541346192359924,
-0.019708959385752678,
-0.052823323756456375,
0.028520910069346428,
-0.016978364437818527,
0.029965590685606003,
0.02680569514632225,
-0.01798989623785019,
0.0605136975646019,
-0.020250611007213593,
0.034832604229450226,
-0.006016246043145657,
-0.07046690583229065,
0.05989201366901398,
-0.04757622629404068,
-0.008663528598845005,
-0.08036313951015472,
-0.016732925549149513,
0.054591987282037735,
0.0767977312207222,
0.006633279379457235,
-0.003954790066927671,
0.027622707188129425,
0.019343402236700058,
0.018255459144711494,
-0.07186282426118851,
0.03881169855594635,
0.017581136897206306,
0.017589859664440155
] | 0 | 0 | 0 | 0 |
|
PN-GAIL: Leveraging Non-optimal Information from Imperfect Demonstrations | https://openreview.net/forum?id=0e2pcSxQJS | [
"Qiang Liu",
"Huiqiao Fu",
"Kaiqiang Tang",
"Chunlin Chen",
"Daoyi Dong"
] | Poster | Imitation learning aims at constructing an optimal policy by emulating expert demonstrations. However, the prevailing approaches in this domain typically presume that the demonstrations are optimal, an assumption that seldom holds true in the complexities of real-world applications. The data collected in practical scenarios often contains imperfections, encompassing both optimal and non-optimal examples. In this study, we propose Positive-Negative Generative Adversarial Imitation Learning (PN-GAIL), a novel approach that falls within the framework of Generative Adversarial Imitation Learning (GAIL). PN-GAIL innovatively leverages non-optimal information from imperfect demonstrations, allowing the discriminator to comprehensively assess the positive and negative risks associated with these demonstrations. Furthermore, it requires only a small subset of labeled confidence scores. Theoretical analysis indicates that PN-GAIL deviates from the non-optimal data while mimicking imperfect demonstrations. Experimental results demonstrate that PN-GAIL surpasses conventional baseline methods in dealing with imperfect demonstrations, thereby significantly augmenting the practical utility of imitation learning in real-world contexts. Our codes are available at https://github.com/QiangLiuT/PN-GAIL. | Generative adversarial imitation learning, imperfect demonstrations, reinforcement learning | We introduce a Positive-Negative Generative Adversarial Imitation Learning (PN-GAIL) method within the framework of Generative Adversarial Imitation Learning (GAIL) to leverage non-optimal information from imperfect demonstrations. | 13,343 | null | [
-0.11756054311990738,
-0.043365299701690674,
0.0967775359749794,
-0.00004723335223388858,
-0.004934694152325392,
-0.0027109882794320583,
0.018514450639486313,
-0.01667642965912819,
-0.001160921179689467,
0.008647306822240353,
0.009756822139024734,
-0.03550276160240173,
0.07668034732341766,
-0.028142563998699188,
-0.07737120240926743,
-0.02657928504049778,
0.1474154144525528,
0.08401574194431305,
-0.031350258737802505,
-0.013782762922346592,
-0.014470779336988926,
-0.02776946872472763,
0.04935651645064354,
-0.02269282005727291,
-0.0020104353316128254,
-0.034055449068546295,
-0.006166508886963129,
0.01075613871216774,
0.023739567026495934,
-0.03381027653813362,
0.004075253382325172,
0.024713478982448578,
0.02954445593059063,
-0.029457364231348038,
-0.08848327398300171,
0.08235689997673035,
-0.15606573224067688,
0.04981476441025734,
0.02038470096886158,
-0.05518666282296181,
-0.01916460134088993,
-0.04657188802957535,
-0.029971612617373466,
0.002776880282908678,
0.0037385858595371246,
-0.0009220708743669093,
0.04436607286334038,
-0.010660912841558456,
0.09170069545507431,
0.0018770976457744837,
-0.03182746097445488,
-0.014879174530506134,
0.04054870456457138,
-0.04238327592611313,
0.056379321962594986,
0.05551653727889061,
0.06075084209442139,
0.006176121067255735,
0.008764388971030712,
0.037478577345609665,
0.051056135445833206,
-0.052890196442604065,
-0.02969253808259964,
-0.04370347782969475,
0.017360897734761238,
-0.02724979817867279,
-0.054005566984415054,
0.08119869977235794,
0.004502178635448217,
0.08836501091718674,
0.04222222790122032,
0.021602770313620567,
-0.0340568944811821,
-0.006985894404351711,
-0.003753995755687356,
0.03801816701889038,
-0.05416354537010193,
-0.06448403745889664,
-0.05107498541474342,
-0.03873000293970108,
0.040291331708431244,
-0.06423971056938171,
0.00022216365323401988,
0.08471132814884186,
0.07160969078540802,
0.011339106597006321,
-0.026686301454901695,
0.0902520939707756,
0.14174915850162506,
0.02296018786728382,
-0.0661071315407753,
-0.013693447224795818,
-0.05417413264513016,
0.02541469596326351,
0.0012179288314655423,
-0.023472735658288002,
-0.017009470611810684,
-0.05659009888768196,
-0.018715757876634598,
0.10647657513618469,
-0.007830017246305943,
0.07860158383846283,
0.02173553965985775,
-0.018795466050505638,
0.010527580045163631,
-0.06224186345934868,
0.010776251554489136,
-0.03320692479610443,
0.06862183660268784,
-0.015054543502628803,
0.040841761976480484,
0.02402591146528721,
0.04937334358692169,
-0.01875394582748413,
-0.00506596639752388,
-0.017994923517107964,
-0.05059707909822464,
0.05673734471201897,
-0.1107621118426323,
0.05103519186377525,
-0.042621493339538574,
0.0011700986651703715,
0.051543571054935455,
0.02027827315032482,
0.011887205764651299,
-0.04065040126442909,
-0.00661882059648633,
-1.5405496315999184e-34,
0.0924060046672821,
0.008497670292854309,
0.07041984796524048,
0.028806332498788834,
0.0525650829076767,
0.02860512025654316,
-0.03715154901146889,
-0.05369889736175537,
0.014961634762585163,
-0.014279847964644432,
0.06936032325029373,
-0.04135054349899292,
-0.06599881500005722,
0.09057655185461044,
0.03772139549255371,
0.03716852143406868,
-0.08101353049278259,
-0.04850590601563454,
-0.020069539546966553,
0.00191532913595438,
0.0624411404132843,
-0.034009065479040146,
-0.061243996024131775,
-0.04665239155292511,
0.0076751732267439365,
0.0016625749412924051,
-0.006500911433249712,
-0.00357528286986053,
0.013918650336563587,
-0.0072946492582559586,
-0.07913167774677277,
-0.021932443603873253,
0.0023325413931161165,
0.002136925933882594,
0.04175807163119316,
0.031813785433769226,
-0.014673191122710705,
-0.08919736742973328,
-0.06555506587028503,
0.023305276408791542,
-0.04742458835244179,
-0.02233138680458069,
0.05096696317195892,
0.016884194687008858,
-0.049718454480171204,
-0.0897907167673111,
0.07200707495212555,
0.03975560888648033,
-0.07478082180023193,
-0.0071142129600048065,
0.021419227123260498,
0.05233434587717056,
-0.0024537830613553524,
-0.09327145665884018,
-0.05469556152820587,
-0.02700415626168251,
0.0012665016110986471,
0.0634765550494194,
-0.018366843461990356,
0.020717447623610497,
-0.011529789306223392,
-0.01108604483306408,
-0.01217825710773468,
0.01941709779202938,
-0.061860255897045135,
0.13296814262866974,
-0.05345849692821503,
0.0038558675441890955,
0.07192578911781311,
0.0512353740632534,
-0.03368296101689339,
0.02615026757121086,
-0.058035869151353836,
-0.05829811468720436,
0.07886960357427597,
0.024509001523256302,
0.009066110476851463,
-0.059645287692546844,
0.06930053979158401,
-0.014897736720740795,
-0.008434263058006763,
-0.034745730459690094,
-0.06988387554883957,
-0.011211675591766834,
-0.029905647039413452,
-0.016590721905231476,
0.05263422429561615,
-0.04189994931221008,
-0.0129138994961977,
0.007848176173865795,
-0.08964735269546509,
-0.009948348626494408,
-0.07510432600975037,
0.0583028681576252,
0.04893040284514427,
-2.350225548671502e-33,
-0.005525706335902214,
0.0018819097895175219,
0.023373408243060112,
0.022393912076950073,
-0.10172360390424728,
-0.00004131365130888298,
0.009568517096340656,
-0.010815504938364029,
-0.03460196405649185,
-0.044864341616630554,
-0.07161138206720352,
0.037483759224414825,
0.060535792261362076,
-0.0485268160700798,
-0.04885132983326912,
-0.034798700362443924,
-0.00888041127473116,
0.09946402907371521,
0.02781612239778042,
-0.03539710491895676,
0.0463004931807518,
0.1067407950758934,
-0.06370134651660919,
0.09354954212903976,
-0.03506862744688988,
0.02158137410879135,
-0.004922890104353428,
0.10433635860681534,
-0.009149812161922455,
-0.008399678394198418,
0.02649315632879734,
0.004874708596616983,
-0.023747937753796577,
0.046699658036231995,
-0.05397532135248184,
0.11730244755744934,
0.06345340609550476,
0.11245419830083847,
-0.09943272173404694,
0.042658526450395584,
0.015415824949741364,
-0.02107124961912632,
0.00036868936149403453,
0.0016670550685375929,
-0.07455772906541824,
-0.06214407831430435,
-0.012558300979435444,
-0.042309802025556564,
0.03816017508506775,
-0.02023714780807495,
-0.03169184550642967,
0.06408379971981049,
-0.049602825194597244,
-0.004218304064124823,
-0.019657889381051064,
-0.06902427971363068,
-0.014635752886533737,
-0.01001087948679924,
0.14365842938423157,
0.019667187705636024,
-0.0855439081788063,
-0.03183221071958542,
-0.020101100206375122,
0.04065341129899025,
-0.10506360977888107,
0.023203590884804726,
-0.009433617815375328,
0.07861308753490448,
0.047692324966192245,
-0.05116364732384682,
0.012716674245893955,
0.057622551918029785,
0.005308788735419512,
-0.06679557263851166,
0.02434953674674034,
-0.001498117228038609,
-0.0067178355529904366,
0.08194758743047714,
-0.046263188123703,
-0.021022379398345947,
0.07083012908697128,
-0.067570261657238,
0.08183437585830688,
0.008511769585311413,
0.030266674235463142,
0.04998977482318878,
-0.05171963945031166,
0.04306909441947937,
0.014632380567491055,
0.048077791929244995,
-0.005287820473313332,
0.07086662203073502,
0.018137620761990547,
-0.024545818567276,
0.030488066375255585,
-4.167630507367903e-8,
-0.09849217534065247,
0.058166127651929855,
0.15120001137256622,
-0.013113315217196941,
0.05482596904039383,
-0.013866263441741467,
-0.0734754279255867,
-0.0015656620962545276,
-0.06512366235256195,
-0.06863619387149811,
0.05053170770406723,
0.024622078984975815,
-0.020812567323446274,
-0.07289440929889679,
-0.04134450480341911,
0.06403302401304245,
0.018613595515489578,
-0.010360412299633026,
0.000968608190305531,
0.04936477169394493,
-0.008877124637365341,
-0.021839939057826996,
-0.032194942235946655,
-0.10420721024274826,
-0.022302277386188507,
-0.03901052474975586,
-0.023405268788337708,
0.02505306713283062,
-0.01578313484787941,
0.008026600815355778,
0.0338452123105526,
-0.02340606227517128,
0.032995980232954025,
-0.0045859115198254585,
-0.01786559261381626,
0.04474157094955444,
0.025320762768387794,
-0.09110301733016968,
-0.01106963213533163,
-0.07970060408115387,
-0.034984417259693146,
0.011627685278654099,
0.0038999374955892563,
-0.027549030259251595,
0.001072658458724618,
0.0007097666384652257,
-0.007701520342379808,
-0.1422092616558075,
0.06661943346261978,
-0.018012287095189095,
0.08471966534852982,
-0.056028444319963455,
-0.006660096347332001,
0.09938949346542358,
0.08378354460000992,
0.05267738550901413,
-0.030338456854224205,
-0.004871853161603212,
-0.009781592525541782,
0.11124944686889648,
0.05432162433862686,
0.0377754382789135,
-0.06849119812250137,
-0.06764528155326843
] | 0 | 0 | 0 | 0 |
|
Attributing Culture-Conditioned Generations to Pretraining Corpora | https://openreview.net/forum?id=XrsOu4KgDE | [
"Huihan Li",
"Arnav Goel",
"Keyu He",
"Xiang Ren"
] | Poster | In open-ended generative tasks like narrative writing or dialogue, large language models often exhibit cultural biases, showing limited knowledge and generating templated outputs for less prevalent cultures. Recent works show that these biases may stem from uneven cultural representation in pretraining corpora. This work investigates how pretraining leads to biased culture-conditioned generations
by analyzing how models associate entities with cultures based on pretraining data patterns. We propose the MEMOED framework (MEMOrization from prEtraining Document) to determine whether a generation for a culture arises from memorization. Using MEMOED on culture-conditioned generations about food and clothing for 110 cultures, we find that high-frequency cultures in pretraining data yield more generations with memorized symbols, while some low-frequency cultures produce none. Additionally, the model favors generating entities with extraordinarily high frequency regardless of the conditioned culture, reflecting biases toward frequent pretraining terms irrespective of relevance. We hope that the MEMOED framework and our insights will inspire more works on attributing model performance on pretraining data. | culture bias, pretraining data, memorization, generalization | We propose the MEMOED framework to determine whether a biased culture-conditioned generation arises from memorization of pretraining data patterns, or results from other associations made due to pretraining data. | 13,342 | 2412.20760 | [
0.004278500098735094,
-0.04880707710981369,
-0.004820032976567745,
0.09031862765550613,
0.03433011472225189,
0.04570205509662628,
-0.028597068041563034,
-0.07911461591720581,
0.08028100430965424,
-0.04440873861312866,
0.0145699018612504,
-0.059904854744672775,
0.07077772170305252,
-0.001732866046950221,
0.009451168589293957,
-0.022390924394130707,
0.06600286811590195,
0.008716341108083725,
-0.06953190267086029,
-0.10241181403398514,
0.031317468732595444,
0.04031612351536751,
0.09869162738323212,
0.018338970839977264,
0.020290382206439972,
-0.07081475853919983,
0.012839708477258682,
-0.024468695744872093,
0.06738672405481339,
0.014438251964747906,
0.039179813116788864,
0.12374475598335266,
0.0031834575347602367,
0.10345549881458282,
-0.03493587300181389,
0.08471549302339554,
-0.05645101144909859,
0.05332285910844803,
0.017908059060573578,
0.0013948953710496426,
-0.015333213843405247,
-0.00723672891035676,
-0.00253119133412838,
0.028278658166527748,
0.059301115572452545,
-0.0004839976318180561,
-0.05918814614415169,
-0.0019748657941818237,
-0.07405908405780792,
0.051848217844963074,
-0.07494080066680908,
0.041137952357530594,
0.052773140370845795,
-0.02777479588985443,
0.022453950718045235,
0.02814481593668461,
0.021848656237125397,
0.0732639729976654,
-0.020979255437850952,
-0.034101273864507675,
-0.12898816168308258,
-0.1377357542514801,
-0.09324689209461212,
-0.016513178125023842,
-0.05488545820116997,
-0.046274833381175995,
0.007260642014443874,
0.10033582895994186,
-0.046838924288749695,
-0.006207518745213747,
-0.021507296711206436,
0.09543240815401077,
0.02484552003443241,
0.10288627445697784,
0.002355356700718403,
0.08426380157470703,
-0.003461394226178527,
-0.0865178108215332,
0.002160667674615979,
-0.10975483059883118,
-0.0015793710481375456,
0.04455414414405823,
0.06759708374738693,
-0.04343476891517639,
0.056599970906972885,
0.022545402869582176,
0.07701633870601654,
0.05642222985625267,
0.0007136865169741213,
0.04667193070054054,
-0.04936211556196213,
-0.07023724168539047,
0.1274530589580536,
0.026154762133955956,
-0.020349187776446342,
0.11541254073381424,
0.03268410637974739,
-0.0390494279563427,
0.018309317529201508,
0.03644167259335518,
0.04446805268526077,
0.05341667681932449,
-0.006790681276470423,
0.058121081441640854,
0.006673275958746672,
-0.1024600937962532,
0.020217198878526688,
0.017718037590384483,
-0.06436403840780258,
-0.008929635398089886,
0.022501539438962936,
0.07177667319774628,
-0.03766554221510887,
0.02294718660414219,
0.022167302668094635,
-0.0721229836344719,
0.03719402849674225,
-0.099825419485569,
-0.032173916697502136,
0.02771979384124279,
0.032243456691503525,
0.03041337989270687,
-0.011486695148050785,
0.026396747678518295,
-0.07335534691810608,
-0.05220974236726761,
-0.016672927886247635,
2.821296277226895e-33,
0.06870763748884201,
0.020950501784682274,
0.012028021737933159,
0.06446215510368347,
0.024717634543776512,
-0.021435732021927834,
-0.021483158692717552,
0.044667694717645645,
-0.019275566563010216,
-0.02046726830303669,
-0.026566555723547935,
0.04766310006380081,
-0.09332132339477539,
0.09066927433013916,
0.021061202511191368,
-0.005390274338424206,
-0.13331246376037598,
0.029161619022488594,
0.04546080157160759,
0.044405460357666016,
0.004737762734293938,
0.03039330057799816,
0.05974223464727402,
-0.058305975049734116,
0.015363757498562336,
0.014089209958910942,
0.05614061281085014,
-0.07411971688270569,
-0.09425843507051468,
0.015565529465675354,
-0.03873453289270401,
-0.08093638718128204,
0.04585304483771324,
-0.011335591785609722,
-0.022827014327049255,
-0.05687182396650314,
0.027265308424830437,
-0.03764460235834122,
0.014643924310803413,
0.014294032007455826,
-0.02913159318268299,
0.010474976152181625,
0.0420415885746479,
-0.007653696928173304,
-0.05855744332075119,
-0.06429045647382736,
0.04659869149327278,
-0.05246903374791145,
-0.05363189056515694,
0.05454563349485397,
0.022386573255062103,
0.03614120930433273,
0.041227780282497406,
-0.041033193469047546,
-0.018761109560728073,
0.03559572622179985,
0.030016563832759857,
-0.030748024582862854,
0.013848589733242989,
-0.030794411897659302,
-0.015207226388156414,
0.06539954990148544,
0.01322658360004425,
0.08560670912265778,
0.01970098726451397,
-0.0006742887781001627,
0.003764492692425847,
0.015539687126874924,
0.054981932044029236,
0.002289680764079094,
0.02205522172152996,
-0.025211231783032417,
-0.06802807748317719,
-0.06190136447548866,
0.03743689879775047,
0.04945831745862961,
0.028340913355350494,
-0.0363469235599041,
0.008675597608089447,
-0.04052937775850296,
0.02358337491750717,
-0.028071129694581032,
-0.060308147221803665,
-0.0673530101776123,
-0.058054015040397644,
-0.012018387205898762,
0.055896393954753876,
-0.049533914774656296,
0.029379459097981453,
0.018009807914495468,
0.039705000817775726,
-0.051146335899829865,
0.020613381639122963,
-0.012396596372127533,
0.02415463700890541,
-3.015176988992746e-33,
0.013577267527580261,
0.0027866605669260025,
0.0031103254295885563,
0.02495085634291172,
-0.02339455857872963,
-0.0772637277841568,
-0.006909781601279974,
0.06702978163957596,
-0.10184401273727417,
-0.07465919852256775,
-0.0066930875182151794,
-0.06190143898129463,
0.0168714988976717,
0.05351138859987259,
0.02205161564052105,
-0.052580006420612335,
0.05743655934929848,
0.07572098821401596,
0.01140157412737608,
0.03731778636574745,
-0.014032132923603058,
0.09938393533229828,
-0.12014895677566528,
0.03368261456489563,
-0.03725331276655197,
0.014545001089572906,
-0.04890693724155426,
0.029188204556703568,
-0.06657422333955765,
-0.016465818509459496,
-0.043953511863946915,
-0.019677715376019478,
0.0489141009747982,
-0.011533007957041264,
-0.07831181585788727,
0.032042939215898514,
0.039421677589416504,
-0.004210584331303835,
-0.05171710625290871,
0.03189706429839134,
0.006913009565323591,
-0.027434349060058594,
-0.12175966054201126,
0.028846560046076775,
-0.09940885752439499,
-0.0610915943980217,
-0.1048971563577652,
0.002247601980343461,
0.06764136254787445,
-0.02459254302084446,
-0.045094046741724014,
-0.015951503068208694,
-0.010720831342041492,
-0.04978537559509277,
-0.04774189367890358,
-0.12088827788829803,
-0.011018285527825356,
-0.0960158109664917,
0.011205832473933697,
0.0016346813645213842,
-0.07258647680282593,
0.021523194387555122,
0.030567282810807228,
-0.12156243622303009,
0.01448079477995634,
-0.10183398425579071,
-0.010185634717345238,
-0.0007538069039583206,
0.0060539040714502335,
0.010214021429419518,
0.04924803227186203,
-0.028762279078364372,
-0.04370301961898804,
0.01572572998702526,
-0.02129460871219635,
-0.002873701509088278,
-0.028783472254872322,
-0.010870957747101784,
0.01992892473936081,
-0.051109831780195236,
-0.02836773544549942,
0.0013801298337057233,
-0.024183427914977074,
0.0609387643635273,
0.04173087328672409,
0.07662232965230942,
-0.01550920121371746,
0.09214231371879578,
-0.0030561338644474745,
0.07008787989616394,
-0.02840050682425499,
-0.013569706119596958,
-0.02157680317759514,
0.07793530076742172,
-0.048433054238557816,
-3.742792387129157e-8,
-0.11269563436508179,
-0.02380041964352131,
0.04426325485110283,
0.12197142094373703,
0.04392486810684204,
-0.03394697979092598,
-0.03310638293623924,
0.08299951255321503,
-0.02242450974881649,
-0.043068718165159225,
-0.026549063622951508,
0.07489130645990372,
-0.03624942898750305,
-0.004903221502900124,
0.010078182443976402,
0.07254646718502045,
0.011543286964297295,
0.0567530132830143,
-0.012946506962180138,
0.024679895490407944,
0.08780995011329651,
0.055153220891952515,
-0.00430761044844985,
-0.02514876425266266,
0.004920568782836199,
-0.0595233328640461,
-0.05688539147377014,
0.09391859173774719,
0.05403483659029007,
-0.025006910786032677,
0.0005879693781025708,
0.00782703049480915,
-0.10270480066537857,
-0.016642997041344643,
0.05848005786538124,
0.031974684447050095,
0.011868358589708805,
-0.031063787639141083,
0.039489105343818665,
-0.03819769248366356,
0.07760835438966751,
0.00959844421595335,
-0.020522277802228928,
0.052600085735321045,
0.029738925397396088,
-0.01637791097164154,
-0.07816328853368759,
-0.04842512309551239,
0.018839016556739807,
-0.04283691197633743,
-0.030586158856749535,
-0.04662159085273743,
-0.00525337178260088,
0.0364927239716053,
0.02779921144247055,
-0.039725784212350845,
0.023755677044391632,
0.02953428588807583,
0.08094576746225357,
0.06603983044624329,
0.0189403984695673,
0.07075818628072739,
-0.013604596257209778,
-0.0312831774353981
] | https://github.com/huihanlhh/CultureGenAttr | 3 | 0 | 0 | 0 |
Semi-Parametric Retrieval via Binary Bag-of-Tokens Index | https://openreview.net/forum?id=l0fn10vSyM | [
"Jiawei Zhou",
"Li Dong",
"Furu Wei",
"Lei Chen"
] | Poster | Information retrieval has transitioned from standalone systems into essential components across broader applications, with indexing efficiency, cost-effectiveness, and freshness becoming increasingly critical yet often overlooked. In this paper, we introduce SemI-parametric Disentangled Retrieval (SiDR), a bi-encoder retrieval framework that decouples retrieval index from neural parameters to enable efficient, low-cost, and parameter-agnostic indexing for emerging use cases. Specifically, in addition to using embeddings as indexes like existing neural retrieval methods, SiDR supports a non-parametric tokenization index for search, achieving BM25-like indexing complexity with significantly better effectiveness. Our comprehensive evaluation across 16 retrieval benchmarks demonstrates that SiDR outperforms both neural and term-based retrieval baselines under the same indexing workload: (i) When using an embedding-based index, SiDR exceeds the performance of conventional neural retrievers while maintaining similar training complexity; (ii) When using a tokenization-based index, SiDR drastically reduces indexing cost and time, matching the complexity of traditional term-based retrieval, while consistently outperforming BM25 on all in-domain datasets; (iii) Additionally, we introduce a late parametric mechanism that matches BM25 index preparation time while outperforming other neural retrieval baselines in effectiveness. | information retrieval, efficient retrieval, retrieval-agumented applications, RAG | A semi-parametric neural retrieval system that supports both parametric and non-parametric index | 13,338 | 2405.01924 | [
-0.04684583842754364,
-0.05296339467167854,
0.03444746881723404,
0.06389832496643066,
0.023335734382271767,
0.05394788458943367,
-0.016828173771500587,
-0.007006341125816107,
0.007398426998406649,
-0.10779386013746262,
-0.0606541708111763,
-0.027403105050325394,
0.022034188732504845,
0.033958107233047485,
-0.05349228158593178,
0.05685205012559891,
0.08857032656669617,
0.05511615052819252,
-0.061787158250808716,
0.019061224535107613,
0.04624196141958237,
0.04069433733820915,
0.031263578683137894,
-0.04497043043375015,
0.033422037959098816,
-0.04575848951935768,
-0.014832430519163609,
-0.05813221260905266,
0.05022314190864563,
0.04182610660791397,
-0.0021726475097239017,
0.05905294418334961,
-0.037991687655448914,
0.08418338745832443,
-0.044629812240600586,
0.015837011858820915,
-0.07111341506242752,
-0.0016614776104688644,
0.04868021234869957,
-0.05931750312447548,
-0.013817956671118736,
0.045395657420158386,
-0.0286091435700655,
0.06288684904575348,
0.049618612974882126,
0.0440937802195549,
-0.010690062306821346,
-0.013340854085981846,
0.028715210035443306,
-0.0042712269350886345,
-0.08848107606172562,
0.06509804725646973,
-0.0373009592294693,
0.06141362711787224,
-0.009676739573478699,
-0.04231598600745201,
-0.04216720536351204,
0.026301220059394836,
-0.05082730948925018,
-0.06699816137552261,
-0.05137291178107262,
-0.08900889754295349,
0.009818513877689838,
-0.034907642751932144,
0.003292074194177985,
-0.03597870469093323,
0.04219786822795868,
-0.021328140050172806,
0.04335302114486694,
-0.009532074443995953,
0.0026396564207971096,
0.06520505994558334,
-0.10469842702150345,
0.03024568036198616,
0.009288426488637924,
0.0653679147362709,
0.07733986526727676,
-0.0019978804048150778,
-0.0019317923579365015,
-0.024976857006549835,
0.00322750024497509,
0.010391990654170513,
0.10176894068717957,
0.037493996322155,
0.09304170310497284,
-0.019510731101036072,
0.027737347409129143,
-0.07001587748527527,
-0.005046521779149771,
0.0046144104562699795,
-0.023638075217604637,
-0.052622150629758835,
0.05598318576812744,
-0.07662788778543472,
0.09469803422689438,
-0.049508459866046906,
0.02696043811738491,
0.06807699054479599,
-0.02887216955423355,
0.03331174701452255,
-0.005935998167842627,
0.058243632316589355,
-0.05629745125770569,
-0.08871481567621231,
0.012330337427556515,
0.021707087755203247,
0.07855141907930374,
-0.023149564862251282,
0.06760124117136002,
-0.02296840213239193,
-0.007979260757565498,
0.08586001396179199,
0.05740974843502045,
-0.03528601676225662,
-0.010727332904934883,
0.02139945700764656,
-0.03136686980724335,
-0.0122393649071455,
0.12785837054252625,
0.07977604866027832,
0.022396614775061607,
0.016593478620052338,
-0.04608788713812828,
-0.04847385361790657,
-0.0756046399474144,
-0.024716675281524658,
-0.013152441009879112,
4.656325243465043e-33,
-0.037068385630846024,
0.04264213144779205,
0.015008536167442799,
0.024639032781124115,
-0.1072026714682579,
0.028819814324378967,
-0.003677814267575741,
0.10561462491750717,
-0.0840279832482338,
-0.009256220422685146,
-0.06274516135454178,
0.0901007428765297,
-0.014714297838509083,
0.04473000764846802,
0.05553191155195236,
-0.051196109503507614,
-0.05316615849733353,
0.033513352274894714,
0.0056207492016255856,
-0.04855203628540039,
0.008032440207898617,
-0.02142351306974888,
-0.04011787101626396,
-0.02000775747001171,
-0.028747476637363434,
-0.07842270284891129,
0.09107429534196854,
-0.09399467706680298,
-0.00376714370213449,
0.0011488741729408503,
-0.09868666529655457,
-0.019575700163841248,
0.037147022783756256,
-0.031869444996118546,
0.031486861407756805,
-0.022805413231253624,
0.04772326350212097,
0.044955115765333176,
-0.028892239555716515,
0.00033265232923440635,
-0.014647760428488255,
0.018007634207606316,
-0.003431039396673441,
-0.06551618874073029,
-0.07577459514141083,
-0.14895564317703247,
0.015540354885160923,
-0.07592111825942993,
-0.004677845165133476,
-0.035556335002183914,
0.014538570307195187,
0.015054375864565372,
-0.04612927511334419,
-0.04114895686507225,
0.08626290410757065,
-0.05808810517191887,
0.05490754917263985,
0.03957604989409447,
-0.034020718187093735,
0.08221270889043808,
-0.01835815981030464,
-0.00032958592055365443,
0.03324832394719124,
0.05930838733911514,
0.01929161511361599,
0.026167897507548332,
0.010341362096369267,
0.037240222096443176,
0.01913643628358841,
0.00681735435500741,
0.019531549885869026,
0.019075941294431686,
0.08802125602960587,
-0.10230327397584915,
0.05277906730771065,
0.008951106108725071,
0.020638184621930122,
-0.09357340633869171,
-0.014121909625828266,
-0.038019780069589615,
0.07234484702348709,
-0.0329841673374176,
0.021801328286528587,
0.01657273806631565,
-0.013696937821805477,
-0.04373970255255699,
0.053853049874305725,
-0.09863841533660889,
-0.019377073273062706,
-0.02324344404041767,
-0.005963136907666922,
0.016233820468187332,
-0.089431531727314,
-0.04799628630280495,
0.025578804314136505,
-3.5152941829569575e-33,
0.002511003753170371,
-0.06646963953971863,
0.01908911019563675,
0.02172842249274254,
-0.019479110836982727,
-0.047949012368917465,
-0.0005919846007600427,
-0.04076780378818512,
-0.03150894492864609,
-0.020952174440026283,
-0.03280488774180412,
-0.06981583684682846,
0.06098233163356781,
0.02736338973045349,
0.027812479063868523,
0.0825057104229927,
-0.1210188940167427,
0.04035266116261482,
-0.03868325427174568,
0.06582721322774887,
-0.023158732801675797,
0.04532002657651901,
-0.06733584403991699,
0.04458380118012428,
-0.10452879220247269,
0.1180974543094635,
-0.0057896776124835014,
0.054589059203863144,
0.011112936772406101,
-0.04528553783893585,
-0.06883343309164047,
-0.04007100686430931,
0.001468799659051001,
-0.0546746663749218,
-0.059571824967861176,
0.06307785958051682,
0.017397155985236168,
-0.03057403489947319,
-0.11068187654018402,
0.018169241026043892,
-0.007278332486748695,
0.037016309797763824,
-0.011396797373890877,
-0.002262091264128685,
0.03889274224638939,
-0.05976566672325134,
-0.11673786491155624,
0.022417685016989708,
0.04905363917350769,
0.032612286508083344,
0.06317494064569473,
0.020643264055252075,
-0.10349302738904953,
0.0027634960133582354,
-0.034258779138326645,
-0.08897539228200912,
-0.11506496369838715,
0.0128790820017457,
0.013671211898326874,
0.04458221048116684,
0.025780105963349342,
0.018517285585403442,
0.0185092743486166,
0.004549216944724321,
0.009748512879014015,
-0.02215787023305893,
0.03153274953365326,
-0.0028272958006709814,
-0.0581948384642601,
-0.0027746944688260555,
0.0007775552221573889,
-0.0493972972035408,
0.032898057252168655,
-0.04472299665212631,
-0.004678546916693449,
0.09956209361553192,
0.0790652185678482,
-0.03358494117856026,
-0.07740602642297745,
0.016414757817983627,
-0.03196370601654053,
-0.014143485575914383,
-0.010600454173982143,
0.055564995855093,
0.01305461023002863,
0.10624673962593079,
-0.030016692355275154,
0.07708066701889038,
-0.01666790060698986,
0.02278689667582512,
0.0028314979281276464,
-0.043122462928295135,
0.02031015418469906,
0.06404903531074524,
0.077127605676651,
-4.327486280431003e-8,
-0.022899659350514412,
0.049885526299476624,
-0.05725722387433052,
0.02272503823041916,
-0.047281425446271896,
-0.005441507790237665,
0.007872283458709717,
0.060646820813417435,
-0.0862755998969078,
-0.01944737881422043,
0.06655503064393997,
-0.061673618853092194,
-0.14066602289676666,
0.02004668116569519,
0.06628991663455963,
0.047483399510383606,
0.003850305685773492,
-0.003667869372293353,
0.023243216797709465,
0.007031625602394342,
0.027100402861833572,
0.08382955938577652,
0.03561114892363548,
-0.022723324596881866,
-0.0286052655428648,
-0.003474039025604725,
0.015423482283949852,
0.10340321809053421,
0.08757515996694565,
-0.007209110539406538,
0.021335722878575325,
0.058069128543138504,
0.026499418541789055,
-0.06056223064661026,
-0.04166692867875099,
-0.013586808927357197,
-0.043047498911619186,
-0.02005847357213497,
-0.03673079237341881,
0.043197523802518845,
0.12373651564121246,
0.04089179262518883,
-0.03553982451558113,
0.0399421751499176,
0.008093546144664288,
0.0031237665098160505,
0.06653987616300583,
0.013655140995979309,
0.06590240448713303,
-0.02803117409348488,
-0.021606553345918655,
-0.09915967285633087,
0.0012757716467604041,
0.07416554540395737,
0.047585610300302505,
0.06855389475822449,
-0.067634716629982,
0.005874007940292358,
0.07002811878919601,
0.0476687029004097,
0.04329762980341911,
-0.03348027914762497,
-0.09944496303796768,
0.046322811394929886
] | https://github.com/jzhoubu/vdr | 2 | 0 | 0 | 0 |
RTop-K: Ultra-Fast Row-Wise Top-K Selection for Neural Network Acceleration on GPUs | https://openreview.net/forum?id=PHg4rAXFVH | [
"Xi Xie",
"Yuebo Luo",
"Hongwu Peng",
"Caiwen Ding"
] | Poster | Abstract Top-k selection algorithms are fundamental in a wide range of applications, including high-performance computing, information retrieval, big data processing, and neural network model training. In this paper, we present RTop-K, a highly efficient parallel row-wise top-k selection algorithm specifically designed for GPUs. RTop-K leverages a binary search-based approach to optimize row-wise top-k selection, providing a scalable and accelerated solution.
We conduct a detailed analysis of early stopping in our algorithm, showing that it effectively maintains the testing accuracy of neural network models while substantially improving performance. Our GPU implementation of RTop-K demonstrates superior performance over state-of-the-art row-wise top-k GPU implementations, achieving an average speed-up of up to 11.49× with early stopping and 7.29× without early stopping. Moreover, RTop-K accelerates the overall training workflow of MaxK-GNNs, delivering speed-ups ranging from 11.97% to 33.29% across different models and datasets. | row-wise topk selection, GPU, CUDA | null | 13,337 | null | [
-0.0754946768283844,
-0.017819201573729515,
-0.00157381861936301,
0.023476749658584595,
0.04875972867012024,
-0.029949340969324112,
-0.05871400237083435,
-0.07049603015184402,
-0.0747252032160759,
0.00720842694863677,
-0.08167178183794022,
0.015480296686291695,
-0.06542838364839554,
0.02130407653748989,
-0.10276000946760178,
-0.011657573282718658,
0.10162939876317978,
0.07840225100517273,
-0.05905688554048538,
-0.025563454255461693,
-0.0486152246594429,
-0.007415494881570339,
0.04034047946333885,
0.008240058086812496,
0.023630859330296516,
-0.0025175753980875015,
-0.014108875766396523,
-0.03769470006227493,
0.009469485841691494,
0.004435723647475243,
-0.007185574620962143,
-0.051668621599674225,
-0.00006783418939448893,
0.0222310908138752,
-0.08137214928865433,
0.029026981443166733,
-0.07847834378480911,
-0.07428072392940521,
0.033961642533540726,
-0.09359060972929001,
0.027670899406075478,
-0.05082904174923897,
-0.00810965709388256,
0.02411595731973648,
0.08653007447719574,
0.009984062984585762,
0.0492631234228611,
-0.027861975133419037,
0.0714648962020874,
-0.03942704200744629,
-0.0578162856400013,
-0.017455630004405975,
-0.014363085851073265,
0.05799059569835663,
-0.04107861965894699,
0.07069864869117737,
0.034172363579273224,
-0.07045344263315201,
0.01220694463700056,
0.030397722497582436,
-0.028292754665017128,
-0.07267154008150101,
-0.025008460506796837,
0.007988659664988518,
-0.028573425486683846,
0.0011897836811840534,
0.04106374830007553,
0.021670734509825706,
0.03323904424905777,
0.0218525193631649,
0.07009755074977875,
0.13778012990951538,
-0.03423330560326576,
0.047869835048913956,
0.01854369044303894,
-0.00007408518285956234,
0.07304497808218002,
0.030481450259685516,
0.06758196651935577,
-0.05443735048174858,
0.03048025630414486,
-0.02248191088438034,
0.047912515699863434,
-0.02495131827890873,
0.047766394913196564,
-0.013061858713626862,
0.0024526126217097044,
0.05862860009074211,
0.0033263261429965496,
-0.03383142128586769,
0.014392291195690632,
0.0018659355118870735,
0.0694384053349495,
-0.11198506504297256,
0.053918272256851196,
0.014941752888262272,
-0.052369434386491776,
-0.061227038502693176,
-0.06903064250946045,
0.01811099238693714,
-0.02604994922876358,
-0.007255190052092075,
0.024165818467736244,
0.036220405250787735,
-0.007825181819498539,
-0.0017723344499245286,
0.045431867241859436,
0.047716062515974045,
0.04763412103056908,
-0.06110171228647232,
0.01516698207706213,
0.06714373081922531,
-0.021395104005932808,
-0.0037824157625436783,
-0.006501561030745506,
0.07874921709299088,
-0.06484397500753403,
-0.020541926845908165,
0.031053489074110985,
0.06812442094087601,
-0.09154275059700012,
-0.00464241998270154,
-0.049558330327272415,
0.004976087249815464,
0.012492864392697811,
-0.004364244174212217,
-0.0592229887843132,
1.8461575916625737e-33,
0.01250623818486929,
0.023237893357872963,
0.09602195769548416,
-0.13859784603118896,
0.04122327268123627,
-0.04344401881098747,
0.025667376816272736,
0.04769887402653694,
-0.000538516731467098,
0.08145570755004883,
-0.15146727859973907,
0.01739061437547207,
-0.04534714296460152,
0.0706748366355896,
0.028996355831623077,
-0.01025574840605259,
-0.09419982135295868,
0.052803151309490204,
-0.09446839988231659,
0.021149147301912308,
0.10582443326711655,
-0.0349825881421566,
-0.023136291652917862,
0.0069360253401100636,
-0.03284504637122154,
0.021750114858150482,
-0.01516882423311472,
-0.040939055383205414,
-0.00866602547466755,
0.028504136949777603,
-0.09727399796247482,
-0.045705266296863556,
-0.026217292994260788,
0.06830061972141266,
0.004359617363661528,
-0.04284283518791199,
-0.05289491266012192,
-0.028550663962960243,
0.07318785786628723,
0.07057265192270279,
-0.03400282934308052,
0.017679700627923012,
0.02982916869223118,
-0.032231610268354416,
-0.1155196949839592,
0.02162034437060356,
0.027101483196020126,
0.034580450505018234,
-0.00510424142703414,
0.0063179354183375835,
0.04613345116376877,
0.007819772697985172,
-0.05664452910423279,
0.0016332903178408742,
0.013461390510201454,
0.010048635303974152,
0.1028352826833725,
0.14901182055473328,
0.10761407017707825,
0.1073872372508049,
-0.020390911027789116,
-0.032745566219091415,
-0.01785234734416008,
0.13658452033996582,
-0.09309130907058716,
-0.015901129692792892,
-0.04587801545858383,
0.05029628798365593,
0.01791064813733101,
0.024782873690128326,
-0.018835457041859627,
-0.0017676533898338675,
0.047766104340553284,
-0.022826308384537697,
0.09568268805742264,
0.01390175148844719,
0.0438709519803524,
-0.09531952440738678,
-0.04291156679391861,
-0.011830057017505169,
0.004699074663221836,
0.04409579187631607,
0.015536417253315449,
-0.05781014636158943,
-0.008080814965069294,
-0.04499874264001846,
-0.008045760914683342,
-0.04981318116188049,
0.007427997887134552,
-0.023929696530103683,
-0.10161971300840378,
-0.007838393561542034,
0.04032941162586212,
-0.04346756637096405,
-0.04628266021609306,
-6.310345712091822e-34,
0.082752525806427,
-0.019900936633348465,
0.040996238589286804,
0.13062678277492523,
0.003062462667003274,
-0.013773267157375813,
0.046062957495450974,
-0.030929947271943092,
-0.04419010132551193,
-0.002583087421953678,
0.04741654172539711,
0.0040814862586557865,
0.09793549031019211,
0.05166790634393692,
0.058816052973270416,
-0.04314951226115227,
-0.069491446018219,
0.036528944969177246,
-0.02032349817454815,
-0.04626401886343956,
0.011058369651436806,
0.023333704099059105,
-0.031578510999679565,
-0.046853527426719666,
-0.06548606604337692,
-0.03561171889305115,
-0.04811198636889458,
0.07650761306285858,
0.052114780992269516,
-0.003717303043231368,
-0.050237029790878296,
0.06932371109724045,
-0.0675172507762909,
0.06179652363061905,
0.09583635628223419,
0.11014162003993988,
0.05781296640634537,
-0.018161902204155922,
-0.003397670341655612,
0.09733855724334717,
0.10437460243701935,
-0.02827247604727745,
0.011142171919345856,
0.07252289354801178,
0.00033956600236706436,
0.055601272732019424,
-0.12216376513242722,
0.007866879925131798,
-0.0038111188914626837,
-0.0333910807967186,
-0.0305264201015234,
0.06276926398277283,
-0.0029368691612035036,
0.06603113561868668,
0.038497161120176315,
-0.03878060728311539,
-0.033236775547266006,
0.042891714721918106,
-0.010965184308588505,
0.016042985022068024,
-0.005699165631085634,
-0.06209329888224602,
0.020026085898280144,
-0.01807805523276329,
0.022276781499385834,
0.0031188938301056623,
0.05634206160902977,
-0.007307566702365875,
-0.016473932191729546,
0.050476059317588806,
-0.047613393515348434,
-0.022122301161289215,
0.05367273837327957,
0.01233753003180027,
-0.10999361425638199,
0.0479070208966732,
0.01739455945789814,
-0.059818241745233536,
0.03208019584417343,
-0.027266040444374084,
-0.00813623983412981,
0.030371367931365967,
-0.032223936170339584,
0.026547739282250404,
0.05422186851501465,
0.05270952358841896,
-0.02099091373383999,
0.022170811891555786,
0.04585590586066246,
-0.05385222285985947,
-0.04084612429141998,
0.05138809233903885,
-0.0015868047485128045,
0.02017248421907425,
-0.04884166643023491,
-3.918410484970991e-8,
-0.021243499591946602,
-0.011871451511979103,
0.015237079933285713,
0.035811103880405426,
0.07284189760684967,
-0.01910279132425785,
0.03385971114039421,
0.06235065683722496,
-0.03220117837190628,
-0.06939053535461426,
0.07144629955291748,
-0.021040653809905052,
0.02993885427713394,
-0.05321483314037323,
0.04236912354826927,
-0.01460591796785593,
-0.04299883916974068,
-0.09170503169298172,
0.05328487977385521,
0.004232374485582113,
-0.009018277749419212,
0.047792840749025345,
0.06528239697217941,
-0.051793720573186874,
-0.025300655514001846,
-0.06730487942695618,
-0.020465191453695297,
0.04872903227806091,
0.10141190141439438,
-0.04066525399684906,
0.007587682455778122,
-0.03594203665852547,
0.016776639968156815,
-0.02738070860505104,
0.08273720741271973,
0.0570109523832798,
-0.03621071204543114,
0.12182167172431946,
0.032090649008750916,
-0.06581348925828934,
0.024444488808512688,
-0.04086817428469658,
0.030023248866200447,
-0.028955776244401932,
-0.08658311516046524,
0.00017483255942352116,
-0.04721260443329811,
0.008878991939127445,
0.026802701875567436,
0.01965613104403019,
-0.007051139138638973,
-0.009233429096639156,
-0.0210312157869339,
0.041829224675893784,
0.0209488607943058,
0.0023448215797543526,
-0.08945567905902863,
-0.07528050243854523,
0.0312504805624485,
-0.0024454400409013033,
-0.0005449351738207042,
-0.04315144196152687,
-0.09112699329853058,
0.06338503956794739
] | 0 | 0 | 0 | 0 |
|
DelTA: An Online Document-Level Translation Agent Based on Multi-Level Memory | https://openreview.net/forum?id=hoYFLRNbhc | [
"Yutong Wang",
"Jiali Zeng",
"Xuebo Liu",
"Derek F. Wong",
"Fandong Meng",
"Jie Zhou",
"Min Zhang"
] | Poster | Large language models (LLMs) have achieved reasonable quality improvements in machine translation (MT).
However, most current research on MT-LLMs still faces significant challenges in maintaining translation consistency and accuracy when processing entire documents.
In this paper, we introduce DelTA, a Document-levEL Translation Agent designed to overcome these limitations.
DelTA features a multi-level memory structure that stores information across various granularities and spans, including Proper Noun Records, Bilingual Summary, Long-Term Memory, and Short-Term Memory, which are continuously retrieved and updated by auxiliary LLM-based components.
Experimental results indicate that DelTA significantly outperforms strong baselines in terms of translation consistency and quality across four open/closed-source LLMs and two representative document translation datasets, achieving an increase in consistency scores by up to 4.58 percentage points and in COMET scores by up to 3.16 points on average.
DelTA employs a sentence-by-sentence translation strategy, ensuring no sentence omissions and offering a memory-efficient solution compared to the mainstream method.
Furthermore, DelTA improves pronoun and context-dependent translation accuracy, and the summary component of the agent also shows promise as a tool for query-based summarization tasks.
The code and data of our approach are released at https://github.com/YutongWang1216/DocMTAgent. | Document-Level Translation, Large Language Models, Autonomous Agents, Natural Language Processing | We develop a document-level translation agent based on multi-level memory to address the challenges associated with the DocMT task. | 13,334 | 2410.08143 | [
0.006829835940152407,
-0.0178250502794981,
-0.020316481590270996,
0.0326949879527092,
0.031011192128062248,
0.01719803921878338,
-0.0561574287712574,
0.03287487104535103,
0.06948751956224442,
-0.055126696825027466,
0.01798507384955883,
-0.006297656800597906,
0.039818476885557175,
-0.053643785417079926,
0.07672903686761856,
0.012448536232113838,
0.06483970582485199,
0.0944967195391655,
-0.05415305495262146,
-0.09149143844842911,
0.04944692179560661,
0.05166667699813843,
0.09408358484506607,
0.01895059458911419,
0.03861934691667557,
0.028704751282930374,
-0.07225185632705688,
-0.039880745112895966,
0.03718450665473938,
-0.06169362738728523,
0.010858121328055859,
0.08351187407970428,
-0.08685395121574402,
0.06218498572707176,
0.030636828392744064,
0.065536729991436,
-0.14764457941055298,
0.008368691429495811,
-0.024942321702837944,
-0.0042399135418236256,
0.014798780903220177,
-0.007608444429934025,
0.007889028638601303,
0.0008407103596255183,
0.06557702273130417,
0.012027093209326267,
-0.028858045116066933,
-0.007591498550027609,
0.008373284712433815,
0.04679511487483978,
-0.10953192412853241,
0.01634286902844906,
0.010141859762370586,
0.05771566554903984,
-0.020566871389746666,
0.04899096116423607,
0.00037399952998384833,
0.041256554424762726,
-0.02322804369032383,
-0.0691068023443222,
-0.09679927676916122,
-0.01688752882182598,
-0.022449450567364693,
-0.005855710711330175,
-0.02425462007522583,
-0.02321293205022812,
0.019513720646500587,
0.03856245428323746,
-0.047824677079916,
0.04604912921786308,
-0.04159536212682724,
0.03507447615265846,
0.0028331701178103685,
0.06443018466234207,
-0.047665074467659,
-0.041770514100790024,
0.010776045732200146,
-0.00938986986875534,
0.12212743610143661,
0.0487029030919075,
0.011854974552989006,
-0.03090830147266388,
0.0434047132730484,
0.024640046060085297,
0.03161164000630379,
0.017529314383864403,
0.06569784134626389,
0.04389956593513489,
-0.03714771568775177,
0.000060485890571726486,
0.04181007295846939,
-0.06092088297009468,
0.10370699316263199,
0.02256862074136734,
-0.03844509273767471,
0.031159592792391777,
0.009280978702008724,
-0.03695538267493248,
0.007468003313988447,
0.06965427100658417,
0.04704088345170021,
0.11455885320901871,
-0.00871429406106472,
-0.0362580344080925,
-0.0901796966791153,
-0.036998189985752106,
0.10746144503355026,
0.07967109978199005,
0.01801130920648575,
-0.06629759818315506,
0.05817708373069763,
-0.023165231570601463,
0.009316361509263515,
-0.008255159482359886,
0.04297909513115883,
-0.09357041865587234,
0.05033909156918526,
-0.030291840434074402,
0.008105119690299034,
-0.007341477088630199,
-0.0281343013048172,
-0.0031075449660420418,
0.10903442651033401,
-0.006799346301704645,
-0.026819400489330292,
0.022136695683002472,
0.04220327362418175,
-3.2730698884810976e-34,
0.07110698521137238,
0.04577843099832535,
-0.00478067621588707,
0.06272411346435547,
0.0033450936898589134,
-0.0018060309812426567,
-0.007029536180198193,
0.04263380914926529,
-0.1122443675994873,
-0.01717996783554554,
0.0031329926569014788,
-0.008181356824934483,
-0.06557870656251907,
0.04758041352033615,
0.007538138888776302,
-0.027680788189172745,
-0.022440437227487564,
0.0374586395919323,
0.016762837767601013,
0.01877756230533123,
0.05946880578994751,
-0.001132810371927917,
0.08289295434951782,
-0.03474980592727661,
0.06333377212285995,
0.042131051421165466,
0.041474416851997375,
-0.06294191628694534,
-0.034830767661333084,
-0.005254102870821953,
-0.1157783642411232,
-0.023722149431705475,
0.06180913746356964,
0.014135104604065418,
0.004144073463976383,
-0.07749215513467789,
-0.0778152272105217,
-0.10328608751296997,
-0.028999950736761093,
-0.09131437540054321,
-0.033945053815841675,
0.02218880131840706,
0.028950989246368408,
-0.03030444122850895,
-0.07057039439678192,
-0.07058600336313248,
0.00646998593583703,
0.022573424503207207,
0.02915511652827263,
-0.005874804221093655,
-0.030788296833634377,
0.05725886672735214,
-0.047882530838251114,
0.022050894796848297,
0.059733327478170395,
0.0744461715221405,
-0.0031174784526228905,
0.055360618978738785,
-0.02246897853910923,
0.004031557589769363,
0.0007977854693308473,
0.027704328298568726,
0.007639527320861816,
0.13761453330516815,
0.11419688910245895,
0.07393786311149597,
-0.002215262968093157,
0.13657256960868835,
0.017257902771234512,
-0.05189470574259758,
-0.023198600858449936,
-0.06214863061904907,
0.027733562514185905,
0.005147835239768028,
0.029084349051117897,
-0.01110958494246006,
0.013753946870565414,
-0.10264817625284195,
0.03482862934470177,
-0.027936238795518875,
0.01840697042644024,
-0.043660588562488556,
0.03327690437436104,
-0.027543608099222183,
-0.04247185215353966,
-0.03989662602543831,
0.025807658210396767,
0.014967839233577251,
0.01960132271051407,
0.07161277532577515,
-0.011230171658098698,
-0.005074775777757168,
-0.008404580876231194,
0.019462494179606438,
-0.020930450409650803,
1.1283130381516541e-33,
-0.011961715295910835,
-0.0771435797214508,
-0.030535070225596428,
0.07845727354288101,
-0.033929985016584396,
-0.09036733210086823,
-0.02252575382590294,
0.04131487384438515,
-0.008774167858064175,
-0.042030368000268936,
-0.07335410267114639,
-0.033677179366350174,
0.027295377105474472,
-0.012609216384589672,
0.04233424365520477,
-0.039629977196455,
0.06749890744686127,
0.027946405112743378,
-0.021756134927272797,
0.02553698979318142,
-0.0214241910725832,
0.04302513971924782,
-0.03787609189748764,
0.042348798364400864,
0.03921370953321457,
0.02280314266681671,
-0.040515072643756866,
0.031880661845207214,
-0.06376264989376068,
-0.03863643482327461,
0.053533412516117096,
0.013910865411162376,
-0.012003004550933838,
-0.03265182301402092,
-0.0864316076040268,
0.031204354017972946,
0.044889405369758606,
0.006408799439668655,
-0.018993979319930077,
0.04307766258716583,
0.056155502796173096,
-0.02769707702100277,
-0.023566825315356255,
-0.09110383689403534,
0.0018375227227807045,
0.019360315054655075,
-0.14779441058635712,
0.007346774451434612,
-0.010665702633559704,
0.03478459268808365,
-0.10633313655853271,
-0.009833664633333683,
-0.12295087426900864,
0.00556863471865654,
0.029075203463435173,
-0.07450903207063675,
-0.05291316658258438,
-0.06554680317640305,
-0.0188002809882164,
-0.13099540770053864,
-0.031374700367450714,
-0.027806909754872322,
-0.043323274701833725,
-0.026327496394515038,
0.06423371285200119,
-0.021793071180582047,
0.04127126559615135,
0.02441156469285488,
0.01870778575539589,
0.03389750048518181,
0.06714261323213577,
-0.08671300858259201,
-0.016734164208173752,
0.026139747351408005,
0.02068965882062912,
-0.06343010812997818,
-0.027355561032891273,
-0.03390092775225639,
0.05987522751092911,
-0.10598108917474747,
-0.0483090877532959,
0.03614673390984535,
0.001727053546346724,
0.0360911600291729,
0.0037951532285660505,
0.025866150856018066,
-0.020395977422595024,
0.06649619340896606,
0.028143681585788727,
0.0825519934296608,
-0.026784686371684074,
0.014379080384969711,
-0.07411429286003113,
0.013874893076717854,
-0.1099882572889328,
-3.7484685577737764e-8,
-0.059373218566179276,
0.027244942262768745,
-0.054401565343141556,
0.029892325401306152,
0.011009975336492062,
-0.0904129296541214,
-0.021173080429434776,
0.03956233337521553,
0.06439733505249023,
-0.01783682405948639,
0.01699809543788433,
-0.04370073229074478,
-0.08366164565086365,
-0.0732741579413414,
0.0007762236637063324,
0.07469664514064789,
0.030424397438764572,
0.08997268974781036,
0.005152284167706966,
-0.03900722786784172,
0.0021200478076934814,
0.05599359795451164,
0.05403224006295204,
-0.0015216725878417492,
0.025947866961359978,
-0.00817929208278656,
-0.019366394728422165,
0.11414319276809692,
0.04901885613799095,
-0.10768800973892212,
-0.027151327580213547,
-0.0025118531193584204,
0.016532357782125473,
-0.03182320296764374,
0.011777061969041824,
0.079983651638031,
0.06354089826345444,
0.04814491048455238,
0.07493704557418823,
-0.0009955280693247914,
0.0570649690926075,
0.013685150071978569,
-0.03284473717212677,
0.06933828443288803,
-0.005041324067860842,
-0.03328288346529007,
-0.07017046958208084,
-0.050254713743925095,
0.0718589648604393,
-0.12946856021881104,
0.03514770790934563,
-0.046479541808366776,
-0.012450266629457474,
0.04118477553129196,
0.0474482960999012,
0.03143175318837166,
0.04824792966246605,
-0.03555157408118248,
0.13333068788051605,
-0.010471737012267113,
0.03253090754151344,
0.0042152851819992065,
-0.0334000438451767,
-0.03314422816038132
] | https://github.com/yutongwang1216/docmtagent | 40 | 0 | 0 | 0 |
Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs | https://openreview.net/forum?id=eENHKMTOfW | [
"Aldo Pareja",
"Nikhil Shivakumar Nayak",
"Hao Wang",
"Krishnateja Killamsetty",
"Shivchander Sudalairaj",
"Wenlong Zhao",
"Seungwook Han",
"Abhishek Bhandwaldar",
"Guangxuan Xu",
"Kai Xu",
"Ligong Han",
"Luke Inglis",
"Akash Srivastava"
] | Poster | The rise of large language models (LLMs) has created a significant disparity: industrial research labs with their computational resources, expert teams, and advanced infrastructures, can effectively fine-tune LLMs, while individual developers and small organizations face barriers due to limited resources to effectively explore the experiment space. In this paper, we aim to bridge this gap by presenting a comprehensive study on supervised fine-tuning of LLMs using instruction-tuning datasets spanning diverse knowledge domains and skills. We focus on small-sized LLMs (3B to 7B parameters) for their cost-efficiency and accessibility. We explore various training configurations and strategies across four open-source pre-trained models. We provide detailed documentation of these configurations, revealing findings that challenge several common training practices, including hyperparameter recommendations from TULU and phased training recommended by Orca. The code used for the experiments can be found here: https://github.com/instructlab/training.
Key insights from our work include: (i) larger batch sizes paired with lower learning rates lead to improved model performance on benchmarks such as MMLU, MTBench, and Open LLM Leaderboard; (ii) early-stage training dynamics, such as lower gradient norms and higher loss values, are strong indicators of better final model performance, allowing for early termination of sub-optimal runs and significant computational savings; (iii) through a thorough exploration of hyperparameters like warmup steps and learning rate schedules, we provide guidance for practitioners and find that certain simplifications do not compromise performance; and (iv) we observe no significant difference in performance between phased (sequentially training on data divided into phases) and stacked (training on the entire dataset at once) strategies, but stacked training is simpler and more sample efficient. With these findings holding robustly across datasets as well as model families and sizes, we hope this study serves as a guide for practitioners fine-tuning small LLMs and promotes a more inclusive research environment for LLM development. | Machine Learning, Generative Models, Large Language Models, Natural Language Processing, Transformers, Fine-Tuning, Instruction Tuning, Synthetic Data Generation, Knowledge Data, Skills Data, Model Generalization, Batch Size, Hyperparameter Optimization, Gradient Norm, MMLU, MTBench, Stacked Training, Phased Training, Compute Efficiency, Sample Efficiency, Flash Attention, Multipack Bucketing | We provide a guide for customizing small LLMs (3B-7B parameters) through instruction tuning on diverse knowledge and skills data, offering insights on training strategies, hyperparameters, and efficient tuning methods to improve performance. | 13,333 | 2412.13337 | [
0.00921750906854868,
-0.09047199040651321,
0.04346887767314911,
0.10032013058662415,
0.04451550170779228,
-0.0012751524336636066,
-0.04647167772054672,
-0.015590306371450424,
-0.05463281273841858,
-0.06726059317588806,
-0.06266900897026062,
-0.054723672568798065,
0.019700709730386734,
-0.06653008610010147,
0.01798940636217594,
-0.00956390518695116,
0.11014522612094879,
-0.013300452381372452,
-0.09803809225559235,
-0.09654109179973602,
0.0351504310965538,
0.021219894289970398,
0.0651424303650856,
-0.013813686557114124,
0.07803572714328766,
-0.048595815896987915,
-0.01923030987381935,
-0.02573658898472786,
0.011814153753221035,
-0.05893459543585777,
-0.007997273467481136,
0.03739435225725174,
-0.033300548791885376,
0.03663560003042221,
0.029343118891119957,
0.015049898996949196,
-0.005241530016064644,
0.021649854257702827,
0.0026287275832146406,
-0.012999468483030796,
-0.029012365266680717,
-0.00988852046430111,
0.04672819375991821,
0.055632464587688446,
0.07558150589466095,
-0.022361502051353455,
0.029012205079197884,
-0.06241053342819214,
-0.08196748793125153,
-0.010621204972267151,
-0.07817856967449188,
-0.09884967654943466,
0.0015722255920991302,
0.03462440147995949,
-0.06447146087884903,
-0.007997374050319195,
0.03792215511202812,
0.016943583264946938,
-0.017273060977458954,
-0.00026748882373794913,
-0.08163174241781235,
-0.14148543775081635,
-0.11457357555627823,
-0.010099227540194988,
-0.05225897207856178,
-0.04341396316885948,
0.009080472402274609,
0.04360481724143028,
-0.0004271755169611424,
-0.016143862158060074,
-0.01560644619166851,
0.047198336571455,
-0.009008128196001053,
0.08092746138572693,
0.020846223458647728,
0.02547929435968399,
0.05429745838046074,
0.016570737585425377,
0.09643800556659698,
-0.08579620718955994,
-0.0010666527086868882,
-0.011944657191634178,
0.028324685990810394,
0.025734223425388336,
0.05102350935339928,
-0.03671213984489441,
0.06009405851364136,
0.059196703135967255,
0.06788758933544159,
0.008226451463997364,
0.02835727669298649,
-0.012006371282041073,
-0.011024737730622292,
0.010211017914116383,
0.0042416914366185665,
0.04208355024456978,
0.024947896599769592,
-0.0761936753988266,
-0.04168014973402023,
0.0765237882733345,
0.03197292238473892,
0.1263248175382614,
0.045838721096515656,
-0.01793074421584606,
-0.04503826051950455,
0.02746383100748062,
0.05832286924123764,
0.07046233862638474,
0.02572052739560604,
-0.03661628067493439,
0.048917997628450394,
0.10649598389863968,
0.015354432165622711,
-0.01704942062497139,
0.01994568482041359,
-0.004951031878590584,
-0.0209667831659317,
-0.11809226125478745,
-0.0050340029411017895,
0.09120132029056549,
-0.10625598579645157,
0.07133094221353531,
0.0010992895113304257,
-0.01679380051791668,
-0.00584066566079855,
0.044373929500579834,
-0.1197558268904686,
3.3179856194615134e-33,
0.10406392812728882,
0.06698446720838547,
-0.04609999433159828,
0.02430899068713188,
0.0332392118871212,
-0.010741589590907097,
0.10863282531499863,
0.0830208882689476,
-0.10077676177024841,
0.0349595844745636,
-0.025703541934490204,
-0.0023032580502331257,
-0.06235946714878082,
0.062219180166721344,
0.08137128502130508,
-0.024907344952225685,
-0.030066221952438354,
0.05757351592183113,
-0.07337217777967453,
-0.023671608418226242,
0.02138504572212696,
-0.012275416404008865,
0.001325584831647575,
-0.04557161405682564,
0.01754866912961006,
0.020440973341464996,
0.05704985186457634,
-0.030103864148259163,
-0.013137341476976871,
0.033304184675216675,
-0.047096557915210724,
0.005859378259629011,
0.03348306939005852,
0.01675453595817089,
0.04850700870156288,
-0.04329657182097435,
-0.04285075515508652,
-0.006703457795083523,
0.07326614111661911,
-0.033525217324495316,
-0.03350813686847687,
0.04126517102122307,
0.07786090672016144,
-0.05247070640325546,
-0.09615559130907059,
-0.05558890849351883,
-0.010107561945915222,
-0.0020331209525465965,
-0.0007485317764803767,
-0.018570253625512123,
0.021329717710614204,
0.04845622554421425,
-0.00012735843483824283,
0.024880142882466316,
0.03035043179988861,
0.038305819034576416,
0.05839596316218376,
0.03133489564061165,
-0.017457840964198112,
0.07404880225658417,
-0.014771156013011932,
0.055522236973047256,
-0.02367684803903103,
0.10475435853004456,
0.02954820916056633,
0.008195145055651665,
-0.01072806678712368,
0.07636480778455734,
0.0508122593164444,
-0.049460846930742264,
-0.06231863796710968,
-0.04744066670536995,
0.007327838335186243,
-0.04889186471700668,
0.026889288797974586,
-0.025792943313717842,
0.06291411817073822,
-0.05865931138396263,
0.013855003751814365,
-0.03001275286078453,
0.07799283415079117,
0.053471341729164124,
0.01309578400105238,
-0.03251202777028084,
-0.05229621008038521,
-0.09035848826169968,
0.015323569066822529,
-0.04746457561850548,
0.008591351099312305,
0.03691762685775757,
-0.02843274176120758,
-0.03319939970970154,
-0.057300519198179245,
-0.007814187556505203,
-0.021109169349074364,
-2.447966618245429e-33,
-0.001243659178726375,
-0.013078239746391773,
-0.02204405702650547,
0.12643218040466309,
-0.017331093549728394,
-0.04137413948774338,
0.020700912922620773,
0.04697207733988762,
-0.023568684235215187,
-0.026459001004695892,
-0.022246602922677994,
-0.06955462694168091,
0.015082436613738537,
-0.005588700529187918,
0.014210155233740807,
-0.0383974090218544,
-0.05269515886902809,
0.0225811880081892,
0.09663975238800049,
0.0368826687335968,
-0.041245102882385254,
0.0770162045955658,
-0.07822120189666748,
0.04579251632094383,
-0.035940904170274734,
-0.0029235035181045532,
-0.05766218900680542,
0.018775399774312973,
-0.029226651415228844,
0.04088423028588295,
-0.04668659716844559,
0.009133470244705677,
0.01812257245182991,
0.009993834421038628,
-0.04469449073076248,
0.04706641286611557,
0.05558720603585243,
-0.011664687655866146,
0.03574416786432266,
0.07868324965238571,
0.0534374974668026,
0.005582965444773436,
-0.06872735172510147,
-0.028932536020874977,
-0.0483677014708519,
0.029717741534113884,
-0.13681519031524658,
-0.03728879615664482,
-0.014971944503486156,
-0.0209511611610651,
-0.020995117723941803,
-0.014640935696661472,
-0.00034077628515660763,
-0.053968146443367004,
-0.029387719929218292,
-0.10679305344820023,
-0.007144805043935776,
-0.09248263388872147,
0.005683903582394123,
-0.012272108346223831,
-0.057428255677223206,
0.019234921783208847,
-0.007660768926143646,
-0.08942322432994843,
-0.004630726762115955,
-0.02980896458029747,
-0.041917067021131516,
0.04430864751338959,
-0.042558494955301285,
-0.022872840985655785,
-0.018759747967123985,
-0.04846932366490364,
-0.0037927920930087566,
-0.019685305655002594,
-0.0598326250910759,
0.0047611091285943985,
-0.011056889779865742,
-0.0545332245528698,
-0.05343537777662277,
-0.0214763842523098,
0.045817453414201736,
0.016548026353120804,
-0.008602838031947613,
0.0671684741973877,
0.05286698788404465,
0.11893092840909958,
0.08969961851835251,
0.08883491903543472,
-0.01786678656935692,
0.04642174392938614,
-0.06366706639528275,
0.02327371947467327,
0.05569971352815628,
0.06539566814899445,
-0.023457542061805725,
-4.962873489944286e-8,
-0.024171262979507446,
0.025576744228601456,
0.007662708405405283,
0.10685326904058456,
-0.01098624523729086,
-0.03055420331656933,
-0.03808285668492317,
0.1416262686252594,
-0.004916733130812645,
0.09358802437782288,
0.08091147243976593,
0.013625992462038994,
-0.06538458913564682,
0.002581662265583873,
0.035821352154016495,
0.08873333036899567,
-0.021066343411803246,
0.03546066954731941,
-0.03957980126142502,
-0.037816956639289856,
0.04466615244746208,
0.08419238030910492,
0.061170708388090134,
-0.0017640104051679373,
0.08683688193559647,
-0.07882118225097656,
-0.013519393280148506,
0.14714601635932922,
0.06850060075521469,
0.0027468751650303602,
-0.01628703437745571,
-0.03898097574710846,
-0.0347224660217762,
-0.06795857101678848,
-0.0029944097623229027,
0.06694000214338303,
-0.003184850560501218,
-0.05644994601607323,
0.0664711594581604,
0.008785624988377094,
-0.011226076632738113,
0.031188804656267166,
-0.048015039414167404,
0.020790372043848038,
-0.004906360991299152,
-0.029601996764540672,
-0.0980447381734848,
-0.09200406819581985,
-0.024535179138183594,
-0.054477885365486145,
0.00011380862997611985,
-0.0178699791431427,
-0.005759098567068577,
0.032356277108192444,
0.030726933851838112,
0.08896873891353607,
-0.061430033296346664,
-0.01535938773304224,
-0.000964573526289314,
0.05659247934818268,
0.024709651246666908,
-0.007340795826166868,
-0.08080989867448807,
0.013870634138584137
] | 0 | 0 | 0 | 0 |
|
ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks | https://openreview.net/forum?id=6bKEWevgSd | [
"Arth Shukla",
"Stone Tao",
"Hao Su"
] | Poster | High-quality benchmarks are the foundation for embodied AI research, enabling significant advancements in long-horizon navigation, manipulation and rearrangement tasks. However, as frontier tasks in robotics get more advanced, they require faster simulation speed, more intricate test environments, and larger demonstration datasets. To this end, we present MS-HAB, a holistic benchmark for low-level manipulation and in-home object rearrangement. First, we provide a GPU-accelerated implementation of the Home Assistant Benchmark (HAB). We support realistic low-level control and achieve over 3x the speed of prior magical grasp implementations at a fraction of the GPU memory usage. Second, we train extensive reinforcement learning (RL) and imitation learning (IL) baselines for future work to compare against. Finally, we develop a rule-based trajectory filtering system to sample specific demonstrations from our RL policies which match predefined criteria for robot behavior and safety. Combining demonstration filtering with our fast environments enables efficient, controlled data generation at scale. | benchmark, dataset, simulation, reinforcement learning, imitation learning, robotics | We provide a GPU-accelerated implementation of the HAB which supports realistic low-level control, run extensive RL and IL baselines, and develop a rule-based trajectory filtering system which enables efficient, controlled data generation at scale. | 13,307 | null | [
-0.0036314392928034067,
-0.07988607883453369,
0.03274638578295708,
-0.02146502025425434,
-0.011299027130007744,
0.015986917540431023,
-0.01720549911260605,
-0.06716838479042053,
-0.03534919396042824,
0.03275968134403229,
-0.010324517264962196,
0.009951025247573853,
-0.010591148398816586,
0.04026733711361885,
0.008271031081676483,
0.021889420226216316,
0.10762989521026611,
0.05100691691040993,
-0.07596548646688461,
0.044125016778707504,
0.012636607512831688,
0.03814836964011192,
0.06913991272449493,
0.050041716545820236,
0.03366689383983612,
0.03453318402171135,
0.005329885054379702,
-0.09762445092201233,
-0.006284666713327169,
-0.036219123750925064,
0.0509880892932415,
-0.05956533923745155,
-0.015773475170135498,
-0.11609085649251938,
-0.0469791479408741,
0.0628451332449913,
-0.04618676006793976,
-0.033092524856328964,
0.006003618706017733,
-0.1085938960313797,
-0.004595394246280193,
-0.006257099099457264,
0.059538427740335464,
0.0471067912876606,
0.056163765490055084,
0.08244427293539047,
0.045720864087343216,
-0.06654191017150879,
0.09135551750659943,
0.01757139340043068,
-0.06418415904045105,
0.005835642106831074,
0.023536141961812973,
-0.01927061937749386,
-0.017057493329048157,
0.1021893322467804,
0.04356754943728447,
-0.013573316857218742,
0.09337494522333145,
-0.012487374246120453,
0.01938697136938572,
-0.10081367939710617,
0.04892633110284805,
-0.03623078018426895,
-0.06341041624546051,
-0.09698870033025742,
0.012516234070062637,
-0.019755130633711815,
0.07177966088056564,
0.037504278123378754,
0.010352722369134426,
0.006347660440951586,
0.005147021729499102,
0.03897164762020111,
0.03473759442567825,
-0.019802885130047798,
-0.024408118799328804,
-0.03481656312942505,
0.015939926728606224,
-0.0897369235754013,
0.0017741515766829252,
-0.06276687979698181,
0.024767154827713966,
0.1127629205584526,
0.024551786482334137,
-0.009230821393430233,
0.00020948576275259256,
0.0540236234664917,
0.04845940321683884,
-0.0008163083111867309,
0.09975804388523102,
-0.04431907460093498,
0.032790787518024445,
-0.059553083032369614,
0.021759459748864174,
-0.10173773765563965,
-0.026233570650219917,
-0.11379940807819366,
-0.027086863294243813,
0.07875019311904907,
-0.034815382212400436,
0.05963990092277527,
0.008464602753520012,
0.03870146721601486,
-0.008121544495224953,
-0.004758918192237616,
0.00668817525729537,
-0.01936439983546734,
-0.05143166333436966,
0.0020530172623693943,
0.048358991742134094,
-0.013443302363157272,
0.03466436266899109,
0.04399501532316208,
0.0402350127696991,
-0.04768622666597366,
0.006012699566781521,
-0.05434343218803406,
-0.02582627721130848,
0.005273899994790554,
-0.03185950592160225,
-0.05874558165669441,
0.04784668982028961,
0.07899566739797592,
0.04940300062298775,
-0.026572518050670624,
-0.04117567837238312,
1.5751242265354635e-33,
0.053200043737888336,
0.04129961505532265,
0.05323643237352371,
-0.02995520830154419,
0.011408161371946335,
-0.002525792922824621,
0.034726936370134354,
0.03670556843280792,
-0.02948828600347042,
0.04339320957660675,
0.030609644949436188,
0.008166813291609287,
-0.0615532360970974,
0.08526875823736191,
0.0355571024119854,
-0.015364688821136951,
-0.04508992284536362,
0.013690998777747154,
-0.08303286135196686,
0.03761306777596474,
0.01647147722542286,
-0.03457631170749664,
-0.05145622789859772,
-0.02678608149290085,
-0.03454957529902458,
-0.01755201257765293,
-0.02852882817387581,
0.031923457980155945,
0.047211598604917526,
-0.020231598988175392,
-0.02667844481766224,
0.002150721848011017,
-0.05816502124071121,
0.04211188852787018,
-0.011880184523761272,
0.023329421877861023,
-0.03243629261851311,
-0.09456846863031387,
0.0038079237565398216,
-0.005273249465972185,
0.010414466261863708,
0.005148191004991531,
0.012963901273906231,
0.0022140108048915863,
-0.07509500533342361,
-0.10550421476364136,
0.013634847477078438,
0.07171723246574402,
-0.06814220547676086,
-0.03087591752409935,
-0.00621686689555645,
0.07097110152244568,
-0.02892182767391205,
-0.05619719624519348,
-0.039468396455049515,
-0.07624521851539612,
0.0403381809592247,
0.07387471944093704,
0.054057441651821136,
0.07202930748462677,
0.044873688369989395,
-0.012506035156548023,
-0.048411402851343155,
0.134068563580513,
-0.009268428198993206,
0.05689583346247673,
0.02829078584909439,
0.13212591409683228,
0.08438364416360855,
0.005037414841353893,
0.000221026930375956,
-0.023012079298496246,
-0.0011125081218779087,
-0.031150124967098236,
0.06004243716597557,
0.03518284857273102,
0.014470652677118778,
-0.10196498781442642,
0.0022666051518172026,
-0.030813371762633324,
-0.03029434010386467,
0.05083160474896431,
0.028376230970025063,
0.05106806755065918,
0.026478856801986694,
-0.11006008088588715,
-0.015882594510912895,
-0.053811877965927124,
-0.051691893488168716,
-0.0005394296022132039,
-0.12144091725349426,
-0.10474124550819397,
-0.012181644327938557,
0.05245165899395943,
-0.011868540197610855,
-3.881681629441856e-34,
0.06291715055704117,
-0.0589827224612236,
0.05239095166325569,
0.05606561899185181,
-0.0451384112238884,
-0.07538234442472458,
0.005516346078366041,
-0.024921409785747528,
-0.00502875866368413,
-0.056095849722623825,
-0.06574051082134247,
0.004417129792273045,
0.06925616413354874,
0.010848135687410831,
0.03744145482778549,
-0.005968528334051371,
-0.02330167591571808,
-0.022163107991218567,
0.003503179643303156,
-0.028545964509248734,
-0.0166452769190073,
0.09423967450857162,
0.002592208096757531,
0.011005530133843422,
-0.07750919461250305,
0.01890738308429718,
0.03785092011094093,
0.05116133764386177,
0.014971710741519928,
0.04072660580277443,
-0.049637701362371445,
-0.02358395978808403,
0.05905400216579437,
0.036273159086704254,
-0.04228260740637779,
0.03714846074581146,
0.021700004115700722,
0.0714341402053833,
-0.018894465640187263,
0.024764614179730415,
0.042037080973386765,
-0.04268926754593849,
0.03335181623697281,
-0.0004866183444391936,
0.060106225311756134,
0.03748299181461334,
-0.09728892147541046,
0.04638848826289177,
-0.07185135781764984,
0.040456172078847885,
-0.03507760167121887,
0.11093536764383316,
-0.09717700630426407,
-0.08765486627817154,
-0.052353061735630035,
-0.05864332243800163,
-0.044914811849594116,
-0.013336648233234882,
0.09475581347942352,
-0.010694148018956184,
-0.059326790273189545,
0.05458950996398926,
-0.026890603825449944,
0.06831561028957367,
0.01356267649680376,
0.05606341361999512,
-0.01788119412958622,
0.03569990396499634,
-0.02040039747953415,
-0.004356753546744585,
0.004709552973508835,
-0.015043859370052814,
0.09525499492883682,
-0.017519021406769753,
0.0308942049741745,
-0.01870221085846424,
0.017789868637919426,
-0.010057824663817883,
0.06256122887134552,
-0.020810823887586594,
0.028443893417716026,
-0.04721932113170624,
0.07240980863571167,
0.10376723110675812,
-0.016811005771160126,
0.07615786045789719,
-0.08831357210874557,
0.04000470042228699,
-0.0019913986325263977,
0.07365448772907257,
-0.025121565908193588,
0.008187493309378624,
0.040983546525239944,
0.004047638736665249,
-0.02176467888057232,
-4.095815597793262e-8,
-0.06669198721647263,
0.06588947027921677,
0.14618243277072906,
0.026011165231466293,
0.020866844803094864,
0.05891558900475502,
0.02742195501923561,
0.010728277266025543,
-0.054495133459568024,
-0.020492738112807274,
0.03235766664147377,
-0.013420960865914822,
0.042979735881090164,
-0.0698661208152771,
-0.028049971908330917,
0.05469648167490959,
0.031038371846079826,
-0.036269161850214005,
-0.007695178501307964,
0.00024701395886950195,
0.029880791902542114,
0.0017421557568013668,
0.08496557921171188,
-0.04618117958307266,
-0.014268472790718079,
-0.09819810092449188,
-0.10305267572402954,
0.05565246567130089,
0.07659772038459778,
-0.052772149443626404,
0.06405507028102875,
-0.037943121045827866,
0.07845339924097061,
0.048657678067684174,
0.04953278601169586,
0.014696016907691956,
-0.030085906386375427,
0.0036894597578793764,
-0.014389538206160069,
-0.04988953471183777,
-0.01711868681013584,
0.033154964447021484,
-0.009552651084959507,
-0.04646749049425125,
-0.033715780824422836,
-0.010396444238722324,
-0.09319737553596497,
-0.18108433485031128,
0.036453213542699814,
-0.0479392372071743,
0.0020570855122059584,
0.0019245976582169533,
-0.029115311801433563,
0.06401120871305466,
0.07016921043395996,
0.09110138565301895,
-0.018282050266861916,
-0.06277990341186523,
0.008111629635095596,
0.08997902274131775,
0.029673077166080475,
0.00243793404661119,
-0.07126500457525253,
-0.0002688937820494175
] | 0 | 0 | 0 | 0 |
|
How Gradient descent balances features: A dynamical analysis for two-layer neural networks | https://openreview.net/forum?id=25j2ZEgwTj | [
"Zhenyu Zhu",
"Fanghui Liu",
"Volkan Cevher"
] | Poster | This paper investigates the fundamental regression task of learning $k$ neurons (\emph{a.k.a.} teachers) from Gaussian input, using two-layer ReLU neural networks with width $m$ (\emph{a.k.a.} students) and $m, k= \mathcal{O}(1)$, trained via gradient descent under proper initialization and a small step-size. Our analysis follows a three-phase structure: \emph{alignment} after weak recovery, \emph{tangential growth}, and \emph{local convergence}, providing deeper insights into the learning dynamics of gradient descent (GD). We prove the global convergence at the rate of $\mathcal{O}(T^{-3})$ for the zero loss of excess risk. Additionally, our results show that GD automatically groups and balances student neurons, revealing an implicit bias toward achieving the minimum ``balanced'' $\ell_2$-norm in the solution. Our work extends beyond previous studies in exact-parameterization setting ($m = k = 1$, (Yehudai and Ohad, 2020)) and single-neuron setting ($m \geq k = 1$, (Xu and Du, 2023)). The key technical challenge lies in handling the interactions between multiple teachers and students during training, which we address by refining the alignment analysis in Phase 1 and introducing a new dynamic system analysis for tangential components in Phase 2. Our results pave the way for further research on optimizing neural network training dynamics and understanding implicit biases in more complex architectures. | learning theory, over-parameterization, learning dynamics | null | 13,306 | null | [
-0.09865108132362366,
-0.084524005651474,
0.03139861673116684,
0.057356011122465134,
0.055389851331710815,
0.015700316056609154,
0.00963848177343607,
-0.019401218742132187,
0.08225856721401215,
-0.018212150782346725,
0.06057055667042732,
0.052739404141902924,
0.03376210480928421,
0.037420522421598434,
-0.10083634406328201,
0.022531650960445404,
0.010806193575263023,
0.11365163326263428,
-0.09346030652523041,
0.015701329335570335,
0.05170939862728119,
-0.009005668573081493,
-0.01027278695255518,
0.05586225166916847,
0.030835384503006935,
0.026601361110806465,
-0.017887726426124573,
0.022973859682679176,
-0.020167097449302673,
-0.017709335312247276,
0.02655334398150444,
-0.04075856879353523,
-0.0217619389295578,
0.04028483107686043,
-0.07006064057350159,
0.015379153192043304,
0.01385410688817501,
0.03988393023610115,
-0.056481458246707916,
-0.004268265329301357,
0.03177396208047867,
0.06806033104658127,
-0.015636898577213287,
0.035470567643642426,
0.11892779916524887,
-0.024959063157439232,
-0.014853113330900669,
-0.11286486685276031,
-0.001732697244733572,
-0.02668166160583496,
0.016850823536515236,
-0.05005080997943878,
-0.0692330002784729,
0.058898989111185074,
0.00007020205521257594,
0.029922621324658394,
0.006753173656761646,
0.06955631822347641,
-0.0684097409248352,
0.058422841131687164,
0.011984300799667835,
-0.11058209091424942,
0.012104056775569916,
-0.015900488942861557,
0.03539310023188591,
-0.009776118211448193,
-0.00286104716360569,
0.03580096736550331,
-0.05151708796620369,
0.09038569033145905,
-0.015722481533885002,
0.057303667068481445,
-0.08406801521778107,
-0.029811382293701172,
0.07156715542078018,
0.012582652270793915,
0.052854347974061966,
0.03829538822174072,
0.037193987518548965,
-0.051480483263731,
0.011185904033482075,
-0.02535218559205532,
0.03565298393368721,
-0.03299243003129959,
-0.0058156088925898075,
-0.06191406399011612,
-0.06934025138616562,
0.016586117446422577,
0.05636877566576004,
-0.008677778765559196,
-0.041020724922418594,
0.061817917972803116,
-0.08856956660747528,
0.03897519409656525,
0.12757357954978943,
-0.04115572199225426,
0.011631404981017113,
-0.06170019879937172,
-0.01875917799770832,
0.09903199970722198,
-0.07286445051431656,
0.06028580665588379,
0.05700792744755745,
0.004646639339625835,
0.01973079703748226,
-0.06583383679389954,
0.04546113312244415,
-0.018511220812797546,
0.06883314996957779,
-0.05124395713210106,
0.08338174223899841,
0.016226662322878838,
0.049915045499801636,
0.057953521609306335,
-0.007775955833494663,
-0.022710386663675308,
0.059158552438020706,
-0.0493832603096962,
-0.05867619439959526,
0.08914758265018463,
-0.021502932533621788,
-0.012172587215900421,
-0.0849619209766388,
-0.007841150276362896,
0.016868585720658302,
0.00005557531403610483,
-0.05092081427574158,
-1.3172999511216159e-34,
0.000035648867196869105,
0.06176629662513733,
-0.026351571083068848,
0.00636539701372385,
-0.017749303951859474,
-0.029361849650740623,
0.0062215435318648815,
0.020542657002806664,
-0.0013457972090691328,
0.009966257959604263,
-0.08743330091238022,
-0.005404285155236721,
-0.0459134504199028,
0.07162930071353912,
0.045022301375865936,
-0.0102520901709795,
-0.02653694711625576,
0.015124557539820671,
0.035824913531541824,
-0.054517678916454315,
0.051130056381225586,
-0.058957379311323166,
0.011266790330410004,
-0.042579032480716705,
-0.06826567649841309,
-0.055873673409223557,
0.016798831522464752,
-0.008788147941231728,
-0.04642697423696518,
0.015735145658254623,
0.0017047852743417025,
-0.0363728366792202,
-0.11773981153964996,
-0.022367320954799652,
0.009410485625267029,
0.008781688287854195,
0.017214929684996605,
0.000466064375359565,
0.042191002517938614,
0.035155072808265686,
-0.032042551785707474,
0.07626282423734665,
0.04753568395972252,
-0.014228307642042637,
-0.004083681385964155,
-0.09181978553533554,
0.055631544440984726,
0.017403051257133484,
0.0025566341355443,
-0.12243495136499405,
-0.09275583922863007,
0.012604797258973122,
-0.07186758518218994,
-0.058657411485910416,
-0.019863445311784744,
0.0931018739938736,
-0.035722047090530396,
0.04866337403655052,
-0.002719050506129861,
0.002408988308161497,
0.0023714317940175533,
0.032503023743629456,
-0.069264717400074,
0.06954596191644669,
-0.01496855914592743,
0.05402517318725586,
-0.042078159749507904,
0.04611491411924362,
0.07986175268888474,
-0.011895814910531044,
0.03188341483473778,
0.03871270269155502,
0.006052309647202492,
-0.09728214144706726,
0.0842883363366127,
0.02856970764696598,
0.09947222471237183,
-0.1002335473895073,
-0.0048241401091217995,
0.02345110848546028,
-0.013300835154950619,
-0.008804414421319962,
-0.02409348450601101,
0.006870852783322334,
-0.10245370119810104,
-0.05612201243638992,
0.08490709960460663,
-0.10495225340127945,
0.00933038629591465,
0.040388185530900955,
-0.09109403938055038,
-0.012554111890494823,
0.02024928107857704,
0.014039513655006886,
0.015361766330897808,
1.966230574343728e-34,
-0.012049891985952854,
0.027389315888285637,
-0.06957346200942993,
0.012491913512349129,
0.006317595019936562,
0.026579955592751503,
-0.06395688652992249,
0.05184510350227356,
-0.016815273091197014,
-0.046042997390031815,
0.03280337527394295,
0.007143803872168064,
-0.015459731221199036,
0.03331909328699112,
0.06940070539712906,
-0.037468355149030685,
-0.08278452605009079,
0.060652244836091995,
0.08068505674600601,
-0.06176849454641342,
0.04231097921729088,
0.012510934844613075,
-0.0681491270661354,
0.09504904597997665,
-0.004655256401747465,
0.013584859669208527,
-0.06649162620306015,
0.0781325101852417,
-0.035552408546209335,
0.04780839383602142,
-0.09085501730442047,
-0.05212022364139557,
-0.01628580316901207,
0.07171746343374252,
-0.029973004013299942,
0.09001520276069641,
-0.04572717100381851,
-0.028250031173229218,
-0.014212816022336483,
0.0860648900270462,
0.03163182735443115,
-0.014305546879768372,
-0.04389624297618866,
-0.05793256312608719,
0.06200442463159561,
-0.058437347412109375,
0.032729458063840866,
0.045536868274211884,
-0.018054431304335594,
0.014360162429511547,
-0.005431929603219032,
-0.016974763944745064,
-0.09646101295948029,
0.007724641356617212,
-0.036823224276304245,
0.010774479247629642,
0.047263551503419876,
0.014890434220433235,
0.017752351239323616,
0.02371404506266117,
-0.045788221061229706,
-0.05320632830262184,
-0.057619210332632065,
0.02892344817519188,
-0.059447772800922394,
0.0005411505699157715,
0.022036058828234673,
0.0387270487844944,
0.07151021808385849,
0.05195871368050575,
0.01690395548939705,
0.06048804149031639,
0.05199652910232544,
-0.027792034670710564,
-0.013614517636597157,
-0.07652690261602402,
-0.02137063257396221,
-0.012752417474985123,
-0.05942491069436073,
-0.053383562713861465,
0.033752068877220154,
0.04293835163116455,
0.022089678794145584,
0.0544205866754055,
0.07731356471776962,
-0.0006393263465724885,
0.007270144764333963,
0.12119465321302414,
0.03147870674729347,
-0.05044678598642349,
-0.007534978445619345,
0.016405142843723297,
-0.012687763199210167,
-0.05930481478571892,
-0.026899296790361404,
-5.152537241315258e-8,
-0.03494339436292648,
0.05883638933300972,
0.07210210710763931,
0.07207129150629044,
0.05360615625977516,
-0.029128214344382286,
0.06497706472873688,
0.062105536460876465,
0.0008492959314025939,
0.0652666836977005,
0.0893944799900055,
0.027881670743227005,
-0.013057425618171692,
-0.11666497588157654,
-0.03248977288603783,
0.12372086942195892,
-0.07164078205823898,
0.014662901870906353,
-0.015943914651870728,
-0.02864251472055912,
0.021990487352013588,
0.005628302227705717,
0.04367247223854065,
-0.025968316942453384,
0.01129806786775589,
-0.12890781462192535,
-0.03415989875793457,
0.07111721485853195,
0.04151410982012749,
-0.012885120697319508,
-0.02137324959039688,
0.008352330885827541,
0.03332042321562767,
0.014434846118092537,
-0.009590396657586098,
0.05287688970565796,
0.03768399730324745,
0.04700489714741707,
0.027401931583881378,
0.018420767039060593,
-0.03180506080389023,
0.06720847636461258,
0.08639486879110336,
-0.02269965596497059,
0.051152992993593216,
-0.0507131852209568,
-0.059096887707710266,
-0.02401018887758255,
0.06029248237609863,
-0.0669131726026535,
0.03438223525881767,
0.04470635950565338,
-0.0013828056398779154,
0.12674227356910706,
0.035244669765233994,
-0.05087205022573471,
-0.08815184980630875,
-0.08424106240272522,
-0.04446762427687645,
0.025342466309666634,
0.0044259121641516685,
0.039922043681144714,
-0.09799547493457794,
-0.03281385451555252
] | 0 | 0 | 0 | 0 |
|
A Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement | https://openreview.net/forum?id=YaBiGjuDiC | [
"Hui Yuan",
"Yifan Zeng",
"Yue Wu",
"Huazheng Wang",
"Mengdi Wang",
"Liu Leqi"
] | Poster | Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for aligning language models (LMs) to be more helpful and less harmful.
At its core, RLHF uses a margin-based loss for preference optimization, which specifies the ideal LM behavior only in terms of the difference between preferred and dispreferred responses. In this paper, we identify a common pitfall of margin-based methods---the under-specification of ideal LM behavior on preferred and dispreferred responses individually, which results in two unintended consequences as the margin increases:
(1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures.
(2) The probability of preferred responses may decrease, even when those responses are ideal.
We demystify the reasons behind these problematic behaviors: margin-based losses couple the change in the preferred probability with the gradient of the dispreferred one, and vice versa, often preventing the preferred probability from increasing while the dispreferred one decreases, and thus causing a synchronized increase or decrease in both probabilities. We term this effect, inherent in margin-based objectives, gradient entanglement.
Formally, we derive conditions for general margin-based alignment objectives under which gradient entanglement becomes concerning: the inner product between the gradient of preferred log-probability and the gradient of dispreferred log-probability is large relative to the individual gradient norms. Furthermore, we theoretically investigate why such inner products can be large when aligning language models and empirically validate our findings. Empirical implications of our framework further extend to explaining important differences in the training dynamics of various preference optimization algorithms and suggesting future directions for improvement. | Alignment, Preference Optimization, Large Language Model | null | 13,302 | 2410.13828 | [
-0.017370130866765976,
-0.07585329562425613,
0.01605626568198204,
-0.012822521850466728,
0.007458006031811237,
0.11045892536640167,
-0.0138473529368639,
0.013408566825091839,
0.10720563679933548,
0.00430840440094471,
-0.05081889405846596,
-0.038533661514520645,
0.05291196331381798,
0.011322204023599625,
-0.01738516241312027,
-0.03360600396990776,
0.11415775120258331,
0.04608254134654999,
-0.12930616736412048,
0.0004990193992853165,
-0.044071439653635025,
-0.0009441695874556899,
0.03878580406308174,
0.026491113007068634,
-0.044390685856342316,
-0.023856276646256447,
0.00881656352430582,
0.009775600396096706,
0.008217399939894676,
-0.049225449562072754,
0.03861648216843605,
-0.015370608307421207,
0.05755428597331047,
-0.034698523581027985,
-0.07758644223213196,
0.06314650923013687,
-0.09327735006809235,
-0.029919050633907318,
0.009981561452150345,
-0.05042341724038124,
-0.13120004534721375,
0.026740457862615585,
0.013608275912702084,
-0.0010146589484065771,
0.052101992070674896,
0.014663130976259708,
-0.009957239031791687,
-0.006579788867384195,
-0.024333873763680458,
-0.07306085526943207,
-0.06224830076098442,
0.013439537957310677,
0.024800719693303108,
0.07586131244897842,
0.032793957740068436,
-0.0020891232416033745,
0.058551009744405746,
0.11264253407716751,
-0.053916819393634796,
-0.01564374379813671,
-0.025839222595095634,
-0.10833141207695007,
-0.034676916897296906,
0.03474731743335724,
-0.015597459860146046,
0.0070377434603869915,
-0.021470213308930397,
0.03654635697603226,
-0.07482275366783142,
0.15242192149162292,
-0.05866599455475807,
0.03682290017604828,
-0.03317511826753616,
0.011279815807938576,
0.05706516280770302,
0.07506770640611649,
0.05274096131324768,
-0.00418662978336215,
0.018290333449840546,
-0.02503933385014534,
0.015609418973326683,
0.06007350981235504,
0.07580236345529556,
-0.000968172971624881,
0.11484894901514053,
-0.08355185389518738,
-0.0545763298869133,
0.027664873749017715,
0.1262757033109665,
-0.05428086966276169,
-0.05350002646446228,
-0.03111645206809044,
-0.0017353547737002373,
0.06819122284650803,
0.0012212684378027916,
0.06777683645486832,
-0.06238101050257683,
-0.0027137023862451315,
-0.07046817988157272,
0.05969129130244255,
0.049457646906375885,
0.05297184735536575,
-0.028788825497031212,
-0.03931781277060509,
0.04477367177605629,
-0.05264897271990776,
-0.03984285518527031,
0.01281662005931139,
0.03807596117258072,
-0.04871831834316254,
0.013449400663375854,
0.021061761304736137,
-0.008359430357813835,
0.11926934868097305,
-0.08745794743299484,
-0.0782913938164711,
0.034157413989305496,
-0.06462962925434113,
-0.006314085330814123,
0.01694324240088463,
-0.01269736047834158,
-0.03784790262579918,
0.017561228945851326,
-0.01917157880961895,
-0.04174725338816643,
-0.04667903855443001,
-0.0084541542455554,
6.421281154740752e-33,
0.02036045677959919,
-0.00959700532257557,
-0.021363871172070503,
0.010264418087899685,
0.04778735712170601,
0.03565966710448265,
-0.029679009690880775,
-0.05444680526852608,
-0.0029705532360821962,
-0.028040077537298203,
0.022619202733039856,
-0.05400566756725311,
-0.029696092009544373,
0.030164754018187523,
-0.004245747346431017,
0.0029641836881637573,
-0.09611932188272476,
0.12007193267345428,
-0.1231687068939209,
-0.0014872413594275713,
0.014426033943891525,
0.00861032772809267,
-0.027187474071979523,
-0.031095925718545914,
-0.05254252254962921,
0.002972199348732829,
0.017427681013941765,
-0.028296317905187607,
-0.07696661353111267,
-0.016084227710962296,
-0.09901920706033707,
-0.034467972815036774,
-0.025484902784228325,
-0.02829638682305813,
0.03147372975945473,
-0.023340614512562752,
-0.04359991103410721,
-0.020067688077688217,
-0.02575242891907692,
-0.036857720464468,
-0.018433650955557823,
0.08569161593914032,
0.03782592713832855,
0.011442907154560089,
0.029647057875990868,
-0.03722715377807617,
0.06275828927755356,
-0.05052926763892174,
-0.020700259134173393,
-0.11614876240491867,
0.015873899683356285,
0.02794051356613636,
0.06396409869194031,
-0.004478584509342909,
-0.03691808506846428,
0.029505984857678413,
0.034584153443574905,
0.08609312027692795,
-0.058423228561878204,
0.03645417094230652,
-0.04204026982188225,
0.04342331364750862,
0.01374936942011118,
-0.0392068512737751,
0.07315191626548767,
0.012418076395988464,
-0.03366074711084366,
0.03400298207998276,
-0.01829013042151928,
0.00871540978550911,
0.008382417261600494,
0.002804157556965947,
-0.038519952446222305,
-0.0017744614742696285,
0.03004247508943081,
-0.047435466200113297,
0.010071824304759502,
-0.031989190727472305,
0.06831325590610504,
-0.052295394241809845,
0.0166847612708807,
-0.015976430848240852,
-0.0036262250505387783,
-0.03479471057653427,
-0.10636726021766663,
-0.13259555399417877,
0.06724518537521362,
-0.06096366420388222,
0.0015035716351121664,
0.059506677091121674,
-0.03191632032394409,
-0.012465610168874264,
0.09712687134742737,
0.07421784102916718,
0.0455218181014061,
-7.005374012088777e-33,
-0.12601374089717865,
-0.014816953800618649,
-0.01340994518250227,
0.05174408107995987,
-0.11689527332782745,
0.0019306485773995519,
0.02956010028719902,
-0.015227307565510273,
-0.003913393244147301,
-0.029432125389575958,
0.00343399029225111,
-0.036730822175741196,
0.018310794606804848,
0.04572039842605591,
0.01700834184885025,
-0.013960388489067554,
0.00639753695577383,
0.018550077453255653,
0.0726528987288475,
-0.008736100047826767,
0.09846507757902145,
0.0336349755525589,
-0.1325158327817917,
0.07541188597679138,
0.001186925102956593,
0.03117913194000721,
-0.015911510214209557,
0.032122038304805756,
-0.0019290882628411055,
0.003822905011475086,
-0.07081501185894012,
0.01438236702233553,
-0.0915369912981987,
0.02627771720290184,
-0.03416116163134575,
0.024112967774271965,
-0.04842974990606308,
-0.016931911930441856,
-0.02576446160674095,
0.08666533976793289,
0.010264533571898937,
-0.0066885147243738174,
-0.06508769094944,
-0.02045729197561741,
0.035970497876405716,
-0.039053890854120255,
-0.0762663409113884,
-0.02141607366502285,
-0.006774585228413343,
-0.022167176008224487,
-0.04370029643177986,
0.05854887142777443,
-0.04725940525531769,
0.023308996111154556,
-0.12455566972494125,
-0.02999088354408741,
-0.018169062212109566,
-0.018409451469779015,
0.08950760215520859,
0.05631162226200104,
-0.035708919167518616,
0.03097637929022312,
0.04346626251935959,
-0.011935043148696423,
0.02520584501326084,
0.05583404377102852,
0.004444400314241648,
0.007425542455166578,
0.06679903715848923,
-0.02640342339873314,
0.08132690191268921,
0.043167997151613235,
0.0447772853076458,
0.03614548593759537,
0.05488310381770134,
0.03288188576698303,
-0.005439723376184702,
-0.0703766793012619,
-0.021748768165707588,
-0.01041700504720211,
0.050551123917102814,
0.04509952664375305,
0.048701416701078415,
0.08244336396455765,
-0.02826838009059429,
0.034445393830537796,
-0.05019752308726311,
0.08237411081790924,
0.005197509191930294,
-0.05340341851115227,
0.0067117055878043175,
-0.028892097994685173,
0.04613093286752701,
0.03359882906079292,
-0.08505988121032715,
-6.215322656544231e-8,
-0.0969523936510086,
-0.03758127987384796,
0.01829458586871624,
0.061155036091804504,
-0.05226079374551773,
-0.02414602041244507,
-0.010973983444273472,
-0.05628197267651558,
0.02155090682208538,
-0.020357083529233932,
-0.002840596018359065,
0.03637317940592766,
-0.06405674666166306,
-0.008115533739328384,
-0.039845697581768036,
0.0923590362071991,
0.007253712974488735,
0.012024791911244392,
-0.025291230529546738,
0.048727817833423615,
0.0861310288310051,
0.01253708079457283,
-0.014876202680170536,
0.016576819121837616,
0.010160580277442932,
-0.04776156321167946,
-0.014342911541461945,
0.15160749852657318,
0.04429997131228447,
-0.0025315850507467985,
-0.001584450015798211,
-0.05352870374917984,
0.018051013350486755,
0.06910836696624756,
-0.014761719852685928,
0.13850955665111542,
-0.04740796238183975,
-0.04482243210077286,
0.057045526802539825,
0.06405630707740784,
0.04262322187423706,
0.09937597066164017,
0.011037689633667469,
0.024927690625190735,
0.044537462294101715,
0.037975240498781204,
-0.008344962261617184,
-0.08709398657083511,
0.05717283859848976,
-0.0775153636932373,
0.01793026365339756,
-0.054050296545028687,
0.0006267123389989138,
0.04038351774215698,
0.034312281757593155,
0.004832419566810131,
-0.06887660175561905,
-0.024789752438664436,
0.0927666574716568,
0.04865289106965065,
0.000729122431948781,
0.05346481502056122,
-0.044448137283325195,
-0.01249707117676735
] | https://github.com/humainlab/understand_marginpo | 5 | 0 | 0 | 0 |
Model Editing as a Robust and Denoised variant of DPO: A Case Study on Toxicity | https://openreview.net/forum?id=lOi6FtIwR8 | [
"Rheeya Uppaal",
"Apratim Dey",
"Yiting He",
"Yiqiao Zhong",
"Junjie Hu"
] | Poster | Recent alignment algorithms such as direct preference optimization (DPO) have been developed to improve the safety of large language models (LLMs) by training these models to match human behaviors exemplified by preference data. However, these methods are
both computationally intensive and lacking in controllability and transparency, inhibiting their widespread use. Furthermore, these tuning-based methods require large-scale preference data for training and are susceptible to noisy preference data. In this paper, we introduce a tuning-free alignment alternative, ProFS (Projection Filter for Subspaces), and demonstrate its effectiveness under the use case of toxicity reduction. Grounded on theory from factor analysis, ProFS is a sample-efficient model editing approach that identifies a toxic subspace in the model parameter space and reduces model toxicity by projecting away the detected subspace. The toxic subspace is identified by extracting preference data embeddings from the language model, and removing non-toxic information from these embeddings. We show that ProFS is more sample-efficient than DPO, further showcasing greater robustness to noisy data. Finally, we attempt to connect tuning based alignment with editing, by establishing both theoretical and empirical connections between ProFS and DPO, showing that ProFS can be interpreted as a denoised version of a single DPO step. | model editing, mechanistic interpretability, ai safety, alignment, toxicity, llms | We present model editing based sample-efficient and noise robust replacement to DPO, for reducing model toxicity. | 13,296 | 2405.13967 | [
0.002956787124276161,
-0.11199219524860382,
-0.00791831873357296,
0.019577719271183014,
0.06214636191725731,
0.06985243409872055,
0.01117620524019003,
-0.00573698990046978,
0.04082171246409416,
-0.0036245055962353945,
-0.026012634858489037,
-0.008123829960823059,
0.0572175607085228,
0.04399150237441063,
-0.055771730840206146,
-0.012485749088227749,
0.11572346836328506,
0.0690741017460823,
-0.09738335013389587,
-0.019864648580551147,
0.004179740324616432,
0.06001617759466171,
0.026701347902417183,
-0.016090398654341698,
-0.0044199381954967976,
0.04745848476886749,
-0.0057751890271902084,
-0.02027827315032482,
-0.00009556171426083893,
-0.025027265772223473,
0.036420270800590515,
0.04669160023331642,
0.022851519286632538,
-0.03538329899311066,
-0.05919533967971802,
0.004283396061509848,
-0.061686236411333084,
0.01809273473918438,
-0.045614078640937805,
-0.04998941347002983,
-0.05262342095375061,
0.033682696521282196,
0.011924870312213898,
0.028955988585948944,
0.010458502918481827,
-0.03579835593700409,
-0.03453481197357178,
-0.07258878648281097,
-0.014554696157574654,
0.03399926424026489,
-0.09512503445148468,
-0.05454147234559059,
0.0005537814577110112,
0.09582361578941345,
0.012620489113032818,
-0.116595059633255,
0.09658107161521912,
-0.016028765588998795,
-0.010395209304988384,
-0.03783608227968216,
-0.0639413595199585,
-0.03181663155555725,
-0.018151015043258667,
0.020171279087662697,
0.029666369780898094,
0.03229045867919922,
-0.06366798281669617,
0.07531090825796127,
-0.08746630698442459,
0.08923687785863876,
-0.11160638928413391,
0.06784408539533615,
0.014474921859800816,
-0.008389064110815525,
-0.022431258112192154,
0.08121640980243683,
-0.010603797622025013,
-0.025300338864326477,
0.12036793678998947,
-0.047157734632492065,
0.07809112966060638,
0.008198125287890434,
0.04715310409665108,
0.0400887094438076,
0.08037541806697845,
0.027466043829917908,
0.038611624389886856,
-0.07975279539823532,
0.07315073162317276,
-0.005606204736977816,
-0.04569539427757263,
0.03443361446261406,
0.11911617964506149,
0.043225470930337906,
-0.0553886704146862,
0.039006054401397705,
-0.011348778381943703,
-0.006250273436307907,
0.029986727982759476,
0.08625832945108414,
0.03702434524893761,
0.06629174947738647,
-0.06816882640123367,
-0.0063568237237632275,
-0.026592478156089783,
-0.04802906885743141,
-0.007155268918722868,
0.08468560129404068,
0.0033539042342454195,
-0.06165594980120659,
-0.0065094418823719025,
0.028905175626277924,
-0.05834184214472771,
0.0056541538797318935,
0.018034322187304497,
-0.06442826986312866,
0.10025995224714279,
-0.07531054317951202,
0.04303920641541481,
-0.0007070500869303942,
-0.034442853182554245,
0.01854524202644825,
-0.0016213132767006755,
0.0022010209504514933,
0.06235470622777939,
0.03675643727183342,
-0.059084754437208176,
2.5646719014469736e-33,
0.05781342834234238,
0.07569868862628937,
0.001497593242675066,
0.005685474723577499,
0.058702707290649414,
-0.0169790368527174,
-0.03693321719765663,
0.04002327471971512,
-0.06132424995303154,
-0.03181145340204239,
0.01607677899301052,
-0.030679041519761086,
-0.05875011160969734,
0.06487578898668289,
0.041683927178382874,
-0.06833909451961517,
0.03619919344782829,
0.08273841440677643,
-0.06349900364875793,
-0.04946120083332062,
0.0658549889922142,
-0.007939716801047325,
-0.01402714941650629,
-0.061481259763240814,
-0.024439089000225067,
-0.009672965854406357,
0.0365925095975399,
-0.008627929724752903,
-0.08435563743114471,
0.004920164123177528,
-0.11610748618841171,
-0.00993451476097107,
0.025618072599172592,
0.030322490260004997,
-0.010672389529645443,
-0.0697159543633461,
-0.02511838637292385,
-0.036736853420734406,
0.030691368505358696,
-0.017507217824459076,
0.05038655549287796,
0.025954045355319977,
0.02354717068374157,
-0.014293615706264973,
-0.026028858497738838,
-0.014878570102155209,
-0.0009453534730710089,
0.020923670381307602,
-0.019046209752559662,
-0.038806695491075516,
0.056127481162548065,
0.0258767232298851,
0.0299211572855711,
-0.007997715845704079,
-0.12976236641407013,
0.05028454214334488,
0.057188019156455994,
-0.06071796640753746,
0.023251166567206383,
-0.012242215685546398,
-0.0807400718331337,
0.05027403309941292,
0.002497789915651083,
0.005483625922352076,
0.07842245697975159,
-0.03736741840839386,
0.00791811104863882,
-0.01538053248077631,
0.08376893401145935,
-0.009855576790869236,
-0.0899365097284317,
0.01913543976843357,
-0.042196426540613174,
0.02630762942135334,
-0.0366385243833065,
-0.0781591385602951,
0.020434873178601265,
-0.026812361553311348,
0.011109364219009876,
0.002408114029094577,
0.009929622523486614,
0.035291388630867004,
-0.025320174172520638,
-0.053870074450969696,
-0.07409626990556717,
-0.17786061763763428,
0.030753592029213905,
-0.018841257318854332,
-0.0021298781502991915,
-0.030242450535297394,
0.02616046369075775,
-0.010973185300827026,
-0.03462229669094086,
0.04045531526207924,
0.027408745139837265,
-1.7471346822332735e-33,
-0.04819972813129425,
0.001982741756364703,
0.04478156939148903,
0.052916958928108215,
0.0006100677419453859,
-0.09950660914182663,
0.033607881516218185,
0.08563664555549622,
0.07972215861082077,
-0.043219178915023804,
-0.010946986265480518,
-0.08038122951984406,
0.045206502079963684,
-0.033807627856731415,
0.04039528965950012,
0.05142166092991829,
-0.027516966685652733,
0.037133414298295975,
-0.011010336689651012,
-0.03361567109823227,
-0.0045445701107382774,
0.038421373814344406,
-0.08324800431728363,
0.10655505210161209,
-0.05363691225647926,
0.05822297930717468,
0.034403204917907715,
0.02096206322312355,
0.03423745930194855,
0.023531027138233185,
-0.08052437752485275,
0.08374185860157013,
-0.10041899234056473,
0.009627205319702625,
-0.07056920975446701,
0.04363088309764862,
-0.04551158472895622,
-0.044951580464839935,
0.012975804507732391,
0.045653410255908966,
-0.0009806531015783548,
0.07655669003725052,
-0.13593994081020355,
0.022571450099349022,
0.04029401019215584,
-0.0197772104293108,
-0.06434830278158188,
-0.06414337456226349,
0.05598880350589752,
-0.04205236956477165,
-0.02895090915262699,
0.03469906374812126,
-0.036599300801754,
0.0548984594643116,
0.016156593337655067,
-0.09823637455701828,
0.027941348031163216,
-0.05016202852129936,
0.0964956060051918,
-0.02521895244717598,
0.029106905683875084,
0.034466955810785294,
0.019576197490096092,
-0.04471354931592941,
-0.018462153151631355,
-0.04530155286192894,
-0.006602047011256218,
0.03638818860054016,
0.010743923485279083,
-0.06564246863126755,
0.02634209394454956,
-0.05192042514681816,
-0.033779799938201904,
0.0037794490344822407,
-0.029369840398430824,
-0.09513112902641296,
0.014058978296816349,
-0.0016104590613394976,
-0.010758908465504646,
0.028167497366666794,
0.0699598416686058,
0.024393996223807335,
0.0394538976252079,
0.0338888019323349,
-0.019073843955993652,
0.09358736127614975,
0.011920713819563389,
0.1297832429409027,
-0.004859759472310543,
-0.03576850891113281,
-0.027874482795596123,
-0.048938650637865067,
-0.010014143772423267,
0.07279515266418457,
-0.04068126156926155,
-4.7859771257208195e-8,
-0.09838856011629105,
-0.021853409707546234,
0.007484025787562132,
0.06771392375230789,
-0.07381702959537506,
-0.07145895063877106,
0.04915761575102806,
0.06629467010498047,
0.026374852284789085,
-0.004061114974319935,
0.008492310531437397,
0.06279940903186798,
-0.058528460562229156,
-0.05157807096838951,
-0.04774995893239975,
0.08263301104307175,
-0.030279258266091347,
0.06476219743490219,
-0.012713512405753136,
0.09430129081010818,
-0.024839317426085472,
0.0769646018743515,
-0.011582750827074051,
-0.009836077690124512,
0.10537057369947433,
-0.02928713522851467,
-0.017279665917158127,
0.053364504128694534,
0.1060173511505127,
0.07597092539072037,
0.0035339328460395336,
0.004955674055963755,
-0.0189627967774868,
-0.01721833646297455,
-0.03835868462920189,
0.05911419913172722,
0.03355056047439575,
-0.07886729389429092,
-0.06842818856239319,
0.031167607754468918,
0.03762231767177582,
0.08324583619832993,
-0.05435342341661453,
0.010399079881608486,
0.027399245649576187,
0.03752554953098297,
0.04189285635948181,
-0.07594411075115204,
0.023739196360111237,
-0.05689268186688423,
0.031783852726221085,
-0.05367976427078247,
-0.03241722658276558,
-0.0021536319982260466,
-0.029243228957057,
0.008450539782643318,
-0.009741481393575668,
-0.013542511500418186,
0.036452073603868484,
-0.018766062334179878,
0.05524337291717529,
0.014541164971888065,
0.10086328536272049,
0.027997953817248344
] | https://github.com/uppaal/detox-edit | 4 | 0 | 0 | 0 |
Should VLMs be Pre-trained with Image Data? | https://openreview.net/forum?id=Pj4Aid3XqL | [
"Sedrick Keh",
"Jean Mercat",
"Samir Yitzhak Gadre",
"Kushal Arora",
"Igor Vasiljevic",
"Benjamin Burchfiel",
"Shuran Song",
"Russ Tedrake",
"Thomas Kollar",
"Ludwig Schmidt",
"Achal Dave"
] | Poster | Pre-trained LLMs that are further trained with image data perform well on vision-language tasks.
While adding images during a second training phase effectively unlocks this capability, it is unclear how much of a gain or loss this two-step pipeline gives over VLMs which integrate images earlier into the training process.
To investigate this, we train models spanning various datasets, scales, image-text ratios, and amount of pre-training done before introducing vision tokens.
We then fine-tune these models and evaluate their downstream performance on a suite of vision-language and text-only tasks.
We find that pre-training with a mixture of image and text data allows models to perform better on vision-language tasks while maintaining strong performance on text-only evaluations.
On an average of 6 diverse tasks, we find that for a 1B model, introducing visual tokens 80\% of the way through pre-training results in a 2\% average improvement over introducing visual tokens to a fully pre-trained model. | vision language models, pre-training, fine-tuning | We vary amount of image data used in pre-training VLMs, questioning the conventional formula of fine-tuning pre-trained LLMs into VLMs | 13,294 | 2503.07603 | [
0.025012636557221413,
-0.037239205092191696,
0.023887744173407555,
0.04550287127494812,
0.08979522436857224,
0.013589156791567802,
-0.0049045332707464695,
-0.0033564683981239796,
0.05544600635766983,
-0.05832238867878914,
-0.0019502246286720037,
-0.039655912667512894,
0.017099643126130104,
0.05926752835512161,
0.03832447901368141,
0.009488213807344437,
0.10131260752677917,
0.0557243786752224,
-0.10739213973283768,
-0.02138134092092514,
0.019715242087841034,
-0.0058884876780211926,
0.09477268904447556,
-0.0010487082181498408,
0.058836985379457474,
0.00011039297532988712,
-0.06301514059305191,
-0.060009125620126724,
0.06067632511258125,
-0.05203373730182648,
-0.026383662596344948,
-0.010580377653241158,
0.027803651988506317,
0.09059049189090729,
-0.0706581398844719,
0.05404185876250267,
0.002620378974825144,
0.014454100281000137,
0.027411209419369698,
-0.051458168774843216,
-0.027416614815592766,
-0.045192983001470566,
0.020650820806622505,
-0.011404708959162235,
0.08208242058753967,
0.025557689368724823,
0.05384565517306328,
-0.048043858259916306,
-0.027187319472432137,
-0.048721298575401306,
-0.09145710617303848,
-0.05043449625372887,
-0.02489028498530388,
-0.013081852346658707,
-0.05409533157944679,
0.056341297924518585,
0.03333854302763939,
0.009178971871733665,
0.015949271619319916,
-0.00405490817502141,
-0.03795415163040161,
-0.07077164202928543,
-0.02535169944167137,
0.023161083459854126,
-0.03477233275771141,
-0.03538244590163231,
0.04505857080221176,
-0.0021221598144620657,
0.0005056123482063413,
-0.024011583998799324,
0.022617194801568985,
0.09474828094244003,
0.017382783815264702,
0.01564050279557705,
0.003927606623619795,
0.05050504580140114,
0.038247548043727875,
-0.015969764441251755,
0.07514660805463791,
-0.09339115023612976,
0.035939883440732956,
0.012947367504239082,
0.05308839678764343,
0.0022093330044299364,
0.12415215373039246,
0.030326368287205696,
-0.010733013041317463,
0.011030109599232674,
-0.022245274856686592,
-0.0347822830080986,
-0.07296396046876907,
-0.07708553224802017,
0.07124555855989456,
0.007111470680683851,
0.015212404541671276,
-0.02536611445248127,
-0.032844558358192444,
-0.11182259768247604,
0.00004168751911493018,
0.05125877633690834,
0.0028884096536785364,
0.040270522236824036,
0.08126077055931091,
0.046717818826436996,
0.0438736192882061,
0.012313841842114925,
0.10894521325826645,
0.09069143235683441,
0.03795261308550835,
-0.026630287989974022,
0.10708404332399368,
0.045155394822359085,
-0.0180999506264925,
0.004657319746911526,
-0.020447414368391037,
-0.02190346270799637,
-0.03230631351470947,
-0.049297403544187546,
0.03609896078705788,
0.06563419848680496,
-0.0319703072309494,
-0.018951868638396263,
0.021041614934802055,
-0.03467230126261711,
-0.006833544932305813,
-0.08679318428039551,
-0.0356081947684288,
1.4373931146530508e-33,
0.016051629558205605,
0.035967640578746796,
0.05174417048692703,
-0.009790943004190922,
0.048853449523448944,
-0.02999553643167019,
-0.010970032773911953,
-0.010109358467161655,
-0.045742202550172806,
-0.028558893129229546,
-0.05110472813248634,
-0.018505781888961792,
-0.03478943184018135,
0.16818834841251373,
0.011122371070086956,
-0.04105069488286972,
-0.006703051272779703,
0.041202325373888016,
-0.016806844621896744,
0.024733547121286392,
0.02082536555826664,
-0.07110854983329773,
-0.004807902965694666,
-0.03407784551382065,
0.04411681368947029,
0.005499894265085459,
0.0670640841126442,
-0.09166314452886581,
-0.057223107665777206,
0.023192116990685463,
-0.12532095611095428,
0.019726095721125603,
0.08663487434387207,
0.05321832373738289,
0.03595852479338646,
0.0018743498949334025,
0.01994432508945465,
0.014732060953974724,
0.0504697784781456,
0.007605158258229494,
0.004383384715765715,
0.09458145499229431,
0.000327645888319239,
-0.08361375331878662,
-0.07635925710201263,
-0.03217983990907669,
-0.023166721686720848,
0.06935836374759674,
-0.09720170497894287,
-0.02076263539493084,
0.04005052521824837,
-0.005869895685464144,
-0.07957357913255692,
-0.02065209485590458,
-0.0028656357899308205,
0.01624246872961521,
0.026422642171382904,
0.027378667145967484,
0.0024444798473268747,
-0.013717100024223328,
-0.021983126178383827,
0.012081719003617764,
-0.020501820370554924,
0.10443707555532455,
0.001752333133481443,
-0.019005289301276207,
0.026798950508236885,
-0.016122527420520782,
-0.01965544931590557,
-0.05933050811290741,
-0.07372613251209259,
0.03357649967074394,
-0.04776062071323395,
-0.08056049793958664,
0.09322268515825272,
-0.0032510128803551197,
0.05909257009625435,
-0.09841731935739517,
0.019852224737405777,
0.10453611612319946,
0.02413896843791008,
-0.02559211105108261,
0.030274737626314163,
-0.09114304929971695,
0.006311073899269104,
-0.03101535513997078,
0.025690540671348572,
-0.14183473587036133,
-0.01360885240137577,
-0.027966344729065895,
-0.039530955255031586,
-0.06135837361216545,
-0.028354179114103317,
-0.02906775288283825,
0.029500063508749008,
-1.1503434553249946e-33,
0.009141084738075733,
0.06808354705572128,
-0.02567366696894169,
0.07370293885469437,
-0.024050330743193626,
-0.020294656977057457,
0.11128438264131546,
0.02025452069938183,
-0.05237707123160362,
-0.07913564890623093,
-0.019528158009052277,
-0.012312815524637699,
-0.0495922714471817,
-0.026779524981975555,
-0.0346546471118927,
-0.07085493206977844,
0.006281968671828508,
-0.013291193172335625,
0.05469898879528046,
0.022831952199339867,
0.03162738308310509,
0.09592321515083313,
-0.06403641402721405,
0.02694600448012352,
-0.05700407549738884,
0.07236560434103012,
-0.08189742267131805,
0.06540598720312119,
-0.05684898421168327,
0.01220272108912468,
-0.010459186509251595,
0.0001239713456016034,
0.06473936885595322,
-0.052546076476573944,
-0.049964480102062225,
0.048987969756126404,
0.015950649976730347,
-0.02384433150291443,
-0.018035458400845528,
0.119588702917099,
0.10517489165067673,
-0.0646827444434166,
-0.04700599983334541,
-0.01107548363506794,
-0.0793905034661293,
-0.06909018754959106,
-0.011032802052795887,
-0.061807844787836075,
-0.0048587266355752945,
0.030842944979667664,
-0.024721914902329445,
0.03308166190981865,
-0.12205930054187775,
0.004080331418663263,
-0.07323270291090012,
-0.12687569856643677,
-0.022743865847587585,
0.00012181252532172948,
-0.008857821114361286,
-0.05240922048687935,
-0.029043905436992645,
-0.005742221139371395,
-0.01660195365548134,
-0.022258244454860687,
0.016828540712594986,
-0.05531252548098564,
-0.04422781616449356,
0.004213350359350443,
0.0446743406355381,
-0.021683933213353157,
0.04727803170681,
-0.0629410445690155,
-0.030331959947943687,
0.0247353482991457,
-0.009133336134254932,
-0.019231805577874184,
-0.02576596476137638,
-0.005812750197947025,
0.02490902505815029,
-0.042765311896800995,
-0.03967937454581261,
-0.03919468820095062,
-0.008290870115160942,
0.08961604535579681,
0.09211822599172592,
0.15312005579471588,
0.012901895679533482,
-0.035330262035131454,
0.011113636195659637,
0.05410298332571983,
-0.08727088570594788,
0.0009701185626909137,
0.005327851045876741,
0.05855565890669823,
-0.01938982494175434,
-3.9422847208925305e-8,
-0.05856456235051155,
0.024402836337685585,
-0.021449532359838486,
0.04469669982790947,
0.011002994142472744,
-0.008587786927819252,
-0.07030350714921951,
0.14431099593639374,
-0.04825172945857048,
-0.03532617539167404,
0.038888197392225266,
0.04012609273195267,
-0.07126842439174652,
-0.08472158759832382,
0.010226189158856869,
0.04025254398584366,
0.05244799703359604,
0.016406839713454247,
-0.006346014328300953,
-0.014732535928487778,
0.05301238223910332,
0.04589923098683357,
0.010223858058452606,
-0.08808524906635284,
0.015876956284046173,
-0.07052052021026611,
-0.03143637999892235,
0.07759375125169754,
0.014615827240049839,
-0.03478888422250748,
0.02649148367345333,
0.024802902713418007,
-0.028334764763712883,
0.014839994721114635,
0.06513114273548126,
-0.0163776483386755,
0.01157583761960268,
-0.016920659691095352,
-0.0031087228562682867,
-0.010154442861676216,
0.06094994395971298,
0.07370804250240326,
0.001985873794183135,
0.020801881328225136,
-0.01240930613130331,
0.012135759927332401,
-0.005414288491010666,
-0.07404478639364243,
0.0030942317098379135,
-0.06892050802707672,
-0.0014334486331790686,
0.002077197190374136,
-0.035690709948539734,
0.14779260754585266,
0.021829139441251755,
-0.023232709616422653,
0.07802402228116989,
-0.02278938703238964,
0.024263208732008934,
0.14404571056365967,
0.020621003583073616,
-0.020687825977802277,
-0.016055719926953316,
-0.06243463233113289
] | 0 | 0 | 0 | 0 |
|
DynaMath: A Dynamic Visual Benchmark for Evaluating Mathematical Reasoning Robustness of Vision Language Models | https://openreview.net/forum?id=VOAMTA8jKu | [
"Chengke Zou",
"Xingang Guo",
"Rui Yang",
"Junyu Zhang",
"Bin Hu",
"Huan Zhang"
] | Poster | The rapid advancements in Vision-Language Models (VLMs) have shown great potential in tackling mathematical reasoning tasks that involve visual context. Unlike humans who can reliably apply solution steps to similar problems with minor modifications, we found that state-of-the-art VLMs like GPT-4o can consistently fail in these scenarios, revealing limitations in their mathematical reasoning capabilities. In this paper, we investigate the **mathematical reasoning robustness** in VLMs and evaluate how well these models perform under different variants of the same question, such as changes in visual numerical values or function graphs.
While several vision-based math benchmarks have been developed to assess VLMs' problem-solving capabilities, these benchmarks contain only static sets of problems and cannot easily evaluate mathematical reasoning robustness.
To fill this gap, we introduce **DynaMath**, a dynamic visual math benchmark designed for in-depth assessment of VLMs. **DynaMath** includes 501 high-quality, multi-topic *seed* questions, *each represented as a Python program*. Those programs are carefully designed and annotated to enable the automatic generation of a much larger set of *concrete* questions, including many different types of visual and textual variations.
**DynaMath** allows us to evaluate the generalization ability of VLMs, by assessing their performance under varying input conditions of a seed question. We evaluated 14 state-of-the-art VLMs with 5,010 generated concrete questions (10 per seed question). Our results show that the worst-case model accuracy, defined as the percentage of correctly answered seed questions in all 10 variants, is significantly lower than the average-case accuracy. In addition, many models show high consistency in answering these questions -- the incorrectness of a certain variant of a seed question is not only due to inherent randomness. Our analysis emphasizes the need to study the robustness of VLMs' reasoning abilities, and **DynaMath** provides valuable insights to guide the development of more reliable models for mathematical reasoning. | Visual Mathematical Benchmark, Vision Language Models | We introduce DYNAMATH, a dynamic visual math benchmark designed for an in-depth assessment of the reasoning robustness of VLMs. | 13,293 | 2411.00836 | [
-0.02277504652738571,
-0.020354123786091805,
-0.028950907289981842,
-0.03435433655977249,
0.04593810811638832,
-0.07862627506256104,
-0.010130883194506168,
0.07087896019220352,
-0.03973887488245964,
-0.03753857687115669,
-0.09946827590465546,
-0.08690116554498672,
0.05660748481750488,
0.14822691679000854,
0.004710645414888859,
0.00617183605208993,
0.049802422523498535,
0.04733934625983238,
-0.10452546924352646,
-0.04566257447004318,
-0.005306436214596033,
-0.030232904478907585,
-0.011875308118760586,
-0.021054381504654884,
0.043818358331918716,
0.028983376920223236,
0.02040199749171734,
-0.048775821924209595,
0.01749437488615513,
0.00002062631210719701,
-0.060053594410419464,
-0.034060876816511154,
-0.039400070905685425,
0.020589176565408707,
-0.05447107180953026,
0.06644945591688156,
-0.009849738329648972,
0.03889505937695503,
-0.02161569707095623,
0.01679019071161747,
-0.09367415308952332,
0.006073262542486191,
0.049602169543504715,
-0.041008781641721725,
0.059088919311761856,
0.015214593149721622,
-0.026600714772939682,
0.032519225031137466,
0.004775418434292078,
-0.04117046296596527,
-0.15644437074661255,
-0.05690816044807434,
-0.0786471962928772,
-0.05251603573560715,
0.020298203453421593,
0.07444539666175842,
0.06553085893392563,
-0.04368192330002785,
0.04488253965973854,
-0.04958094656467438,
0.02572864107787609,
-0.017573446035385132,
0.030037427321076393,
0.04051395133137703,
-0.033149752765893936,
0.03655751794576645,
0.04958401620388031,
-0.017814109101891518,
0.015466384589672089,
0.05620087683200836,
-0.05279772728681564,
0.12246353924274445,
-0.039913780987262726,
-0.005327392835170031,
-0.011432597413659096,
0.0857120081782341,
0.032807573676109314,
-0.07698444277048111,
0.034356992691755295,
-0.10001160949468613,
0.07215224951505661,
-0.021401621401309967,
-0.03888691961765289,
0.0625152736902237,
0.13838030397891998,
0.019070113077759743,
-0.058605607599020004,
0.0318727046251297,
0.06075608357787132,
0.002894874196499586,
-0.05698926001787186,
-0.0789441391825676,
0.003997485153377056,
0.021202603355050087,
0.06760887801647186,
0.0191540215164423,
0.032777395099401474,
-0.07844778895378113,
-0.04764274135231972,
0.06919675320386887,
-0.0011415638728067279,
0.004483197815716267,
0.041427113115787506,
-0.04959285631775856,
0.020707255229353905,
0.02830399014055729,
0.13127675652503967,
0.019371211528778076,
0.026251977309584618,
-0.10040369629859924,
0.06912467628717422,
0.005386230535805225,
0.04657047986984253,
0.03536555543541908,
-0.05375927314162254,
0.001221284968778491,
0.007430040743201971,
-0.03297567740082741,
0.0364551767706871,
-0.008803782984614372,
0.03958903253078461,
0.03653223440051079,
0.05790615826845169,
0.014564328826963902,
0.11618709564208984,
0.0033674058504402637,
-0.023991554975509644,
1.0523742663786112e-33,
0.0009348673629574478,
0.019054720178246498,
0.09920188784599304,
0.04113304987549782,
0.08069109171628952,
-0.005338544026017189,
-0.02741909585893154,
-0.013932956382632256,
-0.013918052427470684,
0.02753995545208454,
-0.04415879771113396,
0.007643816061317921,
-0.09225131571292877,
0.11018145829439163,
0.033949919044971466,
-0.017567938193678856,
0.01980789378285408,
0.02175295725464821,
-0.12099434435367584,
0.006523536518216133,
0.017704954370856285,
-0.03326411545276642,
-0.0623113177716732,
-0.0032521048560738564,
0.020565006881952286,
0.10896467417478561,
0.02779228985309601,
-0.0012568538077175617,
-0.007752466015517712,
0.0018078818684443831,
-0.10877609252929688,
0.009296978823840618,
-0.035935595631599426,
0.08550063520669937,
0.02802063338458538,
-0.018612679094076157,
-0.028418375179171562,
-0.0878828689455986,
0.043459851294755936,
-0.0022455628495663404,
-0.062121447175741196,
0.042386360466480255,
0.005167398136109114,
-0.01658468507230282,
-0.03336234763264656,
0.04376981407403946,
0.0022548921406269073,
0.09352744370698929,
-0.06401699781417847,
-0.03867373242974281,
-0.021897215396165848,
0.033712662756443024,
-0.04156579077243805,
-0.07864440977573395,
-0.05627751350402832,
-0.03038439154624939,
-0.005658207926899195,
0.01485699787735939,
0.02660193480551243,
0.05064473673701286,
-0.01393228117376566,
0.06797152757644653,
0.008082224987447262,
0.0214395672082901,
-0.05065781995654106,
0.013287712819874287,
-0.049908384680747986,
0.05991009995341301,
0.03207778185606003,
0.09302656352519989,
-0.05216927453875542,
0.026770133525133133,
-0.010730362497270107,
-0.07734844833612442,
0.07088128477334976,
0.008564939722418785,
0.05257624760270119,
-0.03851943090558052,
-0.012299511581659317,
0.0038828919641673565,
-0.00938876811414957,
-0.029735321179032326,
-0.0035251257941126823,
-0.06704289466142654,
-0.007916027680039406,
-0.04309476912021637,
0.0713939368724823,
-0.03844718262553215,
-0.07670633494853973,
-0.12355004251003265,
-0.03798673301935196,
-0.04751479625701904,
0.028547929599881172,
-0.07440869510173798,
0.017354892566800117,
-3.050031682191573e-33,
-0.06450237333774567,
0.008401281200349331,
0.006345526315271854,
0.07454574108123779,
-0.059250906109809875,
-0.06996607035398483,
0.018183130770921707,
-0.00985683687031269,
0.022997092455625534,
-0.09412046521902084,
-0.02432693913578987,
-0.03314937278628349,
-0.05935026332736015,
0.03145581856369972,
-0.007367902435362339,
-0.03522650897502899,
-0.08601957559585571,
-0.06481073796749115,
-0.06326985359191895,
0.02739672176539898,
0.005575940012931824,
0.11658076196908951,
-0.09039885550737381,
-0.019257577136158943,
-0.09320509433746338,
0.07403465360403061,
-0.016874171793460846,
-0.03580016642808914,
-0.04921650141477585,
0.040679264813661575,
0.0005179962026886642,
-0.011329476721584797,
-0.04963066801428795,
-0.03268749639391899,
-0.02243303693830967,
0.03608599677681923,
0.074446901679039,
-0.076442651450634,
-0.03720980882644653,
0.08485949039459229,
0.04126943275332451,
-0.04919140040874481,
0.016128649935126305,
0.014628255739808083,
0.002571901772171259,
0.10830304771661758,
0.004293771460652351,
-0.02464931085705757,
-0.024161050096154213,
0.030601127073168755,
-0.0607101134955883,
-0.01159656886011362,
-0.017793940380215645,
0.02979854680597782,
-0.022106748074293137,
-0.013926778919994831,
-0.0630166232585907,
0.05121294781565666,
0.03897091746330261,
0.029948560521006584,
-0.08234872668981552,
-0.02585061639547348,
-0.004664420150220394,
-0.0021300248336046934,
-0.029311014339327812,
0.0050376118160784245,
-0.1078931987285614,
-0.0655626729130745,
0.039532262831926346,
0.006672882940620184,
0.038270629942417145,
-0.07247695326805115,
-0.03797926753759384,
0.030859164893627167,
0.010628367774188519,
0.006971490569412708,
0.008129267953336239,
0.025079185143113136,
0.031286850571632385,
0.022285006940364838,
0.08368916809558868,
-0.08760246634483337,
0.09257927536964417,
0.07555852085351944,
-0.043749962002038956,
0.008017540909349918,
0.020260177552700043,
0.07451002299785614,
0.0679759755730629,
0.019155310466885567,
-0.07090431451797485,
0.03549322113394737,
-0.031095826998353004,
0.06231407821178436,
-0.029177727177739143,
-4.125163854951097e-8,
-0.03353193774819374,
0.08393441140651703,
0.01015014573931694,
-0.05714027211070061,
-0.011487089097499847,
-0.04497117921710014,
-0.0243944451212883,
0.07957066595554352,
-0.0330732986330986,
-0.010732736438512802,
0.04690001904964447,
-0.0372946634888649,
-0.034826572984457016,
-0.041231654584407806,
-0.017931301146745682,
0.0992441400885582,
-0.031700845807790756,
-0.010857805609703064,
-0.007004386745393276,
-0.002483353717252612,
-0.014538871124386787,
0.007116643711924553,
-0.04295998439192772,
0.04805387929081917,
-0.0475478433072567,
-0.03960954770445824,
-0.0313103012740612,
0.0949997827410698,
0.0603761225938797,
-0.01852354407310486,
0.0799620971083641,
0.05603829398751259,
0.059366073459386826,
0.029223410412669182,
0.017362341284751892,
-0.011822674423456192,
-0.06693807989358902,
0.05203891918063164,
0.02079383283853531,
-0.003209272865206003,
-0.025891657918691635,
0.08289166539907455,
-0.000430578162195161,
0.03337329626083374,
0.08113003522157669,
0.052183449268341064,
0.000883683969732374,
-0.05516880005598068,
-0.038720663636922836,
0.0007045592647045851,
-0.12471439689397812,
0.00047294923570007086,
-0.05914558470249176,
0.060113806277513504,
0.0177183635532856,
-0.040699321776628494,
-0.025566931813955307,
-0.036010630428791046,
-0.009297512471675873,
0.08000368624925613,
-0.023851139470934868,
0.07466089725494385,
0.008672171272337437,
-0.031003478914499283
] | 0 | 0 | 0 | 0 |
|
MetaDesigner: Advancing Artistic Typography through AI-Driven, User-Centric, and Multilingual WordArt Synthesis | https://openreview.net/forum?id=Mv3GAYJGcW | [
"Jun-Yan He",
"Zhi-Qi Cheng",
"Chenyang Li",
"Jingdong Sun",
"Qi He",
"Wangmeng Xiang",
"Hanyuan Chen",
"Jin-Peng Lan",
"Xianhui Lin",
"kang zhu",
"Bin Luo",
"Yifeng Geng",
"Xuansong Xie",
"Alexander G Hauptmann"
] | Poster | MetaDesigner introduces a transformative framework for artistic typography synthesis, powered by Large Language Models (LLMs) and grounded in a user-centric design paradigm. Its foundation is a multi-agent system comprising the Pipeline, Glyph, and Texture agents, which collectively orchestrate the creation of customizable WordArt, ranging from semantic enhancements to intricate textural elements. A central feedback mechanism leverages insights from both multimodal models and user evaluations, enabling iterative refinement of design parameters. Through this iterative process, MetaDesigner dynamically adjusts hyperparameters to align with user-defined stylistic and thematic preferences, consistently delivering WordArt that excels in visual quality and contextual resonance. Empirical evaluations underscore the system's versatility and effectiveness across diverse WordArt applications, yielding outputs that are both aesthetically compelling and context-sensitive. | MetaDesigner | null | 13,291 | 2406.19859 | [
0.0651380866765976,
-0.0576024055480957,
0.03675057366490364,
-0.02704811468720436,
0.01537191029638052,
-0.015795275568962097,
-0.03920173645019531,
0.01661970652639866,
0.043285269290208817,
-0.04132835566997528,
-0.09176090359687805,
-0.004096431192010641,
0.10073774307966232,
-0.024805480614304543,
0.009501463733613491,
0.009608092717826366,
0.07798295468091965,
0.012201079167425632,
-0.05622353032231331,
-0.027645913884043694,
0.05483928322792053,
0.015962228178977966,
0.03187203034758568,
-0.08191774785518646,
0.04404202103614807,
0.06525534391403198,
-0.06945928186178207,
-0.0031943009234964848,
0.06717635691165924,
-0.0431564524769783,
0.025253454223275185,
0.024961423128843307,
-0.033165499567985535,
0.02250347100198269,
-0.06292611360549927,
0.08584766834974289,
-0.015823347494006157,
0.07887441664934158,
-0.017101973295211792,
0.007976192981004715,
-0.09524450451135635,
-0.05269429460167885,
-0.06774918735027313,
0.017597028985619545,
0.055484481155872345,
-0.017215237021446228,
-0.08473756909370422,
-0.010373832657933235,
-0.024956997483968735,
0.06549627333879471,
-0.12454775720834732,
-0.13554340600967407,
-0.036765627562999725,
-0.018423570320010185,
-0.013878407888114452,
0.05461740121245384,
0.032269544899463654,
-0.020539812743663788,
0.0054334732703864574,
-0.06585584580898285,
-0.01760636828839779,
0.027678608894348145,
0.02306370809674263,
0.018697571009397507,
0.04673197492957115,
0.015663594007492065,
0.028753122314810753,
-0.016342323273420334,
0.010552815161645412,
-0.050857070833444595,
-0.025558700785040855,
-0.01226355042308569,
0.006077969912439585,
0.039485011249780655,
0.00450108852237463,
0.083712637424469,
-0.006440881174057722,
-0.029651625081896782,
0.02595997042953968,
-0.10946246981620789,
0.05798547342419624,
0.036387648433446884,
0.016748711466789246,
0.04644786939024925,
0.021183140575885773,
0.0002473857603035867,
0.03312792256474495,
-0.012794360518455505,
0.030536923557519913,
-0.005424341652542353,
0.0018956635612994432,
-0.02341494709253311,
0.08251357823610306,
-0.04173435643315315,
-0.00723292026668787,
0.03264842554926872,
-0.0048909918405115604,
-0.10280415415763855,
-0.01872854493558407,
0.07297006994485855,
-0.01982780732214451,
-0.04421680048108101,
0.03089677356183529,
-0.020490044727921486,
-0.027040161192417145,
-0.07362674921751022,
0.058193109929561615,
0.03471558541059494,
-0.07118384540081024,
-0.04231984168291092,
0.042397793382406235,
0.005036861170083284,
-0.009758954867720604,
-0.06290669739246368,
0.04787924140691757,
-0.04418310523033142,
-0.007528604939579964,
-0.02972649224102497,
0.11448600888252258,
-0.012243807315826416,
-0.0018701604567468166,
0.020213916897773743,
-0.018782177940011024,
0.02183659002184868,
0.01778537593781948,
-0.024570580571889877,
-0.0284553375095129,
1.876709240853422e-33,
0.07180573046207428,
0.10127508640289307,
0.07208463549613953,
0.06991193443536758,
0.08036007732152939,
-0.05213021859526634,
-0.04634765535593033,
-0.003884135512635112,
-0.058270011097192764,
-0.06655188649892807,
0.06493647396564484,
0.004438349977135658,
-0.0922873243689537,
0.1694633960723877,
0.023163745179772377,
-0.008975150063633919,
-0.018712887540459633,
0.0000770798942539841,
0.0037368277553468943,
-0.022340385243296623,
-0.040020473301410675,
-0.027733320370316505,
0.07794387638568878,
-0.021147536113858223,
-0.10334839671850204,
0.05702144280076027,
0.018207920715212822,
-0.03306381031870842,
-0.03736710548400879,
0.010322765447199345,
-0.12615205347537994,
-0.006104364525526762,
0.11270399391651154,
0.02630280703306198,
-0.08115937560796738,
-0.06213054060935974,
-0.019510453566908836,
-0.04071469604969025,
0.07995083928108215,
0.06727403402328491,
-0.07737435400485992,
0.0150578198954463,
0.021282970905303955,
0.013707966543734074,
-0.07026948034763336,
0.07786252349615097,
0.025905437767505646,
0.016282420605421066,
-0.0029525512363761663,
-0.0005161876906640828,
0.054299626499414444,
0.07310869544744492,
0.03664659708738327,
0.110823854804039,
0.0019872107077389956,
-0.026410173624753952,
0.02442462183535099,
-0.0326293483376503,
0.009438032284379005,
-0.05681047961115837,
0.007873504422605038,
0.09189645946025848,
0.05122845619916916,
0.0702260211110115,
0.057542432099580765,
0.03981554135680199,
0.06692683696746826,
0.03572184219956398,
0.012655124068260193,
0.042016174644231796,
-0.09944914281368256,
0.020191483199596405,
0.015946242958307266,
0.02282167784869671,
-0.052467480301856995,
-0.05775923281908035,
0.005796113051474094,
-0.0854434072971344,
0.029437042772769928,
0.0013265636516734958,
-0.08865027129650116,
0.024985089898109436,
-0.046867333352565765,
-0.04204126447439194,
-0.02046561986207962,
-0.07633817195892334,
-0.0054192873649299145,
-0.07289886474609375,
0.022023268043994904,
-0.035526636987924576,
-0.05226176977157593,
0.00003350974293425679,
0.012951403856277466,
0.010688760317862034,
-0.037539590150117874,
-1.4982715079046563e-33,
-0.025134844705462456,
-0.04080821946263313,
-0.05604211241006851,
0.10634323954582214,
-0.0010438343742862344,
-0.003058746922761202,
-0.036265213042497635,
0.005073786247521639,
0.03574888035655022,
-0.03740014135837555,
0.044036656618118286,
0.01614976115524769,
-0.02766829915344715,
-0.05665028095245361,
-0.04140274226665497,
-0.05167166516184807,
0.07256756722927094,
-0.02293892204761505,
-0.016651298850774765,
0.008874977938830853,
0.06188984215259552,
0.02531077340245247,
-0.05167762190103531,
-0.019381845369935036,
-0.07301332801580429,
0.06939664483070374,
-0.03857020288705826,
-0.038155630230903625,
0.015295467339456081,
0.04046160355210304,
0.07160481810569763,
-0.030444176867604256,
0.050322525203228,
-0.05370365455746651,
-0.00690465746447444,
0.05573343485593796,
-0.035147931426763535,
-0.05858156457543373,
-0.006195629481226206,
0.12149999290704727,
0.01597677357494831,
-0.027539938688278198,
-0.007730149198323488,
-0.015586315654218197,
-0.045627303421497345,
-0.03365129604935646,
-0.11260689795017242,
-0.053798872977495193,
0.06954678893089294,
0.011654997244477272,
0.008813153021037579,
0.00199502962641418,
-0.026142584159970284,
-0.06575804203748703,
-0.0634700134396553,
-0.11464022099971771,
-0.01462243776768446,
-0.012131549417972565,
0.03941641375422478,
0.05353687331080437,
-0.05922410637140274,
-0.026569746434688568,
-0.002354311989620328,
-0.14235103130340576,
-0.024631917476654053,
-0.01231496874243021,
-0.05306074395775795,
-0.027874218299984932,
0.013695420697331429,
0.023342223837971687,
0.046521302312612534,
-0.03474524989724159,
-0.042146168649196625,
0.05072058364748955,
0.0366479717195034,
-0.028819214552640915,
0.059118058532476425,
-0.00352070783264935,
-0.00002814592698996421,
-0.09145424515008926,
-0.005900491960346699,
0.0407283641397953,
0.0010199553798884153,
-0.0025948958937078714,
0.012818770483136177,
0.10697124898433685,
-0.0970299020409584,
0.13783347606658936,
0.020110594108700752,
0.03110482171177864,
0.06656097620725632,
0.06115896627306938,
0.03725060820579529,
0.09518714994192123,
-0.04343273490667343,
-3.8764103038602116e-8,
-0.01635001227259636,
0.011709209531545639,
0.03926628828048706,
0.009921574033796787,
-0.069292813539505,
-0.017621217295527458,
0.04164489731192589,
-0.004935156088322401,
-0.0081040533259511,
-0.06878404319286346,
0.1245397999882698,
-0.03806430846452713,
-0.023498065769672394,
-0.022806840017437935,
0.019904443994164467,
0.04193468764424324,
-0.005260813981294632,
0.1021277979016304,
-0.05222569778561592,
-0.13444271683692932,
0.0034154951572418213,
0.07331618666648865,
-0.05335798114538193,
-0.07890396565198898,
0.028833884745836258,
0.03630656749010086,
-0.044064152985811234,
0.01140914298593998,
-0.05452526733279228,
0.01879286952316761,
0.060136981308460236,
0.07075521349906921,
0.0016930648125708103,
0.027858443558216095,
-0.028184298425912857,
0.01158800907433033,
-0.027690408751368523,
-0.013256862759590149,
-0.012663125991821289,
-0.01949675753712654,
0.0478660948574543,
0.06960630416870117,
-0.024565564468503,
0.04726322740316391,
-0.0019382389727979898,
0.024799229577183723,
0.04544544965028763,
-0.03570013493299484,
-0.004753539338707924,
0.005918398965150118,
-0.029708947986364365,
-0.05943289026618004,
0.022234084084630013,
0.05419418588280678,
0.04234011843800545,
-0.03418901190161705,
-0.007050845772027969,
0.1340474635362625,
0.0957614853978157,
0.02843547612428665,
0.031701914966106415,
0.03086920641362667,
0.014795653522014618,
-0.027531128376722336
] | 0 | 0 | 0 | 0 |
|
Adaptive Rank Allocation: Speeding Up Modern Transformers with RaNA Adapters | https://openreview.net/forum?id=uAtDga3q0r | [
"Roberto Garcia",
"Jerry Weihong Liu",
"Daniel Sorvisto",
"Sabri Eyuboglu"
] | Poster | Large Language Models (LLMs) are computationally intensive, particularly during inference. Neuron-adaptive techniques, which selectively activate neurons in Multi-Layer Perceptron (MLP) layers, offer some speedups but suffer from limitations in modern Transformers. These include reliance on sparse activations, incompatibility with attention layers, and the use of costly neuron masking techniques. To address these issues, we propose the Adaptive Rank Allocation framework and introduce the Rank and Neuron Allocator (RaNA) adapter. RaNA adapters leverage rank adapters, which operate on linear layers by applying both low-rank matrix decompositions and adaptive masking to efficiently allocate compute without depending on activation sparsity. This enables RaNA to be generally applied to MLPs and linear components of attention modules, while eliminating the need for expensive maskers found in neuron-adaptive methods. Notably, when compared to neuron adapters, RaNA improves perplexity by up to 7 points and increases accuracy by up to 8 percentage-points when reducing FLOPs by $\sim$44\% in state-of-the-art Transformer architectures. These results position RaNA as a robust solution for improving inference efficiency in modern Transformer architectures. | Large Language Models, Adaptive compute, Rank adapters, Neuron adapters | null | 13,289 | 2503.18216 | [
-0.04072653129696846,
-0.0838000476360321,
-0.001496007782407105,
-0.004879603162407875,
-0.011875918135046959,
0.08140937983989716,
-0.018588673323392868,
-0.006549191195517778,
0.03438488766551018,
-0.028736375272274017,
0.008439202792942524,
-0.017444150522351265,
0.01240705419331789,
0.03587352856993675,
-0.07473573088645935,
0.03575744479894638,
0.05762878805398941,
0.1141626164317131,
-0.040679819881916046,
-0.07854931801557541,
0.005218680016696453,
-0.022181883454322815,
0.030597100034356117,
-0.029019242152571678,
0.05781110003590584,
-0.018247274681925774,
-0.10392646491527557,
-0.07505311816930771,
0.0002685508516151458,
-0.03836304694414139,
0.022691182792186737,
-0.023861294612288475,
0.040853191167116165,
0.027408035472035408,
-0.11625225096940994,
0.010503235273063183,
-0.10333950072526932,
0.038189951330423355,
0.0010235722875222564,
0.004509574268013239,
0.03634095937013626,
0.0187514778226614,
-0.06896568834781647,
-0.005971906241029501,
0.06284422427415848,
-0.0034087623935192823,
0.06602386385202408,
-0.06718508899211884,
0.0056064557284116745,
-0.07184379547834396,
-0.03352745622396469,
-0.04275061935186386,
-0.014496579766273499,
0.04922672361135483,
-0.0802089050412178,
-0.004104156047105789,
0.01489577628672123,
0.029317917302250862,
-0.013357128947973251,
-0.02269880287349224,
-0.10090494155883789,
-0.02379663661122322,
0.09312738478183746,
-0.03080017864704132,
-0.0008211949025280774,
-0.01530932541936636,
0.08269675076007843,
0.009443082846701145,
-0.04586666822433472,
-0.004153242334723473,
-0.024801094084978104,
0.059193436056375504,
0.007939745672047138,
-0.02466326579451561,
0.026277843862771988,
-0.01417524553835392,
0.042153216898441315,
-0.04912399500608444,
0.11762596666812897,
0.0032295244745910168,
0.016067329794168472,
-0.013165399432182312,
-0.013932903297245502,
-0.008407877758145332,
0.14370198547840118,
0.01489660982042551,
-0.0034219499211758375,
-0.008031297475099564,
-0.0535469613969326,
-0.062288131564855576,
-0.005907837301492691,
-0.09259919077157974,
0.006566743366420269,
-0.008592436090111732,
-0.009752440266311169,
-0.023794570937752724,
-0.01276816613972187,
0.017773455008864403,
-0.03467487916350365,
0.06167113408446312,
-0.010458930395543575,
-0.04142892733216286,
-0.03196415677666664,
-0.001929104095324874,
0.008902186527848244,
-0.008311808109283447,
0.0190056711435318,
0.03488476946949959,
-0.00973825715482235,
-0.11465627700090408,
0.044000294059515,
0.06443417817354202,
0.0002013599150814116,
0.04911369830369949,
0.022793708369135857,
-0.030317047610878944,
0.055385444313287735,
-0.06205059587955475,
0.0846894159913063,
0.0219874270260334,
-0.0024450889322906733,
0.012359727174043655,
0.020561642944812775,
0.05853390321135521,
0.021291807293891907,
0.04300475865602493,
-0.08749336749315262,
4.376807013176904e-33,
-0.009325124323368073,
0.08103542029857635,
-0.009619523771107197,
0.015908008441329002,
0.009297001175582409,
0.006891414523124695,
0.08601181954145432,
0.00456206314265728,
0.037506960332393646,
0.002068002475425601,
-0.06690361350774765,
0.034742746502161026,
-0.032079100608825684,
0.08169612288475037,
0.034569598734378815,
-0.00283384183421731,
-0.003564249724149704,
0.04316650331020355,
-0.02786354348063469,
-0.13685187697410583,
0.09009337425231934,
-0.009527711197733879,
0.00984118226915598,
-0.124042809009552,
0.0017314781434834003,
-0.03973870724439621,
-0.006581325549632311,
-0.10911004990339279,
-0.043226275593042374,
-0.005416077096015215,
-0.06664474308490753,
-0.0007096821209415793,
0.020344488322734833,
-0.005902132485061884,
0.04302491247653961,
-0.08171219378709793,
0.015929684042930603,
-0.04381195455789566,
0.0783415138721466,
-0.050717614591121674,
-0.043055981397628784,
0.05480556935071945,
-0.03447374328970909,
-0.0751420184969902,
-0.05146003141999245,
0.02322510816156864,
0.012776699848473072,
0.06534726172685623,
-0.0327935554087162,
-0.023393820971250534,
0.061718620359897614,
-0.028659384697675705,
-0.03737340867519379,
0.011349309235811234,
0.032557498663663864,
-0.03580565005540848,
0.06386516243219376,
-0.013749302364885807,
0.08276382088661194,
0.09484507888555527,
-0.07996516674757004,
-0.004338902421295643,
-0.08252063393592834,
0.053486406803131104,
0.08888684213161469,
-0.033791955560445786,
0.0015444409800693393,
0.04835404083132744,
0.0708840936422348,
-0.03585885837674141,
-0.014717202633619308,
0.021565832197666168,
0.027915995568037033,
-0.05311289802193642,
-0.053421176970005035,
-0.05396158620715141,
0.013346625491976738,
-0.04099585488438606,
-0.04733220860362053,
-0.015356105752289295,
0.029866967350244522,
0.06172770634293556,
-0.02276710420846939,
-0.07947448641061783,
-0.01938573084771633,
-0.03133204206824303,
0.0995517447590828,
-0.02488548308610916,
0.02294686809182167,
0.003523689229041338,
0.03736870363354683,
-0.02600134164094925,
0.02740423008799553,
-0.10388544946908951,
-0.009248758666217327,
-2.8984733588301323e-33,
-0.03625679016113281,
0.037549879401922226,
-0.07524136453866959,
0.06522940844297409,
0.014871585182845592,
-0.06657195836305618,
0.04963909462094307,
0.01804850436747074,
-0.031581323593854904,
-0.05443010851740837,
-0.0955653116106987,
-0.02313757687807083,
0.05456751957535744,
-0.006734061986207962,
0.1359301656484604,
0.005072806961834431,
-0.10303167253732681,
0.014579017646610737,
0.02603880688548088,
-0.021827787160873413,
-0.05084013566374779,
0.10655532032251358,
-0.04458492621779442,
0.0465058907866478,
-0.04608112573623657,
0.03161964565515518,
-0.08504929393529892,
0.0970931351184845,
0.08386670053005219,
-0.03784947842359543,
-0.0723915696144104,
0.017669901251792908,
-0.020129093900322914,
0.03519221022725105,
-0.014015214517712593,
0.05278332158923149,
-0.030097566545009613,
-0.0004470158601179719,
0.03628288581967354,
0.04316166415810585,
0.03437413275241852,
-0.033503517508506775,
-0.05014127865433693,
0.006032477132976055,
-0.006847888696938753,
0.023871174082159996,
-0.07629651576280594,
-0.030811147764325142,
-0.0561414510011673,
-0.0018494428368285298,
-0.008689744397997856,
-0.012356889434158802,
-0.05011109635233879,
-0.02496189810335636,
-0.012392280623316765,
0.0035031349398195744,
-0.028176603838801384,
-0.011347990483045578,
0.09821130335330963,
-0.07504048943519592,
0.016699928790330887,
-0.051937639713287354,
0.030315522104501724,
-0.08265656977891922,
-0.04166203364729881,
0.04352256655693054,
0.04539411887526512,
-0.011417822912335396,
0.028716538101434708,
0.020024890080094337,
0.034877222031354904,
0.0027653968427330256,
0.036095358431339264,
-0.04606587812304497,
-0.037020180374383926,
0.04914140701293945,
-0.038765132427215576,
0.036598388105630875,
0.013469316996634007,
-0.04943264648318291,
-0.059473928064107895,
0.02441572956740856,
-0.019664036110043526,
0.0005140448920428753,
0.05972895398736,
0.13323193788528442,
0.08734804391860962,
0.045532211661338806,
0.029344098642468452,
-0.041308194398880005,
0.008717834018170834,
0.044416312128305435,
0.04820535331964493,
0.014280924573540688,
-0.005515012424439192,
-4.169013934074428e-8,
-0.04932413622736931,
0.024960262700915337,
-0.051287516951560974,
0.054143551737070084,
0.052477944642305374,
-0.14570340514183044,
0.04357866942882538,
0.06377562135457993,
-0.008854725398123264,
0.021458754315972328,
0.12945011258125305,
-0.07160365581512451,
0.02364119328558445,
-0.01571403071284294,
0.0942976251244545,
0.09441637992858887,
0.01146456878632307,
0.04907101020216942,
-0.024768749251961708,
-0.06967409700155258,
-0.08651947975158691,
0.16137684881687164,
-0.006203135941177607,
0.004220640752464533,
0.025091487914323807,
-0.03213440254330635,
-0.004534472245723009,
0.08071143925189972,
0.08702593296766281,
0.012959525920450687,
0.02437429688870907,
0.0003539356985129416,
0.05138994753360748,
-0.031062232330441475,
0.10240456461906433,
0.11003623902797699,
0.0302441343665123,
0.0004738173447549343,
-0.04225052893161774,
-0.02847667597234249,
0.008201099000871181,
-0.00044961634557694197,
0.000027678299375111237,
0.05903993174433708,
0.024025781080126762,
-0.02268710732460022,
-0.032841265201568604,
-0.15828360617160797,
0.00046726211439818144,
-0.03912004083395004,
0.05534420162439346,
0.004518712870776653,
-0.03449142351746559,
0.03445456922054291,
-0.0020894058980047703,
0.04496931657195091,
0.0018078340217471123,
-0.01797308400273323,
-0.019158391281962395,
0.047892551869153976,
0.023094506934285164,
-0.003422392765060067,
-0.07109472155570984,
0.0313350073993206
] | https://github.com/roberto09/rana | 1 | 0 | 0 | 0 |
EFFICIENT JAILBREAK ATTACK SEQUENCES ON LARGE LANGUAGE MODELS VIA MULTI-ARMED BANDIT-BASED CONTEXT SWITCHING | https://openreview.net/forum?id=jCDF7G3LpF | [
"Aditya Ramesh",
"Shivam Bhardwaj",
"Aditya Saibewar",
"Manohar Kaul"
] | Poster | Content warning: This paper contains examples of harmful language and content.
Recent advances in large language models (LLMs) have made them increasingly vulnerable to jailbreaking attempts, where malicious users manipulate models into generating harmful content. While existing approaches rely on either single-step attacks that trigger immediate safety responses or multi-step methods that inefficiently iterate prompts using other LLMs, we introduce ``Sequence of Context" (SoC) attacks that systematically alter conversational context through strategically crafted context-switching queries (CSQs). We formulate this as a multi-armed bandit (MAB) optimization problem, automatically learning optimal sequences of CSQs that gradually weaken the model's safety boundaries. Our theoretical analysis provides tight bounds on both the expected sequence length until successful jailbreak and the convergence of cumulative rewards. Empirically, our method achieves a 95\% attack success rate, surpassing PAIR by 63.15\%, AutoDAN by 60\%, and ReNeLLM by 50\%. We evaluate our attack across multiple open-source LLMs including Llama and Mistral variants. Our findings highlight critical vulnerabilities in current LLM safeguards and emphasize the need for defenses that consider sequential attack patterns rather than relying solely on static prompt filtering or iterative refinement. | JailBreak, AI Security, LLM Vunlnerability | We propose a novel multi-step attack methodology to automatically generate an optimal sequence of prompts that gradually steers the LLM towards eliciting harmful/unsafe responses using a Multi Armed Bandit Framework. | 13,287 | null | [
-0.08682641386985779,
-0.038828641176223755,
-0.02390880137681961,
0.0003352042695041746,
0.039219822734594345,
0.03711686655879021,
0.11577165126800537,
-0.02202763594686985,
0.04546230658888817,
-0.016507698222994804,
-0.02571265771985054,
-0.07703495770692825,
0.11488761007785797,
-0.04452246055006981,
0.058528803288936615,
0.04439790919423103,
0.061422571539878845,
0.06954557448625565,
-0.04426214471459389,
0.005449134856462479,
-0.039735693484544754,
0.03489060327410698,
0.02775132656097412,
-0.0498172827064991,
0.0022345813922584057,
-0.017216628417372704,
-0.06590566784143448,
-0.027899665758013725,
0.012034745886921883,
0.029132777824997902,
0.0008524327422492206,
0.023185141384601593,
-0.005517361219972372,
-0.0069223977625370026,
-0.028217697516083717,
-0.008581836707890034,
-0.07182513177394867,
-0.05149669200181961,
0.018270418047904968,
-0.06048012897372246,
-0.07021830230951309,
0.05623704940080643,
0.0018822228303179145,
0.033847179263830185,
0.01786061003804207,
-0.08811091631650925,
-0.03503657877445221,
0.05248193070292473,
0.029465198516845703,
-0.03436685726046562,
-0.09461899846792221,
-0.0006484119221568108,
0.03389272093772888,
0.021718522533774376,
-0.037261586636304855,
-0.04098382592201233,
0.007721936795860529,
0.0832744613289833,
0.04502955824136734,
0.058658357709646225,
-0.027019144967198372,
-0.07884290814399719,
-0.06920339167118073,
-0.009373203851282597,
-0.016166090965270996,
-0.036816418170928955,
-0.0034550498239696026,
0.06716970354318619,
0.0006939743761904538,
0.0852627083659172,
-0.02207656018435955,
0.010265949182212353,
-0.03747854381799698,
0.05037452280521393,
0.007574449758976698,
0.07778295874595642,
0.008451768197119236,
-0.03977195918560028,
0.024628140032291412,
-0.07757989317178726,
0.014651092700660229,
-0.036375727504491806,
0.07132066786289215,
-0.02691388688981533,
0.06727512925863266,
0.03944973275065422,
-0.00539005734026432,
-0.0017720912583172321,
0.14024342596530914,
0.012345620431005955,
0.0146033288910985,
-0.021863682195544243,
0.05252368748188019,
0.04556853324174881,
0.06083366274833679,
0.03091566264629364,
-0.036846134811639786,
-0.06612654030323029,
-0.023700008168816566,
0.09995648264884949,
0.048001352697610855,
0.06261809170246124,
0.012418217957019806,
-0.122261181473732,
0.01158718392252922,
-0.004979835823178291,
0.03452683240175247,
0.01912582293152809,
0.012145278975367546,
-0.05431550741195679,
0.018263239413499832,
0.10981716960668564,
0.08177098631858826,
0.026709934696555138,
0.0062405019998550415,
-0.020804986357688904,
0.01567731611430645,
-0.005887623410671949,
0.007119391579180956,
0.12214961647987366,
0.04700278118252754,
0.04316983371973038,
0.025447390973567963,
-0.01844840496778488,
-0.052640631794929504,
-0.02807389199733734,
-0.03008546307682991,
1.1668317839761927e-33,
0.07534532994031906,
-0.0105363167822361,
-0.030096178874373436,
0.03374205157160759,
0.04206003248691559,
-0.024920640513300896,
0.0498955212533474,
0.05183866620063782,
-0.08314871788024902,
0.04362152889370918,
0.024175619706511497,
-0.006961043458431959,
-0.03318670392036438,
0.03178638219833374,
0.12167464196681976,
0.003205301007255912,
-0.03446571156382561,
-0.025547867640852928,
0.03539072349667549,
-0.0853910893201828,
0.07322385907173157,
-0.06174304708838463,
0.017699118703603745,
-0.02366527169942856,
0.0490424744784832,
0.03102262318134308,
-0.00006484485493274406,
-0.026095548644661903,
-0.016266746446490288,
0.046503789722919464,
-0.12731334567070007,
-0.02312607876956463,
-0.04025065898895264,
0.06849910318851471,
0.037332937121391296,
-0.0022446811199188232,
-0.02997772954404354,
-0.020096750929951668,
-0.0318792462348938,
0.0061032455414533615,
-0.06607276201248169,
0.013047311455011368,
0.03539357706904411,
0.04075075313448906,
-0.034133244305849075,
-0.0628572553396225,
-0.08376877754926682,
-0.01363285630941391,
-0.04599054157733917,
-0.078797347843647,
-0.02451421320438385,
0.043375369161367416,
-0.001192515599541366,
0.02219153195619583,
-0.0739717110991478,
0.020596029236912727,
-0.05617918074131012,
0.028454400599002838,
0.009744769893586636,
0.10950319468975067,
0.042086537927389145,
-0.01765981689095497,
-0.04916556179523468,
0.03598954156041145,
0.0053481957875192165,
-0.013425562530755997,
-0.05214216932654381,
0.030742648988962173,
0.007239676546305418,
0.01056357566267252,
-0.00014718706370331347,
-0.05419166386127472,
0.02037276327610016,
-0.019786423072218895,
-0.015627505257725716,
0.004540981259196997,
0.07578198611736298,
0.026933297514915466,
0.03816580772399902,
-0.06400987505912781,
0.04989885911345482,
-0.027519341558218002,
0.044284638017416,
0.0006720106466673315,
-0.07241157442331314,
-0.036193184554576874,
0.003299955977126956,
-0.13007356226444244,
-0.043694011867046356,
0.01532542984932661,
-0.04585694894194603,
-0.003525400534272194,
-0.00952835287898779,
0.004340172279626131,
-0.008905080147087574,
-2.717093845814284e-33,
-0.05375194177031517,
-0.07315365225076675,
-0.010480860248208046,
0.0380621962249279,
-0.01942780427634716,
-0.10358450561761856,
0.024841737002134323,
0.022539613768458366,
-0.03961576521396637,
-0.057132091373205185,
-0.09362713992595673,
0.04571385681629181,
0.11200780421495438,
-0.024528363719582558,
0.0326431542634964,
-0.044861868023872375,
0.061316922307014465,
0.03918235003948212,
-0.042136453092098236,
0.07425536960363388,
0.004717361647635698,
0.058717142790555954,
-0.10627839714288712,
0.00244738906621933,
-0.01346709206700325,
-0.03274986520409584,
-0.04963365197181702,
0.03149528056383133,
0.03248108550906181,
0.06339016556739807,
-0.025459308177232742,
0.013603759929537773,
-0.031077321618795395,
-0.02871198020875454,
-0.08323469012975693,
0.01719842664897442,
0.08976109325885773,
-0.010547801852226257,
-0.03532014042139053,
0.12337539345026016,
0.12787184119224548,
-0.015287630259990692,
-0.044452596455812454,
-0.026895813643932343,
0.023241333663463593,
0.02591862343251705,
-0.07445613294839859,
-0.013138148933649063,
0.008152118884027004,
-0.08844515681266785,
0.04540000110864639,
-0.01401482243090868,
0.005304360296577215,
0.022425219416618347,
-0.02971741370856762,
-0.10733282566070557,
-0.0015396871604025364,
-0.07013053447008133,
0.033957745879888535,
0.0026138974353671074,
-0.06074410676956177,
0.02094458043575287,
-0.008944402448832989,
-0.0407443605363369,
0.08597635477781296,
-0.07037577033042908,
-0.04440221190452576,
0.01974637061357498,
0.025247633457183838,
-0.05564504861831665,
0.04256633296608925,
0.0010159468511119485,
-0.0524556003510952,
-0.0260478463023901,
-0.021396519616246223,
0.020466752350330353,
-0.05174194648861885,
-0.11034820973873138,
-0.007692093960940838,
-0.014055688865482807,
0.05550025776028633,
0.046843040734529495,
0.02142600528895855,
0.031758666038513184,
0.0249179620295763,
0.03903232142329216,
0.0742073580622673,
0.07817117124795914,
-0.0008780379430390894,
0.009373772889375687,
0.007680155802518129,
-0.06810739636421204,
-0.011214627884328365,
0.06010827049612999,
-0.0456802137196064,
-4.2141209632973187e-8,
-0.030833518132567406,
0.03079153411090374,
-0.0024227441754192114,
0.0540352500975132,
0.0968940258026123,
0.10061902552843094,
-0.10871545225381851,
0.009515918791294098,
0.033907074481248856,
-0.022618837654590607,
0.06706816703081131,
0.05914241448044777,
0.01931760460138321,
-0.043989527970552444,
-0.1358635574579239,
0.06206978112459183,
-0.04617803171277046,
-0.051239531487226486,
-0.020261947065591812,
-0.059347398579120636,
0.007384666707366705,
0.07625778764486313,
0.04710301756858826,
-0.019325174391269684,
-0.004272326361387968,
-0.056491706520318985,
-0.028156165033578873,
0.04646741598844528,
0.10346779227256775,
0.04309438541531563,
-0.0792613998055458,
-0.05327680706977844,
-0.05255172401666641,
-0.028509341180324554,
-0.053522299975156784,
0.09540285915136337,
-0.06122966110706329,
-0.08806530386209488,
0.02520616538822651,
-0.036771051585674286,
0.0372747927904129,
-0.017092131078243256,
0.014519469812512398,
0.013960950076580048,
-0.07844651490449905,
-0.010088764131069183,
-0.016323061659932137,
-0.0663352683186531,
0.09771707653999329,
-0.0645536407828331,
0.0674811378121376,
-0.09288904815912247,
-0.0029024279210716486,
0.07579553127288818,
0.12092050909996033,
0.022799769416451454,
-0.0410720594227314,
-0.028302673250436783,
0.04382997751235962,
0.08192954212427139,
0.005745620001107454,
-0.024698084220290184,
-0.017734264954924583,
-0.024816179648041725
] | 0 | 0 | 0 | 0 |
|
A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts | https://openreview.net/forum?id=TrKRpaOk8y | [
"Suyu Ge",
"Xihui Lin",
"Yunan Zhang",
"Jiawei Han",
"Hao Peng"
] | Poster | Training and serving long-context large language models (LLMs) incurs substantial overhead.
To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving.
This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance.
This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension.
LongGen builds on three key insights:
(1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well.
(2) It is essential for the model to have direct access to all tokens.
A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance.
(3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K.
We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales.
During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline.
During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup.
Compared to baselines that apply KV-cache reduction techniques to full-attention long-context LLMs, LongGen achieves substantially stronger performance not only on the Needle-in-a-Haystack retrieval task, but also on more challenging long-context reasoning tasks, including BABILong and RULER. | Long-Context LLM, Efficient LLM, Context Extension, KV Cache Reduction | This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training and inference overhead, but also achieve better long-context performance. | 13,286 | 2410.01485 | [
0.023267308250069618,
-0.011051367036998272,
0.016598384827375412,
-0.0055417087860405445,
0.06900481134653091,
0.004188593942672014,
-0.015899045392870903,
0.012878885492682457,
-0.00227357167750597,
-0.08979948610067368,
-0.04729894921183586,
-0.017010508105158806,
-0.03416650742292404,
-0.002129186876118183,
0.06653688848018646,
-0.022835662588477135,
0.12352213263511658,
0.08848121762275696,
-0.10699966549873352,
-0.03711890056729317,
0.036837268620729446,
-0.021710503846406937,
0.04919969663023949,
-0.006230001337826252,
0.011178944259881973,
0.015577923506498337,
-0.04015415906906128,
-0.07607681304216385,
0.1022031381726265,
0.040443096309900284,
0.07461579889059067,
0.060352735221385956,
0.05801919475197792,
0.0698300376534462,
-0.07134544104337692,
0.051698923110961914,
-0.10476233810186386,
-0.029828837141394615,
-0.024032631888985634,
-0.06967536360025406,
0.027468763291835785,
0.014553088694810867,
-0.04123522713780403,
0.10496886819601059,
0.06215858459472656,
-0.02994927018880844,
0.07820280641317368,
-0.019401123747229576,
-0.024509256705641747,
-0.03153473138809204,
-0.0466722808778286,
-0.03423900902271271,
-0.006362625397741795,
0.04961296543478966,
-0.06600748747587204,
0.01230059377849102,
-0.019032983109354973,
0.04906834289431572,
-0.03702535852789879,
0.0614011324942112,
-0.03788229823112488,
-0.13255664706230164,
-0.04685324430465698,
-0.024690289050340652,
-0.0626290962100029,
-0.022222274914383888,
0.09208626300096512,
0.039966266602277756,
-0.004466879181563854,
-0.021661413833498955,
0.008990622125566006,
0.07878901809453964,
-0.03231560438871384,
0.03739691153168678,
-0.013510268181562424,
0.04087836295366287,
0.016007186844944954,
-0.03511614352464676,
0.09783349186182022,
-0.11185984313488007,
0.04197121784090996,
-0.034601811319589615,
0.08330092579126358,
-0.049747247248888016,
0.10211978107690811,
0.006446158047765493,
0.034730374813079834,
-0.004780868533998728,
-0.03431215137243271,
0.0047432491555809975,
0.0394732840359211,
-0.11456765234470367,
0.05796420946717262,
-0.026070525869727135,
0.05823936313390732,
-0.01318080723285675,
0.005314879585057497,
-0.04653209447860718,
-0.0669536218047142,
0.0704624131321907,
0.030709128826856613,
0.04099120572209358,
-0.022622844204306602,
0.06009165942668915,
-0.03238476440310478,
0.03956753760576248,
-0.032103486359119415,
0.11662465333938599,
-0.05632002651691437,
-0.02894960343837738,
0.03824153542518616,
0.09362060576677322,
-0.004030956886708736,
0.014837827533483505,
0.027566274628043175,
-0.032153964042663574,
-0.024243304505944252,
-0.08354383707046509,
0.06642234325408936,
0.11941136419773102,
-0.04102512076497078,
-0.0179411880671978,
-0.05602099373936653,
0.012318483553826809,
-0.035532016307115555,
-0.056841086596250534,
-0.03976449742913246,
5.1189561062525245e-33,
0.018947314471006393,
0.03299485146999359,
-0.04532701149582863,
-0.0038923234678804874,
0.016948377713561058,
0.040060341358184814,
0.033306486904621124,
0.04200007766485214,
-0.037435464560985565,
0.019556395709514618,
0.04464631527662277,
0.0395733080804348,
-0.02137618698179722,
0.1091211661696434,
0.012441561557352543,
-0.002348077716305852,
-0.026281429454684258,
0.06642496585845947,
-0.007768808864057064,
-0.010497196577489376,
0.0022450233809649944,
0.010226677171885967,
0.040500838309526443,
-0.03043757565319538,
0.03813161700963974,
-0.00849073100835085,
0.0424862876534462,
-0.06242746859788895,
-0.025420209392905235,
0.0254964679479599,
-0.1163787841796875,
-0.039725735783576965,
-0.058153606951236725,
0.0649648979306221,
0.012430384755134583,
-0.04916927218437195,
0.02297055535018444,
-0.05630842223763466,
0.03630516305565834,
-0.04400169104337692,
-0.02052411623299122,
0.05493488162755966,
0.07784314453601837,
-0.0788496881723404,
-0.15753884613513947,
-0.07571141421794891,
0.014047138392925262,
0.03054565005004406,
-0.049089010804891586,
0.04557723179459572,
0.08347849547863007,
-0.012271992862224579,
-0.06639230251312256,
-0.020582107827067375,
-0.030142892152071,
-0.01390159223228693,
0.04809616133570671,
0.0771661326289177,
0.054407376796007156,
0.09909161180257797,
-0.028554856777191162,
0.01730726845562458,
-0.022720566019415855,
0.17090338468551636,
0.01879892311990261,
0.022193415090441704,
-0.07870488613843918,
0.08885014057159424,
0.00993435364216566,
-0.041516948491334915,
-0.02441958151757717,
-0.02441880665719509,
0.03050919994711876,
0.0236191563308239,
0.012696320191025734,
-0.01639586128294468,
0.014169037342071533,
-0.0597873218357563,
-0.060673318803310394,
0.030517401173710823,
0.04623870179057121,
0.03806867077946663,
0.014906301163136959,
-0.014144506305456161,
-0.05943184345960617,
-0.03871011361479759,
-0.013226200826466084,
-0.1323908120393753,
-0.02836987003684044,
-0.0579332672059536,
-0.05184800177812576,
-0.042631328105926514,
0.05644122138619423,
-0.004560382571071386,
-0.07384835928678513,
-3.948598298401346e-33,
0.04417416825890541,
-0.016014402732253075,
-0.02968101017177105,
0.05655162036418915,
0.009942281991243362,
0.0023518060334026814,
0.019834598526358604,
0.044078730046749115,
-0.11483126133680344,
-0.06174594163894653,
-0.03594345599412918,
-0.03321896493434906,
0.06990721821784973,
0.03374088555574417,
0.039279643446207047,
-0.013981216587126255,
0.006344994064420462,
-0.04954511299729347,
-0.06611685454845428,
0.040372058749198914,
0.022418929263949394,
0.022409602999687195,
-0.05686710774898529,
-0.00424010306596756,
-0.02954641543328762,
0.003937467001378536,
-0.07909213751554489,
-0.0022152322344481945,
0.002181921387091279,
0.0013473090948536992,
-0.07200076431035995,
0.023640232160687447,
-0.02515038661658764,
-0.07451583445072174,
-0.009569304063916206,
0.056372858583927155,
0.060300312936306,
0.03331570699810982,
-0.02512683905661106,
0.05165286734700203,
0.13101337850093842,
-0.05559448152780533,
0.039500121027231216,
0.03545315936207771,
-0.010737703181803226,
0.05371939763426781,
-0.10993470996618271,
-0.016359509900212288,
-0.012309038080275059,
0.027972877025604248,
-0.029047055169939995,
0.019588856026530266,
-0.018313147127628326,
0.0020851159933954477,
-0.03550077602267265,
-0.126102015376091,
-0.03928525000810623,
-0.011001774109899998,
0.028769100084900856,
-0.1104116216301918,
0.03934584930539131,
-0.05472864583134651,
-0.008836778812110424,
-0.09147313982248306,
0.06702276319265366,
-0.04958377406001091,
-0.0115679781883955,
-0.032402154058218,
0.020900024101138115,
-0.041272275149822235,
0.005643346346914768,
-0.005047024227678776,
-0.07933614403009415,
-0.005061340052634478,
-0.05205618217587471,
0.0945538580417633,
0.028090447187423706,
-0.15508636832237244,
0.021492836996912956,
-0.004632577300071716,
-0.04157274216413498,
0.0014949137112125754,
-0.016343258321285248,
-0.016114968806505203,
0.0642174631357193,
0.045730870217084885,
0.027886681258678436,
0.06325970590114594,
0.025280175730586052,
-0.000454548338893801,
-0.00668670330196619,
0.07425393909215927,
0.024342045187950134,
0.040106456726789474,
-0.02429530955851078,
-4.1713498433182394e-8,
-0.0128287049010396,
0.01449651550501585,
-0.029623281210660934,
0.07455166429281235,
0.031340111047029495,
-0.047239597886800766,
-0.00028680532705038786,
0.12344721704721451,
0.008965232409536839,
0.038596950471401215,
0.036237627267837524,
-0.04302193224430084,
-0.024416707456111908,
-0.06877398490905762,
0.014909110963344574,
0.04224369302392006,
0.011933845467865467,
-0.019582943990826607,
-0.01758546568453312,
-0.012610827572643757,
0.007723455782979727,
0.06863201409578323,
0.016480281949043274,
-0.0017924755811691284,
0.016745110973715782,
-0.05785766988992691,
0.02702287584543228,
0.09232805669307709,
0.08819030970335007,
-0.06761910766363144,
-0.007060270290821791,
0.04429526627063751,
-0.07794097810983658,
-0.06394446641206741,
-0.007675934582948685,
0.01960347779095173,
0.04809897765517235,
0.040405645966529846,
0.040041714906692505,
0.015600334852933884,
0.005801483988761902,
-0.0194380022585392,
-0.034897059202194214,
0.0446249321103096,
-0.020623620599508286,
0.02588346041738987,
-0.04080556333065033,
-0.08666571229696274,
0.0016478481702506542,
0.002939812373369932,
0.027520904317498207,
-0.016681702807545662,
-0.011436413042247295,
0.045844271779060364,
-0.009054647758603096,
0.04171523079276085,
0.02924460545182228,
0.00582559360191226,
0.03812854364514351,
0.03191613405942917,
-0.006803007330745459,
-0.0017424835823476315,
-0.08632885664701462,
0.012964400462806225
] | 0 | 0 | 0 | 0 |
|
What Secrets Do Your Manifolds Hold? Understanding the Local Geometry of Generative Models | https://openreview.net/forum?id=etif9j1CnG | [
"Ahmed Imtiaz Humayun",
"Ibtihel Amara",
"Cristina Nader Vasconcelos",
"Deepak Ramachandran",
"Candice Schumann",
"Junfeng He",
"Katherine A Heller",
"Golnoosh Farnadi",
"Negar Rostamzadeh",
"Mohammad Havaei"
] | Poster | Deep Generative Models are frequently used to learn continuous representations of complex data distributions using a finite number of samples. For any generative model, including pre-trained foundation models with GAN, Transformer or Diffusion architectures, generation performance can vary significantly based on which part of the learned data manifold is sampled. In this paper we study the post-training local geometry of the learned manifold and its relationship to generation outcomes for models ranging from toy settings to the latent decoder of the near state-of-the-art Stable Diffusion 1.4 Text-to-Image model. Building on the theory of continuous piecewise-linear (CPWL) generators, we characterize the local geometry in terms of three geometric descriptors - scaling ($\psi$), rank ($\nu$), and complexity ($\delta$). We provide quantitative and qualitative evidence showing that for a given latent, the local descriptors are indicative of generation aesthetics, artifacts, diversity, and memorization. Finally we demonstrate that training a reward model using the local geometry allows us to control the log-likelihood of a generated sample under the learned distribution, and improve the qualitative aspects of an image. | Geometry, Diffusion models, VAE, Generative Models, Guidance, Memorization, Out-of-Distribution Detection | We show that the local geometry of generative models is indicative of generation aesthetics, artifacts, diversity, and memorization. | 13,284 | 2408.08307 | [
-0.07169017940759659,
-0.07565652579069138,
0.03158225491642952,
0.03136960789561272,
0.05019206926226616,
0.024058567360043526,
-0.0823865458369255,
-0.053457994014024734,
0.06202845647931099,
-0.042258065193891525,
0.004545788746327162,
-0.031148990616202354,
0.03568531572818756,
0.033491600304841995,
-0.03176512196660042,
-0.014423268847167492,
0.056007225066423416,
0.03449589014053345,
-0.07188460975885391,
-0.06685919314622879,
-0.0283434446901083,
0.06864939630031586,
0.042153775691986084,
0.03695518895983696,
0.07841842621564865,
0.028316916897892952,
0.03456028178334236,
-0.010083232074975967,
0.09183993190526962,
-0.041750501841306686,
0.07337476313114166,
0.05806918814778328,
-0.05171729624271393,
-0.022406738251447678,
-0.04639941826462746,
0.14764173328876495,
-0.06322111189365387,
0.07674995809793472,
0.06759224086999893,
-0.037885237485170364,
0.023111915215849876,
0.03563397005200386,
0.04887039586901665,
0.031673092395067215,
0.08692057430744171,
-0.03790454566478729,
-0.05167720466852188,
-0.0514521598815918,
-0.026087770238518715,
-0.014325438998639584,
-0.01738465391099453,
0.007531294133514166,
-0.07209082692861557,
0.004560097586363554,
0.014514761976897717,
0.031668007373809814,
0.08831233531236649,
0.017834922298789024,
0.024866467341780663,
-0.028327133506536484,
-0.05927155911922455,
-0.09542377293109894,
-0.06459099799394608,
-0.052601322531700134,
0.006565257906913757,
-0.042813390493392944,
0.01507563702762127,
0.031203070655465126,
-0.026498252525925636,
-0.03499990329146385,
-0.02601642906665802,
0.03645775839686394,
-0.06678758561611176,
-0.012389319017529488,
0.04657963663339615,
0.021430132910609245,
-0.0034273522906005383,
-0.07091886550188065,
0.00391359580680728,
-0.05274083837866783,
0.04132438451051712,
0.07104896008968353,
0.11459505558013916,
0.004337257705628872,
0.014760959893465042,
0.031159082427620888,
0.04201630875468254,
0.030665896832942963,
0.03851502761244774,
-0.023287687450647354,
-0.04116327688097954,
0.04877090826630592,
-0.10512475669384003,
-0.00044898182386532426,
0.017856355756521225,
0.018979664891958237,
-0.03926292434334755,
-0.031496211886405945,
0.05145709961652756,
0.08368942886590958,
0.06028271093964577,
0.018554052338004112,
0.011217854917049408,
0.01490701362490654,
0.06049950793385506,
-0.047450143843889236,
0.04566730931401253,
0.07584886997938156,
0.03277454525232315,
-0.08842062205076218,
0.07136079668998718,
0.028586063534021378,
-0.04496241360902786,
0.05188677832484245,
-0.023591650649905205,
-0.05905497446656227,
0.012365452945232391,
-0.022196147590875626,
0.016482552513480186,
0.04110456630587578,
-0.0324365571141243,
0.04706883803009987,
-0.05512876436114311,
-0.005551544483751059,
-0.037128422409296036,
-0.048274245113134384,
-0.03480894863605499,
-9.878512896264796e-34,
0.02931504137814045,
0.01658489927649498,
0.028744589537382126,
0.10527832806110382,
0.08390358090400696,
0.011832932010293007,
0.004643548745661974,
-0.008315588347613811,
-0.004534950014203787,
0.004973491653800011,
-0.08627178519964218,
0.017871160060167313,
-0.09061097353696823,
0.1536099910736084,
0.06396260112524033,
-0.024203244596719742,
-0.04918584227561951,
-0.01424051821231842,
-0.05403434857726097,
0.014598753303289413,
0.04693686589598656,
0.009625035338103771,
-0.00029024231480434537,
-0.03413688391447067,
-0.011368421837687492,
0.03159487247467041,
0.001038397429510951,
-0.013686385937035084,
-0.12465762346982956,
-0.014208915643393993,
-0.09803894907236099,
0.007405733224004507,
0.021311618387699127,
-0.018085718154907227,
0.03709055483341217,
0.002999856835231185,
-0.020082086324691772,
-0.020755581557750702,
-0.049437034875154495,
-0.06485284119844437,
-0.023605581372976303,
0.032130662351846695,
-0.049008335918188095,
-0.024038683623075485,
-0.14238105714321136,
0.0018379975808784366,
0.06919687241315842,
-0.0072241611778736115,
-0.07316683232784271,
-0.06611212342977524,
0.024743404239416122,
0.019816195592284203,
-0.030650760978460312,
-0.10048796981573105,
-0.004373381845653057,
-0.01513348612934351,
0.03507150337100029,
-0.028796706348657608,
0.11620005965232849,
0.044340845197439194,
0.023829633370041847,
0.045088812708854675,
-0.019256887957453728,
0.07144348323345184,
0.01973719522356987,
-0.0028650248423218727,
-0.0006438969285227358,
0.024221789091825485,
0.08673549443483353,
0.02099689468741417,
-0.0024100167211145163,
0.04838535934686661,
-0.10837764292955399,
-0.12684248387813568,
0.10494645684957504,
-0.04581895470619202,
0.011073368601500988,
-0.030561530962586403,
0.002759296912699938,
0.06625440716743469,
-0.02761637046933174,
-0.07195264101028442,
-0.046275410801172256,
-0.09889060258865356,
-0.024101469665765762,
-0.01708037778735161,
0.005860817153006792,
-0.11761269718408585,
0.002610139548778534,
-0.03574911877512932,
-0.05884025618433952,
-0.042345188558101654,
-0.014748423360288143,
-0.0000019015955103895976,
0.004743205849081278,
-6.906858585611246e-34,
-0.02280281111598015,
0.01374309603124857,
0.03477555885910988,
0.032770272344350815,
-0.02302398346364498,
0.006155029404908419,
-0.023319320753216743,
0.08999866247177124,
-0.045290034264326096,
-0.07981056720018387,
-0.054551150649785995,
0.07104728370904922,
-0.020671995356678963,
-0.07865995168685913,
0.002427786821499467,
-0.004689088556915522,
0.0016984755638986826,
-0.02877379208803177,
-0.008924185298383236,
0.050899192690849304,
0.042823296040296555,
0.08898352831602097,
-0.12728479504585266,
0.015256569720804691,
-0.04706137627363205,
0.09021981805562973,
-0.03961610421538353,
0.01593667082488537,
-0.011844879947602749,
0.03342653065919876,
0.025263020768761635,
-0.09137215465307236,
-0.012012657709419727,
0.010478155687451363,
-0.10027077794075012,
0.09188877046108246,
0.07824913412332535,
0.02673589065670967,
-0.06199432536959648,
0.09915785491466522,
-0.007987910881638527,
-0.015237919054925442,
-0.029883598908782005,
0.07389436662197113,
-0.04776740074157715,
-0.03539904206991196,
-0.022373033687472343,
-0.023259788751602173,
-0.019774606451392174,
0.06038898602128029,
-0.0006803307915106416,
0.10379675775766373,
-0.02071395516395569,
0.011542975902557373,
-0.06292162835597992,
-0.10001251101493835,
-0.08222505450248718,
-0.043512001633644104,
0.08596954494714737,
-0.01695273257791996,
-0.06213993951678276,
-0.09324507415294647,
-0.018917975947260857,
0.005496460944414139,
-0.07512009143829346,
-0.06282322108745575,
-0.047590915113687515,
-0.01114876288920641,
-0.004695176612585783,
0.096795953810215,
-0.01987006887793541,
-0.01702830195426941,
0.00021972513059154153,
-0.018856825307011604,
-0.010225857608020306,
-0.08443976938724518,
0.025837430730462074,
0.03716344013810158,
0.01160336472094059,
-0.04086409509181976,
0.055572204291820526,
-0.013747463934123516,
0.032287608832120895,
0.034765999764204025,
0.11905168741941452,
0.019195670261979103,
0.023430634289979935,
-0.03411135450005531,
0.03706855699419975,
0.031173521652817726,
0.003440474160015583,
0.001912964042276144,
-0.030669620260596275,
0.0495435893535614,
0.0017005844274535775,
-5.001296798923249e-8,
-0.07990904152393341,
0.03144855797290802,
-0.007768154609948397,
-0.041560642421245575,
0.01362600177526474,
-0.007495191413909197,
0.0063712215051054955,
0.09929671138525009,
-0.040737513452768326,
0.01372346468269825,
0.09034247696399689,
-0.03212744742631912,
-0.06601177155971527,
-0.06349430233240128,
0.02375779114663601,
0.07374471426010132,
0.07782796770334244,
0.03818792104721069,
0.02847636304795742,
0.006701851263642311,
0.01341948565095663,
0.031053461134433746,
0.03770066797733307,
-0.06173023208975792,
0.01910516433417797,
-0.04211553558707237,
-0.011692062020301819,
0.044520627707242966,
0.03531073406338692,
0.0003657311317510903,
0.04031369462609291,
-0.0033160392194986343,
0.03224615380167961,
-0.0016480883350595832,
0.018378088250756264,
0.034770090132951736,
-0.004071167670190334,
-0.08188167214393616,
0.01843029260635376,
-0.05015740916132927,
0.049839578568935394,
0.03740127757191658,
-0.006858389358967543,
0.017369018867611885,
0.010931679047644138,
0.09406568109989166,
0.05301104485988617,
-0.049398861825466156,
0.011929618194699287,
0.07008778303861618,
0.03421933948993683,
-0.07333555817604065,
-0.0004480204079300165,
0.07921569049358368,
0.10099559277296066,
-0.011604772880673409,
0.044332802295684814,
-0.04641038924455643,
0.06390488147735596,
-0.010206733830273151,
0.016345106065273285,
0.0675249993801117,
-0.02960905432701111,
-0.045586660504341125
] | 0 | 0 | 0 | 0 |
|
L-WISE: Boosting Human Visual Category Learning Through Model-Based Image Selection and Enhancement | https://openreview.net/forum?id=AoIKgHu9Si | [
"Morgan Bruce Talbot",
"Gabriel Kreiman",
"James J. DiCarlo",
"Guy Gaziv"
] | Poster | The currently leading artificial neural network models of the visual ventral stream - which are derived from a combination of performance optimization and robustification methods - have demonstrated a remarkable degree of behavioral alignment with humans on visual categorization tasks. We show that image perturbations generated by these models can enhance the ability of humans to accurately report the ground truth class. Furthermore, we find that the same models can also be used out-of-the-box to predict the proportion of correct human responses to individual images, providing a simple, human-aligned estimator of the relative difficulty of each image. Motivated by these observations, we propose to augment visual learning in humans in a way that improves human categorization accuracy at test time. Our learning augmentation approach consists of (i) selecting images based on their model-estimated recognition difficulty, and (ii) applying image perturbations that aid recognition for novice learners. We find that combining these model-based strategies leads to categorization accuracy gains of 33-72% relative to control subjects without these interventions, on unmodified, randomly selected held-out test images. Beyond the accuracy gain, the training time for the augmented learning group was also shortened by 20-23%, despite both groups completing the same number of training trials. We demonstrate the efficacy of our approach in a fine-grained categorization task with natural images, as well as two tasks in clinically relevant image domains - histology and dermoscopy - where visual learning is notoriously challenging. To the best of our knowledge, our work is the first application of artificial neural networks to increase visual learning performance in humans by enhancing category-specific image features. | Human-aligned models, robust neural networks, visual perception, perceptual learning, medical machine learning | Robust artificial neural networks can be used to design effective training strategies for enhancing visual category learning in human subjects, resulting in significant improvements in human accuracy and efficiency across multiple image domains. | 13,281 | null | [
0.005249351263046265,
-0.11243096739053726,
0.038831304758787155,
0.06523768603801727,
0.040565650910139084,
0.041061777621507645,
-0.01393688004463911,
0.018102748319506645,
-0.03780094161629677,
-0.022251682355999947,
-0.020684286952018738,
-0.10518621653318405,
-0.008542034775018692,
0.06658708304166794,
-0.04926052689552307,
-0.03789835795760155,
0.09235861897468567,
0.1275336742401123,
-0.0579812154173851,
0.06165290251374245,
-0.020839175209403038,
-0.05458957701921463,
0.06302038580179214,
0.0005681805778294802,
0.02082785777747631,
-0.039010223001241684,
0.015599112957715988,
-0.02131432667374611,
-0.010880845598876476,
-0.11494093388319016,
-0.012056376785039902,
-0.07368242740631104,
0.025609102100133896,
-0.017320243641734123,
-0.0703790932893753,
0.028917253017425537,
0.027000105008482933,
0.02489842288196087,
0.01923496648669243,
-0.028094422072172165,
-0.029545148834586143,
-0.03281134366989136,
0.05044287443161011,
-0.03990359976887703,
0.026619384065270424,
0.08348776400089264,
0.020186705514788628,
-0.0514661967754364,
-0.014960378408432007,
-0.03513671085238457,
0.006852210033684969,
0.007139342371374369,
-0.035456616431474686,
0.03342388942837715,
-0.1225338876247406,
0.04506658390164375,
0.06621286273002625,
-0.037030480802059174,
-0.0071966988034546375,
0.04766003414988518,
0.004320806358009577,
-0.029507046565413475,
0.015652816742658615,
0.060412030667066574,
-0.004420363809913397,
-0.019466975703835487,
0.014145289547741413,
-0.030352100729942322,
-0.03153826296329498,
0.008033055812120438,
-0.0064764730632305145,
0.050237543880939484,
-0.060001060366630554,
0.010352061130106449,
0.052626073360443115,
-0.0007068272097967565,
0.046785395592451096,
0.003387200180441141,
0.10915997624397278,
-0.08965560048818588,
-0.0001371225225739181,
0.046200163662433624,
-0.03435927629470825,
0.007220768835395575,
0.11728855967521667,
0.006100935861468315,
-0.03509827330708504,
0.06117154657840729,
-0.022869670763611794,
-0.005809707101434469,
-0.007586490362882614,
-0.02414861135184765,
-0.021757133305072784,
0.0025810939259827137,
0.07488204538822174,
0.01759149506688118,
0.052841391414403915,
-0.047339942306280136,
-0.07937657833099365,
0.028600599616765976,
-0.0019444968784227967,
0.02851223014295101,
-0.04225969687104225,
-0.026623617857694626,
0.03659530356526375,
-0.008392158895730972,
0.08593519032001495,
0.016099272295832634,
0.08280621469020844,
-0.12430688738822937,
0.010438717901706696,
-0.023058613762259483,
-0.03514629229903221,
0.014957312494516373,
-0.0261703934520483,
0.020208215340971947,
-0.03088679350912571,
-0.02760423719882965,
0.044827185571193695,
-0.012547697871923447,
0.018760528415441513,
0.022946514189243317,
0.019674750044941902,
-0.0075986068695783615,
0.12344978004693985,
-0.0742398351430893,
-0.06040598079562187,
1.579177110651908e-33,
0.022662831470370293,
0.015560761094093323,
0.09170880168676376,
0.021781476214528084,
0.04980170354247093,
-0.03141158074140549,
-0.06885366886854172,
0.014145433902740479,
-0.007244005799293518,
0.059529613703489304,
-0.05967952683568001,
-0.009287616237998009,
-0.06525545567274094,
0.14762820303440094,
0.03464625030755997,
0.0449456050992012,
-0.06715507805347443,
0.0440380796790123,
-0.10408186167478561,
0.025197338312864304,
0.021118242293596268,
-0.062289416790008545,
0.0020598843693733215,
-0.08785715699195862,
-0.035795364528894424,
0.022702258080244064,
0.042420804500579834,
-0.019825750961899757,
-0.06703399121761322,
-0.005206869915127754,
-0.025715230032801628,
0.030680570751428604,
0.03747401386499405,
0.05160003900527954,
0.030948098748922348,
-0.04641168564558029,
0.07497034966945648,
-0.00658041425049305,
0.025714432820677757,
0.08196575939655304,
0.017543699592351913,
0.08951824903488159,
-0.0015656144823879004,
-0.011278972961008549,
-0.02786514349281788,
0.02731345407664776,
0.06895208358764648,
0.013358508236706257,
-0.07656967639923096,
-0.04869883880019188,
0.05433151125907898,
0.056386999785900116,
-0.087068110704422,
-0.05480213090777397,
-0.062021248042583466,
0.05393732339143753,
-0.03397269546985626,
0.009931804612278938,
-0.056446123868227005,
-0.005798793397843838,
-0.01621609926223755,
0.04814817011356354,
0.08174451440572739,
0.0878545492887497,
-0.011619488708674908,
-0.038075532764196396,
0.05781441926956177,
0.009501022286713123,
-0.013662912882864475,
0.05477646738290787,
-0.039397429674863815,
0.024806950241327286,
-0.01281765103340149,
-0.11292997002601624,
0.0873093530535698,
-0.0013834956334903836,
0.07303783297538757,
-0.05543682724237442,
-0.010477427393198013,
0.008985067717730999,
0.03313567489385605,
-0.04923202097415924,
-0.021221112459897995,
-0.06300020217895508,
-0.05480153113603592,
-0.04002067819237709,
0.06716711819171906,
-0.06079604849219322,
0.025477604940533638,
0.020159795880317688,
-0.055998384952545166,
0.0229184553027153,
0.011532552540302277,
-0.005464540794491768,
0.012539539486169815,
-1.4951816108009245e-33,
-0.002259453758597374,
-0.019345257431268692,
0.008951907977461815,
-0.028457682579755783,
-0.04832436144351959,
-0.023957503959536552,
0.10081777721643448,
0.028523830696940422,
-0.03662196546792984,
-0.031414587050676346,
-0.02609122171998024,
0.017963720485568047,
-0.020847460255026817,
-0.029234010726213455,
-0.013803975656628609,
-0.04853089153766632,
-0.0564553402364254,
0.0648827776312828,
0.007718742825090885,
-0.0033342004753649235,
0.0175862368196249,
0.0719413161277771,
-0.03167795017361641,
0.00290930294431746,
-0.07751263678073883,
0.04837219789624214,
-0.010808516293764114,
0.08347783237695694,
0.01253814809024334,
-0.045604556798934937,
-0.016756156459450722,
0.008242588490247726,
-0.04968583211302757,
-0.053654056042432785,
-0.020832959562540054,
0.13080613315105438,
0.019253404811024666,
-0.016421673819422722,
-0.053367748856544495,
0.13267256319522858,
0.007534009404480457,
0.050783999264240265,
-0.14715221524238586,
-0.01973954401910305,
0.027489913627505302,
-0.0694454088807106,
0.010600073263049126,
-0.05222107470035553,
-0.007016533054411411,
0.07144299149513245,
-0.05776115134358406,
0.00044469255954027176,
-0.10354118049144745,
0.05915788188576698,
-0.00507300766184926,
-0.005122967064380646,
-0.04302385821938515,
-0.03183290734887123,
0.10095123946666718,
0.07500576972961426,
-0.08527020364999771,
0.04122486710548401,
-0.04454673454165459,
-0.011920341290533543,
-0.05291087180376053,
-0.004643684718757868,
-0.04666878655552864,
0.0290218498557806,
0.08130380511283875,
0.031245023012161255,
0.014532697387039661,
-0.047370266169309616,
0.03554253652691841,
0.008376176469027996,
0.035343561321496964,
-0.020121939480304718,
0.05038889870047569,
0.021229354664683342,
-0.010879839770495892,
-0.07217075675725937,
0.020012537017464638,
-0.107045479118824,
-0.038878269493579865,
0.07933653146028519,
0.04577470198273659,
0.07342954725027084,
0.07145041972398758,
0.05756920203566551,
0.06507368385791779,
-0.01422024704515934,
-0.07739312946796417,
0.024730218574404716,
-0.012688570655882359,
0.023632895201444626,
-0.07878539711236954,
-4.476007475773258e-8,
-0.11246692389249802,
-0.0304137971252203,
0.016226643696427345,
0.013627906329929829,
0.01883712038397789,
0.03304588049650192,
-0.08976592123508453,
0.0694897472858429,
-0.055069081485271454,
-0.07720774412155151,
0.04512329399585724,
0.03660322725772858,
0.0444939099252224,
-0.030205821618437767,
0.03703291341662407,
0.056973136961460114,
0.004754671361297369,
0.09887760877609253,
0.012799079529941082,
0.07801194489002228,
0.0409247949719429,
-0.003079822286963463,
0.042923957109451294,
-0.06580539792776108,
-0.027636079117655754,
-0.05128134414553642,
-0.04399764537811279,
0.056758783757686615,
-0.03273218125104904,
-0.00105942168738693,
0.06733746826648712,
0.039652373641729355,
0.07075505703687668,
-0.05495685711503029,
0.10128968954086304,
0.06213664263486862,
-0.04064081609249115,
-0.05612761899828911,
-0.04811250418424606,
-0.018196264281868935,
-0.003035754431039095,
0.05953723192214966,
0.03721379488706589,
0.026453936472535133,
-0.02526141330599785,
-0.01734791323542595,
0.02035844884812832,
-0.08546272665262222,
0.017384810373187065,
-0.028650138527154922,
-0.01847703382372856,
-0.0008996051619760692,
-0.020422369241714478,
0.09196789562702179,
0.04087848961353302,
-0.055130310356616974,
0.007415941916406155,
-0.09819244593381882,
-0.013989183120429516,
0.08115742355585098,
0.08271706104278564,
0.03948475047945976,
-0.06610670685768127,
-0.055067867040634155
] | 0 | 0 | 0 | 0 |
|
Predictive Uncertainty Quantification for Bird's Eye View Segmentation: A Benchmark and Novel Loss Function | https://openreview.net/forum?id=k3y0oyK7sn | [
"Linlin Yu",
"Bowen Yang",
"Tianhao Wang",
"Kangshuo Li",
"Feng Chen"
] | Poster | The fusion of raw sensor data to create a Bird's Eye View (BEV) representation is critical for autonomous vehicle planning and control. Despite the growing interest in using deep learning models for BEV semantic segmentation, anticipating segmentation errors and enhancing the explainability of these models remain underexplored. This paper introduces a comprehensive benchmark for predictive uncertainty quantification in BEV segmentation, evaluating multiple uncertainty quantification methods across three popular datasets with three representative network architectures. Our study focuses on the effectiveness of quantified uncertainty in detecting misclassified and out-of-distribution (OOD) pixels while also improving model calibration. Through empirical analysis, we uncover challenges in existing uncertainty quantification methods and demonstrate the potential of evidential deep learning techniques, which capture both aleatoric and epistemic uncertainty. To address these challenges, we propose a novel loss function, Uncertainty-Focal-Cross-Entropy (UFCE), specifically designed for highly imbalanced data, along with a simple uncertainty-scaling regularization term that improves both uncertainty quantification and model calibration for BEV segmentation. | Uncertainty Quantification, Evidential Deep Learning, Bird's Eye View (BEV) Segmentation | null | 13,279 | 2405.20986 | [
0.03899117559194565,
-0.07681111991405487,
0.06489686667919159,
0.050729140639305115,
0.12332813441753387,
-0.009842447005212307,
0.04506941884756088,
-0.09327062219381332,
0.04681851714849472,
-0.06523741036653519,
-0.06128182262182236,
-0.13217926025390625,
0.0012434530071914196,
0.06612731516361237,
-0.09761474281549454,
-0.042341265827417374,
0.03053992986679077,
0.07920282334089279,
-0.12866201996803284,
0.025124697014689445,
0.05416424572467804,
0.0401221364736557,
0.012824848294258118,
0.06623638421297073,
0.04106209799647331,
0.02815568633377552,
0.006220123264938593,
-0.0324769988656044,
-0.023976709693670273,
-0.023636367172002792,
0.03931684046983719,
-0.018519235774874687,
-0.024016248062253,
0.021398287266492844,
-0.004766147118061781,
0.016696259379386902,
-0.011789177544414997,
-0.0026030424050986767,
0.050735946744680405,
-0.05651462450623512,
0.012639128603041172,
-0.02931322157382965,
-0.011661040596663952,
0.035370878875255585,
0.059986662119627,
0.08132924139499664,
0.0202445387840271,
-0.049000371247529984,
0.07021591812372208,
-0.009958652779459953,
-0.008715508505702019,
-0.049345482140779495,
-0.01533852331340313,
-0.017955971881747246,
-0.022301042452454567,
0.07147803157567978,
0.007696359418332577,
-0.0018360777758061886,
0.015914777293801308,
0.00044224673183634877,
0.017480604350566864,
-0.0524795837700367,
0.029089344665408134,
0.04152107238769531,
-0.06879321485757828,
-0.006580226588994265,
-0.02952369675040245,
-0.03323199599981308,
0.0036818194203078747,
-0.03596498817205429,
0.0075525701977312565,
0.08158247917890549,
-0.002141515724360943,
0.02685016207396984,
0.0052123116329312325,
0.04525553062558174,
0.09792384505271912,
-0.047063376754522324,
0.049295730888843536,
-0.05033141002058983,
0.03068813681602478,
0.036527298390865326,
0.029610563069581985,
0.026061801239848137,
0.18305449187755585,
-0.008685516193509102,
-0.06888281553983688,
0.03736591711640358,
-0.0324629507958889,
-0.01955438405275345,
-0.10118506848812103,
-0.061916641891002655,
0.004063678905367851,
0.01893555000424385,
0.06251568347215652,
0.028668144717812538,
-0.07111460715532303,
-0.1586066484451294,
0.026284288614988327,
0.04755150154232979,
-0.041873808950185776,
-0.02264752984046936,
0.03254332393407822,
-0.00788097269833088,
0.026468243449926376,
-0.0009526032372377813,
0.05405399203300476,
0.01347291562706232,
0.0655876100063324,
0.0026509277522563934,
0.0826527327299118,
-0.022768700495362282,
0.04863537475466728,
0.0048471842892467976,
0.02143278531730175,
-0.01761723682284355,
-0.05043577775359154,
0.0064580366015434265,
0.014507966116070747,
-0.003749522380530834,
-0.09332747757434845,
-0.013492724858224392,
0.03452404960989952,
0.05223557725548744,
0.08019743859767914,
-0.023058317601680756,
-0.00477376626804471,
2.2084923377020203e-33,
0.005110498517751694,
-0.024501413106918335,
0.05791757255792618,
0.0029590632766485214,
0.06091882660984993,
0.025026623159646988,
-0.04627358168363571,
0.0003052637039218098,
-0.09503728896379471,
0.023160479962825775,
-0.022863978520035744,
-0.029034076258540154,
-0.09717193245887756,
0.02547990158200264,
0.11582821607589722,
0.02370932698249817,
-0.05021192133426666,
-0.00822502002120018,
-0.05495011806488037,
-0.052544400095939636,
-0.06657977402210236,
-0.07311351597309113,
-0.04481261968612671,
0.0332336500287056,
0.05020701885223389,
0.008962237276136875,
0.01917947456240654,
0.01579909212887287,
0.0176737941801548,
-0.00004236568929627538,
-0.07737228274345398,
0.01842155121266842,
0.05770627409219742,
0.014240797609090805,
-0.007986590266227722,
-0.05018499866127968,
-0.04023722931742668,
0.01065844390541315,
-0.0180436372756958,
0.0005929418839514256,
0.05456077679991722,
0.04489913955330849,
0.013437829911708832,
-0.024612637236714363,
-0.03335447236895561,
-0.010772272944450378,
0.05333236977458,
0.016441870480775833,
-0.09940522164106369,
-0.0027582491748034954,
0.07010409981012344,
-0.03671335056424141,
-0.006436381954699755,
-0.06192786991596222,
-0.05602947622537613,
-0.039099227637052536,
0.06502968817949295,
0.030538665130734444,
-0.0133031802251935,
-0.04058554768562317,
-0.0733557641506195,
-0.009201398119330406,
-0.01579573191702366,
0.050139009952545166,
0.010709942318499088,
0.015230459161102772,
0.020765826106071472,
0.08274143934249878,
-0.06880870461463928,
0.037248607724905014,
-0.04070565477013588,
0.0073823630809783936,
0.034208111464977264,
-0.08066336065530777,
0.020973481237888336,
0.02817751094698906,
-0.02778012678027153,
0.0348728708922863,
0.03548678755760193,
0.04522089660167694,
-0.01951492950320244,
0.0825105607509613,
-0.032530978322029114,
0.03744976967573166,
-0.02677302621304989,
0.019397059455513954,
0.05056693032383919,
-0.02901030331850052,
-0.03317854553461075,
0.07736470550298691,
-0.020035358145833015,
-0.03915948048233986,
-0.022660357877612114,
-0.03205132484436035,
-0.04575135186314583,
-2.679966407756539e-33,
-0.07306443899869919,
0.07015394419431686,
0.006180993746966124,
0.03796558082103729,
-0.07272099703550339,
-0.06005723774433136,
0.011781283654272556,
0.04527480527758598,
0.05639096722006798,
-0.12881547212600708,
-0.054297495633363724,
-0.017812425270676613,
0.051087651401758194,
0.008188402280211449,
-0.016511503607034683,
-0.00449036993086338,
-0.0510772168636322,
0.01966535672545433,
-0.04430990666151047,
0.004871295765042305,
0.06634732335805893,
0.04795493558049202,
-0.07157976925373077,
-0.0026526811998337507,
-0.010526937432587147,
0.04415004700422287,
-0.021073859184980392,
0.02713611349463463,
-0.010133967734873295,
-0.05781245976686478,
-0.018503030762076378,
-0.05189032107591629,
-0.029079532250761986,
-0.04438750818371773,
-0.019060634076595306,
0.03932211175560951,
0.013755650259554386,
-0.05926361307501793,
-0.0830710157752037,
0.036750227212905884,
0.017632102593779564,
-0.06387736648321152,
-0.09218151122331619,
-0.044000737369060516,
-0.0485423244535923,
-0.05684048309922218,
0.08408615738153458,
0.021251147612929344,
-0.03438976779580116,
0.0888507142663002,
-0.08045091480016708,
0.015275641344487667,
-0.10221734642982483,
0.13722266256809235,
-0.07843728363513947,
-0.007817134261131287,
-0.04742230847477913,
0.033014677464962006,
0.07689839601516724,
0.016841726377606392,
-0.06030780449509621,
-0.033766504377126694,
-0.02986595593392849,
0.025730453431606293,
-0.02779672108590603,
-0.029495159164071083,
-0.0797300934791565,
0.01274862326681614,
0.08801195025444031,
0.011342325247824192,
0.03826992213726044,
-0.032971933484077454,
0.015343138948082924,
0.03237605467438698,
-0.051293279975652695,
-0.04984040930867195,
0.041534505784511566,
-0.0001171214462374337,
0.07069319486618042,
0.04253171756863594,
0.10466207563877106,
-0.03947402164340019,
0.08960355818271637,
0.14647267758846283,
0.07364631444215775,
0.06278648972511292,
0.031982168555259705,
0.021364036947488785,
0.060868170112371445,
0.04798966646194458,
-0.06852719187736511,
-0.016035744920372963,
0.0021747013088315725,
0.03930916264653206,
-0.05277033522725105,
-3.9332242351974855e-8,
-0.1119656190276146,
0.1036204919219017,
0.044736649841070175,
-0.06841515004634857,
0.05466924607753754,
-0.022457752376794815,
-0.004849028307944536,
0.04849408194422722,
-0.05349599942564964,
0.005286079365760088,
0.054678719490766525,
-0.0512998141348362,
-0.07924997806549072,
-0.0453047901391983,
-0.042424049228429794,
0.0736512541770935,
0.04916417598724365,
0.06714937090873718,
0.027191096916794777,
-0.007993585430085659,
-0.0465703085064888,
0.019793828949332237,
-0.0038171231281012297,
-0.03286832198500633,
0.039980895817279816,
-0.11488969624042511,
-0.06703125685453415,
0.1025211438536644,
0.010759865865111351,
-0.022684356197714806,
-0.027499357238411903,
0.011976443231105804,
0.10571406781673431,
0.021029360592365265,
0.006318052764981985,
0.06609310954809189,
-0.01875513419508934,
0.015784988179802895,
0.04680721461772919,
-0.03010781668126583,
0.011732895858585835,
0.02847851999104023,
-0.007831557653844357,
0.027312327176332474,
-0.023086857050657272,
0.06996503472328186,
0.06449360400438309,
-0.010092245414853096,
0.03172650560736656,
-0.04151483252644539,
-0.013444367796182632,
0.007590757217258215,
0.003419772954657674,
0.12185174971818924,
0.0312824472784996,
-0.021190157160162926,
0.01084812544286251,
-0.08391333371400833,
-0.009017234668135643,
0.03879006206989288,
0.012471631169319153,
0.041563618928194046,
-0.11053595691919327,
-0.04208281636238098
] | 0 | 0 | 0 | 0 |
|
DiscoveryBench: Towards Data-Driven Discovery with Large Language Models | https://openreview.net/forum?id=vyflgpwfJW | [
"Bodhisattwa Prasad Majumder",
"Harshit Surana",
"Dhruv Agarwal",
"Bhavana Dalvi Mishra",
"Abhijeetsingh Meena",
"Aryan Prakhar",
"Tirth Vora",
"Tushar Khot",
"Ashish Sabharwal",
"Peter Clark"
] | Poster | Can the rapid advances in code generation, function calling, and data analysis using large language models (LLMs) help automate the search and verification of hypotheses purely from a set of provided datasets? To evaluate this question, we present DiscoveryBench, the first comprehensive benchmark that formalizes the multi-step process of data-driven discovery. The benchmark is designed to systematically assess current model capabilities in discovery tasks and provide a useful resource for improving them. Our benchmark contains 264 tasks collected across 6 diverse domains, such as sociology and engineering, by manually deriving discovery workflows from published papers to approximate the real-world challenges faced by researchers, where each task is defined by a dataset, its metadata, and a discovery goal in natural language. We additionally provide 903 synthetic tasks to conduct controlled evaluations on data-driven workflows that are not covered in the manually collected split. Furthermore, our structured formalism of data-driven discovery enables a facet-based evaluation that provides useful insights into different failure modes. We evaluate several popular LLM-based reasoning frameworks using both open and closed LLMs as baselines on DiscoveryBench and find that even the best system scores only 25%. Our benchmark, thus, illustrates the challenges in autonomous data-driven discovery and serves as a valuable resource for the community to make progress. | scientific discovery, data-driven discovery, data analysis, large language models, hypothesis generation, hypothesis verification | We present DiscoveryBench, the first comprehensive benchmark that formalizes the multi-step process of data-driven discovery designed to systematically assess current LLM capabilities in discovery tasks. | 13,276 | 2407.01725 | [
-0.11620894819498062,
-0.030362380668520927,
-0.019548045471310616,
0.06695565581321716,
0.08225441724061966,
-0.06683792918920517,
-0.024977494031190872,
-0.019419332966208458,
-0.024838263168931007,
0.024971328675746918,
-0.05809631943702698,
-0.12848089635372162,
0.015732455998659134,
0.03486604988574982,
0.035913676023483276,
0.08546348661184311,
0.022535812109708786,
-0.0072080232203006744,
-0.02382642589509487,
-0.03065682388842106,
-0.029692184180021286,
0.06994540244340897,
0.035784684121608734,
0.014530053362250328,
0.01278737187385559,
0.04106737673282623,
0.022131426259875298,
-0.037540167570114136,
0.029438432306051254,
0.07184042781591415,
0.016821276396512985,
0.07694140076637268,
-0.013413461856544018,
0.038749828934669495,
0.006205110810697079,
0.08739393204450607,
0.05539310351014137,
0.026360802352428436,
-0.0282913688570261,
-0.010570027865469456,
-0.05198899284005165,
-0.020417502149939537,
0.018299905583262444,
-0.0062301442958414555,
0.032003141939640045,
-0.015424632467329502,
-0.030559903010725975,
-0.006997088901698589,
-0.05791902169585228,
-0.030535012483596802,
-0.07715649157762527,
-0.07813237607479095,
0.04895661398768425,
0.050702717155218124,
0.01840103231370449,
0.027058234438300133,
0.05102597549557686,
-0.02524367719888687,
-0.002218664390966296,
-0.05296860635280609,
-0.03688882291316986,
-0.12574909627437592,
-0.054840054363012314,
-0.02728481963276863,
-0.016446156427264214,
-0.006865769624710083,
-0.07219172269105911,
0.03621871769428253,
-0.0014918590895831585,
0.010909564793109894,
-0.07480057328939438,
0.07995925098657608,
-0.024008367210626602,
0.046645086258649826,
0.004526897333562374,
0.003588749561458826,
0.04486282169818878,
-0.04933954402804375,
0.07033128291368484,
-0.11231299489736557,
-0.03306332975625992,
0.011156871914863586,
-0.04622941091656685,
0.09905511885881424,
0.06179535761475563,
-0.004856455605477095,
0.06468244642019272,
0.07658593356609344,
0.03679364174604416,
0.10894579440355301,
-0.054062362760305405,
-0.045655135065317154,
0.059963490813970566,
-0.005571259651333094,
0.02892935648560524,
0.04332052543759346,
0.0017460924573242664,
-0.08750230073928833,
0.015949822962284088,
0.03150717914104462,
-0.04412826895713806,
0.13738034665584564,
0.05210000276565552,
-0.12417367845773697,
-0.04455950856208801,
-0.011760875582695007,
-0.01543727982789278,
0.04235420376062393,
0.01617695949971676,
-0.06098899245262146,
0.008753555826842785,
0.06409388035535812,
0.10403253883123398,
-0.015162763185799122,
0.01906498521566391,
-0.030661476776003838,
-0.03082050383090973,
-0.03906915709376335,
-0.010941958986222744,
0.0497707724571228,
0.01696384698152542,
0.08538325130939484,
0.028589976951479912,
-0.026560820639133453,
0.04397060349583626,
0.021651800721883774,
-0.06069522351026535,
1.6129077373777612e-33,
0.10890164971351624,
0.015863342210650444,
0.025174777954816818,
0.13947820663452148,
0.04549830034375191,
-0.07266881316900253,
0.007877936586737633,
0.05262334644794464,
-0.1084069162607193,
-0.010084343142807484,
-0.0008104122243821621,
0.07275794446468353,
-0.07419342547655106,
0.01037562545388937,
0.016623560339212418,
0.03379063308238983,
-0.044045589864254,
0.05236872285604477,
-0.09372630715370178,
0.0068793585523962975,
0.08089908212423325,
-0.06721615046262741,
-0.00025799989816732705,
-0.04237598925828934,
0.06710366159677505,
0.010244496166706085,
0.042121391743421555,
-0.039721887558698654,
0.013606846332550049,
0.028607739135622978,
-0.12018905580043793,
-0.04361525923013687,
0.06371428817510605,
0.10971569269895554,
-0.015021564438939095,
0.02234618365764618,
-0.06635268032550812,
-0.08113246411085129,
0.038771893829107285,
0.024643516167998314,
-0.010397889651358128,
0.035649754106998444,
-0.007533782161772251,
0.029175424948334694,
-0.03967418894171715,
-0.06932052224874496,
-0.04438118264079094,
0.027703039348125458,
0.07666022330522537,
-0.0021014874801039696,
0.09988070279359818,
-0.00183109228964895,
0.031886644661426544,
0.0069901649840176105,
-0.01561881136149168,
0.017663443461060524,
-0.07163643091917038,
0.0018827158492058516,
0.054027218371629715,
0.07388301193714142,
-0.024022679775953293,
0.03113553114235401,
-0.02911349944770336,
0.09252738952636719,
0.008072208613157272,
0.028466101735830307,
-0.04366469383239746,
0.06872720271348953,
0.04956921190023422,
-0.0036842720583081245,
0.006614602170884609,
0.033220648765563965,
0.02198491431772709,
-0.007108157500624657,
0.09957679361104965,
-0.004748772829771042,
0.07670596241950989,
-0.07586615532636642,
-0.00694769574329257,
-0.022893832996487617,
0.016694195568561554,
-0.0014393995516002178,
0.018623117357492447,
-0.03834303095936775,
0.014872352592647076,
-0.05171280726790428,
-0.0000016634453459118959,
-0.041750792413949966,
-0.08389519900083542,
-0.03669508919119835,
-0.09076672047376633,
-0.03020278364419937,
-0.04230588674545288,
-0.028645996004343033,
0.0481811948120594,
-1.8969466617997738e-33,
-0.02459096536040306,
-0.06513869017362595,
-0.003773269709199667,
0.03510962426662445,
-0.041334155946969986,
-0.09943139553070068,
-0.04896942898631096,
-0.010916512459516525,
0.027096783742308617,
-0.05791967362165451,
-0.0009374979417771101,
-0.10723177343606949,
-0.0003071730025112629,
-0.022688552737236023,
0.04582471400499344,
-0.03990045562386513,
-0.032609935849905014,
-0.0610010139644146,
-0.008998297154903412,
0.15414094924926758,
-0.05114789307117462,
0.0558859221637249,
-0.15537060797214508,
0.017590835690498352,
0.003582635661587119,
0.0397254154086113,
-0.005399791058152914,
0.005578984040766954,
0.026970479637384415,
0.0239737369120121,
-0.04729489982128143,
0.013834438286721706,
-0.02505507320165634,
-0.06466725468635559,
-0.053543172776699066,
0.009267427958548069,
0.031038733199238777,
-0.026730621233582497,
-0.006837303750216961,
0.035964835435152054,
0.03527257964015007,
0.02838166616857052,
-0.0902119129896164,
0.03413112089037895,
0.02923053689301014,
0.028458181768655777,
-0.09117395430803299,
0.03633563965559006,
-0.003930646926164627,
-0.088376984000206,
0.000154076173203066,
-0.03918733820319176,
0.02240564674139023,
0.005850519053637981,
0.028422558680176735,
-0.05904010683298111,
-0.03539023548364639,
-0.054106611758470535,
-0.06992960721254349,
0.022600242868065834,
-0.08662696927785873,
0.028194613754749298,
0.05925532057881355,
0.0452895313501358,
0.0060254354029893875,
-0.0856250748038292,
-0.054388318210840225,
-0.020429328083992004,
-0.026932410895824432,
-0.009556530974805355,
0.04372161254286766,
-0.0021045960020273924,
-0.012714042328298092,
-0.0247416403144598,
0.03652035444974899,
-0.019014956429600716,
0.011868110857903957,
-0.06617989391088486,
0.003189232898876071,
0.049505580216646194,
0.04184199869632721,
0.021278919652104378,
0.0671234279870987,
-0.009081420488655567,
0.0056440322659909725,
0.04500609263777733,
0.01755734719336033,
0.12282441556453705,
-0.04916432499885559,
0.03865675628185272,
-0.09796655923128128,
-0.04827134311199188,
-0.0646752119064331,
0.05177254229784012,
0.0023121065460145473,
-4.165066513905913e-8,
-0.06792975217103958,
0.055705517530441284,
0.000770751794334501,
-0.00001781560968083795,
0.008716107346117496,
0.033059410750865936,
-0.0594039112329483,
0.09019847214221954,
0.008086293004453182,
-0.041834644973278046,
0.04045484960079193,
0.027658967301249504,
0.022289125248789787,
-0.028700798749923706,
-0.03452480211853981,
0.06362764537334442,
0.013824868015944958,
-0.051684170961380005,
-0.05725615471601486,
-0.005586429964751005,
0.06281014531850815,
0.0244679544121027,
0.01611386239528656,
0.03735736757516861,
0.04845883697271347,
-0.1020316407084465,
-0.01585383340716362,
0.07553017139434814,
0.036371637135744095,
-0.08061470836400986,
-0.03245633840560913,
-0.0260007381439209,
0.0193839892745018,
-0.030523568391799927,
0.08573810011148453,
0.054180845618247986,
0.0020807194523513317,
-0.0352516770362854,
0.05049227923154831,
-0.04734846577048302,
0.07816527038812637,
0.07234452664852142,
-0.03820132091641426,
0.0014167058980092406,
-0.02941187657415867,
-0.022341765463352203,
-0.05264637991786003,
0.04451551288366318,
-0.011272618547081947,
-0.0333138033747673,
-0.020129837095737457,
-0.0393039807677269,
0.0017601746367290616,
0.04406922683119774,
0.018034201115369797,
0.11012787371873856,
-0.08409509062767029,
0.03469720855355263,
0.039716850966215134,
-0.025276077911257744,
0.012145177461206913,
-0.03502447158098221,
0.03893310949206352,
0.07431182265281677
] | https://github.com/allenai/discoverybench | 76 | 0 | 0 | 0 |
Optimizing (L0,L1)-Smooth Functions by Gradient Methods | https://openreview.net/forum?id=GQ1Tc3vHbt | [
"Daniil Vankov",
"Anton Rodomanov",
"Angelia Nedich",
"Lalitha Sankar",
"Sebastian U Stich"
] | Poster | We study gradient methods for optimizing $(L_0, L_1)$-smooth functions, a
class that generalizes Lipschitz-smooth functions and has gained attention for
its relevance in machine learning.
We provide new insights into the structure of this function class and develop
a principled framework for analyzing optimization methods in this setting.
While our convergence rate estimates recover existing results for minimizing
the gradient norm in nonconvex problems, our approach significantly improves
the best-known complexity bounds for convex objectives.
Moreover, we show that the gradient method with Polyak stepsizes and the
normalized gradient method achieve nearly the same complexity guarantees as
methods that rely on explicit knowledge of $(L_0, L_1)$.
Finally, we demonstrate that a carefully designed accelerated gradient
method can be applied to $(L_0, L_1)$-smooth functions, further improving all
previous results. | $(L_0, L_1)$-smoothness, gradient methods, convex optimization, worst-case complexity bounds, acceleration, Polyak stepsizes, nonconvex optimization | null | 13,268 | null | [
-0.07668408006429672,
-0.0773656889796257,
-0.014793853275477886,
-0.025300931185483932,
0.03377722576260567,
-0.06367454677820206,
-0.04396383464336395,
-0.08175686746835709,
-0.07589107006788254,
0.024324333295226097,
-0.06730224937200546,
0.03986146301031113,
0.01268128864467144,
-0.005282669328153133,
-0.041401758790016174,
0.04238029941916466,
0.002629271475598216,
0.04908658191561699,
-0.04129591956734657,
-0.02585844323039055,
0.003368113888427615,
0.056424349546432495,
-0.026715943589806557,
-0.02578585408627987,
0.0372396744787693,
-0.048881519585847855,
-0.026962224394083023,
0.008008313365280628,
0.023010753095149994,
0.020861415192484856,
0.02355906553566456,
-0.05239003151655197,
0.0029077150393277407,
0.003197144018486142,
-0.03760029748082161,
0.07122619450092316,
-0.008629688993096352,
0.08034501224756241,
-0.008580484427511692,
0.028034888207912445,
-0.036006879061460495,
0.03180031478404999,
0.04033029451966286,
-0.004734003450721502,
0.10070394724607468,
0.0047349752858281136,
-0.01494421437382698,
-0.10463669151067734,
0.03326236084103584,
-0.0005671015242114663,
-0.010551080107688904,
-0.03270644694566727,
-0.0859851986169815,
-0.0055454084649682045,
-0.0728483498096466,
0.02765759639441967,
0.012063843198120594,
0.02550981007516384,
0.0161910317838192,
0.04602915048599243,
-0.06217283010482788,
-0.16699758172035217,
0.028645014390349388,
0.06411586701869965,
-0.030436353757977486,
-0.05925494804978371,
0.0705500915646553,
0.008564820513129234,
0.0023600661661475897,
0.10407651215791702,
-0.04863011837005615,
0.0672791451215744,
-0.08162836730480194,
0.010601929388940334,
-0.007891380228102207,
0.01416790671646595,
0.08895368129014969,
0.07294727861881256,
-0.028333110734820366,
-0.10318584740161896,
0.02248879335820675,
-0.003212451469153166,
0.011217358522117138,
0.02283410355448723,
0.07143814861774445,
0.008413692004978657,
-0.07730740308761597,
-0.02105269953608513,
0.05970467999577522,
0.03679240122437477,
-0.05673149228096008,
0.06796775758266449,
-0.10331122577190399,
-0.0026021571829915047,
-0.03346405550837517,
-0.0220525823533535,
0.03974343091249466,
-0.022203493863344193,
-0.02108045667409897,
0.12969979643821716,
-0.08183687180280685,
0.045159898698329926,
-0.0025516257155686617,
-0.03174954280257225,
0.07775722444057465,
-0.05820392444729805,
0.04890170320868492,
0.0037113039288669825,
0.00975784845650196,
-0.07200318574905396,
0.03619333729147911,
0.04410378262400627,
0.031251516193151474,
0.003930774517357349,
0.03141633793711662,
-0.047232721000909805,
0.03818786144256592,
-0.07224789261817932,
0.06737993657588959,
0.0442998930811882,
-0.04682149365544319,
-0.011936228722333908,
0.04975960776209831,
0.03654862567782402,
-0.037008922547101974,
0.040827758610248566,
-0.06054018810391426,
3.4865243260625667e-33,
-0.04630468413233757,
0.010441410355269909,
0.05428081005811691,
-0.0010424500796943903,
-0.008772657252848148,
0.041617266833782196,
0.024938534945249557,
-0.020173659548163414,
-0.016780661419034004,
0.06035676226019859,
0.0014054278144612908,
-0.004070700611919165,
-0.06004319339990616,
0.04461994022130966,
0.03268980234861374,
0.00646359845995903,
-0.02607688494026661,
0.03823249042034149,
-0.003625444369390607,
-0.03894840553402901,
0.0677327811717987,
-0.09217868745326996,
-0.029878320172429085,
-0.003256183350458741,
-0.04338929429650307,
-0.022053176537156105,
0.10949867218732834,
0.004227929282933474,
-0.032417427748441696,
0.006481250282377005,
-0.07832752913236618,
0.013022596016526222,
-0.009057598188519478,
0.00812529120594263,
0.05123627930879593,
0.05147544667124748,
0.042286813259124756,
-0.047098882496356964,
-0.013078032992780209,
0.021257184445858,
-0.05355376750230789,
0.03591805696487427,
0.018042810261249542,
-0.06135747954249382,
-0.1332901567220688,
-0.07139161974191666,
-0.0208551287651062,
0.024134261533617973,
0.016333002597093582,
-0.05699486657977104,
-0.05644892528653145,
0.03239159658551216,
0.013973020948469639,
-0.031122904270887375,
0.023351840674877167,
-0.02039409428834915,
0.04584083706140518,
0.021847186610102654,
0.019090862944722176,
0.041164036840200424,
-0.004702000413089991,
0.0024403862189501524,
-0.011750121600925922,
-0.005165940150618553,
-0.06741742044687271,
0.06781916320323944,
-0.01464806031435728,
0.06528645008802414,
0.036716021597385406,
-0.066569484770298,
0.02643396519124508,
0.022495150566101074,
0.02158001810312271,
-0.11003118753433228,
0.1149277538061142,
0.03677259385585785,
0.02642184868454933,
-0.07908295840024948,
-0.003158977720886469,
-0.04915965348482132,
0.0062778708525002,
0.08005888015031815,
0.019724862650036812,
-0.052175913006067276,
-0.03792892396450043,
-0.09380339831113815,
0.038713373243808746,
-0.042548153549432755,
-0.046447254717350006,
-0.016935240477323532,
-0.2006121575832367,
0.04937276989221573,
-0.01681678742170334,
0.005476487800478935,
-0.04059884697198868,
-4.225246118421571e-33,
-0.03597443550825119,
-0.011050203815102577,
-0.0213546771556139,
0.058855246752500534,
0.02811727672815323,
-0.004473441280424595,
-0.07458659261465073,
-0.0650813952088356,
0.03089033253490925,
-0.0025804457254707813,
0.06962014734745026,
0.005996511783450842,
0.042715754359960556,
-0.041979528963565826,
0.023577788844704628,
0.026530450209975243,
-0.07866412401199341,
0.03864492475986481,
0.00042569945799186826,
-0.0009879719000309706,
-0.0019372247625142336,
0.050977494567632675,
-0.028176717460155487,
0.06099855154752731,
-0.012168248184025288,
0.0056510199792683125,
-0.08792564272880554,
0.05741441249847412,
-0.06955485790967941,
-0.03244861587882042,
-0.06998607516288757,
0.021347032859921455,
-0.052360329777002335,
-0.051432088017463684,
0.0019198250956833363,
0.0431072935461998,
-0.03148622810840607,
0.04540757089853287,
0.017522308975458145,
0.07972470670938492,
0.05722356215119362,
0.010437911376357079,
0.01167348027229309,
-0.032537590712308884,
0.021570725366473198,
-0.017980745062232018,
-0.005971681792289019,
-0.08401788026094437,
0.005695804487913847,
0.04837377369403839,
-0.01382129080593586,
0.0018041922012344003,
-0.023287976160645485,
0.061959024518728256,
-0.04369257763028145,
0.025211075320839882,
-0.06950604170560837,
-0.03384236618876457,
0.11518976092338562,
-0.02903762459754944,
-0.015911897644400597,
0.016560163348913193,
-0.11795199662446976,
0.10369982570409775,
0.05617523565888405,
-0.02807934582233429,
0.01855328306555748,
0.015253733843564987,
0.027849648147821426,
0.07988346368074417,
0.0063722580671310425,
0.07941660284996033,
-0.0011374815367162228,
0.048181287944316864,
-0.07310467213392258,
-0.005808301270008087,
0.08492757380008698,
-0.08212418109178543,
-0.05258900299668312,
-0.006075381301343441,
0.17617638409137726,
0.048391781747341156,
0.09685606509447098,
-0.015665709972381592,
0.027124807238578796,
0.02281254343688488,
0.061825405806303024,
0.0302299614995718,
0.0007891279528848827,
-0.02681421861052513,
-0.04783746600151062,
-0.00888922717422247,
-0.018804611638188362,
-0.048938799649477005,
-0.056140828877687454,
-4.0744676965687177e-8,
-0.075404092669487,
-0.0457925871014595,
-0.008769181556999683,
-0.06627681106328964,
0.06214059144258499,
-0.04328000918030739,
-0.043971847742795944,
0.06325972080230713,
0.0002950163616333157,
-0.05988248810172081,
0.05823918431997299,
0.020498333498835564,
-0.020469386130571365,
-0.052098821848630905,
-0.007862172089517117,
0.06427855044603348,
-0.026650063693523407,
0.024427535012364388,
-0.007612563204020262,
0.0038444800302386284,
-0.05581838637590408,
0.06983454525470734,
0.025806088000535965,
-0.0460890531539917,
0.040421273559331894,
-0.09119061380624771,
0.01946733333170414,
0.048928286880254745,
0.15251027047634125,
-0.03211254999041557,
-0.12080708146095276,
0.05821990966796875,
0.031561557203531265,
0.01233669463545084,
0.012038403190672398,
0.10291333496570587,
0.004668454173952341,
0.03497645631432533,
-0.04596814513206482,
-0.0067473528906702995,
-0.04782950505614281,
0.10524896532297134,
-0.03953484073281288,
-0.025341549888253212,
0.012766887433826923,
-0.026652222499251366,
-0.008948869071900845,
-0.04948725923895836,
-0.047438960522413254,
0.02404133416712284,
0.05129940062761307,
0.07374504208564758,
0.012713480740785599,
0.08448236435651779,
0.04315860942006111,
-0.023891732096672058,
-0.05078830569982529,
-0.03359800949692726,
0.005952015519142151,
0.07472241669893265,
-0.044824376702308655,
0.06798066943883896,
0.01643345132470131,
0.003512920346111059
] | 0 | 0 | 0 | 0 |
|
Compute-Optimal LLMs Provably Generalize Better with Scale | https://openreview.net/forum?id=MF7ljU8xcf | [
"Marc Anton Finzi",
"Sanyam Kapoor",
"Diego Granziol",
"Anming Gu",
"Christopher De Sa",
"J Zico Kolter",
"Andrew Gordon Wilson"
] | Poster | Why do larger language models generalize better? To explore this question, we develop generalization bounds on the pretraining objective of large language models (LLMs) in the compute-optimal regime, as described by the Chinchilla scaling laws. We introduce a novel, fully empirical Freedman-type martingale concentration inequality that tightens existing bounds by accounting for the variance of the loss function. The generalization bound can be broken into three contributions: the number of parameters per token, the loss variance, and the quantization error at a fixed bitrate. As language models are scaled up, the number of parameters per data point stays constant; however, both the loss variance and the quantization error decrease, implying that larger models should have \emph{smaller} generalization gaps. We examine why larger models tend to be more quantizable from an information theoretic perspective, showing that the rate at which they can integrate new information grows slower than their capacity on the compute optimal frontier. From these findings we produce a scaling law for the generalization gap, showing that our bounds decrease in a predictable way. | generalization bounds, language models, scaling laws | We construct a generalization bound for LLMs which gets better as the models get bigger. | 13,266 | null | [
-0.0040020449087023735,
-0.0572354681789875,
0.049568381160497665,
0.029924185946583748,
0.06589911878108978,
0.043771352618932724,
0.005016693379729986,
0.04265236482024193,
0.13010402023792267,
-0.012075943872332573,
-0.023715419694781303,
-0.004779136274009943,
0.029790084809064865,
-0.047691114246845245,
0.04309523105621338,
-0.03422997146844864,
0.0669129267334938,
0.08113311976194382,
-0.1582348793745041,
-0.07908021658658981,
0.047089893370866776,
0.021106820553541183,
0.0153021439909935,
-0.0065647270530462265,
0.06928496062755585,
-0.0833420529961586,
-0.023896867409348488,
-0.009053030051290989,
0.050335343927145004,
-0.005960620008409023,
-0.015034900978207588,
0.03455568477511406,
0.07252485305070877,
0.045563969761133194,
-0.01262843981385231,
0.03227740153670311,
-0.024825438857078552,
-0.015647148713469505,
0.011295647360384464,
0.03137357532978058,
0.028820408508181572,
0.03791993483901024,
0.038315415382385254,
0.02939528226852417,
0.05968989431858063,
0.034132082015275955,
0.04068461060523987,
-0.0012392703210934997,
-0.08916112035512924,
-0.0443708673119545,
-0.010014018043875694,
-0.016591941937804222,
-0.03474290668964386,
0.030621757730841637,
-0.06058396026492119,
-0.10553668439388275,
0.0383349284529686,
0.011106300167739391,
-0.04802964627742767,
0.04414835944771767,
-0.054330386221408844,
-0.09545189887285233,
-0.04076055437326431,
-0.017828835174441338,
-0.005084729287773371,
-0.010862520895898342,
0.07693702727556229,
0.009472067467868328,
-0.047672998160123825,
0.12426717579364777,
-0.059801135212183,
-0.0023792763240635395,
-0.12442458420991898,
0.09875704348087311,
0.014179285615682602,
-0.04487331956624985,
0.08116570115089417,
0.034938160330057144,
0.03520973399281502,
-0.07225554436445236,
-0.014786982908844948,
0.004657723940908909,
0.04691820964217186,
-0.01511932909488678,
0.06990718096494675,
-0.020405126735568047,
-0.01404774934053421,
0.0036384607665240765,
0.10445906221866608,
-0.06419660896062851,
-0.062368445098400116,
-0.030356107279658318,
-0.06838778406381607,
0.0773361548781395,
0.06624161452054977,
-0.013748865574598312,
0.0045990534126758575,
-0.07056477665901184,
0.011316807009279728,
0.08583780378103256,
0.0667644664645195,
0.02269096113741398,
0.06281539797782898,
-0.008116009645164013,
0.03526687249541283,
-0.0890183299779892,
0.02104746177792549,
0.05431756749749184,
0.022466382011771202,
-0.01474945992231369,
0.05350716784596443,
0.09680717438459396,
0.025115057826042175,
0.053983937948942184,
-0.003300097305327654,
-0.015177521854639053,
0.02424420230090618,
0.0036786156706511974,
-0.0227971114218235,
0.0421692356467247,
-0.05651995539665222,
0.017150089144706726,
-0.01468120701611042,
-0.02113630250096321,
-0.052389249205589294,
0.01947987638413906,
-0.07153841108083725,
1.435773779349297e-33,
0.08881093561649323,
0.00935358926653862,
0.02593577653169632,
0.02383808046579361,
0.026874123141169548,
0.08380670100450516,
0.025292977690696716,
-0.0575045570731163,
0.007916031405329704,
0.02047792263329029,
-0.027666065841913223,
0.042606089264154434,
-0.06294094026088715,
0.02203184738755226,
0.05666076019406319,
0.04061397910118103,
-0.041190486401319504,
0.06913593411445618,
0.004413650371134281,
-0.09640442579984665,
0.1083909347653389,
-0.051123060286045074,
-0.010593226179480553,
-0.056325122714042664,
-0.008363530971109867,
-0.011913635767996311,
0.007586842868477106,
-0.08622586727142334,
-0.05632537230849266,
0.013465022668242455,
-0.1729956865310669,
-0.017831331118941307,
0.046627227216959,
-0.047832898795604706,
0.023628463968634605,
-0.03901094198226929,
0.008622316643595695,
-0.01452375203371048,
-0.01928522065281868,
-0.07698354125022888,
0.008278949186205864,
0.060648683458566666,
0.044008538126945496,
-0.030699657276272774,
-0.07674824446439743,
-0.10091765224933624,
-0.005738771054893732,
-0.02731943316757679,
-0.014009119011461735,
-0.12005586922168732,
0.0730380192399025,
-0.011921370401978493,
-0.05601118877530098,
0.023669647052884102,
0.061803560703992844,
0.0071702515706419945,
0.011679726652801037,
0.011115030385553837,
-0.04368172958493233,
0.06413349509239197,
-0.049472875893116,
0.02811693400144577,
0.05504833161830902,
0.08076051622629166,
0.09492126852273941,
0.03349192813038826,
-0.04112860560417175,
0.07653064280748367,
0.03161372244358063,
-0.010918005369603634,
-0.005530606023967266,
-0.029554538428783417,
-0.01633135974407196,
-0.051145825535058975,
0.03297539800405502,
-0.042272258549928665,
0.07123960554599762,
-0.09666295349597931,
0.018775811418890953,
0.06460218131542206,
-0.004868136253207922,
-0.04338927939534187,
0.01636703684926033,
-0.06691443920135498,
-0.06439079344272614,
-0.09623945504426956,
0.05469568446278572,
-0.025781916454434395,
-0.01814419776201248,
-0.03359944000840187,
-0.06251078844070435,
-0.013061858713626862,
-0.0070697227492928505,
-0.01960366778075695,
-0.04643801227211952,
-3.483988564342652e-33,
-0.09263034164905548,
-0.02300226129591465,
-0.012424937449395657,
0.1204468384385109,
-0.023124869912862778,
-0.10540496557950974,
-0.015663830563426018,
0.0631890818476677,
-0.06687390804290771,
-0.09035642445087433,
-0.07652036100625992,
-0.014808310195803642,
0.03552549332380295,
0.03441113606095314,
0.006972077768296003,
-0.00619027903303504,
0.015876207500696182,
0.07233921438455582,
0.06807664036750793,
0.04127083346247673,
0.0027518034912645817,
-0.0019705055747181177,
-0.024256879463791847,
0.016514936462044716,
0.032995905727148056,
-0.04020145162940025,
-0.11596714705228806,
0.10778316110372543,
-0.0306122787296772,
-0.03167396038770676,
-0.054411619901657104,
-0.04356035962700844,
-0.06710677593946457,
-0.00536783691495657,
-0.054766569286584854,
-0.00596140930429101,
0.035012274980545044,
0.11716807633638382,
-0.025099536404013634,
0.11881668120622635,
0.08544355630874634,
-0.06850764900445938,
-0.05857427790760994,
-0.012573414482176304,
0.06233884394168854,
-0.031964849680662155,
-0.07020357996225357,
-0.017532045021653175,
0.0019061079947277904,
-0.04123140871524811,
0.05070370063185692,
0.018684741109609604,
0.026386326178908348,
0.04687992110848427,
-0.05106443539261818,
-0.02778335101902485,
-0.04008530452847481,
-0.0667824000120163,
-0.012605449184775352,
0.020992035046219826,
-0.072476826608181,
-0.009497258812189102,
-0.062357790768146515,
-0.03623410314321518,
-0.03620152920484543,
-0.017073260620236397,
-0.027896663174033165,
-0.01836216263473034,
0.07134804129600525,
-0.026266878470778465,
0.004788014106452465,
-0.03478598594665527,
-0.05196016654372215,
0.045846570283174515,
-0.08935552090406418,
0.03851792961359024,
0.048554111272096634,
-0.04602372646331787,
0.021755021065473557,
-0.004302632994949818,
0.042876120656728745,
0.10530310869216919,
0.04229428991675377,
-0.030009189620614052,
0.023627588525414467,
0.020346682518720627,
0.07259970158338547,
0.048172932118177414,
0.03954590857028961,
-0.03003014251589775,
-0.049709029495716095,
-0.003576381132006645,
-0.01229360792785883,
0.01808561012148857,
-0.11767231673002243,
-4.5326629560804577e-8,
-0.01886284351348877,
0.06771460920572281,
-0.02348133735358715,
0.04499417170882225,
0.06293472647666931,
-0.013445315882563591,
-0.07591605931520462,
0.042087338864803314,
0.0454307459294796,
0.00431688642129302,
0.06671411544084549,
0.029207365587353706,
-0.06729728728532791,
0.016267040744423866,
-0.015632687136530876,
0.06687536835670471,
-0.047463756054639816,
-0.03214143216609955,
0.0249785128980875,
-0.009246696718037128,
0.008843449875712395,
0.03564101457595825,
-0.0000027334856440575095,
-0.07400199770927429,
0.05499393492937088,
-0.055441588163375854,
0.017618563026189804,
0.11257310956716537,
0.009696980006992817,
0.020923176780343056,
-0.01714208908379078,
0.011887118220329285,
-0.03227894380688667,
0.06468620896339417,
0.026613455265760422,
0.018772197887301445,
-0.07637009769678116,
-0.03566510230302811,
0.03779732063412666,
-0.07651060819625854,
0.032270364463329315,
0.02027830109000206,
-0.029259121045470238,
0.025594675913453102,
0.044087402522563934,
-0.007601284887641668,
-0.03814128786325455,
-0.021213125437498093,
0.04094774276018143,
-0.019811265170574188,
0.043253347277641296,
0.008951704017817974,
0.03417671099305153,
0.0285587590187788,
0.0382973849773407,
0.05917353555560112,
-0.008606513030827045,
-0.02045508287847042,
-0.028011871501803398,
-0.010551034472882748,
-0.023010894656181335,
-0.011864949949085712,
-0.07578159123659134,
-0.004160537850111723
] | 0 | 0 | 0 | 0 |
|
Safety-Prioritizing Curricula for Constrained Reinforcement Learning | https://openreview.net/forum?id=f3QR9TEERH | [
"Cevahir Koprulu",
"Thiago D. Simão",
"Nils Jansen",
"ufuk topcu"
] | Poster | Curriculum learning aims to accelerate reinforcement learning (RL) by generating curricula, i.e., sequences of tasks of increasing difficulty.
Although existing curriculum generation approaches provide benefits in sample efficiency, they overlook safety-critical settings where an RL agent must adhere to safety constraints.
Thus, these approaches may generate tasks that cause RL agents to violate safety constraints during training and behave suboptimally after.
We develop a safe curriculum generation approach (SCG) that aligns the objectives of constrained RL and curriculum learning: improving safety during training and boosting sample efficiency.
SCG generates sequences of tasks where the RL agent can be safe and performant by initially generating tasks with minimum safety violations over high-reward ones.
We empirically show that compared to the state-of-the-art curriculum learning approaches and their naively modified safe versions, SCG achieves optimal performance and the lowest amount of constraint violations during training. | curriculum learning, constrained reinforcement learning | We propose a safe curriculum generation method that reduces safety constraint violations during training while boosting the learning speed of constrained RL agents. | 13,261 | null | [
-0.06614730507135391,
0.0003629099519457668,
-0.07343901693820953,
0.017803678289055824,
0.05133945494890213,
0.09861990064382553,
-0.037376612424850464,
-0.023192130029201508,
-0.03554282337427139,
0.04787208139896393,
-0.0639326199889183,
-0.025702249258756638,
0.048684168606996536,
0.04772891104221344,
0.01953226886689663,
-0.022286908701062202,
0.08286937326192856,
0.0696193277835846,
-0.07414029538631439,
-0.006970901973545551,
0.009501918219029903,
0.037652939558029175,
0.12109080702066422,
0.02590961568057537,
-0.04124545305967331,
-0.03512955456972122,
0.015676654875278473,
-0.021357597783207893,
0.03655821084976196,
-0.038014739751815796,
0.056089695543050766,
-0.10169101506471634,
0.003335231216624379,
-0.018305450677871704,
0.01174948550760746,
0.038928139954805374,
-0.09211478382349014,
0.027937186881899834,
-0.01540489960461855,
0.0015463182935491204,
-0.10152223706245422,
0.02891393192112446,
0.007364135235548019,
-0.0028337270487099886,
-0.007176003884524107,
-0.034698233008384705,
-0.009000844322144985,
-0.038442958146333694,
0.005173124372959137,
-0.05091559514403343,
-0.04024019464850426,
-0.008122246712446213,
0.0014244074700400233,
0.02088700421154499,
-0.01689271628856659,
0.04766005650162697,
0.025175880640745163,
-0.012363284826278687,
-0.03846972808241844,
-0.004730321001261473,
-0.12008481472730637,
-0.08985938876867294,
-0.07413731515407562,
-0.04182124510407448,
-0.08364897966384888,
0.0030569760128855705,
-0.02870832569897175,
0.07108595967292786,
0.03633297607302666,
0.08484672009944916,
-0.06426481902599335,
0.03177745267748833,
0.005033604335039854,
-0.013275464065372944,
0.04996470734477043,
0.08854705840349197,
0.009366718120872974,
-0.041797738522291183,
0.020421048626303673,
-0.032343413680791855,
0.020522048696875572,
-0.022571861743927002,
0.021039564162492752,
0.060070183128118515,
0.10646332055330276,
0.0006861472502350807,
0.009988624602556229,
0.04570717364549637,
0.10015463829040527,
0.023007722571492195,
0.03866206482052803,
0.044170282781124115,
0.10738585889339447,
0.06261063367128372,
0.0259623434394598,
0.00500351469963789,
-0.06044916808605194,
-0.08731906861066818,
-0.04391824081540108,
0.02858084999024868,
-0.008974933996796608,
0.03989597409963608,
-0.024315359070897102,
0.058412816375494,
0.019348347559571266,
-0.07783154398202896,
-0.0023875010665506124,
-0.004937303252518177,
-0.00017016989295370877,
-0.04134361073374748,
0.09562727063894272,
0.004612633027136326,
0.05162647366523743,
0.00088357157073915,
-0.032075315713882446,
0.07058928161859512,
-0.03829067945480347,
0.013426556251943111,
-0.03814215958118439,
0.0933140218257904,
0.008660833351314068,
-0.031031055375933647,
0.04075342416763306,
0.005012646783143282,
0.012295510619878769,
-0.10948028415441513,
0.02052825130522251,
2.4175944156140736e-33,
0.09188073128461838,
0.040222398936748505,
0.01913364790380001,
-0.012064117006957531,
-0.005760327447205782,
0.01483333483338356,
0.011399438604712486,
0.024808935821056366,
-0.05946454033255577,
0.04251425713300705,
0.04745930805802345,
0.0160124059766531,
-0.09397260844707489,
0.051733966916799545,
0.09948492795228958,
-0.020101036876440048,
-0.11326701194047928,
0.06919285655021667,
-0.09391672909259796,
-0.02958185411989689,
0.015110025182366371,
-0.07442719489336014,
0.002946301130577922,
-0.06774850189685822,
0.007436948828399181,
0.0730506032705307,
0.033466652035713196,
0.024850087240338326,
0.025818223133683205,
0.0009298108052462339,
-0.06698141992092133,
0.06338251382112503,
-0.002413645153865218,
0.05141841247677803,
0.004624904599040747,
0.07135085761547089,
-0.09507472813129425,
-0.007913004606962204,
0.04688968136906624,
0.09059964865446091,
-0.010428576730191708,
-0.012363490648567677,
0.16542385518550873,
0.05420943722128868,
-0.001450781594030559,
-0.07730979472398758,
0.02207004278898239,
0.005444152746349573,
-0.07516337931156158,
-0.08491431176662445,
-0.07705257833003998,
0.023350737988948822,
0.03564458712935448,
-0.06460157781839371,
-0.07273377478122711,
0.04852695390582085,
0.07665122300386429,
0.04659799858927727,
-0.06854647397994995,
0.030246809124946594,
-0.02399424836039543,
0.019839966669678688,
0.009880904108285904,
0.07269471138715744,
-0.009869653731584549,
-0.008827934972941875,
-0.0490424744784832,
0.04017801955342293,
0.09075190871953964,
-0.04456121474504471,
-0.037844475358724594,
-0.011206956580281258,
-0.02730788104236126,
-0.08132174611091614,
0.04034248739480972,
0.029050400480628014,
0.08094383776187897,
-0.05678056553006172,
-0.026778053492307663,
-0.0643477514386177,
0.0652291476726532,
0.04254128411412239,
-0.019697437062859535,
-0.0027482337318360806,
0.028980981558561325,
-0.08346863090991974,
-0.0712163895368576,
-0.02076680026948452,
0.03685745224356651,
0.022491954267024994,
-0.04971012473106384,
-0.04885777831077576,
-0.028028765693306923,
0.09744156152009964,
0.0066821034997701645,
-2.4545515906619417e-33,
0.025489231571555138,
-0.011062880046665668,
-0.12004723399877548,
0.06370392441749573,
0.012396808713674545,
-0.055708467960357666,
0.0321536660194397,
-0.08464938402175903,
-0.012861443683505058,
-0.05892400071024895,
-0.06910031288862228,
-0.02523619309067726,
-0.00018339662346988916,
-0.014389891177415848,
-0.003348193597048521,
-0.10548603534698486,
-0.030053116381168365,
0.0741594210267067,
-0.017289267852902412,
-0.028189893811941147,
0.049240317195653915,
0.03363509848713875,
-0.09890127927064896,
0.060898806899785995,
0.009553351439535618,
-0.015693597495555878,
-0.021995676681399345,
0.046690914779901505,
0.051291871815919876,
0.08129984885454178,
0.003091182792559266,
-0.017246654257178307,
0.034239862114191055,
0.029990503564476967,
-0.05879419669508934,
0.03991500660777092,
0.07399287819862366,
0.02884708344936371,
-0.06936093419790268,
0.0753144919872284,
0.07893160730600357,
-0.030602367594838142,
-0.03211190178990364,
-0.018354184925556183,
0.011843554675579071,
-0.02401367388665676,
0.006011992692947388,
-0.0012930165976285934,
-0.03542128950357437,
0.0233137384057045,
-0.07661779969930649,
-0.00610866816714406,
-0.07401113212108612,
-0.05210892856121063,
0.01613755337893963,
-0.07107488065958023,
-0.003682356793433428,
-0.06047310680150986,
0.1050371453166008,
0.012592783197760582,
-0.03792571276426315,
0.05669640004634857,
-0.05590859428048134,
0.06205088272690773,
-0.007047905120998621,
0.02809201553463936,
-0.045900896191596985,
0.05407414957880974,
0.03365732729434967,
-0.008521494455635548,
0.029796017333865166,
0.06789398193359375,
0.06622438132762909,
-0.08197052031755447,
-0.001210079644806683,
-0.04065767675638199,
-0.052966102957725525,
-0.010496051050722599,
-0.019662462174892426,
-0.02402735874056816,
0.04921005293726921,
-0.01891484297811985,
0.022673752158880234,
0.0861242488026619,
0.013560504652559757,
0.06713379919528961,
0.028118519112467766,
0.07048690319061279,
0.045866597443819046,
-0.031039489433169365,
-0.03175397962331772,
-0.007433723658323288,
0.017103835940361023,
0.017324769869446754,
-0.09574109315872192,
-3.78257318800479e-8,
-0.007985579781234264,
0.0057546221651136875,
-0.002686032559722662,
0.03548431023955345,
0.09453674405813217,
-0.01430958416312933,
-0.029448261484503746,
-0.03092627041041851,
-0.02842688374221325,
-0.006491697859019041,
0.06701968610286713,
0.03002794459462166,
0.06013317033648491,
-0.056173503398895264,
-0.06239400431513786,
0.07834966480731964,
0.03516722843050957,
-0.01533515565097332,
0.00843852385878563,
-0.006729250308126211,
0.001027274178341031,
0.014492152258753777,
0.004461898002773523,
-0.04513268545269966,
-0.0035438952036201954,
-0.06694638729095459,
-0.011778181418776512,
0.0827452763915062,
0.10629385709762573,
0.06017513573169708,
-0.008542240597307682,
-0.08815643936395645,
0.05229562148451805,
0.0016763329040259123,
0.0014634847175329924,
0.10163822770118713,
0.005018025636672974,
-0.05008415877819061,
0.024938810616731644,
-0.09876970201730728,
0.014304806478321552,
0.05528269335627556,
0.021763449534773827,
0.021617887541651726,
-0.051672451198101044,
-0.0029788336250931025,
-0.09818482398986816,
-0.04241112992167473,
0.08698296546936035,
-0.06268326938152313,
-0.06936563551425934,
-0.036904022097587585,
-0.03444783389568329,
0.016748467460274696,
0.07106522470712662,
0.007356270216405392,
-0.024446148425340652,
-0.10148458927869797,
0.010599080473184586,
0.06317058950662613,
0.00436375429853797,
0.030412867665290833,
-0.022969774901866913,
-0.0006239049253053963
] | 0 | 0 | 0 | 0 |
|
Generalized Behavior Learning from Diverse Demonstrations | https://openreview.net/forum?id=Q7EjHroO1w | [
"Varshith Sreeramdass",
"Rohan R Paleja",
"Letian Chen",
"Sanne van Waveren",
"Matthew Gombolay"
] | Poster | Diverse behavior policies are valuable in domains requiring quick test-time adaptation or personalized human-robot interaction. Human demonstrations provide rich information regarding task objectives and factors that govern individual behavior variations, which can be used to characterize \textit{useful} diversity and learn diverse performant policies.
However, we show that prior work that builds naive representations of demonstration heterogeneity fails in generating successful novel behaviors that generalize over behavior factors.
We propose Guided Strategy Discovery (GSD), which introduces a novel diversity formulation based on a learned task-relevance measure that prioritizes behaviors exploring modeled latent factors.
We empirically validate across three continuous control benchmarks for generalizing to in-distribution (interpolation) and out-of-distribution (extrapolation) factors that GSD outperforms baselines in novel behavior discovery by $\sim$21\%.
Finally, we demonstrate that GSD can generalize striking behaviors for table tennis in a virtual testbed while leveraging human demonstrations collected in the real world.
Code is available at https://github.com/CORE-Robotics-Lab/GSD. | Behavior Discovery, Demonstrator Heterogeneity | We propose an imitation learning approach that utilizes a new diversity formulation to generate novel behaviors that generalize over demonstrators' latent preferences. | 13,253 | null | [
-0.006528046447783709,
-0.08609028905630112,
0.051411621272563934,
-0.024206241592764854,
0.08576694875955582,
0.03077942132949829,
-0.011977167800068855,
-0.03242935985326767,
0.009228934533894062,
0.046667732298374176,
-0.0226888507604599,
-0.04325823485851288,
0.0029238788411021233,
0.04300675913691521,
0.07393474131822586,
-0.060789089649915695,
0.13382384181022644,
0.057456690818071365,
-0.08281558007001877,
0.022006912156939507,
0.00701757799834013,
0.013870696537196636,
0.12246127426624298,
-0.03499358892440796,
-0.05215131863951683,
-0.019922584295272827,
0.04863167926669121,
-0.09459401667118073,
0.016550811007618904,
-0.026651114225387573,
0.011451521888375282,
0.012253488413989544,
-0.036856599152088165,
-0.07053696364164352,
-0.07930591702461243,
0.0793340802192688,
-0.06400050222873688,
0.014232125133275986,
-0.01840129867196083,
-0.017376994714140892,
-0.012152397073805332,
-0.023886334151029587,
0.06924949586391449,
0.034387193620204926,
0.04093063995242119,
0.049906592816114426,
-0.009265545755624771,
-0.09003058075904846,
0.03267335146665573,
-0.0029600183479487896,
-0.00752414483577013,
0.035567451268434525,
0.11235152184963226,
-0.0697261244058609,
0.06063385307788849,
-0.030486879870295525,
0.06445292383432388,
0.007619138341397047,
0.09006935358047485,
-0.014215935952961445,
0.04204247146844864,
-0.12470544129610062,
-0.008755971677601337,
-0.05078384280204773,
-0.0295695923268795,
-0.025638720020651817,
0.00013654952635988593,
0.011044450104236603,
-0.0006471282104030252,
-0.021638808771967888,
-0.031114526093006134,
0.04701576754450798,
0.030035225674510002,
0.03903636336326599,
0.07112850993871689,
0.05872761085629463,
-0.03342624753713608,
-0.08013448119163513,
-0.0372830294072628,
-0.032930951565504074,
-0.00205003609880805,
-0.032840631902217865,
-0.07298608869314194,
0.06494787335395813,
0.12332816421985626,
-0.019313666969537735,
0.038463789969682693,
0.05422227829694748,
0.015461361035704613,
0.045929934829473495,
-0.007284792605787516,
-0.00564949307590723,
-0.013440988026559353,
-0.07623182237148285,
-0.13494756817817688,
-0.023001912981271744,
-0.03053307719528675,
-0.1158459261059761,
0.013796791434288025,
0.05586931109428406,
0.014863773249089718,
0.050355397164821625,
0.020920660346746445,
0.0351371131837368,
-0.043334510177373886,
-0.08389144390821457,
-0.019597722217440605,
0.03799768164753914,
0.07714550197124481,
-0.026573797687888145,
-0.02968808263540268,
-0.05979461967945099,
0.03681498393416405,
0.05355051904916763,
0.03849826753139496,
-0.004657485522329807,
0.011505045928061008,
-0.008264734409749508,
-0.006738436408340931,
0.067876435816288,
0.05806906148791313,
-0.035933829843997955,
0.02526436187326908,
-0.0008777237962931395,
0.027485331520438194,
0.026969071477651596,
-0.04830433428287506,
1.4933130339605979e-33,
0.08416127413511276,
-0.004748739767819643,
0.10432847589254379,
0.05414928123354912,
0.025427894666790962,
-0.022359680384397507,
0.03855181857943535,
-0.001331773237325251,
-0.03488258272409439,
0.014496419578790665,
-0.01210626307874918,
0.028367873281240463,
-0.04594035446643829,
0.0882154032588005,
-0.01145516149699688,
0.08642732352018356,
-0.09389810264110565,
0.041142381727695465,
-0.041534069925546646,
0.029057612642645836,
0.0904344916343689,
0.01694440469145775,
-0.02401946485042572,
-0.028570106253027916,
0.02477659471333027,
-0.029775818809866905,
0.03184768185019493,
-0.011078394018113613,
-0.03399863466620445,
-0.0008144046878442168,
-0.12877197563648224,
0.012338369153439999,
-0.04147600010037422,
0.08641182631254196,
0.05127902701497078,
0.03608177229762077,
-0.05203815922141075,
-0.07027682662010193,
0.01831171102821827,
0.002626789500936866,
-0.049393683671951294,
-0.06156982481479645,
-0.0001664867450017482,
-0.03219258785247803,
-0.06454813480377197,
-0.06714371591806412,
0.06611992418766022,
-0.005150947719812393,
-0.05816921591758728,
-0.017628798261284828,
0.007958400063216686,
0.08108700066804886,
0.0635327398777008,
-0.07160503417253494,
0.018024278804659843,
-0.01632850244641304,
0.018840376287698746,
-0.004047596361488104,
0.007335038390010595,
0.009100375697016716,
0.011113575659692287,
0.008050460368394852,
-0.04952333867549896,
0.10546673089265823,
-0.022153541445732117,
0.06106620281934738,
0.0013426756486296654,
0.04991462826728821,
0.08371256291866302,
0.04142841696739197,
0.01662421226501465,
0.030400017276406288,
-0.0207620058208704,
-0.0569952018558979,
0.07765117287635803,
-0.03657542169094086,
0.053657568991184235,
-0.07050837576389313,
0.034248847514390945,
-0.04527236893773079,
-0.012725301086902618,
0.019284015521407127,
-0.07109469175338745,
-0.031323082745075226,
-0.06310044229030609,
-0.04616326093673706,
0.019237739965319633,
-0.06737252324819565,
-0.05612100660800934,
-0.016026372089982033,
-0.08274321258068085,
-0.05343570560216904,
-0.10433442145586014,
0.07079186290502548,
0.03657388314604759,
-3.3191552363405816e-33,
-0.029909130185842514,
0.0002619474835228175,
0.019404569640755653,
-0.016516828909516335,
0.027302619069814682,
-0.028239605948328972,
-0.0033116501290351152,
-0.004093825817108154,
-0.026532357558608055,
-0.04160338267683983,
-0.06725144386291504,
-0.03723495081067085,
0.013754109852015972,
-0.008078495971858501,
-0.015413944609463215,
0.035532739013433456,
0.015589767135679722,
0.045125775039196014,
0.035894159227609634,
-0.027068305760622025,
0.04993995651602745,
0.06239934265613556,
-0.031217442825436592,
0.02356317639350891,
-0.024689385667443275,
-0.017045697197318077,
-0.010811618529260159,
0.054384514689445496,
0.003136151237413287,
-0.0006066823843866587,
-0.043371766805648804,
-0.012401660904288292,
-0.0063627976924180984,
0.008054734207689762,
-0.023589519783854485,
0.04350076615810394,
0.03511076420545578,
0.011812121607363224,
-0.015926524996757507,
0.10035045444965363,
-0.009776708669960499,
-0.06320956349372864,
-0.008050371892750263,
0.04556027799844742,
0.05114714801311493,
0.017007902264595032,
-0.06883281469345093,
-0.011944734491407871,
-0.03405340388417244,
0.025577722117304802,
-0.05109152942895889,
0.026507923379540443,
-0.0798255056142807,
-0.05056183785200119,
-0.06588470935821533,
-0.06727535277605057,
0.008139744400978088,
0.002321419073268771,
0.02840745821595192,
0.008029655553400517,
-0.050472959876060486,
0.04331021010875702,
-0.015531628392636776,
0.07940366864204407,
-0.04580168053507805,
-0.005933975335210562,
-0.011206980794668198,
-0.006523946300148964,
-0.0038578566163778305,
-0.039003245532512665,
-0.0374194011092186,
-0.0130843510851264,
0.051022179424762726,
-0.0372559055685997,
0.02337881363928318,
-0.028477748855948448,
-0.010381216183304787,
-0.009873918257653713,
-0.005525889806449413,
0.01932530477643013,
0.05177868530154228,
0.012561462819576263,
0.08024575561285019,
0.035189371556043625,
-0.032908376306295395,
0.22222130000591278,
-0.027111727744340897,
0.16642072796821594,
-0.016957879066467285,
-0.009914571419358253,
-0.018769843503832817,
0.04745123162865639,
-0.036184024065732956,
0.03361755982041359,
-0.039905231446027756,
-4.47120420687952e-8,
-0.08529685437679291,
0.09740583598613739,
0.06938236206769943,
-0.006886343006044626,
-0.006558074615895748,
0.047963131219148636,
-0.049873340874910355,
0.007772179786115885,
-0.017116690054535866,
-0.016062399372458458,
0.0618281252682209,
-0.0070885904133319855,
0.06729217618703842,
-0.005040204618126154,
0.03077257238328457,
0.027723457664251328,
-0.005284841172397137,
-0.04182552918791771,
-0.026151735335588455,
0.04923639073967934,
0.02882104553282261,
-0.015744106844067574,
0.04976365342736244,
-0.05832509696483612,
-0.02985488437116146,
-0.073038749396801,
-0.06362178176641464,
0.039886243641376495,
0.022112838923931122,
-0.008057883940637112,
0.036142073571681976,
-0.04849400371313095,
0.003624690230935812,
-0.010330921038985252,
-0.007114849518984556,
0.08846909552812576,
-0.018128443509340286,
-0.09440454095602036,
0.0391484871506691,
0.012727858498692513,
-0.08586540073156357,
0.0655757486820221,
-0.04332008212804794,
0.03354896977543831,
-0.04833096265792847,
-0.0031590645667165518,
-0.08540184050798416,
-0.12196114659309387,
-0.012539870105683804,
-0.0525171160697937,
-0.007516353856772184,
-0.09042348712682724,
-0.025317315012216568,
0.06729492545127869,
0.17022225260734558,
0.07654069364070892,
-0.017046315595507622,
-0.07067646086215973,
0.04503058269619942,
0.06286492943763733,
0.032804686576128006,
-0.02337467670440674,
-0.11936959624290466,
0.0505644716322422
] | 0 | 0 | 0 | 0 |
|
Neural Stochastic Differential Equations for Uncertainty-Aware Offline RL | https://openreview.net/forum?id=hxUMQ4fic3 | [
"Cevahir Koprulu",
"Franck Djeumou",
"ufuk topcu"
] | Poster | Offline model-based reinforcement learning (RL) offers a principled approach to using a learned dynamics model as a simulator to optimize a control policy.
Despite the near-optimal performance of existing approaches on benchmarks with high-quality datasets, most struggle on datasets with low state-action space coverage or suboptimal demonstrations.
We develop a novel offline model-based RL approach that particularly shines in low-quality data regimes while maintaining competitive performance on high-quality datasets.
Neural Stochastic Differential Equations for Uncertainty-aware, Offline RL (NUNO) learns a dynamics model as neural stochastic differential equations (SDE),
where its drift term can leverage prior physics knowledge as inductive bias.
In parallel, its diffusion term provides distance-aware estimates of model uncertainty by matching the dynamics' underlying stochasticity near the training data regime while providing high but bounded estimates beyond it.
To address the so-called model exploitation problem in offline model-based RL, NUNO builds on existing studies by penalizing and adaptively truncating neural SDE's rollouts according to uncertainty estimates.
Our empirical results in D4RL and NeoRL MuJoCo benchmarks evidence that NUNO outperforms state-of-the-art methods in low-quality datasets by up to 93% while matching or surpassing their performance by up to 55% in some high-quality counterparts. | neural stochastic differential equations, offline reinforcement learning, physics-informed machine learning | We develop an uncertainty-aware, offline model-based reinforcement learning approach with neural stochastic differential equations that outperforms the state-of-the-art in continuous control benchmarks, particularly in low-quality datasets. | 13,251 | null | [
-0.08933000266551971,
-0.09792279452085495,
0.02814461663365364,
0.10579638183116913,
-0.026895711198449135,
-0.01876687817275524,
-0.10007081180810928,
0.040902670472860336,
0.04421898350119591,
-0.034088846296072006,
-0.0377354733645916,
0.01917175017297268,
0.026655925437808037,
-0.03287015110254288,
-0.05413352698087692,
0.023474840447306633,
0.05476558953523636,
0.07960440963506699,
-0.1194039061665535,
-0.03008977696299553,
-0.020762020722031593,
0.03279765695333481,
0.03773321956396103,
0.034284401684999466,
0.04145964980125427,
0.012038230895996094,
0.03179675340652466,
0.013833754695951939,
-0.03398926556110382,
-0.016310211271047592,
0.06108512356877327,
-0.09260294586420059,
-0.035316478461027145,
-0.08481854200363159,
0.02439565397799015,
0.035673756152391434,
-0.13394564390182495,
-0.016585160046815872,
-0.009325683116912842,
-0.03720703348517418,
0.00948886014521122,
0.031484875828027725,
0.020597748458385468,
0.02141393907368183,
0.08452880382537842,
-0.04024438187479973,
0.00864934828132391,
-0.08610863238573074,
0.04263986647129059,
-0.007449017837643623,
-0.033600542694330215,
-0.053374890238046646,
-0.00022998977510724217,
-0.00710716238245368,
0.0449896976351738,
0.061674900352954865,
0.04511813446879387,
0.007374431937932968,
-0.0044297571294009686,
-0.011784500442445278,
0.007052842061966658,
-0.09799956530332565,
-0.0015211431309580803,
-0.008727745153009892,
-0.11615682393312454,
0.03186637908220291,
-0.01535386685281992,
0.11761633306741714,
0.08342112600803375,
-0.01393505372107029,
-0.02845534309744835,
0.08597418665885925,
-0.02060127817094326,
0.01757439412176609,
0.004320419859141111,
-0.07151414453983307,
0.07895900309085846,
-0.02986413612961769,
0.03906286135315895,
-0.03170457482337952,
-0.011417143978178501,
-0.008581658825278282,
0.009327868930995464,
0.005880700424313545,
0.019159160554409027,
0.009125477634370327,
0.03070475533604622,
0.0750114694237709,
0.1364256888628006,
-0.03510993346571922,
0.013756136409938335,
0.044959571212530136,
-0.0021261454094201326,
-0.0238330215215683,
-0.011993804015219212,
0.03304530680179596,
-0.015359491109848022,
-0.06365982443094254,
-0.036898840218782425,
0.03341637924313545,
0.031209902837872505,
0.07776787132024765,
0.004170737229287624,
0.041136033833026886,
0.014850336126983166,
-0.05585513636469841,
0.03996099904179573,
0.00666083674877882,
0.025968698784708977,
-0.04693015292286873,
0.0692296251654625,
0.002953993622213602,
-0.00011332717258483171,
0.06279424577951431,
0.011819927953183651,
0.014121907763183117,
0.018432026728987694,
-0.09855209290981293,
-0.05727795511484146,
0.04362954944372177,
-0.05792830139398575,
-0.05837375298142433,
-0.014633603394031525,
0.012995285913348198,
0.027475783601403236,
-0.07346402853727341,
-0.01901598833501339,
1.850412146528581e-33,
0.018176259472966194,
-0.07493309676647186,
0.04403916373848915,
-0.027588380500674248,
0.07995579391717911,
-0.06103987246751785,
0.049421072006225586,
0.0012069707736372948,
-0.07737590372562408,
0.023840406909585,
-0.0634760931134224,
-0.01623876765370369,
-0.06636402010917664,
0.013337579555809498,
0.07210110872983932,
-0.014812925830483437,
-0.07592644542455673,
-0.012582075782120228,
0.010340849868953228,
-0.032231587916612625,
0.08622091263532639,
-0.09626206755638123,
-0.03118913620710373,
-0.030293360352516174,
0.030411507934331894,
0.022645041346549988,
0.03135526925325394,
0.03762098774313927,
0.02355024963617325,
0.016184745356440544,
-0.026335764676332474,
-0.004330407828092575,
-0.012487480416893959,
0.04236957058310509,
0.056259892880916595,
0.03016791306436062,
-0.07653209567070007,
-0.012681315653026104,
-0.06692951917648315,
0.025217706337571144,
0.04267781227827072,
-0.012773305177688599,
-0.04636489227414131,
0.08216989785432816,
-0.09124670922756195,
-0.1252157986164093,
0.06376811116933823,
-0.0991029441356659,
-0.02952677011489868,
-0.04495138302445412,
-0.038907989859580994,
0.0662817731499672,
-0.009179675951600075,
-0.06925652921199799,
-0.0149008147418499,
-0.0020112115889787674,
0.022112872451543808,
0.027437303215265274,
-0.012416447512805462,
0.060485728085041046,
-0.07050210982561111,
-0.022788025438785553,
0.04519428685307503,
0.03959549963474274,
0.0836586132645607,
0.09111037105321884,
-0.059861499816179276,
0.0703471451997757,
-0.024332312867045403,
-0.027589671313762665,
0.016171010211110115,
-0.05228012800216675,
0.028133030980825424,
-0.0694916695356369,
0.12097733467817307,
-0.05323755741119385,
0.07810508459806442,
-0.017250942066311836,
0.030725769698619843,
0.0007265463937073946,
0.019359715282917023,
-0.03958289325237274,
-0.09078729152679443,
-0.015700090676546097,
-0.01725190132856369,
-0.06110360845923424,
-0.024555975571274757,
-0.04341006651520729,
-0.042039863765239716,
-0.007361574098467827,
-0.06930231302976608,
-0.06228187307715416,
-0.017849063500761986,
-0.002077254233881831,
0.02613278292119503,
-2.8897150075615446e-33,
0.019147399812936783,
-0.0527007021009922,
-0.0022643725387752056,
0.039345238357782364,
-0.00897494051605463,
-0.0157272107899189,
-0.051876913756132126,
0.016434360295534134,
0.05488737300038338,
-0.03180292248725891,
-0.0784788653254509,
0.003484956454485655,
-0.023134497925639153,
-0.0012219945201650262,
0.052996475249528885,
-0.07123810797929764,
-0.0038588910829275846,
-0.011141868308186531,
0.06678064167499542,
0.039636094123125076,
-0.007540649268776178,
-0.022719616070389748,
-0.059899333864450455,
0.06426463276147842,
-0.009154981933534145,
0.04780978709459305,
-0.041932109743356705,
0.062026381492614746,
-0.009851369075477123,
0.013594580814242363,
-0.01679866574704647,
-0.06972334533929825,
-0.02037675864994526,
-0.017498064786195755,
-0.07475116848945618,
0.11909005045890808,
0.06173362210392952,
0.030511822551488876,
-0.005294358357787132,
0.048110317438840866,
0.019052697345614433,
-0.06499925255775452,
-0.03642105311155319,
-0.031590238213539124,
0.04036441445350647,
0.001140234642662108,
-0.05895388871431351,
-0.0110117606818676,
-0.02458685077726841,
-0.011604788713157177,
0.025884926319122314,
0.1001821905374527,
-0.03095708228647709,
0.012366152368485928,
-0.02635585144162178,
-0.04583124443888664,
-0.05250050127506256,
-0.015704119578003883,
0.014526727609336376,
0.02631823904812336,
0.013428915292024612,
-0.050713468343019485,
-0.036125097423791885,
0.0399906300008297,
-0.07436077296733856,
0.022146349772810936,
-0.06937768310308456,
0.060465458780527115,
0.039687637239694595,
0.014841975644230843,
-0.04429584741592407,
-0.022197961807250977,
0.00044783123303204775,
-0.015492958948016167,
0.010150279849767685,
-0.06633621454238892,
0.003440198488533497,
-0.02060375176370144,
0.04831743240356445,
0.035380229353904724,
0.056766439229249954,
0.03992934897542,
0.07323706895112991,
0.037103209644556046,
0.05515827611088753,
0.06557193398475647,
-0.025150064378976822,
0.07209857553243637,
0.007605941034853458,
-0.0006557735032401979,
-0.0018523207399994135,
-0.011033355258405209,
0.011377743445336819,
0.05004904046654701,
-0.0552552305161953,
-4.1708844378263166e-8,
-0.05214652046561241,
0.08468212932348251,
0.203629732131958,
0.004620397929102182,
0.12626904249191284,
0.01049414649605751,
0.016593733802437782,
-0.0347413532435894,
0.02388928085565567,
-0.032020289450883865,
0.04789595305919647,
0.03142053633928299,
-0.02602931670844555,
-0.08509814739227295,
-0.03006935678422451,
0.10799889266490936,
-0.010528997518122196,
-0.04375379532575607,
0.04188751056790352,
0.013690740801393986,
0.06078267842531204,
0.03434433415532112,
-0.04078467935323715,
-0.050320953130722046,
0.08388026058673859,
-0.023384151980280876,
-0.0017795844469219446,
0.1385420560836792,
0.010307800024747849,
-0.016335569322109222,
0.031015390530228615,
-0.06012341007590294,
0.08349045366048813,
0.009718300774693489,
0.07506757229566574,
0.07302936166524887,
0.1375211626291275,
-0.07030362635850906,
-0.03290766850113869,
-0.031100235879421234,
-0.057967495173215866,
0.06067727878689766,
-0.002014689613133669,
0.008068574592471123,
-0.0327107310295105,
0.02454374171793461,
-0.04036031663417816,
-0.07016424834728241,
0.09059549868106842,
-0.02819567546248436,
0.03380326181650162,
-0.007054205983877182,
-0.08143862336874008,
0.07667552679777145,
0.04819771647453308,
-0.0039299544878304005,
0.029408540576696396,
-0.10346774011850357,
-0.0462033785879612,
0.06437669694423676,
-0.029635732993483543,
-0.029544852674007416,
-0.03767402470111847,
0.0493914932012558
] | 0 | 0 | 0 | 0 |
|
Towards Optimal Multi-draft Speculative Decoding | https://openreview.net/forum?id=9KxnxWOBA5 | [
"Zhengmian Hu",
"Tong Zheng",
"Vignesh Viswanathan",
"Ziyi Chen",
"Ryan A. Rossi",
"Yihan Wu",
"Dinesh Manocha",
"Heng Huang"
] | Poster | Large Language Models (LLMs) have become an indispensable part of natural language processing tasks. However, autoregressive sampling has become an efficiency bottleneck. Multi-Draft Speculative Decoding (MDSD) is a recent approach where, when generating each token, a small draft model generates multiple drafts, and the target LLM verifies them in parallel, ensuring that the final output conforms to the target model distribution. The two main design choices in MDSD are the draft sampling method and the verification algorithm. For a fixed draft sampling method, the optimal acceptance rate is a solution to an optimal transport problem, but the complexity of this problem makes it difficult to solve for the optimal acceptance rate and measure the gap between existing verification algorithms and the theoretical upper bound. This paper discusses the dual of the optimal transport problem, providing a way to efficiently compute the optimal acceptance rate. For the first time, we measure the theoretical upper bound of MDSD efficiency for vocabulary sizes in the thousands and quantify the gap between existing verification algorithms and this bound. We also compare different draft sampling methods based on their optimal acceptance rates. Our results show that the draft sampling method strongly influences the optimal acceptance rate, with sampling without replacement outperforming sampling with replacement. Additionally, existing verification algorithms do not reach the theoretical upper bound for both without replacement and with replacement sampling. Our findings suggest that carefully designed draft sampling methods can potentially improve the optimal acceptance rate and enable the development of verification algorithms that closely match the theoretical upper bound. | speculative sampling | we compute maximal acceptance rate for multi-draft speculative decoding | 13,244 | 2502.18779 | [
-0.05097262188792229,
-0.05175909399986267,
-0.021808339282870293,
0.011944526806473732,
0.07133153825998306,
-0.03632323816418648,
0.004758452530950308,
0.0281078964471817,
0.09255429357290268,
0.041413940489292145,
-0.05837677791714668,
-0.010022061876952648,
0.032896313816308975,
-0.018416697159409523,
-0.013514583930373192,
-0.02211739309132099,
0.07564077526330948,
0.013555624522268772,
-0.06522424519062042,
-0.040958013385534286,
-0.01391039602458477,
0.0228392593562603,
0.015151437371969223,
-0.0063575152307748795,
0.07958720624446869,
-0.05325040593743324,
-0.056556910276412964,
-0.04481884092092514,
0.022017695009708405,
0.027741536498069763,
0.03707362711429596,
0.08529680222272873,
-0.016344845294952393,
0.0037095730658620596,
0.01638929173350334,
-0.015336320735514164,
-0.05190347880125046,
-0.048070311546325684,
0.05184641107916832,
0.0007129713776521385,
-0.06387251615524292,
0.008241250179708004,
0.01391605008393526,
0.034495603293180466,
0.017146816477179527,
-0.0425347164273262,
-0.016519401222467422,
0.02398599684238434,
0.014267872087657452,
0.03329668194055557,
0.0065544601529836655,
-0.01733827404677868,
-0.03823528066277504,
0.03984915092587471,
-0.10323327779769897,
-0.012514309957623482,
-0.0037191284354776144,
-0.0314570814371109,
-0.06098277121782303,
0.008603757247328758,
-0.13277019560337067,
-0.11633844673633575,
-0.08080556243658066,
-0.025139208883047104,
-0.03292717784643173,
0.03228403255343437,
0.0009342132252641022,
0.0183105431497097,
0.003244086867198348,
0.0643850713968277,
-0.07821106910705566,
0.06870219856500626,
-0.03017364628612995,
0.04896143823862076,
0.0038410876877605915,
0.008344943635165691,
0.027220958843827248,
-0.021998118609189987,
0.05959899723529816,
-0.085860975086689,
-0.0022177482023835182,
0.0334559828042984,
0.025582479313015938,
-0.06899712979793549,
0.05202222615480423,
0.014632213860750198,
0.05112919956445694,
0.07573480904102325,
0.11449764668941498,
-0.0015436992980539799,
-0.024326100945472717,
-0.06391815841197968,
0.03732191398739815,
0.06321101635694504,
-0.009728813543915749,
0.05119955539703369,
-0.01883547380566597,
0.02561337873339653,
0.009699065238237381,
0.06931768357753754,
0.004226010292768478,
0.08151216804981232,
0.07183405756950378,
-0.11869016289710999,
-0.040830690413713455,
-0.05948926508426666,
0.04008447378873825,
0.034493301063776016,
0.007738681044429541,
-0.029750211164355278,
-0.01257420890033245,
0.043630342930555344,
0.053653329610824585,
0.050658151507377625,
0.012300549075007439,
0.00845631118863821,
0.024660399183630943,
-0.01932428404688835,
0.015526476316154003,
0.08672524243593216,
-0.07736475020647049,
0.047871366143226624,
0.08820120245218277,
-0.024465780705213547,
-0.024328622967004776,
-0.01853843964636326,
0.05834173411130905,
9.975832808253343e-34,
-0.009814934805035591,
0.09550552815198898,
0.02316654473543167,
0.02886175364255905,
-0.00970503594726324,
-0.04176205396652222,
0.07871630042791367,
-0.004949024412781,
-0.00875227339565754,
-0.06578285247087479,
0.0027404737193137407,
0.008287499658763409,
-0.0034452201798558235,
0.053056392818689346,
0.024021903052926064,
0.0076656886376440525,
-0.021316951140761375,
0.09749134629964828,
-0.04276205226778984,
-0.005845045670866966,
0.07197599112987518,
-0.048734989017248154,
-0.018260706216096878,
-0.06815183907747269,
-0.01827530562877655,
0.016021905466914177,
-0.0013078238116577268,
-0.03470384329557419,
-0.017755426466464996,
0.042123064398765564,
-0.10262753814458847,
-0.07084701210260391,
0.013186708092689514,
0.04072359949350357,
0.020833544433116913,
-0.031820736825466156,
0.04649179056286812,
-0.02369428053498268,
-0.01361480075865984,
-0.10561435669660568,
-0.007273825351148844,
0.043824702501297,
0.06073525920510292,
0.04999511316418648,
-0.08745130896568298,
-0.036619070917367935,
-0.030725836753845215,
-0.04283418133854866,
-0.0033621564507484436,
-0.04703446850180626,
0.08889614045619965,
0.01058630645275116,
-0.09339311718940735,
0.01796051114797592,
0.009874256327748299,
-0.07170281559228897,
-0.007452015765011311,
-0.028793657198548317,
0.05571804568171501,
0.10283355414867401,
0.004518595989793539,
0.09545984864234924,
0.006646668538451195,
0.0468389168381691,
0.07905106991529465,
0.029765205457806587,
-0.08290018141269684,
0.024967795237898827,
0.07257074117660522,
-0.009968251921236515,
-0.03059256263077259,
-0.06361843645572662,
-0.030706370249390602,
-0.05203425511717796,
0.053056057542562485,
-0.02798050083220005,
0.10293947905302048,
-0.04517750069499016,
-0.012942608445882797,
-0.0020021589007228613,
-0.014452709816396236,
-0.06043890863656998,
-0.018232805654406548,
-0.020761864259839058,
-0.018335523083806038,
-0.08759947866201401,
0.032845526933670044,
-0.0160906370729208,
-0.019434846937656403,
-0.06590049713850021,
-0.08612758666276932,
0.026991860941052437,
-0.056019168347120285,
-0.027207834646105766,
0.06714432686567307,
-1.644316579089748e-33,
0.030720150098204613,
-0.03343553468585014,
0.03730430826544762,
0.09771500527858734,
-0.06433163583278656,
-0.1272750049829483,
-0.03619539737701416,
0.0031253364868462086,
0.04713983088731766,
-0.09055142104625702,
-0.12946096062660217,
-0.05948387831449509,
0.05907301977276802,
-0.016048038378357887,
0.0499623566865921,
-0.0803292840719223,
0.005223797634243965,
-0.01346131693571806,
0.05073779448866844,
0.13512687385082245,
-0.00393548933789134,
0.022405829280614853,
-0.09422996640205383,
0.08213384449481964,
0.04582810774445534,
0.02606016770005226,
-0.05063120648264885,
0.03530624508857727,
-0.010291414335370064,
-0.028963427990674973,
-0.021893642842769623,
0.008608315140008926,
-0.03190897777676582,
-0.0039797197096049786,
-0.04616127163171768,
-0.00527617335319519,
0.03824459761381149,
0.07539727538824081,
0.0181649848818779,
0.13187764585018158,
0.05721210315823555,
-0.023782722651958466,
-0.016466669738292694,
0.00582096679136157,
-0.0494183786213398,
-0.011695974506437778,
-0.08341071754693985,
-0.021761173382401466,
0.03235296159982681,
-0.013653241097927094,
0.03977634757757187,
0.05491206422448158,
0.006362885236740112,
0.06635001301765442,
0.03278721123933792,
-0.07572728395462036,
-0.07370556145906448,
-0.06558682024478912,
0.06963197886943817,
-0.0058187334798276424,
0.0034924601204693317,
0.022520747035741806,
0.017904961481690407,
-0.09150851517915726,
0.057176049798727036,
-0.10438386350870132,
-0.01476262602955103,
-0.0053460849449038506,
-0.001505075371824205,
0.047140881419181824,
0.03314882889389992,
-0.03912149369716644,
-0.007219476159662008,
0.081410713493824,
0.02762146107852459,
0.020258240401744843,
0.08513975888490677,
-0.035853154957294464,
-0.008725777268409729,
0.00989572424441576,
-0.04413127154111862,
0.10502072423696518,
-0.016813628375530243,
0.07027831673622131,
0.06591076403856277,
-0.0042090024799108505,
0.055459197610616684,
-0.01648390106856823,
-0.01200831588357687,
-0.04086710512638092,
0.0107847535982728,
-0.029379483312368393,
0.07535263895988464,
0.0629924088716507,
-0.10091742873191833,
-4.908930861802219e-8,
-0.0482344925403595,
-0.0032936795614659786,
0.004410953260958195,
0.022775335237383842,
0.0840379074215889,
0.03493420407176018,
-0.10638225078582764,
-0.016998425126075745,
0.017762737348675728,
-0.05826409161090851,
0.12196880578994751,
-0.0007757045677863061,
-0.05589716136455536,
-0.0030740101356059313,
-0.10651885718107224,
0.051921453326940536,
-0.0386488176882267,
-0.020846756175160408,
-0.0350014865398407,
-0.04504542425274849,
-0.03648051247000694,
0.052207205444574356,
0.009070435538887978,
-0.05880044400691986,
0.05525938421487808,
0.05496376380324364,
0.05758945271372795,
0.026748698204755783,
0.0422779805958271,
-0.008270926773548126,
0.00550204049795866,
0.0021153846755623817,
-0.027995409443974495,
-0.031023476272821426,
0.034867167472839355,
0.046982672065496445,
0.06911589950323105,
-0.021037999540567398,
-0.0013876637676730752,
-0.032234277576208115,
0.11409790813922882,
0.04380758851766586,
-0.1013668030500412,
-0.0256629828363657,
0.038161855190992355,
-0.05382369086146355,
-0.007340467069298029,
-0.06690896302461624,
-0.040707748383283615,
-0.0445166639983654,
0.03446711599826813,
-0.0875047892332077,
0.01962433010339737,
0.10245808959007263,
-0.013071316294372082,
0.03217506408691406,
-0.06945645809173584,
0.01762586273252964,
0.018044404685497284,
-0.003752554301172495,
0.06270778179168701,
0.06811143457889557,
-0.05897268280386925,
0.05066223070025444
] | 0 | 0 | 0 | 0 |
|
Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images | https://openreview.net/forum?id=FtjLUHyZAO | [
"Sichen Zhu",
"Yuchen Zhu",
"Molei Tao",
"Peng Qiu"
] | Poster | Spatial Transcriptomics (ST) allows a high-resolution measurement of RNA sequence abundance by systematically connecting cell morphology depicted in Hematoxylin and eosin (H\&E) stained histology images to spatially resolved gene expressions. ST is a time-consuming, expensive yet powerful experimental technique that provides new opportunities to understand cancer mechanisms at a fine-grained molecular level, which is critical for uncovering new approaches for disease diagnosis and treatments. Here, we present $\textbf{Stem}$ ($\underline{\textbf{S}}$pa$\underline{\textbf{T}}$ially resolved gene $\underline{\textbf{E}}$xpression inference with diffusion $\underline{\textbf{M}}$odel), a novel computational tool that leverages a conditional diffusion generative model to enable in silico gene expression inference from H&E stained images. Through better capturing the inherent stochasticity and heterogeneity in ST data, $\textbf{Stem}$ achieves state-of-the-art performance on spatial gene expression prediction and generates biologically meaningful gene profiles for new H&E stained images at test time. We evaluate the proposed algorithm on datasets with various tissue sources and sequencing platforms, where it demonstrates clear improvement over existing approaches. $\textbf{Stem}$ generates high-fidelity gene expression predictions that share similar gene variation levels as ground truth data, suggesting that our method preserves the underlying biological heterogeneity. Our proposed pipeline opens up the possibility of analyzing existing, easily accessible H&E stained histology images from a genomics point of view without physically performing gene expression profiling and empowers potential biological discovery from H&E stained histology images. Code is available at: https://github.com/SichenZhu/Stem. | Gene Expression Prediction, Diffusion Model, Spatial Transcriptomics, H&E | null | 13,234 | 2501.15598 | [
-0.01891828142106533,
-0.04306213557720184,
0.07406605035066605,
-0.039548736065626144,
0.05623370409011841,
-0.016036367043852806,
-0.08132320642471313,
0.036420390009880066,
0.05784684792160988,
0.013236688449978828,
-0.016055451706051826,
-0.0018697159830480814,
-0.053358979523181915,
0.061008550226688385,
-0.06033005937933922,
0.005897180642932653,
-0.1258334219455719,
0.05196265131235123,
-0.043869372457265854,
-0.03682783618569374,
-0.001697347965091467,
0.05561170354485512,
-0.030426640063524246,
-0.013134974054992199,
0.039802685379981995,
0.019104022532701492,
0.09158410131931305,
0.020773762837052345,
-0.010369997471570969,
-0.03936159983277321,
0.05308779701590538,
0.04079554229974747,
-0.02179533988237381,
0.027664506807923317,
-0.004798194859176874,
-0.004617790691554546,
-0.04631352797150612,
-0.01957463100552559,
0.002148690167814493,
-0.03160344809293747,
0.09272493422031403,
0.03632836043834686,
0.025703903287649155,
0.028414851054549217,
0.08423179388046265,
-0.0678848996758461,
0.034481436014175415,
-0.03775172308087349,
0.025511249899864197,
0.012104646302759647,
0.011707281693816185,
-0.024124015122652054,
-0.077387236058712,
0.03194521367549896,
-0.022808397188782692,
0.057521190494298935,
0.09179741889238358,
-0.059620652347803116,
0.029825888574123383,
-0.04931708052754402,
0.004395885393023491,
-0.06637750566005707,
-0.012665735557675362,
-0.025487568229436874,
-0.0020807806868106127,
0.021826650947332382,
0.08459404855966568,
0.060165297240018845,
-0.036628298461437225,
-0.012237084098160267,
-0.038923803716897964,
0.0984807163476944,
-0.06952007114887238,
0.05405949428677559,
-0.0008646236383356154,
-0.026674503460526466,
0.016025446355342865,
0.01980004645884037,
0.09665557742118835,
-0.06792891770601273,
0.0160232987254858,
0.05819733440876007,
0.1003415435552597,
-0.06464218348264694,
-0.019523873925209045,
0.07107330113649368,
-0.06022975593805313,
0.08353453129529953,
0.04245812073349953,
0.024986380711197853,
0.040858808904886246,
0.007584380451589823,
-0.11918263882398605,
0.010139025747776031,
0.007127001881599426,
0.041397515684366226,
0.03877858445048332,
0.008094042539596558,
0.08850590884685516,
0.031856950372457504,
0.006139720790088177,
0.03484503552317619,
-0.016014734283089638,
-0.011559550650417805,
0.05149199441075325,
-0.02163076028227806,
0.08185388147830963,
0.05940745398402214,
0.026814807206392288,
0.031885046511888504,
0.035120636224746704,
-0.06073419377207756,
0.040906235575675964,
0.04950626194477081,
0.0016636387445032597,
-0.07981317490339279,
0.03473489359021187,
-0.02397279627621174,
0.06439865380525589,
-0.055386733263731,
-0.030111273750662804,
-0.03220712020993233,
-0.09559202194213867,
0.0406455434858799,
0.046148575842380524,
0.018197575584053993,
-0.07549943774938583,
-6.457390412652593e-34,
0.048699360340833664,
-0.02489815093576908,
0.039260346442461014,
0.06574615091085434,
0.02096213400363922,
0.013802134431898594,
-0.04312378913164139,
-0.07327453047037125,
-0.08725209534168243,
-0.002213595900684595,
-0.13545255362987518,
-0.03698093444108963,
-0.07435889542102814,
0.02227351814508438,
-0.02373422123491764,
-0.021281830966472626,
-0.04392990097403526,
0.11313480138778687,
-0.11215086281299591,
0.02125835232436657,
0.0026021532248705626,
-0.06326165050268173,
-0.08887633681297302,
-0.06331730633974075,
-0.037784986197948456,
-0.011989153921604156,
-0.03157835826277733,
-0.02880050800740719,
0.03264608979225159,
0.006556037813425064,
-0.1501186490058899,
-0.02684883214533329,
-0.0066375453025102615,
0.0386774055659771,
0.07581919431686401,
-0.014332878403365612,
-0.015320497564971447,
-0.030369384214282036,
0.022895880043506622,
0.012435783632099628,
0.014687269926071167,
0.05842055380344391,
0.02818794548511505,
-0.02227964997291565,
-0.07043619453907013,
0.0016534040914848447,
0.004603820852935314,
0.0006099132006056607,
0.025480804964900017,
-0.012875983491539955,
0.06197685748338699,
-0.006744381505995989,
0.059134699404239655,
-0.04245882108807564,
-0.019386975094676018,
0.001093258266337216,
-0.019822629168629646,
-0.049530163407325745,
0.0711086243391037,
0.050140757113695145,
0.059847358614206314,
0.006711948662996292,
0.01215015072375536,
0.10457620024681091,
0.1121784895658493,
-0.021409140899777412,
-0.10863099247217178,
0.07625047862529755,
-0.013904407620429993,
0.11596350371837616,
0.015845712274312973,
0.04679479822516441,
0.04134714603424072,
-0.11729218810796738,
0.042625103145837784,
0.04373018071055412,
0.002441201126202941,
0.0697864443063736,
-0.0387592650949955,
0.040555667132139206,
-0.06774668395519257,
-0.02967536263167858,
-0.1317109912633896,
-0.062464721500873566,
-0.01710878685116768,
-0.009592433460056782,
0.010037067346274853,
-0.04733668640255928,
-0.10650676488876343,
-0.03517445921897888,
0.07994693517684937,
-0.026167359203100204,
-0.007527399808168411,
-0.04155643656849861,
0.028877289965748787,
-7.352676756775099e-34,
-0.02949446626007557,
0.07088206708431244,
0.0539151169359684,
-0.06126905605196953,
-0.07843391597270966,
0.04512675851583481,
0.04247567057609558,
-0.00169319286942482,
0.0916181281208992,
-0.045081306248903275,
0.04444684460759163,
0.07289112359285355,
-0.0697273388504982,
-0.00591756496578455,
0.015127136372029781,
-0.0005380170769058168,
0.009373467415571213,
0.022988982498645782,
-0.034765925258398056,
0.026752404868602753,
-0.06253605335950851,
0.08424228429794312,
-0.053481291979551315,
0.024330712854862213,
-0.07839547097682953,
0.03447340801358223,
-0.029881849884986877,
0.0542137585580349,
-0.08013439178466797,
-0.024067139253020287,
-0.03452128916978836,
-0.03921494260430336,
-0.050680678337812424,
-0.023333165794610977,
-0.07919277250766754,
0.06985121220350266,
0.018029188737273216,
-0.03943236172199249,
0.053758587688207626,
-0.04481325298547745,
0.03149258345365524,
-0.03694065660238266,
-0.12111125141382217,
0.058345306664705276,
0.04315340146422386,
0.05902500078082085,
-0.03817015513777733,
0.03409809619188309,
0.09116112440824509,
0.030322743579745293,
-0.0428621806204319,
0.10552296042442322,
-0.0312325619161129,
0.09478037059307098,
-0.018350079655647278,
-0.04114370048046112,
-0.05559800565242767,
-0.05562078580260277,
-0.04542029649019241,
-0.07552172243595123,
-0.10946020483970642,
-0.06645751744508743,
-0.03959311917424202,
0.013719037175178528,
-0.08521588146686554,
0.00778923137113452,
-0.0015086759813129902,
-0.016681846231222153,
0.025017336010932922,
0.008318973705172539,
-0.020164230838418007,
0.060343798249959946,
0.030455149710178375,
0.022956889122724533,
0.024667058140039444,
-0.07069528102874756,
0.04537736624479294,
-0.01777760498225689,
0.0030179410241544247,
0.00968237780034542,
0.030340062454342842,
-0.03460242226719856,
0.030324170365929604,
-0.0812658742070198,
0.07714197784662247,
-0.010873169638216496,
-0.023565704002976418,
-0.011451523751020432,
0.02577878162264824,
0.006561135873198509,
0.017702147364616394,
0.0035404253285378218,
0.005949453916400671,
-0.044372305274009705,
-0.05979947745800018,
-4.721381330341501e-8,
-0.02598688378930092,
-0.003717835294082761,
0.0484718382358551,
-0.07176923006772995,
0.01873031258583069,
0.00975814275443554,
-0.0117759695276618,
0.11310787498950958,
0.07457089424133301,
0.01220609713345766,
-0.0021441979333758354,
-0.048864591866731644,
-0.07501732558012009,
-0.082310251891613,
0.07939626276493073,
0.052705299109220505,
0.010771466419100761,
-0.025317609310150146,
-0.04752456769347191,
0.022203542292118073,
-0.08072751760482788,
0.04390958324074745,
-0.009217886254191399,
0.06598518043756485,
0.0531151257455349,
-0.009980321861803532,
0.0164854247123003,
-0.004701347090303898,
0.08290477842092514,
-0.022868942469358444,
0.08643783628940582,
0.01414717361330986,
-0.0011500284308567643,
0.05637414753437042,
-0.03570173308253288,
0.03288319334387779,
0.09987559914588928,
0.013477605767548084,
-0.02293161116540432,
-0.013661501929163933,
-0.040230706334114075,
0.008484293706715107,
-0.05723942816257477,
-0.052882567048072815,
-0.004713988862931728,
0.05078895762562752,
0.07767327129840851,
0.048597920686006546,
0.016920339316129684,
-0.03297528252005577,
0.0016454958822578192,
0.01676439307630062,
-0.05795515328645706,
0.05601218715310097,
0.002497062785550952,
-0.017016848549246788,
-0.006268794182687998,
0.030168011784553528,
0.07024919986724854,
0.030736370012164116,
0.034593839198350906,
-0.024376820772886276,
0.017763271927833557,
-0.011710827238857746
] | https://github.com/SichenZhu/Stem | 25 | 0 | 0 | 0 |
Diffusion State-Guided Projected Gradient for Inverse Problems | https://openreview.net/forum?id=kRBQwlkFSP | [
"Rayhan Zirvi",
"Bahareh Tolooshams",
"Anima Anandkumar"
] | Poster | Recent advancements in diffusion models have been effective in learning data priors for solving inverse problems. They leverage diffusion sampling steps for inducing a data prior while using a measurement guidance gradient at each step to impose data consistency. For general inverse problems, approximations are needed when an unconditionally trained diffusion model is used since the measurement likelihood is intractable, leading to inaccurate posterior sampling. In other words, due to their approximations, these methods fail to preserve the generation process on the data manifold defined by the diffusion prior, leading to artifacts in applications such as image restoration. To enhance the performance and robustness of diffusion models in solving inverse problems, we propose Diffusion State-Guided Projected Gradient (DiffStateGrad), which projects the measurement gradient onto a subspace that is a low-rank approximation of an intermediate state of the diffusion process. DiffStateGrad, as a module, can be added to a wide range of diffusion-based inverse solvers to improve the preservation of the diffusion process on the prior manifold and filter out artifact-inducing components. We highlight that DiffStateGrad improves the robustness of diffusion models in terms of the choice of measurement guidance step size and noise while improving the worst-case performance. Finally, we demonstrate that DiffStateGrad improves upon the state-of-the-art on linear and nonlinear image restoration inverse problems. Our code is available at https://github.com/Anima-Lab/DiffStateGrad. | Diffusion models, Inverse problems, Robustness, Subspace, Projection, Box inpainting, Phase retrieval | We can improve the performance and robustness of diffusion-based models in solving inverse problems by adding a Diffusion State-Guided Projected Gradient step. | 13,232 | 2410.03463 | [
-0.04992985352873802,
-0.061543796211481094,
0.09991063177585602,
0.049909137189388275,
0.02962861768901348,
-0.07331659644842148,
-0.0049433778040111065,
-0.08061668276786804,
-0.0030805980786681175,
0.008299837820231915,
-0.02210143767297268,
0.01707177609205246,
0.013363761827349663,
0.03397374600172043,
-0.08719527721405029,
-0.029080551117658615,
0.04185988008975983,
0.05750766023993492,
-0.06457270681858063,
-0.07539577782154083,
-0.06131475418806076,
0.019796637818217278,
-0.061381012201309204,
-0.0176630150526762,
0.06125061959028244,
0.03827403858304024,
0.07400286942720413,
-0.031955536454916,
-0.008281444199383259,
-0.012501451186835766,
0.05060287564992905,
-0.05423780158162117,
-0.04183368757367134,
-0.02084127999842167,
-0.06128077581524849,
0.07193789631128311,
-0.0017671267269179225,
0.03139324486255646,
0.006230644881725311,
-0.02155238203704357,
0.008501861244440079,
0.09046974033117294,
-0.001486757304519415,
-0.00226262747310102,
0.06799217313528061,
-0.053416717797517776,
0.009182434529066086,
-0.07252117991447449,
0.04018187150359154,
0.03346855565905571,
0.02293625846505165,
0.02271895669400692,
-0.06575961410999298,
0.01002353522926569,
0.0069644758477807045,
0.054428793489933014,
0.11200745403766632,
0.00875437818467617,
-0.0038952846080064774,
0.019545799121260643,
-0.052713651210069656,
-0.05535972863435745,
-0.06903162598609924,
-0.03345672786235809,
-0.020850805565714836,
0.03922419622540474,
0.044224318116903305,
0.023652806878089905,
-0.01874404586851597,
0.1040554940700531,
-0.07025842368602753,
0.06313474476337433,
0.005680210422724485,
-0.03340880200266838,
0.0021283808164298534,
-0.054699450731277466,
0.027407020330429077,
0.021400086581707,
0.027857765555381775,
-0.055961914360523224,
0.09815622121095657,
0.0212841909378767,
0.06454428285360336,
-0.025672808289527893,
0.06432682275772095,
0.01166874822229147,
-0.017615336924791336,
-0.008813806809484959,
0.056562136858701706,
0.010747530497610569,
-0.04269162937998772,
-0.04261700063943863,
-0.15654678642749786,
0.007694634143263102,
0.0029751204419881105,
-0.026907118037343025,
0.04397553950548172,
-0.020932845771312714,
0.07077633589506149,
0.07397201657295227,
-0.0019591273739933968,
-0.0006637241458520293,
-0.03155418485403061,
0.021927597001194954,
-0.012320934794843197,
-0.02314356528222561,
0.04199850559234619,
0.06833577156066895,
-0.011413119733333588,
0.019220048561692238,
0.05319624021649361,
-0.02435857057571411,
0.0762493684887886,
0.03886254504323006,
0.07542262971401215,
-0.10576185584068298,
0.04127722233533859,
-0.013743540272116661,
-0.02913227491080761,
0.005246687214821577,
-0.01846214011311531,
-0.012961480766534805,
-0.005128501448780298,
-0.005777814891189337,
0.06521084159612656,
-0.008279346860945225,
0.022044749930500984,
2.1551324246853734e-33,
-0.0002608993381727487,
-0.0373416393995285,
0.023970603942871094,
0.04789349064230919,
0.03719572350382805,
0.03534109145402908,
0.016679929569363594,
-0.031072378158569336,
-0.014437111094594002,
-0.049650389701128006,
-0.010269830003380775,
-0.019952137023210526,
-0.14359962940216064,
0.06713725626468658,
0.011796058155596256,
0.01958906650543213,
-0.000851245247758925,
0.10538903623819351,
-0.06131419911980629,
-0.028912631794810295,
0.0006723401020281017,
-0.03750522807240486,
-0.04878419265151024,
-0.08752426505088806,
0.020980004221200943,
-0.01749899424612522,
0.0641586035490036,
-0.018606262281537056,
-0.026288552209734917,
-0.002092126291245222,
-0.10759302973747253,
-0.04966675862669945,
-0.01576150208711624,
0.07421505451202393,
-0.007652441505342722,
0.03877411410212517,
0.07142798602581024,
-0.0024717419873923063,
-0.03688570857048035,
-0.0456085205078125,
0.058244772255420685,
0.08399627357721329,
0.023804977536201477,
-0.007200170308351517,
-0.13686829805374146,
-0.061240844428539276,
-0.031144345179200172,
-0.008363182656466961,
0.011736549437046051,
-0.08774573355913162,
-0.030742745846509933,
-0.07113434374332428,
0.037174660712480545,
-0.13506096601486206,
-0.08354341238737106,
-0.022294459864497185,
0.052364032715559006,
-0.045186497271060944,
0.05315376818180084,
-0.003569314954802394,
0.08897767961025238,
0.024126585572957993,
-0.04827339947223663,
0.06198810040950775,
0.02240057662129402,
0.042868342250585556,
-0.011670961044728756,
0.030733546242117882,
0.005329667590558529,
0.024949802085757256,
-0.042370423674583435,
-0.015666959807276726,
-0.0028890115208923817,
-0.014394531957805157,
0.12317237257957458,
0.04157349094748497,
0.050043314695358276,
0.016496025025844574,
0.029509078711271286,
-0.049546439200639725,
-0.0640539601445198,
-0.03665183484554291,
-0.048079315572977066,
-0.0348472073674202,
-0.03752877563238144,
-0.0031168796122074127,
0.013602207414805889,
-0.05893217772245407,
-0.06261768192052841,
-0.07533315569162369,
-0.059334997087717056,
-0.030035190284252167,
-0.06430750340223312,
0.05812299996614456,
0.028914807364344597,
-1.6583574917683357e-33,
-0.0010126458946615458,
0.0776638388633728,
0.091628298163414,
0.020119858905673027,
-0.04003935307264328,
-0.04963236674666405,
0.037708740681409836,
0.06377442926168442,
0.0867755264043808,
-0.10290930420160294,
-0.06184829771518707,
-0.03918345272541046,
-0.016256174072623253,
0.01965136080980301,
-0.01700981706380844,
0.0016258099349215627,
-0.02824215032160282,
-0.03922584652900696,
-0.012627819553017616,
0.011705907061696053,
-0.0885239690542221,
0.08775145560503006,
-0.021752163767814636,
0.04728714004158974,
-0.07208096235990524,
0.05915697291493416,
-0.009396563284099102,
0.080269455909729,
0.017290562391281128,
-0.07377501577138901,
-0.0541929267346859,
-0.007739518769085407,
0.038932837545871735,
-0.029406046494841576,
-0.12110961973667145,
0.15411461889743805,
0.010934684425592422,
-0.037027087062597275,
-0.04545438662171364,
0.009218968451023102,
0.03822093456983566,
-0.006362409330904484,
-0.06713423132896423,
0.038852084428071976,
0.06705430895090103,
-0.06090368703007698,
0.04118756577372551,
-0.000688843138050288,
0.009108911268413067,
0.018934445455670357,
-0.037196747958660126,
0.07049842923879623,
-0.04716355726122856,
0.0776854157447815,
-0.0731550008058548,
-0.006475836504250765,
-0.06128190830349922,
-0.000011168081982759759,
0.13095985352993011,
0.024261154234409332,
-0.007978152483701706,
-0.03001871332526207,
0.04134388640522957,
-0.013413992710411549,
-0.04320814087986946,
0.0879799872636795,
-0.02707039937376976,
0.04731359705328941,
0.014759628102183342,
0.06386958807706833,
-0.04654406011104584,
0.06118835508823395,
0.05891413986682892,
-0.018551906570792198,
0.07040354609489441,
-0.11514188349246979,
-0.012867183424532413,
0.0035255535040050745,
0.009154869243502617,
0.016615627333521843,
0.0747828334569931,
-0.015148119069635868,
-0.024311108514666557,
0.027916541323065758,
0.028996648266911507,
-0.012644827365875244,
0.005519961006939411,
0.04157617315649986,
-0.00012912860256619751,
0.002363675506785512,
-0.0015244090463966131,
-0.054763879626989365,
0.006207870785146952,
-0.03180583193898201,
-0.017022712156176567,
-4.144735754607609e-8,
-0.13582845032215118,
0.05916360020637512,
0.07602129876613617,
-0.014851375482976437,
0.0379057377576828,
0.003543515456840396,
0.07215334475040436,
0.08786657452583313,
0.03993481397628784,
-0.10723263770341873,
0.012986504472792149,
-0.032003432512283325,
-0.0521354153752327,
-0.0886516273021698,
-0.08563630282878876,
0.12076325714588165,
0.016662264242768288,
0.04789707437157631,
-0.02276710979640484,
0.00714710121974349,
-0.0016765834297984838,
0.038377366960048676,
0.025430435314774513,
-0.003998504485934973,
0.07929616421461105,
-0.028486361727118492,
-0.03068944625556469,
0.02108556404709816,
0.10470792651176453,
-0.014564266428351402,
-0.013297236524522305,
-0.005549953319132328,
0.0689324140548706,
0.056003015488386154,
-0.04671159386634827,
0.07300079613924026,
-0.004692982882261276,
-0.018137548118829727,
-0.042659927159547806,
-0.10464460402727127,
-0.07445725798606873,
0.032063111662864685,
-0.010255102068185806,
0.013386141508817673,
0.031959474086761475,
0.09519440680742264,
0.02557414211332798,
0.008085867390036583,
0.007658044807612896,
-0.015607516281306744,
0.06489423662424088,
-0.047961678355932236,
0.02929927408695221,
0.03353556990623474,
0.04888506978750229,
-0.044688738882541656,
0.032365988940000534,
-0.0024656972382217646,
0.028154488652944565,
0.04491642117500305,
-0.009296642616391182,
0.016360564157366753,
0.03210429102182388,
0.043067362159490585
] | https://github.com/anima-lab/diffstategrad | 7 | 0 | 0 | 0 |
MADGEN: Mass-Spec attends to De Novo Molecular generation | https://openreview.net/forum?id=78tc3EiUrN | [
"Yinkai Wang",
"Xiaohui Chen",
"Liping Liu",
"Soha Hassoun"
] | Poster | The annotation (assigning structural chemical identities) of MS/MS spectra remains a significant challenge due to the enormous molecular diversity in biological samples and the limited scope of reference databases. Currently, the vast majority of spectral measurements remain in the "dark chemical space" without structural annotations. To improve annotation, we propose MADGEN (Mass-spec Attends to De Novo Molecular GENeration), a scaffold-based method for de novo molecular structure generation guided by mass spectrometry data. MADGEN operates in two stages: scaffold retrieval and spectra-conditioned molecular generation starting with the scaffold. In the first stage, given an MS/MS spectrum, we formulate scaffold retrieval as a ranking problem and employ contrastive learning to align mass spectra with candidate molecular scaffolds. In the second stage, starting from the retrieved scaffold, we employ the MS/MS spectrum to guide an attention-based generative model to generate the final molecule. Our approach constrains the molecular generation search space, reducing its complexity and improving generation accuracy. We evaluate MADGEN on three datasets (NIST23, CANOPUS, and MassSpecGym) and evaluate MADGEN's performance with a predictive scaffold retriever and with an oracle retriever. We demonstrate the effectiveness of using attention to integrate spectral information throughout the generation process to achieve strong results with the oracle retriever. | AI4Science, Biology Discovery, Metabolomics, MS/MS spectra | A novel framework for De Novo molecule generation with classifier-free guidance from mass spectra. | 13,230 | 2501.01950 | [
-0.09741999953985214,
-0.10468114912509918,
-0.035946693271398544,
0.019658584147691727,
0.05017409846186638,
0.02699809893965721,
-0.02403055876493454,
-0.008816178888082504,
0.009759060107171535,
-0.017750557512044907,
-0.007460843771696091,
-0.07298681885004044,
-0.03168226778507233,
0.015434525907039642,
-0.07543028146028519,
0.09096000343561172,
0.005743327550590038,
-0.0036466498859226704,
0.01990610361099243,
0.013469971716403961,
0.10905535519123077,
0.04882071912288666,
-0.006210148334503174,
0.012982230633497238,
0.04008620232343674,
-0.0007799433660693467,
-0.025869270786643028,
0.05641942471265793,
0.003770815208554268,
-0.01012292318046093,
0.13885486125946045,
0.05884099379181862,
0.034164901822805405,
-0.0214775949716568,
0.06231502816081047,
0.023882804438471794,
-0.1278248280286789,
0.07466616481542587,
0.02125590667128563,
-0.05351528152823448,
0.04693495109677315,
0.04011731594800949,
0.0016077463515102863,
0.009127197787165642,
0.042845163494348526,
-0.08347655832767487,
-0.014031630009412766,
-0.02852044440805912,
0.039870984852313995,
0.016296930611133575,
-0.05549948662519455,
-0.02103872410953045,
-0.06347883492708206,
0.0363864004611969,
0.023860635235905647,
0.052515026181936264,
0.013997073285281658,
-0.08333379030227661,
0.0018999457824975252,
-0.002467144513502717,
0.01795588992536068,
-0.11039315909147263,
-0.014276792295277119,
-0.09025673568248749,
0.06812508404254913,
0.012634885497391224,
0.08371482789516449,
0.02521996758878231,
0.041695352643728256,
-0.0203462615609169,
-0.05989495664834976,
0.08760889619588852,
-0.030306970700621605,
0.1423032134771347,
0.04197479411959648,
0.16367396712303162,
0.043524276465177536,
0.0335049107670784,
0.04594931751489639,
-0.03658997640013695,
0.012352081947028637,
0.06397224962711334,
0.06016499921679497,
0.004543515853583813,
0.11231938004493713,
0.03363429754972458,
0.041948352009058,
0.09574814885854721,
-0.0005262453923933208,
0.015339924953877926,
-0.03847572207450867,
-0.05175355449318886,
-0.004059665836393833,
-0.03190578147768974,
-0.06076047569513321,
0.03087383322417736,
0.08345793932676315,
0.04000129923224449,
0.07217452675104141,
0.007222505286335945,
0.06076003983616829,
0.0712529793381691,
0.012077270075678825,
-0.026958346366882324,
-0.03768308833241463,
0.0020631749648600817,
0.05877146124839783,
0.053009532392024994,
0.06386521458625793,
-0.03674681857228279,
0.061811018735170364,
0.05601072683930397,
0.02667599730193615,
-0.0037266756407916546,
0.04010318964719772,
0.029436122626066208,
0.0009168272372335196,
-0.05049411207437515,
0.018385551869869232,
-0.0021361205726861954,
-0.030702004209160805,
0.023586779832839966,
-0.043103303760290146,
0.018733074888586998,
-0.00649087131023407,
0.07961136102676392,
-0.11811815947294235,
2.2673519112415846e-33,
0.05775213614106178,
0.04320328310132027,
0.06403376907110214,
0.10767394304275513,
0.0170857273042202,
-0.11027845740318298,
-0.04220901057124138,
-0.05174586549401283,
-0.07456126809120178,
0.038027651607990265,
-0.018062645569443703,
0.005515349563211203,
-0.0258832648396492,
0.04364308714866638,
-0.03296547383069992,
-0.02874540165066719,
0.004730756860226393,
0.03138619661331177,
-0.04199285805225372,
-0.06953656673431396,
-0.09049514681100845,
0.03649669140577316,
-0.007312373723834753,
-0.11050920933485031,
0.05361441150307655,
0.053048066794872284,
0.007130009587854147,
-0.0010690585477277637,
-0.04384743049740791,
-0.011488941498100758,
-0.08146295696496964,
-0.0281843114644289,
-0.00508455652743578,
0.04066545516252518,
0.039388153702020645,
-0.028441375121474266,
0.043564844876527786,
-0.08958571404218674,
0.011297177523374557,
-0.009278728626668453,
0.038037799298763275,
0.09201312810182571,
0.01256377249956131,
0.04208262264728546,
-0.065301813185215,
-0.028743833303451538,
-0.004270326811820269,
-0.03996031731367111,
0.03708045557141304,
-0.03739447519183159,
-0.05736313760280609,
-0.009819979779422283,
0.05609182268381119,
-0.027549833059310913,
-0.10981709510087967,
0.017151404172182083,
-0.0034152811858803034,
-0.01747896522283554,
0.11282806843519211,
0.01618231274187565,
-0.03848137706518173,
0.03303917497396469,
-0.030608119443058968,
0.10217931866645813,
0.07607311755418777,
0.03253740444779396,
-0.06194283813238144,
0.02940354123711586,
0.04921434074640274,
0.08150465786457062,
0.01882922649383545,
0.026683958247303963,
0.044935062527656555,
-0.05736793950200081,
0.06892639398574829,
0.01420726627111435,
0.014768658205866814,
-0.005429754499346018,
-0.04268721491098404,
0.030810795724391937,
0.015420579351484776,
0.056056853383779526,
-0.11056037992238998,
-0.034483782947063446,
-0.10578212887048721,
-0.07383009046316147,
0.049755264073610306,
-0.08659310638904572,
-0.037744589149951935,
-0.07069402933120728,
-0.036597032099962234,
-0.011468479409813881,
-0.0081478301435709,
-0.050321176648139954,
-0.08674807101488113,
-2.0744180234411226e-33,
0.05384662002325058,
0.03334015607833862,
0.05798610672354698,
-0.0060821217484772205,
-0.005454291123896837,
0.030458034947514534,
-0.003139376873150468,
-0.03439405560493469,
-0.07624224573373795,
-0.042294494807720184,
0.0445706732571125,
-0.038034699857234955,
0.024778008460998535,
-0.13132557272911072,
0.028831642121076584,
-0.033955276012420654,
-0.026873819530010223,
0.021303102374076843,
0.025200581178069115,
0.07343222200870514,
-0.039276283234357834,
0.07570987939834595,
-0.04483518376946449,
0.008637013845145702,
-0.02348204515874386,
0.056279029697179794,
0.04636124148964882,
-0.012288939207792282,
-0.038655124604701996,
-0.02862200327217579,
-0.008677772246301174,
0.030741510912775993,
-0.05233997479081154,
-0.07042723894119263,
-0.04133353754878044,
0.09141756594181061,
0.040908731520175934,
-0.03360346332192421,
-0.012678205966949463,
0.05685826763510704,
-0.0022500355262309313,
0.058091551065444946,
-0.055394455790519714,
0.04029333218932152,
-0.005055648274719715,
0.025127625092864037,
-0.033911485224962234,
0.030511170625686646,
0.05225907266139984,
-0.026195626705884933,
0.042711060494184494,
0.00039792360621504486,
-0.06375692039728165,
-0.06760773062705994,
-0.0028699510730803013,
-0.09328349679708481,
-0.033660732209682465,
-0.03601047769188881,
0.000031446554203284904,
-0.004908502567559481,
-0.049817245453596115,
0.016533944755792618,
-0.04158699885010719,
-0.07455087453126907,
-0.0277889221906662,
0.026205992326140404,
-0.03873706981539726,
-0.04548284411430359,
-0.05087307095527649,
-0.028414040803909302,
0.011911813169717789,
-0.009999238885939121,
0.062174879014492035,
0.0644773542881012,
0.0573975034058094,
-0.025213759392499924,
-0.07467148452997208,
-0.03532889857888222,
-0.05691488832235336,
-0.016224972903728485,
0.07659188657999039,
-0.0382576584815979,
-0.00805273000150919,
0.0355234295129776,
0.0625586211681366,
0.005672168452292681,
0.030661551281809807,
0.06266512721776962,
-0.011071844957768917,
-0.03289784491062164,
-0.038783639669418335,
-0.009538991376757622,
-0.03336917236447334,
0.0851793959736824,
0.000762515643145889,
-4.062136227389601e-8,
0.01780281402170658,
-0.0035093845799565315,
0.05574070289731026,
-0.007090236525982618,
-0.026783520355820656,
0.009661812335252762,
-0.05569799616932869,
0.019600646570324898,
0.002725467551499605,
0.008381844498217106,
0.07317652553319931,
-0.06113368272781372,
-0.07666587829589844,
0.009854835458099842,
-0.0066617997363209724,
0.022844798862934113,
-0.023175787180662155,
-0.021443098783493042,
-0.05031486600637436,
0.001593679771758616,
-0.02753387950360775,
0.0970974937081337,
-0.002173927379772067,
-0.02697405032813549,
0.12490111589431763,
-0.06697321683168411,
-0.022741122171282768,
0.07067710906267166,
0.03585765138268471,
-0.05162118747830391,
0.05172586068511009,
0.022119488567113876,
-0.03977269306778908,
0.07785671949386597,
-0.010151450522243977,
-0.013806493952870369,
-0.07110000401735306,
0.02773510105907917,
0.006457753479480743,
-0.00686621991917491,
-0.030907299369573593,
0.06247159093618393,
-0.1151987686753273,
0.01769830472767353,
-0.07358871400356293,
0.0023767470847815275,
-0.07621602714061737,
0.023813921958208084,
-0.022259173914790154,
-0.027985934168100357,
0.0031154463067650795,
-0.11642219126224518,
-0.014789615757763386,
0.0048217326402664185,
-0.03216348960995674,
0.027570359408855438,
0.010995753109455109,
0.04933277145028114,
0.03143619745969772,
-0.08376402407884598,
0.07425942271947861,
-0.05434505268931389,
-0.05887196213006973,
-0.03340018913149834
] | https://github.com/HassounLab/MADGEN | 4 | 0 | 0 | 0 |
Towards Federated RLHF with Aggregated Client Preference for LLMs | https://openreview.net/forum?id=mqNKiEB6pd | [
"Feijie Wu",
"Xiaoze Liu",
"Haoyu Wang",
"Xingchen Wang",
"Lu Su",
"Jing Gao"
] | Poster | Reinforcement learning with human feedback (RLHF) fine-tunes a pretrained large language model (LLM) using user preference data, enabling it to generate content aligned with human preferences. However, due to privacy concerns, users may be reluctant to share sensitive preference data. To address this, we propose utilizing Federated Learning (FL) techniques, allowing large-scale preference collection from diverse real-world users without requiring them to transmit data to a central server. Our federated RLHF methods (i.e., FedBis and FedBiscuit) encode each client’s preferences into binary selectors and aggregate them to capture common preferences. In particular, FedBiscuit overcomes key challenges, such as preference heterogeneity and reward hacking, through innovative solutions like grouping clients with similar preferences to reduce heterogeneity and using multiple binary selectors to enhance LLM output quality. To evaluate the performance of the proposed methods, we establish the first federated RLHF benchmark with a heterogeneous human preference dataset. Experimental results show that by integrating the LLM with aggregated client preferences, FedBis and FedBiscuit significantly enhance the professionalism and readability of the generated content. | Federated learning, RLHF, LLM | null | 13,225 | 2407.03038 | [
-0.03151808679103851,
-0.08521696925163269,
0.0010488404659554362,
0.023621691390872,
0.02251189388334751,
0.03909148648381233,
-0.02691175974905491,
-0.030109146609902382,
-0.021456368267536163,
-0.015498400665819645,
-0.04610750824213028,
-0.04773910343647003,
0.09254980832338333,
-0.02700706385076046,
0.0739951878786087,
0.03162721171975136,
0.11126888543367386,
0.01688550040125847,
-0.07051758468151093,
-0.05958167836070061,
-0.0715872049331665,
-0.05158955976366997,
0.015775704756379128,
-0.031381793320178986,
-0.016343140974640846,
-0.015505729243159294,
0.017592670395970345,
-0.019440745934844017,
-0.03368902578949928,
-0.021266315132379532,
0.0838877335190773,
0.052341312170028687,
-0.03251168876886368,
-0.04276527464389801,
-0.09053883701562881,
0.06978178769350052,
-0.10840607434511185,
-0.018867645412683487,
-0.0561247281730175,
-0.05587727203965187,
-0.0800166204571724,
-0.059368740767240524,
0.013446284458041191,
0.05165306478738785,
0.002301628701388836,
0.05594320595264435,
-0.009967007674276829,
0.00012365453585516661,
0.0029888951685279608,
0.05458207055926323,
0.01188910473138094,
-0.029053546488285065,
0.03897123783826828,
0.05067628622055054,
0.009310686029493809,
-0.008452821522951126,
-0.05562092736363411,
-0.005560438614338636,
-0.0030684694647789,
-0.004672191105782986,
-0.1059875637292862,
-0.12139319628477097,
-0.02509903721511364,
0.030192451551556587,
-0.049471545964479446,
-0.021427107974886894,
0.008849529549479485,
0.04997863993048668,
0.0059507121331989765,
-0.022475453093647957,
-0.014580294489860535,
0.027040814980864525,
-0.012799998745322227,
-0.018159614875912666,
0.02017146348953247,
0.05915292352437973,
-0.029750697314739227,
-0.027135400101542473,
0.05958239734172821,
-0.03784537687897682,
-0.07947172969579697,
0.07024723291397095,
-0.0045584458857774734,
-0.02511793188750744,
0.08314958959817886,
-0.01941526308655739,
0.03133599832653999,
0.027913056313991547,
-0.05547523871064186,
0.059141017496585846,
-0.015122349373996258,
0.0840887501835823,
0.0713198333978653,
-0.009269539266824722,
-0.0606621690094471,
0.03786352649331093,
-0.03455807641148567,
-0.062459494918584824,
-0.03212364390492439,
0.09926737844944,
-0.016173498705029488,
0.08862686902284622,
-0.06520703434944153,
0.019604654982686043,
-0.011511915363371372,
-0.001911464729346335,
-0.04374776408076286,
0.075456403195858,
0.03388749808073044,
-0.05598275735974312,
-0.024713538587093353,
0.07401295751333237,
-0.025766244158148766,
0.02357552945613861,
-0.005646975245326757,
-0.004185123834758997,
0.03374127671122551,
-0.0035336799919605255,
0.017714792862534523,
0.024102654308080673,
-0.013065574690699577,
-0.0771176740527153,
0.031985338777303696,
0.01487149391323328,
0.00960765965282917,
-0.021154217422008514,
-0.03606854006648064,
2.4516405891047257e-33,
0.08577841520309448,
0.026606544852256775,
-0.014210939407348633,
-0.06795758008956909,
0.014694906771183014,
-0.00816286914050579,
0.05965075641870499,
0.09203086793422699,
-0.14867401123046875,
-0.012114513665437698,
0.10203435271978378,
0.08131381124258041,
0.016856105998158455,
0.09492229670286179,
0.048690810799598694,
-0.05408827215433121,
0.007522427011281252,
0.06332655996084213,
0.07902003079652786,
-0.09695373475551605,
0.0327756442129612,
0.06508773565292358,
0.06102540343999863,
-0.017982877790927887,
0.002706063911318779,
0.00197159918025136,
0.038743216544389725,
0.025563975796103477,
0.054106902331113815,
0.021181775256991386,
0.01800711452960968,
-0.08489974588155746,
0.0190926194190979,
0.062332071363925934,
-0.06467226147651672,
0.004731831140816212,
-0.04176446050405502,
-0.054110363125801086,
-0.018190637230873108,
-0.027340829372406006,
0.0604582317173481,
-0.033001773059368134,
0.06296747922897339,
-0.02362019009888172,
-0.1603686511516571,
0.02004251442849636,
-0.02022312767803669,
-0.00010239686525892466,
-0.0066138990223407745,
-0.034661807119846344,
0.05097603052854538,
-0.009517727419734001,
0.01811549812555313,
0.005644810386002064,
-0.007202159147709608,
0.0011556037934496999,
-0.04886721819639206,
0.06083079054951668,
0.006993085145950317,
0.032331038266420364,
-0.06809349358081818,
0.046964485198259354,
0.029972126707434654,
0.036392875015735626,
0.002605596324428916,
0.014004465192556381,
-0.005674554035067558,
-0.03288446366786957,
0.04533322900533676,
-0.02664273977279663,
0.039139945060014725,
0.08539608120918274,
0.038742151111364365,
-0.008980793878436089,
0.047151561826467514,
-0.04013802111148834,
0.07249587029218674,
-0.043499283492565155,
0.07851605117321014,
-0.04905369505286217,
0.022681210190057755,
0.047159552574157715,
-0.034697167575359344,
-0.015903646126389503,
-0.025036057457327843,
-0.039666589349508286,
-0.0009129474055953324,
-0.0642651692032814,
-0.022181332111358643,
0.037319399416446686,
-0.03309451788663864,
-0.02093743532896042,
0.022491950541734695,
0.013084960170090199,
0.027767561376094818,
-2.7272379947198957e-33,
-0.0662040263414383,
-0.02569163404405117,
-0.01894587278366089,
0.07801990956068039,
0.0221254862844944,
-0.05730380117893219,
-0.04537541791796684,
0.01777178794145584,
-0.01267706323415041,
-0.020223936066031456,
-0.07531726360321045,
-0.0016336559783667326,
0.05279335752129555,
-0.05808103084564209,
-0.022966472432017326,
-0.038755472749471664,
-0.042138274759054184,
-0.036353062838315964,
0.012911620549857616,
0.023840487003326416,
-0.04085613787174225,
0.08177755773067474,
-0.04708308354020119,
0.10800161212682724,
0.06311527639627457,
0.03385843709111214,
-0.03937425836920738,
0.05813117325305939,
-0.0027784451376646757,
-0.021642519161105156,
0.027510862797498703,
-0.049456775188446045,
-0.12593050301074982,
-0.13424764573574066,
-0.018193596974015236,
0.004168076906353235,
-0.015145501121878624,
0.11568783223628998,
-0.09592396020889282,
0.15152540802955627,
-0.014130041934549809,
-0.09858020395040512,
-0.14353425800800323,
-0.0010767882922664285,
0.02855793386697769,
0.009749838151037693,
-0.11915189772844315,
0.03613759204745293,
0.07538866996765137,
-0.009739773347973824,
-0.042743172496557236,
0.057142987847328186,
-0.013212904334068298,
0.01387265045195818,
-0.02215474657714367,
-0.06450386345386505,
0.03400326147675514,
-0.017778050154447556,
0.07649178057909012,
-0.01130968052893877,
0.00960354134440422,
-0.021222373470664024,
0.014298525638878345,
0.053830958902835846,
0.029950983822345734,
-0.06956551223993301,
-0.010759946890175343,
0.03381741791963577,
-0.06433338671922684,
0.047520849853754044,
0.03830226510763168,
-0.03784815967082977,
0.01491359993815422,
-0.01860099285840988,
0.10834752768278122,
0.02737351879477501,
0.02776329778134823,
-0.07928947359323502,
-0.04606181010603905,
0.07569124549627304,
0.06895226985216141,
0.08723366260528564,
0.08073316514492035,
0.03555195406079292,
0.06826391816139221,
0.041702549904584885,
0.06037174165248871,
0.10801250487565994,
0.011951598338782787,
0.01674104854464531,
0.03469773754477501,
0.06079661101102829,
-0.03665802255272865,
0.038176581263542175,
0.01376303005963564,
-4.3113377756753835e-8,
-0.0607563741505146,
-0.033740632236003876,
-0.001656576176173985,
0.04322630167007446,
0.001654455321840942,
0.02312180958688259,
-0.020198745653033257,
0.032892100512981415,
-0.03652098402380943,
-0.010655897669494152,
0.00006382174615282565,
0.05312289670109749,
-0.010886539705097675,
-0.09625811129808426,
-0.019667312502861023,
0.07028310000896454,
0.017530057579278946,
-0.05966568365693092,
-0.016917437314987183,
0.07017271965742111,
0.01804247871041298,
0.014166167005896568,
0.020892443135380745,
-0.08637126535177231,
0.01911243423819542,
-0.027164824306964874,
0.015763070434331894,
0.03608119860291481,
-0.018370691686868668,
0.046722330152988434,
-0.03923684358596802,
-0.04028479382395744,
0.0052330465987324715,
-0.015372022986412048,
0.028145689517259598,
0.07899194210767746,
-0.05881819874048233,
-0.03787125647068024,
-0.017614929005503654,
-0.06276913732290268,
0.045393284410238266,
0.030386582016944885,
-0.039626091718673706,
0.0070012519136071205,
-0.031710751354694366,
-0.037046290934085846,
-0.060060784220695496,
-0.06832516938447952,
0.05381229519844055,
-0.05909248813986778,
-0.04063764959573746,
-0.06101466342806816,
-0.04213056340813637,
0.05058354511857033,
0.12155419588088989,
-0.006074723321944475,
-0.06725345551967621,
-0.05073818191885948,
0.14301352202892303,
0.03294705972075462,
0.043319497257471085,
-0.003641410730779171,
-0.06347102671861649,
-0.005037781782448292
] | 0 | 0 | 0 | 0 |
|
Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control | https://openreview.net/forum?id=w3iM4WLuvy | [
"Devdhar Patel",
"Hava T Siegelmann"
] | Poster | Reinforcement learning (RL) is rapidly reaching and surpassing human-level control capabilities. However, state-of-the-art RL algorithms often require timesteps and reaction times significantly faster than human capabilities, which is impractical in real-world settings and typically necessitates specialized hardware. We introduce Sequence Reinforcement Learning (SRL), an RL algorithm designed to produce a sequence of actions for a given input state, enabling effective control at lower decision frequencies. SRL addresses the challenges of learning action sequences by employing both a model and an actor-critic architecture operating at different temporal scales. We propose a "temporal recall" mechanism, where the critic uses the model to estimate intermediate states between primitive actions, providing a learning signal for each individual action within the sequence. Once training is complete, the actor can generate action sequences independently of the model, achieving model-free control at a slower frequency. We evaluate SRL on a suite of continuous control tasks, demonstrating that it achieves performance comparable to state-of-the-art algorithms while significantly reducing actor sample complexity. To better assess performance across varying decision frequencies, we introduce the Frequency-Averaged Score (FAS) metric. Our results show that SRL significantly outperforms traditional RL algorithms in terms of FAS, making it particularly suitable for applications requiring variable decision frequencies. Furthermore, we compare SRL with model-based online planning, showing that SRL achieves comparable FAS while leveraging the same model during training that online planners use for planning. | Decision Frequency, Action Sequence Generation, Model-Based Training, Model-Free Control, Efficient Learning, Reinforcement Learning | We introduce an algorithm that achieves competitive continuous control at extremely slow control frequencies using action sequences | 13,222 | 2410.08979 | [
-0.06621682643890381,
-0.035760488361120224,
-0.024943172931671143,
0.050817642360925674,
-0.01002456247806549,
0.08794020861387253,
-0.05430004745721817,
-0.04331156238913536,
0.04278174415230751,
0.024595629423856735,
-0.029577216133475304,
-0.05726519227027893,
-0.053454626351594925,
0.07721708714962006,
0.0037234267219901085,
0.0004030167474411428,
0.0699814110994339,
0.04371900111436844,
-0.01362399011850357,
-0.055548056960105896,
0.05541957914829254,
0.01977304369211197,
0.12064375728368759,
-0.000772685743868351,
-0.008563022129237652,
-0.019407356157898903,
0.006587181240320206,
-0.08263292163610458,
0.03924518823623657,
0.007074442226439714,
0.07848318666219711,
-0.045750297605991364,
0.07299710810184479,
-0.07183356583118439,
-0.13914832472801208,
0.05239555239677429,
-0.09021782875061035,
-0.01775847189128399,
-0.03740667179226875,
-0.06968372315168381,
0.0016798999859020114,
0.004944969434291124,
0.053271468728780746,
0.05000264197587967,
0.023618776351213455,
0.02669336646795273,
0.014197686687111855,
-0.06253606826066971,
0.02124735154211521,
0.021047279238700867,
-0.014861705712974072,
-0.0033853279892355204,
0.046703655272722244,
-0.016134392470121384,
0.012366680428385735,
0.028748538345098495,
0.06561516225337982,
0.032282620668411255,
0.005707504227757454,
-0.038469523191452026,
-0.06492578983306885,
-0.13337883353233337,
-0.0534696988761425,
-0.049308739602565765,
-0.11578754335641861,
0.022678695619106293,
0.00406564911827445,
0.025752317160367966,
0.03534593805670738,
0.008467832580208778,
-0.04583710804581642,
0.04487447813153267,
-0.007570210844278336,
0.009644757024943829,
0.017690211534500122,
-0.0513901524245739,
-0.01694527454674244,
-0.0614032968878746,
0.04895477741956711,
0.003525322303175926,
0.0028278608806431293,
-0.0739203467965126,
-0.04960022121667862,
0.005512559786438942,
0.09646975249052048,
-0.0026887794956564903,
0.06134585291147232,
0.08098731935024261,
0.06070110574364662,
0.021589824929833412,
0.004172791726887226,
0.017879076302051544,
0.05795851722359657,
-0.09177642315626144,
-0.02217547409236431,
0.04593677446246147,
-0.01912565715610981,
-0.06017078831791878,
-0.058112263679504395,
0.05698904022574425,
0.017585352063179016,
0.05592162162065506,
-0.08509402722120285,
0.04259689897298813,
-0.040242429822683334,
0.004677686374634504,
0.011393003165721893,
0.02460254542529583,
-0.0003625264507718384,
-0.0724286213517189,
0.02369074337184429,
-0.033578187227249146,
0.0029309529345482588,
0.07390011847019196,
0.04577526077628136,
0.02793879620730877,
-0.009010818786919117,
-0.07654739916324615,
0.007070361170917749,
0.0923445075750351,
-0.020770685747265816,
-0.037189632654190063,
-0.00036281003849580884,
0.047422170639038086,
0.03824377804994583,
-0.03512847423553467,
-0.028877701610326767,
-1.77261666765309e-34,
0.060534246265888214,
0.020783642306923866,
0.06633245944976807,
0.0005362699739634991,
0.008540426380932331,
0.04966776818037033,
0.015330884605646133,
0.02283966727554798,
-0.0065367440693080425,
0.013539097271859646,
0.014405202120542526,
-0.03029814548790455,
-0.07483524829149246,
0.08225184679031372,
0.060728512704372406,
0.02574712410569191,
-0.07102195173501968,
0.030983878299593925,
-0.0035317419096827507,
-0.06337524205446243,
0.03466268256306648,
-0.026705769822001457,
0.003526270855218172,
-0.07019510120153427,
0.030453188344836235,
-0.036893993616104126,
0.03284841403365135,
0.04868362843990326,
0.013043982908129692,
-0.01588999480009079,
-0.0359615832567215,
0.0549977570772171,
-0.03733981400728226,
0.05980239808559418,
0.06560983508825302,
0.021938707679510117,
-0.06548116356134415,
0.01700122468173504,
0.030024634674191475,
0.02549450471997261,
0.0029155306983739138,
-0.017498932778835297,
-0.009574128314852715,
0.06729967147111893,
-0.09571927040815353,
-0.1407821923494339,
0.008890151046216488,
0.011035177856683731,
-0.10480345785617828,
0.026625368744134903,
0.05592846870422363,
0.07501096278429031,
-0.005256235133856535,
-0.09075643122196198,
-0.019737718626856804,
0.0417289100587368,
0.03250807151198387,
0.04370498284697533,
-0.042060330510139465,
0.11457952111959457,
0.01679692603647709,
0.05797780305147171,
-0.019216304644942284,
0.15002863109111786,
-0.03858872130513191,
0.07805925607681274,
-0.06266383826732635,
0.1675393134355545,
0.01940896175801754,
-0.019321473315358162,
0.06901691854000092,
0.013059696182608604,
0.05931587889790535,
-0.06533676385879517,
0.05881984159350395,
-0.027439668774604797,
0.07176858186721802,
-0.11206171661615372,
-0.10288704931735992,
0.05696903169155121,
-0.008706529624760151,
-0.04844757914543152,
-0.04911802336573601,
0.006167442537844181,
0.02406512200832367,
-0.09072937816381454,
0.008702551014721394,
-0.07077434659004211,
0.03852488100528717,
-0.04010574147105217,
-0.12523727118968964,
-0.025718649849295616,
0.01133222971111536,
0.03784716874361038,
0.0013444055803120136,
-1.1963003253253535e-33,
0.07523545622825623,
-0.0749715194106102,
0.00888762529939413,
0.033743809908628464,
0.0301650557667017,
-0.03143872320652008,
-0.05082808807492256,
0.004295370075851679,
0.02099626325070858,
-0.06185128539800644,
-0.054821357131004333,
0.008952710777521133,
0.04315108433365822,
-0.03844456374645233,
0.009466861374676228,
-0.04714575409889221,
-0.00878820475190878,
-0.0037662822287529707,
0.007338631432503462,
0.0075221373699605465,
0.06672516465187073,
0.036134228110313416,
-0.07358720153570175,
0.011562373489141464,
-0.07176399230957031,
-0.015369823202490807,
0.004026023671030998,
0.05485893413424492,
-0.011555003002285957,
0.013241940177977085,
-0.057310108095407486,
-0.03039192594587803,
0.022523580119013786,
-0.009947974234819412,
-0.07389634847640991,
0.02758435718715191,
0.10128285735845566,
0.04959539324045181,
-0.04869508370757103,
0.0920424684882164,
0.07198330014944077,
-0.07824498414993286,
-0.02754727564752102,
0.015676815062761307,
0.009647956117987633,
-0.005408189725130796,
-0.09838210791349411,
-0.0056438022293150425,
-0.05031190440058708,
0.013100034557282925,
-0.06206393614411354,
-0.0005208875518292189,
-0.06402183324098587,
-0.054138172417879105,
-0.06897037476301193,
-0.05078309029340744,
-0.049634553492069244,
0.043763257563114166,
0.10940854251384735,
0.020318331196904182,
-0.03451164439320564,
0.05252087116241455,
-0.03314918279647827,
-0.007619792129844427,
0.03397996351122856,
0.011490138247609138,
0.04021284729242325,
-0.007801956497132778,
0.06373635679483414,
0.044325072318315506,
0.015358243137598038,
0.05090225487947464,
-0.005598242860287428,
0.004498195834457874,
0.026555946096777916,
-0.06321964412927628,
-0.04728463664650917,
-0.02695273980498314,
-0.02748396061360836,
-0.03938966244459152,
0.03520576283335686,
0.015190940350294113,
0.007361601106822491,
0.016118420287966728,
-0.019851624965667725,
0.1033039540052414,
-0.02672428824007511,
0.09438715875148773,
0.03909912705421448,
-0.005224129650741816,
-0.03766029700636864,
0.04382212087512016,
0.03050096519291401,
0.04678602144122124,
-0.025867894291877747,
-4.3621746215194435e-8,
-0.05541694164276123,
0.030710220336914062,
0.0559287965297699,
0.036766961216926575,
0.04688002169132233,
-0.004498159047216177,
-0.010587824508547783,
-0.03943820297718048,
0.056326549500226974,
-0.019740158692002296,
0.081291064620018,
0.043722331523895264,
0.04284875839948654,
-0.09356924891471863,
-0.046797651797533035,
0.03955049067735672,
0.03971336409449577,
-0.05212258920073509,
0.030345145612955093,
-0.03688373789191246,
0.028312882408499718,
-0.015415642410516739,
0.05652393400669098,
-0.03964097425341606,
-0.01595095731317997,
-0.08758775889873505,
-0.004192708525806665,
0.10955646634101868,
0.05499247834086418,
0.004140859469771385,
0.03354281559586525,
-0.032069914042949677,
0.019516102969646454,
-0.0387139692902565,
0.025318240746855736,
-0.0006338513339869678,
0.05898261070251465,
-0.05313601344823837,
0.05373203381896019,
-0.05976004898548126,
-0.03874059394001961,
0.09502170979976654,
-0.05171580985188484,
-0.02193768322467804,
-0.0720389187335968,
-0.03410973399877548,
-0.07489216327667236,
-0.09522107988595963,
0.058175358921289444,
-0.005662987474352121,
0.07244671881198883,
0.018362753093242645,
-0.04168266803026199,
0.05448324605822563,
0.08713626861572266,
0.06803938746452332,
0.04717521741986275,
-0.07074794173240662,
0.010737363249063492,
0.029763950034976006,
0.020808089524507523,
-0.03216436132788658,
-0.07924716919660568,
-0.00750738987699151
] | https://github.com/dee0512/Temporally-Layered-Architecture | 2 | 0 | 0 | 0 |
Physics of Language Models: Part 3.2, Knowledge Manipulation | https://openreview.net/forum?id=oDbiL9CLoS | [
"Zeyuan Allen-Zhu",
"Yuanzhi Li"
] | Poster | Language models can store vast factual knowledge, yet their ability to flexibly use this knowledge for downstream tasks (e.g., via instruction finetuning) remains questionable. This paper investigates four fundamental knowledge manipulation tasks: \textbf{retrieval} (e.g., "What is person A's attribute X?"), \textbf{classification} (e.g., "Is A's attribute X even or odd?"), \textbf{comparison} (e.g., "Is A greater than B in attribute X?"), and \textbf{inverse search} (e.g., "Which person's attribute X equals T?").
We show that language models excel in knowledge retrieval but struggle even in the simplest classification or comparison tasks unless Chain of Thoughts (CoTs) are employed during both training and inference. Moreover, their performance in inverse knowledge search is virtually 0\%, regardless of the prompts.
Our primary contribution is a \emph{controlled, synthetic experiment} that confirms these weaknesses are \emph{inherent} to language models: they cannot efficiently manipulate knowledge from pre-training data, even when such knowledge is perfectly stored in the models, despite adequate training and sufficient model size. Our findings also apply to modern pretrained language models such as GPT-4, thus giving rise to many Turing tests to distinguish Humans from contemporary AIs. | knowledge manipulation, language models, generative models, reversal curse | null | 13,219 | 2309.14402 | [
-0.04366171360015869,
-0.000003042335720238043,
0.05436624214053154,
0.04203375428915024,
-0.007068677805364132,
-0.023039506748318672,
0.060573406517505646,
0.027788443490862846,
0.07733582705259323,
0.04171598330140114,
0.008244022727012634,
-0.03340736776590347,
0.06393514573574066,
0.03912108764052391,
0.00005076974775874987,
0.010961554013192654,
0.10245814919471741,
0.06431498378515244,
-0.1119130402803421,
-0.015255364589393139,
0.017255432903766632,
0.009504069574177265,
0.03505079448223114,
-0.006397072225809097,
-0.0002197994472226128,
-0.08001646399497986,
-0.09321549534797668,
-0.04944320768117905,
-0.0018542357720434666,
-0.050162725150585175,
-0.0026759242173284292,
0.04405452683568001,
0.006435113959014416,
0.09856310486793518,
-0.05150835961103439,
0.08657914400100708,
-0.03820565342903137,
0.022270577028393745,
0.015933794900774956,
-0.04248322546482086,
-0.09882364422082901,
-0.039393581449985504,
0.019980531185865402,
0.01762446016073227,
0.07599914073944092,
0.06306622177362442,
0.028617417439818382,
0.02361622080206871,
-0.07696061581373215,
-0.0023716147989034653,
-0.14435994625091553,
-0.029954805970191956,
-0.027078088372945786,
0.01746220514178276,
-0.004451042506843805,
0.022877207025885582,
0.03355224058032036,
0.009240121580660343,
-0.03782204911112785,
-0.0836496502161026,
-0.07308431714773178,
-0.08195420354604721,
-0.04008403420448303,
-0.02039119228720665,
0.016415925696492195,
-0.034068670123815536,
0.011871905997395515,
-0.006560414098203182,
-0.026441866531968117,
0.04742041602730751,
0.017839139327406883,
0.05851157382130623,
-0.05734037607908249,
0.00815247930586338,
0.07132363319396973,
0.046290501952171326,
-0.006333077326416969,
-0.005472859833389521,
-0.015759319067001343,
-0.044979095458984375,
-0.0031541495118290186,
-0.029176227748394012,
0.023039618507027626,
0.014057323336601257,
0.07234181463718414,
-0.018151670694351196,
0.01707357168197632,
0.02843196503818035,
0.07072551548480988,
-0.027961138635873795,
0.022887224331498146,
-0.06454046070575714,
0.08894830197095871,
-0.03510185703635216,
0.11897572875022888,
0.06119926646351814,
0.0513266995549202,
-0.027225369587540627,
-0.012442991137504578,
0.07449585199356079,
0.061909228563308716,
0.13956917822360992,
-0.03681759536266327,
-0.06802099943161011,
-0.03735250234603882,
-0.02037833444774151,
0.01271834596991539,
-0.013640635646879673,
0.024413615465164185,
-0.1417389214038849,
-0.004116937052458525,
0.013754461891949177,
-0.008043715730309486,
-0.008137308992445469,
-0.021515116095542908,
-0.011484235525131226,
0.03514993563294411,
-0.03749539703130722,
-0.06775102019309998,
0.03555486723780632,
-0.04185093194246292,
0.03943642973899841,
0.07118722051382065,
0.06452001631259918,
0.03000892698764801,
-0.011046267114579678,
-0.07658779621124268,
1.4391931822132436e-33,
0.06573886424303055,
0.06130971759557724,
-0.008825290016829967,
0.03925938159227371,
-0.02253200113773346,
-0.012256697751581669,
0.014475863426923752,
0.017928963527083397,
-0.02969466708600521,
-0.010811517015099525,
-0.019378311932086945,
0.05362720787525177,
-0.021733708679676056,
0.008489120751619339,
0.07888562232255936,
0.047082070261240005,
-0.0247966255992651,
0.009758065454661846,
-0.06801477074623108,
-0.0463053323328495,
0.15330874919891357,
0.013249463401734829,
-0.03443712368607521,
-0.061548035591840744,
-0.04022098332643509,
-0.014403698034584522,
-0.01420644298195839,
-0.03377925232052803,
-0.046580154448747635,
0.010014304891228676,
-0.1306622475385666,
0.0007138947839848697,
0.0652429610490799,
0.07482770830392838,
-0.002569501055404544,
-0.03536960855126381,
0.005410714540630579,
-0.057439371943473816,
0.04259525239467621,
-0.005439571104943752,
-0.028058629482984543,
0.05506238341331482,
0.03712033852934837,
-0.014417700469493866,
-0.06100790202617645,
-0.05996723473072052,
0.026164786890149117,
-0.025257108733057976,
-0.07264722138643265,
-0.015842624008655548,
0.05248325318098068,
0.04815087094902992,
-0.006663799751549959,
-0.054991550743579865,
0.07755883038043976,
0.030885852873325348,
0.0035665526520460844,
0.054135341197252274,
-0.0343787781894207,
0.0767570212483406,
-0.029692266136407852,
0.0231268722563982,
-0.0061438423581421375,
0.039895813912153244,
-0.014239695854485035,
-0.0071975477039813995,
-0.05411224812269211,
0.03449361398816109,
0.12711843848228455,
0.026389311999082565,
-0.04988573119044304,
-0.01750509813427925,
-0.04796965420246124,
-0.07370920479297638,
0.04545065760612488,
-0.009670637547969818,
0.0005316392634995282,
-0.11894125491380692,
0.020392969250679016,
0.011620632372796535,
0.03767683729529381,
-0.07575538009405136,
0.031170321628451347,
-0.027655191719532013,
-0.03718981519341469,
-0.07191120833158493,
0.055676478892564774,
-0.12765948474407196,
0.002639366779476404,
-0.022484881803393364,
0.007654690183699131,
-0.02845899946987629,
-0.06642071902751923,
-0.0016900881892070174,
0.02989269606769085,
-1.8093019866734564e-33,
-0.04625195637345314,
-0.06864279508590698,
-0.08248923718929291,
0.05577649921178818,
-0.0324329137802124,
-0.0581941232085228,
0.033251263201236725,
-0.027450868859887123,
-0.08842004835605621,
-0.13517196476459503,
-0.05957679823040962,
-0.04851716384291649,
0.020396696403622627,
-0.015851667150855064,
0.003201273735612631,
-0.015111292712390423,
-0.01193213276565075,
0.0067934622056782246,
0.04026547074317932,
0.05864470452070236,
-0.07699273526668549,
0.060720473527908325,
-0.08077237755060196,
0.0429404191672802,
-0.019212406128644943,
0.036226026713848114,
-0.01072676945477724,
0.04640745371580124,
-0.006897641345858574,
0.02776184305548668,
-0.09146039932966232,
0.007279586512595415,
-0.03778744488954544,
-0.0007189757307060063,
-0.07315987348556519,
0.036350883543491364,
0.025383925065398216,
-0.0165378637611866,
-0.014120301231741905,
0.07535798102617264,
0.022622913122177124,
0.022984208539128304,
-0.05203107371926308,
-0.04114925488829613,
-0.023772096261382103,
-0.04857005923986435,
-0.14426074922084808,
-0.036535609513521194,
0.08051008731126785,
-0.05646046996116638,
0.060791559517383575,
-0.03995174914598465,
-0.015320350416004658,
-0.0023840293288230896,
-0.022959765046834946,
-0.04379935935139656,
0.00798257626593113,
-0.037773776799440384,
0.03138350695371628,
0.07717407494783401,
-0.027316369116306305,
0.02826363779604435,
0.017414210364222527,
0.04807030409574509,
0.03988342732191086,
-0.000511153950355947,
-0.04852626845240593,
0.05778110772371292,
0.03738118335604668,
-0.06255228072404861,
0.09373148530721664,
-0.03305488079786301,
0.02534598484635353,
-0.03816676884889603,
0.07486484199762344,
0.011179705150425434,
-0.002297207247465849,
-0.019871819764375687,
-0.0598907507956028,
-0.060916755348443985,
-0.044387202709913254,
0.010645652189850807,
0.03283676505088806,
0.015240122564136982,
-0.013947637751698494,
0.05785752087831497,
-0.019959794357419014,
0.13441576063632965,
-0.03177770599722862,
-0.030207887291908264,
-0.006061388645321131,
-0.0027711980510503054,
-0.0023089356254786253,
0.11122352629899979,
0.013404414057731628,
-4.7936033809037326e-8,
-0.08221126347780228,
0.034835077822208405,
0.039665333926677704,
0.09334316104650497,
0.030496535822749138,
0.04076294228434563,
-0.05325179919600487,
0.038555439561605453,
-0.05301357060670853,
-0.026713496074080467,
0.050468847155570984,
0.047073736786842346,
-0.016848048195242882,
-0.006054128985852003,
0.021414760500192642,
0.08248602598905563,
-0.009223162196576595,
0.01105076540261507,
-0.01900925487279892,
-0.012050557881593704,
0.09620922803878784,
0.01879660226404667,
0.017492368817329407,
0.014071985147893429,
0.005331018473953009,
-0.01736733503639698,
-0.0763625055551529,
0.11574364453554153,
0.07158521562814713,
0.036846090108156204,
0.008970747701823711,
-0.04077713564038277,
-0.006246540229767561,
0.013087904080748558,
0.10016827285289764,
0.08170584589242935,
-0.010457123629748821,
-0.0883970782160759,
-0.052554838359355927,
0.035697028040885925,
0.09212581068277359,
0.06694841384887695,
-0.07872974127531052,
0.0513344332575798,
0.09346045553684235,
-0.030386412516236305,
-0.04600914567708969,
-0.08741732686758041,
0.03733410686254501,
-0.0000535939252586104,
0.06220736727118492,
0.003700086148455739,
-0.08373747020959854,
0.08232075721025467,
0.0276128351688385,
0.011898526921868324,
-0.00047139913658611476,
0.04533235728740692,
-0.04097432643175125,
0.0005741062923334539,
0.09228517860174179,
0.025276614353060722,
0.0030524914618581533,
-0.0035821052733808756
] | 0 | 1 | 0 | 0 |
|
Logic-Logit: A Logic-Based Approach to Choice Modeling | https://openreview.net/forum?id=vJgJSrYPe1 | [
"Shuhan Zhang",
"Wendi Ren",
"Shuang Li"
] | Poster | In this study, we propose a novel rule-based interpretable choice model, {\bf Logic-Logit}, designed to effectively learn and explain human choices. Choice models have been widely applied across various domains—such as commercial demand forecasting, recommendation systems, and consumer behavior analysis—typically categorized as parametric, nonparametric, or deep network-based. While recent innovations have favored neural network approaches for their computational power, these flexible models often involve large parameter sets and lack interpretability, limiting their effectiveness in contexts where transparency is essential.
Previous empirical evidence shows that individuals usually use {\it heuristic decision rules} to form their consideration sets, from which they then choose. These rules are often represented as {\it disjunctions of conjunctions} (i.e., OR-of-ANDs). These rules-driven, {\it consider-then-choose} decision processes enable people to quickly screen numerous alternatives while reducing cognitive and search costs. Motivated by this insight, our approach leverages logic rules to elucidate human choices, providing a fresh perspective on preference modeling. We introduce a unique combination of column generation techniques and the Frank-Wolfe algorithm to facilitate efficient rule extraction for preference modeling—a process recognized as NP-hard. Our empirical evaluation, conducted on both synthetic datasets and real-world data from commercial and healthcare domains, demonstrates that Logic-Logit significantly outperforms baseline models in terms of interpretability and accuracy. | Choice Model, Preference Learning, Interpretability, Rule Learning | null | 13,216 | null | [
-0.01668940670788288,
-0.06931421160697937,
-0.08200183510780334,
0.07627978175878525,
0.04891018196940422,
-0.0009101290488615632,
0.031166810542345047,
0.04477424547076225,
0.011429745703935623,
-0.03237610682845116,
-0.019982045516371727,
-0.008125582709908485,
-0.007988988421857357,
0.03234057128429413,
-0.01081776898354292,
-0.026359152048826218,
0.006581824738532305,
0.013749398291110992,
-0.08317233622074127,
-0.017352553084492683,
0.03301527351140976,
-0.00230414024554193,
0.008913563564419746,
-0.05894339084625244,
-0.0409012995660305,
0.016386115923523903,
0.0307066161185503,
0.014079402200877666,
-0.029062233865261078,
0.011529959738254547,
0.02642137184739113,
0.03659462928771973,
-0.07337154448032379,
0.006142569240182638,
-0.05247141793370247,
0.013491244986653328,
-0.08437786251306534,
-0.01043153926730156,
-0.00638774037361145,
-0.034323327243328094,
-0.02927531860768795,
-0.08199018239974976,
-0.06730791181325912,
0.0682414174079895,
0.07661761343479156,
0.0138247637078166,
-0.031445905566215515,
-0.04644486680626869,
-0.0025676225777715445,
0.04928959161043167,
-0.09429895132780075,
0.07163514196872711,
0.009150233119726181,
-0.027209695428609848,
0.0340791791677475,
0.01933344267308712,
-0.05131448805332184,
0.017678916454315186,
-0.055446457117795944,
-0.032669033855199814,
-0.08994115144014359,
-0.13962461054325104,
-0.06309299170970917,
-0.00007317907147808,
-0.020860563963651657,
0.014271672815084457,
-0.09220593422651291,
0.031932055950164795,
-0.016117027029395103,
0.05595087632536888,
-0.020520389080047607,
0.11049073934555054,
-0.05221771448850632,
-0.007238108664751053,
-0.01608729548752308,
0.07516953349113464,
0.11261247098445892,
-0.034891970455646515,
0.05751579999923706,
-0.025948040187358856,
-0.10762257874011993,
-0.01949644461274147,
-0.04432838782668114,
0.0524502657353878,
0.057231467217206955,
-0.01853123866021633,
0.0033124375622719526,
0.022739753127098083,
0.03301425278186798,
0.05830895155668259,
-0.02130543813109398,
0.014009182341396809,
0.07626847177743912,
-0.025093788281083107,
-0.031677037477493286,
0.06781039386987686,
-0.03768951818346977,
-0.03758857026696205,
-0.03841220214962959,
0.06589366495609283,
-0.008772484958171844,
0.10581261664628983,
0.00556604890152812,
0.006002123933285475,
0.02896612137556076,
-0.0682903304696083,
0.05101032555103302,
0.09092996269464493,
0.10598327219486237,
-0.05961577221751213,
-0.03781241551041603,
0.038554172962903976,
0.015470916405320168,
-0.07043340057134628,
0.023380711674690247,
0.0002056109078694135,
-0.012724819593131542,
-0.035139188170433044,
0.06551077961921692,
0.05483756959438324,
-0.05273279547691345,
0.050273142755031586,
-0.010218735784292221,
-0.004236782435327768,
0.005430727731436491,
0.029157094657421112,
-0.08470665663480759,
8.860102243265791e-34,
-0.07862808555364609,
0.04306233674287796,
0.06679516285657883,
-0.029937297105789185,
0.025155065581202507,
0.021294185891747475,
-0.016649020835757256,
0.07140691578388214,
-0.01788194663822651,
0.035169221460819244,
-0.03314848989248276,
0.07961796224117279,
-0.1024712473154068,
0.12756197154521942,
0.08318546414375305,
-0.020571712404489517,
-0.07803258299827576,
0.02781083434820175,
-0.0391840934753418,
-0.05915999040007591,
0.056134793907403946,
-0.011115560308098793,
0.01387561485171318,
0.015975942835211754,
-0.004917432088404894,
-0.04739726334810257,
0.06087396666407585,
0.02198926731944084,
0.018482552841305733,
-0.002492015715688467,
0.019533079117536545,
-0.03764887899160385,
0.027904899790883064,
-0.038248211145401,
0.024691835045814514,
0.0037135244347155094,
-0.05594741553068161,
-0.011362428776919842,
0.03645896539092064,
0.00919303111732006,
-0.06288702040910721,
-0.010364091955125332,
-0.016160912811756134,
0.09692569077014923,
-0.07287366688251495,
-0.0036807095166295767,
0.00001298601182497805,
0.030819742009043694,
-0.052215684205293655,
-0.04484991729259491,
0.0395054891705513,
0.03738873824477196,
0.0795094445347786,
-0.05312380939722061,
-0.051183003932237625,
-0.05007597804069519,
0.01892026513814926,
-0.031405869871377945,
0.03567396104335785,
-0.0077048842795193195,
-0.026139473542571068,
0.007856756448745728,
0.05371244251728058,
0.04405387118458748,
-0.0072612459771335125,
0.06994476914405823,
-0.0482027493417263,
0.028453847393393517,
0.03156577795743942,
-0.03660457581281662,
0.015723608434200287,
0.04551815614104271,
0.0067034219391644,
-0.07011028379201889,
0.052487995475530624,
0.06397350877523422,
0.0009493085090070963,
-0.03803810104727745,
0.015169472433626652,
-0.06352923810482025,
0.07382038980722427,
0.04555828869342804,
-0.03409265726804733,
-0.01732727140188217,
-0.006182919256389141,
-0.007538415491580963,
0.023249030113220215,
-0.10647860169410706,
-0.03170986473560333,
-0.06294767558574677,
-0.04853023961186409,
0.023770488798618317,
-0.010907764546573162,
0.003586443839594722,
0.024754105135798454,
-2.5588796530332968e-33,
-0.08381341397762299,
0.007794986944645643,
-0.033906642347574234,
0.04115533456206322,
0.05446977540850639,
-0.0345795676112175,
-0.08732806146144867,
-0.06891682744026184,
0.12424881011247635,
-0.04666774719953537,
-0.06328218430280685,
-0.047726910561323166,
0.08697259426116943,
-0.028222138062119484,
0.03152666985988617,
0.031524598598480225,
-0.08449501544237137,
-0.001271566143259406,
0.005426373332738876,
0.005837689153850079,
-0.04397215694189072,
0.09303513169288635,
-0.17812566459178925,
0.028305523097515106,
-0.07543473690748215,
0.028974471613764763,
-0.02045748010277748,
0.04958350211381912,
-0.01502476166933775,
-0.04169367253780365,
-0.07545514404773712,
-0.10816486924886703,
-0.017923977226018906,
-0.05653445050120354,
0.06542128324508667,
0.08213260769844055,
0.007841788232326508,
0.026993149891495705,
-0.07612752169370651,
0.05477384105324745,
-0.07104972004890442,
-0.036349911242723465,
-0.019496219232678413,
-0.003531242022290826,
-0.018982643261551857,
0.0597793348133564,
0.01206787396222353,
-0.030883455649018288,
-0.007535092532634735,
0.04978937283158302,
0.04017078876495361,
0.0639616996049881,
-0.06718671321868896,
0.06038083881139755,
-0.06075620651245117,
-0.0066388435661792755,
0.022980323061347008,
-0.0109286829829216,
0.11699983477592468,
0.03174373880028725,
-0.05549270659685135,
0.00593503937125206,
0.10008586943149567,
-0.007239618804305792,
-0.0157307256013155,
-0.10196300595998764,
0.026057187467813492,
0.006344389170408249,
-0.004080281127244234,
-0.06369085609912872,
-0.01725992187857628,
-0.06435400247573853,
0.020338166505098343,
0.011300611309707165,
0.03122306615114212,
-0.010903419926762581,
-0.005429094657301903,
-0.04057743772864342,
-0.01781950518488884,
0.0008429103181697428,
0.0011605877662077546,
-0.0024838601239025593,
0.053508613258600235,
0.023305898532271385,
0.03593792766332626,
-0.015654737129807472,
0.05264144763350487,
0.07193008810281754,
-0.06049641594290733,
0.04015416279435158,
-0.005467922892421484,
0.03792722523212433,
-0.023074574768543243,
0.056437619030475616,
-0.1487165242433548,
-4.9901210275038466e-8,
-0.023561229929327965,
-0.007694322615861893,
0.06686488538980484,
0.09637031704187393,
0.10218492150306702,
-0.06539138406515121,
0.05449020117521286,
0.09375742822885513,
-0.04776407405734062,
0.013041600584983826,
0.022706734016537666,
0.06585010886192322,
-0.015014353208243847,
-0.03371245786547661,
0.03624366596341133,
0.07635476440191269,
0.07110067456960678,
-0.0230100080370903,
-0.010855269618332386,
0.037833426147699356,
0.03747372329235077,
0.03407370299100876,
-0.006126190070062876,
0.03358817100524902,
0.09222793579101562,
-0.02772338315844536,
-0.06829860806465149,
0.0991552323102951,
0.029658440500497818,
0.12252487987279892,
0.06946580857038498,
0.05410102382302284,
-0.013935472816228867,
-0.056799501180648804,
0.01713346689939499,
0.05869773030281067,
-0.027776828035712242,
-0.08195134252309799,
-0.04864669218659401,
-0.03431358560919762,
-0.006311696022748947,
-0.013536406680941582,
0.0070404005236923695,
0.06645649671554565,
-0.01103700790554285,
-0.030446594581007957,
-0.03430032357573509,
-0.07644584774971008,
0.01063891127705574,
0.0264189001172781,
-0.023509206250309944,
-0.085017129778862,
0.053516827523708344,
-0.02886650152504444,
0.0798087865114212,
-0.0534781739115715,
0.009494584985077381,
-0.0469764806330204,
0.06966744363307953,
0.03614813834428787,
0.06991851329803467,
0.010054223239421844,
-0.045760251581668854,
0.03367969021201134
] | 0 | 0 | 0 | 0 |
|
On Calibration of LLM-based Guard Models for Reliable Content Moderation | https://openreview.net/forum?id=wUbum0nd9N | [
"Hongfu Liu",
"Hengguan Huang",
"Xiangming Gu",
"Hao Wang",
"Ye Wang"
] | Poster | Large language models (LLMs) pose significant risks due to the potential for generating harmful content or users attempting to evade guardrails. Existing studies have developed LLM-based guard models designed to moderate the input and output of threat LLMs, ensuring adherence to safety policies by blocking content that violates these protocols upon deployment. However, limited attention has been given to the reliability and calibration of such guard models. In this work, we empirically conduct comprehensive investigations of confidence calibration for 9 existing LLM-based guard models on 12 benchmarks in both user input and model output classification. Our findings reveal that current LLM-based guard models tend to 1) produce overconfident predictions, 2) exhibit significant miscalibration when subjected to jailbreak attacks, and 3) demonstrate limited robustness to the outputs generated by different types of response models. Additionally, we assess the effectiveness of post-hoc calibration methods to mitigate miscalibration. We demonstrate the efficacy of temperature scaling and, for the first time, highlight the benefits of contextual calibration for confidence calibration of guard models, particularly in the absence of validation sets. Our analysis and experiments underscore the limitations of current LLM-based guard models and provide valuable insights for the future development of well-calibrated guard models toward more reliable content moderation. We also advocate for incorporating reliability evaluation of confidence calibration when releasing future LLM-based guard models. | Content Moderation, LLM-based Guard Models, Calibration, Safety | null | 13,206 | 2410.10414 | [
-0.02271493524312973,
-0.06709479540586472,
-0.005938033573329449,
0.02378377504646778,
0.08310099691152573,
0.023296019062399864,
0.04193985462188721,
-0.0409146212041378,
0.04144108667969704,
-0.017058340832591057,
-0.02132357284426689,
-0.07536227256059647,
0.14096200466156006,
-0.01393667608499527,
0.02709294483065605,
-0.022125665098428726,
0.1430959552526474,
0.05490979924798012,
-0.07620833069086075,
-0.04669725149869919,
0.04728326201438904,
0.04105253890156746,
0.013280666433274746,
0.002263641217723489,
-0.04267885908484459,
-0.06780250370502472,
-0.09953833371400833,
0.02160739153623581,
-0.0882464125752449,
-0.014203929342329502,
-0.03506902605295181,
-0.01176017988473177,
-0.014876934699714184,
0.013052938506007195,
-0.04540399834513664,
-0.047083690762519836,
0.009142915718257427,
-0.04087409749627113,
-0.03278566524386406,
-0.06468407064676285,
-0.06770526617765427,
0.00974214542657137,
0.061305977404117584,
0.08059810847043991,
-0.005769494920969009,
-0.07189671695232391,
0.00786674302071333,
-0.0035328641533851624,
-0.11057822406291962,
-0.008174668997526169,
-0.01910688541829586,
0.05091750621795654,
0.047931719571352005,
0.046980120241642,
-0.057375457137823105,
-0.06276960670948029,
0.02953992784023285,
0.09032701700925827,
0.014864393509924412,
0.05134759843349457,
-0.06632248312234879,
-0.056811895221471786,
-0.08089996874332428,
0.01874811016023159,
0.005826340522617102,
0.032541438937187195,
-0.023715004324913025,
0.04993879795074463,
0.017636727541685104,
0.04719795659184456,
-0.07701049000024796,
-0.0076797702349722385,
-0.00668684858828783,
0.014249946922063828,
0.024238217622041702,
0.022332746535539627,
-0.032363153994083405,
0.0020281129982322454,
0.04926251620054245,
-0.09659865498542786,
0.023921465501189232,
-0.02041252888739109,
0.050034377723932266,
0.02037922292947769,
0.05804486200213432,
0.042729705572128296,
-0.0009396867826581001,
0.06226406991481781,
0.07934599369764328,
0.03837466612458229,
0.0033768422435969114,
0.007448271382600069,
0.06706266850233078,
0.053331490606069565,
-0.008043178357183933,
0.025128882378339767,
-0.01313661690801382,
-0.10895107686519623,
-0.040581852197647095,
0.06942567229270935,
0.026281284168362617,
0.0050332797691226006,
0.017283141613006592,
-0.05897020921111107,
0.07109691947698593,
-0.004339701030403376,
0.03706452250480652,
0.004014056641608477,
-0.023684367537498474,
0.01490648090839386,
0.014754541218280792,
0.0611155666410923,
0.0017620897851884365,
-0.019826464354991913,
0.05034441500902176,
-0.03330767899751663,
-0.027297258377075195,
-0.041202131658792496,
0.03849949315190315,
0.0932268425822258,
0.009530238807201385,
0.057243332266807556,
0.10185591876506805,
-0.026521163061261177,
0.05040046572685242,
0.013811939395964146,
-0.05228795111179352,
2.7128653722291855e-33,
0.05304364114999771,
0.020701086148619652,
-0.09667162597179413,
0.029831554740667343,
0.036567412316799164,
-0.023098114877939224,
-0.017087504267692566,
0.0560498908162117,
0.020498869940638542,
-0.030865486711263657,
-0.002237615641206503,
-0.04847755655646324,
-0.004665869753807783,
0.029265623539686203,
0.10757393389940262,
0.10300423204898834,
-0.039007626473903656,
0.05237627401947975,
-0.04091411828994751,
0.01082845963537693,
0.048957426100969315,
-0.08808168768882751,
0.06442657858133316,
-0.023733770474791527,
0.037714309990406036,
0.07533769309520721,
0.01886202208697796,
0.04901885613799095,
-0.045086268335580826,
0.04951373115181923,
-0.08748393505811691,
-0.07184378057718277,
0.04416033625602722,
-0.02421870455145836,
0.05127076804637909,
0.0004152164328843355,
0.0028873009141534567,
-0.014665697701275349,
0.02159111760556698,
-0.061798345297575,
-0.052301809191703796,
0.01356961578130722,
0.013379096984863281,
0.044354937970638275,
-0.036739181727170944,
-0.05367348715662956,
-0.04243475943803787,
-0.06835021078586578,
-0.041052885353565216,
-0.08210534602403641,
0.014777049422264099,
0.02630506083369255,
0.040130384266376495,
0.00619692075997591,
-0.14455263316631317,
-0.024329721927642822,
-0.011327693238854408,
0.003921305760741234,
-0.007634334731847048,
0.0336054190993309,
-0.013260193169116974,
0.06347254663705826,
-0.03332525119185448,
0.005637005902826786,
0.041747480630874634,
-0.051997505128383636,
-0.0769721269607544,
0.05395880341529846,
-0.0208513755351305,
-0.000782059330958873,
-0.06893180310726166,
-0.037022631615400314,
0.009028622880578041,
0.02857937291264534,
-0.0941171646118164,
-0.049531809985637665,
0.06264926493167877,
-0.00567714124917984,
0.048561155796051025,
-0.03519284725189209,
0.07165011018514633,
-0.040106695145368576,
0.0338992215692997,
-0.02480444870889187,
-0.06512022018432617,
-0.07716620713472366,
0.012403810396790504,
-0.046680569648742676,
-0.07733476161956787,
0.06925460696220398,
-0.010883694514632225,
-0.034749533981084824,
0.01912088133394718,
0.057705748826265335,
-0.039150431752204895,
-4.378202545376521e-33,
-0.09639352560043335,
-0.01424251589924097,
-0.001337311347015202,
0.07048770040273666,
-0.11378684639930725,
-0.03334483504295349,
-0.022008024156093597,
0.056902866810560226,
0.005299990065395832,
-0.06842371821403503,
-0.05749218910932541,
0.0195834431797266,
0.06957381218671799,
0.016673943027853966,
0.04243506118655205,
-0.07405798882246017,
-0.025707164779305458,
0.024295691400766373,
0.0026085167191922665,
0.030724702402949333,
0.05229940265417099,
0.015981515869498253,
-0.11562983691692352,
0.018487678840756416,
-0.06367514282464981,
-0.024626797065138817,
-0.05710706114768982,
-0.014775967225432396,
0.029139190912246704,
0.0031845238991081715,
-0.013086041435599327,
0.018158869817852974,
-0.002008309355005622,
-0.002180869225412607,
-0.05235762894153595,
-0.0012814493384212255,
0.013816822320222855,
0.03606932982802391,
-0.01912900246679783,
0.07431921362876892,
0.10105488449335098,
0.00279115024022758,
-0.1217840313911438,
-0.0442386120557785,
-0.01189487986266613,
0.003831004723906517,
0.014791489578783512,
-0.04378074035048485,
0.028222957625985146,
-0.03001220151782036,
-0.0037695886567234993,
-0.023845605552196503,
-0.01763685792684555,
0.002574280137196183,
-0.05907100811600685,
-0.07200068980455399,
-0.03418419137597084,
-0.0730479434132576,
0.007016312796622515,
0.032459478825330734,
-0.059055231511592865,
0.00684417225420475,
-0.011075464077293873,
-0.0706513300538063,
0.07467770576477051,
-0.04157602787017822,
-0.025291351601481438,
0.03686672821640968,
0.04528326541185379,
-0.013988090679049492,
0.10796552151441574,
-0.011988122947514057,
-0.02125817909836769,
0.01862579584121704,
-0.021380875259637833,
-0.061402011662721634,
-0.03464397042989731,
-0.08804795891046524,
-0.07635805755853653,
0.02441118285059929,
0.06903134286403656,
0.04910410940647125,
-0.016301941126585007,
0.05635124444961548,
0.11794613301753998,
0.055267415940761566,
0.07837497442960739,
0.059552550315856934,
0.01986953429877758,
0.0034524104557931423,
-0.02905375137925148,
0.030484644696116447,
-0.08416576683521271,
0.0752720758318901,
-0.04685331881046295,
-4.736151026918378e-8,
-0.013675102032721043,
0.03854825720191002,
-0.020735835656523705,
0.08391567319631577,
0.01911856234073639,
0.10551661252975464,
-0.05468437820672989,
-0.07507196813821793,
0.037724852561950684,
-0.07767900824546814,
0.08867368847131729,
-0.03396179899573326,
-0.047377169132232666,
-0.06597022712230682,
0.003931185696274042,
0.034523166716098785,
-0.0014060208341106772,
0.08636851608753204,
-0.06817338615655899,
-0.03324318304657936,
0.01911398023366928,
0.009412163868546486,
0.004914768505841494,
-0.028149381279945374,
0.04130956158041954,
-0.009877889417111874,
-0.04056999832391739,
0.08774103969335556,
0.05726192891597748,
0.07978886365890503,
-0.039601463824510574,
-0.05770239979028702,
-0.05037417262792587,
-0.015035659074783325,
0.002837076084688306,
0.1559624820947647,
-0.010602626018226147,
-0.047244779765605927,
0.10169881582260132,
0.035690706223249435,
0.021377049386501312,
0.0224774032831192,
-0.03706962242722511,
0.07151499390602112,
0.011728251352906227,
0.003244431223720312,
-0.017223872244358063,
-0.11433561891317368,
0.07512832432985306,
-0.049941252917051315,
0.044366732239723206,
-0.03525976836681366,
-0.04938938841223717,
0.1285901665687561,
0.02697017602622509,
0.040429744869470596,
0.005388353485614061,
-0.05298641696572304,
-0.004126455169171095,
0.05588464066386223,
0.04113021492958069,
-0.02034340426325798,
-0.02390175685286522,
-0.026931755244731903
] | https://github.com/waffle-liu/calibration_guard_model | 5 | 0 | 0 | 0 |
Unified Convergence Analysis for Score-Based Diffusion Models with Deterministic Samplers | https://openreview.net/forum?id=HrdVqFSn1e | [
"Runjia Li",
"Qiwei Di",
"Quanquan Gu"
] | Poster | Score-based diffusion models have emerged as powerful techniques for generating samples from high-dimensional data distributions. These models involve a two-phase process: first, injecting noise to transform the data distribution into a known prior distribution, and second, sampling to recover the original data distribution from noise. Among the various sampling methods, deterministic samplers stand out for their enhanced efficiency. However, analyzing these deterministic samplers presents unique challenges, as they preclude the use of established techniques such as Girsanov's theorem, which are only applicable to stochastic samplers. Furthermore, existing analysis for deterministic samplers usually focuses on specific examples, lacking a generalized approach for general forward processes and various deterministic samplers. Our paper addresses these limitations by introducing a unified convergence analysis framework. To demonstrate the power of our framework, we analyze the variance-preserving (VP) forward process with the exponential integrator (EI) scheme, achieving iteration complexity of $\tilde{O}(d^2/\epsilon)$.
Additionally, we provide a detailed analysis of Denoising Diffusion Implicit Models (DDIM)-type samplers, which have been underexplored in previous research, achieving polynomial iteration complexity. | Diffusion Models, Probability Flow ODEs, Unified Framework, Deterministic Samplers | A unified framework for convergence analysis of diffusion models with deterministic samplers. | 13,205 | 2410.14237 | [
-0.0984419733285904,
-0.10073778033256531,
0.045825377106666565,
0.03688124939799309,
0.05086354538798332,
-0.059291642159223557,
-0.010435476899147034,
-0.07819424569606781,
0.08184188604354858,
-0.010107374750077724,
0.017397945746779442,
0.0495230108499527,
0.019384320825338364,
-0.026241149753332138,
0.0038115070201456547,
-0.09542522579431534,
0.0598989762365818,
0.05899309739470482,
0.009458278305828571,
-0.027225179597735405,
-0.046384915709495544,
0.10129270702600479,
-0.05847315117716789,
0.02110319212079048,
0.062964528799057,
-0.061311863362789154,
0.06959562748670578,
-0.046841152012348175,
0.01567661203444004,
0.0041484832763671875,
0.12218553572893143,
-0.05977746844291687,
-0.036853112280368805,
-0.005007797386497259,
-0.021173499524593353,
0.02255374751985073,
-0.029472526162862778,
0.017977822571992874,
0.011655612848699093,
0.04978550598025322,
0.06937167793512344,
0.00933231133967638,
0.002738772891461849,
0.014318771660327911,
0.06725640594959259,
-0.04583366960287094,
-0.004930611699819565,
-0.06774603575468063,
0.01342845894396305,
0.07131566852331161,
0.06295885890722275,
0.035214491188526154,
-0.0735706314444542,
-0.02091187611222267,
0.028107937425374985,
-0.04370361566543579,
0.05935873091220856,
-0.009249920025467873,
0.016161402687430382,
0.055693842470645905,
-0.015797873958945274,
-0.06814616918563843,
-0.011508113704621792,
0.003960141446441412,
-0.0635853260755539,
0.02306813932955265,
0.08044426888227463,
0.06591296941041946,
0.04215078800916672,
0.03358105570077896,
-0.10420303046703339,
0.051591113209724426,
-0.1344527304172516,
0.014570964500308037,
-0.026335502043366432,
-0.028249703347682953,
-0.0014393262099474669,
-0.013784663751721382,
-0.03994064778089523,
-0.06838526576757431,
0.045935556292533875,
0.05821969732642174,
0.05985071882605553,
-0.0692276805639267,
0.037265222519636154,
-0.008832275867462158,
0.04137958958745003,
0.041752349585294724,
0.08985476940870285,
-0.058638833463191986,
-0.051747653633356094,
0.09028268605470657,
-0.11786534637212753,
0.04957285150885582,
0.055423881858587265,
-0.035715263336896896,
0.02674192003905773,
0.012895848602056503,
0.12850122153759003,
0.060910362750291824,
0.04698579013347626,
0.04602590575814247,
0.00637046480551362,
-0.021208731457591057,
0.03578009083867073,
-0.13726124167442322,
0.07170378416776657,
-0.001720232074148953,
-0.03281376138329506,
0.0155832190066576,
0.021386468783020973,
0.03422228991985321,
0.07735946774482727,
0.059763092547655106,
0.01838044263422489,
-0.12202956527471542,
-0.02431952953338623,
-0.021950209513306618,
0.035661663860082626,
0.033446043729782104,
-0.013279691338539124,
0.0242296252399683,
-0.030246922746300697,
0.023583507165312767,
0.05658258870244026,
0.007073480635881424,
0.01375489216297865,
2.7792378240384428e-33,
-0.028755275532603264,
-0.03849217668175697,
0.07338432222604752,
0.0033796988427639008,
0.02982492186129093,
0.040332552045583725,
0.0036797653883695602,
-0.09276334941387177,
-0.000985312508419156,
0.05325686186552048,
0.00411661621183157,
0.015484506264328957,
-0.13817501068115234,
0.011970662511885166,
0.046294644474983215,
-0.005357310175895691,
-0.04231490194797516,
0.08680204302072525,
0.0650990754365921,
-0.012699533253908157,
0.13878950476646423,
-0.03093028999865055,
-0.023409735411405563,
0.006782257463783026,
-0.002654240233823657,
0.029551563784480095,
0.032427407801151276,
-0.034665290266275406,
0.03712085634469986,
0.019269084557890892,
-0.07597056776285172,
-0.04780972748994827,
-0.06727628409862518,
-0.0036972519010305405,
0.004977211821824312,
-0.0039027524180710316,
0.0030427398160099983,
0.007302902173250914,
-0.034905675798654556,
-0.0683397427201271,
0.004273916594684124,
0.09359930455684662,
0.01134277693927288,
-0.004528329242020845,
-0.15030643343925476,
-0.07033823430538177,
0.018589528277516365,
-0.04201618582010269,
0.1096605435013771,
-0.08938330411911011,
-0.020989177748560905,
-0.066618911921978,
-0.046215154230594635,
-0.07358591258525848,
-0.03959454596042633,
-0.0017491114558652043,
0.03896402195096016,
-0.06268584728240967,
-0.001921884366311133,
0.07921195775270462,
-0.05385032668709755,
0.05981408432126045,
-0.08021943271160126,
-0.018681664019823074,
0.12154785543680191,
0.006927370559424162,
-0.012331895530223846,
0.02885633520781994,
-0.013547347858548164,
-0.061765991151332855,
0.05411989614367485,
0.02112038992345333,
-0.06522872298955917,
-0.08053548634052277,
-0.03179064020514488,
-0.06152496486902237,
0.050457704812288284,
0.007926496677100658,
-0.013993505388498306,
-0.053159281611442566,
-0.09866176545619965,
-0.04873060807585716,
-0.07106055319309235,
-0.03762476518750191,
0.009767466224730015,
0.012336501851677895,
-0.04902574419975281,
-0.050410736352205276,
-0.078379325568676,
-0.07302834838628769,
-0.02935059741139412,
0.014858797192573547,
-0.006389209069311619,
-0.005576788913458586,
-0.05755629017949104,
-3.819877993566459e-33,
-0.03992094472050667,
0.03350457176566124,
0.03869812563061714,
0.08270096033811569,
0.045743200927972794,
-0.043711498379707336,
-0.036895837634801865,
0.017371833324432373,
0.11196485906839371,
-0.06840092688798904,
-0.09374118596315384,
-0.03491808474063873,
0.04439837485551834,
0.0474039688706398,
0.035739462822675705,
-0.0065988460555672646,
0.022401459515094757,
0.06951480358839035,
-0.08325912058353424,
-0.023320943117141724,
-0.027678484097123146,
0.010105704888701439,
0.05284006893634796,
0.008806251920759678,
-0.026426326483488083,
0.008577590808272362,
-0.049477625638246536,
0.1347547322511673,
-0.023481076583266258,
-0.06267649680376053,
-0.08634324371814728,
-0.008836221881210804,
0.021721649914979935,
-0.017660342156887054,
-0.04890722781419754,
0.03972263261675835,
-0.00596811855211854,
0.050584301352500916,
0.02761480025947094,
0.10286732017993927,
0.07296925038099289,
0.013912638649344444,
-0.03894286975264549,
-0.03270691633224487,
0.0714367926120758,
0.028060829266905785,
-0.012298258021473885,
0.01196355652064085,
0.017363635823130608,
-0.005952608771622181,
0.035324227064847946,
0.07817456126213074,
-0.021553872153162956,
0.10765337944030762,
-0.03326838091015816,
-0.04333014041185379,
-0.012852747924625874,
-0.06799842417240143,
0.021327419206500053,
-0.003617587499320507,
0.023715607821941376,
0.019884295761585236,
-0.007898247800767422,
-0.003689879085868597,
-0.05621789023280144,
0.028729921206831932,
-0.04266192391514778,
-0.02101781778037548,
0.010257869027554989,
0.038993578404188156,
-0.02576964534819126,
0.005471305921673775,
-0.021466771140694618,
-0.006766380276530981,
0.011981881223618984,
-0.11371251940727234,
0.031699761748313904,
-0.038891471922397614,
-0.010167940519750118,
-0.03532681614160538,
0.04629424959421158,
0.08020015805959702,
0.006239353679120541,
-0.037859514355659485,
0.0388646237552166,
0.03451363369822502,
0.05597400665283203,
-0.0486169308423996,
0.029605820775032043,
-0.01735469326376915,
0.0077867028303444386,
-0.0164472833275795,
0.07556013762950897,
0.04127795994281769,
-0.09145735204219818,
-4.9002981228341014e-8,
-0.07333112508058548,
0.046786945313215256,
0.04929817095398903,
0.022403107956051826,
0.06705128401517868,
0.0053550167940557,
-0.0018686874536797404,
-0.0353245735168457,
-0.009929838590323925,
-0.1012071818113327,
0.07837938517332077,
-0.01965908706188202,
-0.029436742886900902,
-0.05734522268176079,
0.0014750971458852291,
0.08163605630397797,
-0.002589116571471095,
-0.015110432170331478,
0.005222321953624487,
-0.029403770342469215,
-0.08342938870191574,
0.07291332632303238,
-0.027567388489842415,
-0.0006281475070863962,
0.03364140912890434,
0.04095086082816124,
0.0001659629342611879,
-0.03489946201443672,
0.10960914194583893,
0.0021185025107115507,
0.021496545523405075,
0.030683936551213264,
-0.0059084827080369,
0.02334744669497013,
-0.016163397580385208,
-0.005849170032888651,
0.03164170682430267,
0.030627429485321045,
-0.0007618948584422469,
-0.042569343000650406,
-0.055136825889348984,
0.05764257162809372,
-0.08212511986494064,
-0.005465647205710411,
-0.05848846584558487,
0.04303368553519249,
0.019305286929011345,
-0.033146392554044724,
0.04350369796156883,
-0.004390106536448002,
0.0016178097575902939,
-0.01592208817601204,
0.02232246845960617,
0.03540892153978348,
0.05047354847192764,
-0.001406475086696446,
-0.02029399573802948,
-0.024333564564585686,
0.04013782739639282,
0.006329709198325872,
0.06390698999166489,
-0.015184388495981693,
-0.04937626048922539,
0.059315331280231476
] | 0 | 0 | 0 | 0 |
|
Optimizing Neural Network Representations of Boolean Networks | https://openreview.net/forum?id=1H90Gb9rJ9 | [
"Joshua Russell",
"Ignacio Gavier",
"Devdhar Patel",
"Edward Rietman",
"Hava T Siegelmann"
] | Poster | Neural networks are known to be universal computers for Boolean functions. Recent advancements in hardware have significantly reduced matrix multiplication times, making neural network simulation both fast and efficient. Consequently, functions defined by complex Boolean networks are increasingly viable candidates for simulation through their neural network representation. Prior research has introduced a general method for deriving neural network representations of Boolean networks. However, the resulting neural networks are often suboptimal in terms of the number of neurons and connections, leading to slower simulation performance. Optimizing them while preserving functional equivalence --lossless optimization-- is an NP-hard problem, and current methods only provide lossy solutions. In this paper, we present a deterministic algorithm to optimize such neural networks in terms of neurons and connections while preserving functional equivalence. Moreover, to accelerate the compression of the neural network, we introduce an objective-aware algorithm that exploits representations that are shared among subproblems of the overall optimization. We demonstrate experimentally that we are able to reduce connections and neurons by up to 70% and 60%, respectively, in comparison to state-of-the-art. We also find that our objective-aware algorithm results in consistent speedups in optimization time, achieving up to 34.3x and 5.9x speedup relative to naive and caching solutions, respectively. Our methods are of practical relevance to applications such as high-throughput circuit simulation and placing neurosymbolic systems on the same hardware architecture. | Neural Networks, Boolean Networks, Lossless Optimization, Integer Linear Programming, NPN Classification | null | 13,203 | null | [
-0.06121041998267174,
-0.009466271847486496,
-0.03858942911028862,
0.01617862656712532,
-0.00403065187856555,
-0.03994947299361229,
0.000044448268454289064,
-0.14327462017536163,
0.01011432521045208,
-0.05259469151496887,
-0.041765984147787094,
-0.03771097585558891,
-0.037336211651563644,
0.08559738844633102,
-0.09061514586210251,
0.017750564962625504,
0.008757621981203556,
0.11844543367624283,
-0.08878903836011887,
-0.0003328797174617648,
0.002346200402826071,
-0.03870945796370506,
-0.059022456407547,
0.009834946133196354,
0.03812728449702263,
-0.024228602647781372,
-0.002003180794417858,
-0.060154568403959274,
0.004920050036162138,
-0.008931368589401245,
0.05033193528652191,
-0.06618542224168777,
0.03796592354774475,
-0.044983088970184326,
-0.06126170977950096,
0.0767778754234314,
-0.047234710305929184,
-0.03313359245657921,
0.017304107546806335,
-0.05128420889377594,
0.0016037938185036182,
-0.011103712022304535,
-0.01674763858318329,
0.010254837572574615,
0.058311015367507935,
0.08267154544591904,
0.050550200045108795,
-0.02784418687224388,
0.044267505407333374,
0.041801732033491135,
-0.008440794423222542,
0.0462864525616169,
-0.0659649521112442,
-0.03175431489944458,
0.05209723860025406,
-0.002718282165005803,
0.003927832469344139,
0.04578555002808571,
-0.13247212767601013,
0.04421786218881607,
-0.02505205199122429,
-0.08342666923999786,
0.02054917998611927,
0.037585608661174774,
-0.004812655970454216,
0.041961412876844406,
0.05921311303973198,
0.04598614573478699,
0.048343416303396225,
0.05472062528133392,
0.026771390810608864,
0.0558214969933033,
-0.024590028449892998,
0.016741912811994553,
0.015200204215943813,
0.05073165148496628,
0.15314359962940216,
0.012589139863848686,
0.019974714145064354,
0.01938636042177677,
-0.022193126380443573,
0.0031578633934259415,
-0.005227682180702686,
0.02604742906987667,
0.07386235147714615,
-0.018816089257597923,
-0.0517953559756279,
0.03186613693833351,
0.027399303391575813,
-0.01647823490202427,
-0.029388217255473137,
0.030792808160185814,
0.049416571855545044,
-0.04794648289680481,
0.07066512852907181,
0.001960760448127985,
0.038983870297670364,
0.015084213577210903,
-0.04700465127825737,
0.04711298272013664,
-0.06015779450535774,
0.04828751087188721,
0.06050049141049385,
-0.07367175817489624,
0.07517317682504654,
0.021610336378216743,
0.06578480452299118,
0.02019794099032879,
-0.003505132859572768,
-0.10661634802818298,
0.017975643277168274,
0.029561486095190048,
0.03476467728614807,
-0.004645020235329866,
0.0036618185695260763,
0.037643082439899445,
-0.007257789373397827,
-0.01353180781006813,
0.07231927663087845,
0.08213501423597336,
-0.07320939004421234,
0.03950335085391998,
-0.04756114259362221,
0.06369125843048096,
0.018153876066207886,
0.003356668632477522,
-0.1286374181509018,
2.4886418451907496e-33,
-0.020870160311460495,
0.09764470160007477,
0.017935406416654587,
-0.025270557031035423,
0.0794622078537941,
-0.06318749487400055,
0.050013016909360886,
0.05824148282408714,
0.019207516685128212,
0.039705660194158554,
-0.07451965659856796,
0.02426423504948616,
-0.07626735419034958,
0.04775579646229744,
0.054177362471818924,
-0.03709643334150314,
-0.03078046627342701,
-0.014499963261187077,
0.0002486011653672904,
-0.1297624409198761,
0.001869544037617743,
-0.05106916278600693,
-0.03327597305178642,
-0.009075782261788845,
0.005376413464546204,
-0.06450556963682175,
0.07379258424043655,
-0.016416458413004875,
-0.029784416779875755,
-0.03281530365347862,
-0.07003825902938843,
0.010373508557677269,
-0.0032374742440879345,
-0.030486274510622025,
0.0658872053027153,
-0.03814488276839256,
0.03982791304588318,
-0.006747790612280369,
-0.003678984474390745,
-0.00004713017551694065,
-0.015080686658620834,
0.0180816613137722,
-0.009412229061126709,
-0.026630068197846413,
-0.08843107521533966,
-0.08245046436786652,
-0.012091880664229393,
0.0619000606238842,
0.04716245085000992,
-0.03757680952548981,
0.060860808938741684,
0.0003870412183459848,
0.045317187905311584,
-0.042728304862976074,
0.030309243127703667,
-0.10021872073411942,
0.06650220602750778,
0.0393119640648365,
0.048287954181432724,
0.15484267473220825,
-0.060973990708589554,
-0.015523905865848064,
0.0007111750892363489,
0.054264236241579056,
-0.0297692883759737,
0.06867846846580505,
-0.019237961620092392,
0.027330689132213593,
0.0715952068567276,
-0.04637696221470833,
-0.011289789341390133,
0.037937961518764496,
-0.043429333716630936,
-0.10014347732067108,
0.050448160618543625,
0.028358953073620796,
0.011836892925202847,
-0.09739015251398087,
-0.05603955313563347,
-0.01017979346215725,
-0.023758377879858017,
0.12197156995534897,
-0.05647685378789902,
0.013246309012174606,
-0.008141444995999336,
-0.051869992166757584,
0.053501177579164505,
-0.03299237787723541,
0.0023064373526722193,
-0.021790210157632828,
-0.06677830219268799,
-0.044957756996154785,
0.07011482119560242,
-0.07957824319601059,
0.0112606855109334,
-2.8013007509774232e-33,
-0.07546218484640121,
0.009263969026505947,
-0.028231391683220863,
0.044546306133270264,
-0.012350065633654594,
-0.08094432950019836,
-0.01019237656146288,
-0.019650977104902267,
0.0034589748829603195,
0.06346061825752258,
-0.0071173678152263165,
-0.05976627767086029,
0.04213856905698776,
-0.015330218710005283,
-0.02368306741118431,
0.030700447037816048,
-0.0878932848572731,
-0.03023466467857361,
0.016967635601758957,
-0.03457128256559372,
0.007501354906708002,
0.19315925240516663,
-0.03378307819366455,
-0.04578502103686333,
-0.03341134637594223,
0.01850621961057186,
-0.06138724461197853,
0.0712103620171547,
0.08859295397996902,
0.0645693987607956,
-0.08807511627674103,
0.034018438309431076,
-0.09827613085508347,
0.023818088695406914,
0.05269908159971237,
0.04625639319419861,
0.06470760703086853,
0.043023522943258286,
0.021127894520759583,
-0.03557662293314934,
0.044158853590488434,
-0.03678926080465317,
-0.07937457412481308,
0.032074764370918274,
0.08534426242113113,
-0.009243876673281193,
-0.006909053772687912,
-0.05958367884159088,
-0.0033774604089558125,
0.05505336821079254,
0.01233994122594595,
0.0068138279020786285,
-0.039025019854307175,
0.026028217747807503,
-0.005721654277294874,
-0.05963941663503647,
-0.08254051953554153,
0.023957153782248497,
0.11307846754789352,
-0.021269695833325386,
-0.02373070828616619,
-0.14179806411266327,
-0.005190208554267883,
-0.04046478122472763,
-0.036283206194639206,
-0.00513485400006175,
0.02232276275753975,
0.06168220192193985,
0.04892915114760399,
0.020584827288985252,
-0.01742038130760193,
0.0441630594432354,
0.0207875557243824,
0.0320277214050293,
-0.10288640856742859,
-0.014759328216314316,
0.07655832171440125,
0.013778642751276493,
0.0008695460855960846,
0.06759722530841827,
0.04404057189822197,
0.029243888333439827,
0.020714184269309044,
-0.0014797094045206904,
0.002137579256668687,
0.09233411401510239,
0.033075518906116486,
0.0458429753780365,
0.04347297549247742,
0.021353736519813538,
-0.03494352102279663,
0.07577740401029587,
0.021904652938246727,
0.042062267661094666,
-0.09126457571983337,
-4.213883286752207e-8,
-0.06403777748346329,
0.034680478274822235,
-0.0018698222702369094,
0.005539105739444494,
0.05666752904653549,
-0.03815323859453201,
-0.008741199970245361,
0.0034899923484772444,
-0.03659007325768471,
-0.09178248047828674,
0.05236333981156349,
0.04951980337500572,
-0.0235827025026083,
-0.03788232430815697,
0.0023029849398881197,
-0.0014218938304111362,
-0.027177611365914345,
-0.07552305608987808,
0.044286634773015976,
-0.028553597629070282,
-0.017785118892788887,
0.01655411347746849,
-0.018295053392648697,
0.08074628561735153,
-0.0016893261345103383,
-0.1051015704870224,
-0.055165018886327744,
0.014829476363956928,
0.109683096408844,
0.02083442732691765,
-0.047806255519390106,
0.025468453764915466,
0.05158976465463638,
-0.008168265223503113,
0.042044151574373245,
0.05599912628531456,
0.018111737444996834,
0.06559549272060394,
-0.03377332165837288,
-0.022801373153924942,
-0.053450264036655426,
-0.02870040014386177,
-0.029209598898887634,
0.0041021378710865974,
-0.016090214252471924,
-0.10516072809696198,
0.028955964371562004,
-0.026324013248085976,
-0.03170429542660713,
0.0035949796438217163,
0.014795443043112755,
0.004111137241125107,
-0.07035032659769058,
-0.028288708999753,
-0.009852198883891106,
-0.05351412668824196,
-0.014297832734882832,
-0.021838903427124023,
0.032155342400074005,
0.10172339528799057,
-0.009052962064743042,
0.08038940280675888,
-0.014029399491846561,
-0.04793917015194893
] | 0 | 0 | 0 | 0 |
|
Causal Order: The Key to Leveraging Imperfect Experts in Causal Inference | https://openreview.net/forum?id=9juyeCqL0u | [
"Aniket Vashishtha",
"Abbavaram Gowtham Reddy",
"Abhinav Kumar",
"Saketh Bachu",
"Vineeth N. Balasubramanian",
"Amit Sharma"
] | Poster | Large Language Models (LLMs) have recently been used as experts to infer causal graphs, often by repeatedly applying a pairwise prompt that asks about the causal relationship of each variable pair. However, such experts, including human domain experts, cannot distinguish between direct and indirect effects given a pairwise prompt. Therefore, instead of the graph, we propose that causal order be used as a more stable output interface for utilizing expert knowledge. When querying a perfect expert with a pairwise prompt, we show that the inferred graph can have significant errors whereas the causal order is always correct. In practice, however, LLMs are imperfect experts and we find that pairwise prompts lead to multiple cycles and do not yield a valid order. Hence, we propose a prompting strategy that introduces an auxiliary variable for every variable pair and instructs the LLM to avoid cycles within this triplet. We show, both theoretically and empirically, that such a triplet prompt leads to fewer cycles than the pairwise prompt. Across multiple real-world graphs, the triplet prompt yields a more accurate order using both LLMs and human annotators as experts. By querying the expert with different auxiliary variables for the same variable pair, it also increases robustness---triplet method with much smaller models such as Phi-3 and Llama-3 8B outperforms a pairwise prompt with GPT-4. For practical usage, we show how the estimated causal order from the triplet method can be used to reduce error in downstream discovery and effect inference tasks. | Causal Order, Imperfect Experts, Causal Inference, LLMs | null | 13,202 | null | [
0.015282552689313889,
-0.03956151753664017,
0.03461233153939247,
0.053154315799474716,
0.04434284567832947,
-0.0005260012112557888,
-0.043618157505989075,
-0.02102484181523323,
0.02330242656171322,
0.013997750356793404,
-0.07371779531240463,
-0.010483097285032272,
0.08115513622760773,
0.05099008232355118,
0.021232442930340767,
0.0402105487883091,
0.033278144896030426,
0.006937786471098661,
-0.03002529963850975,
-0.08275744318962097,
0.021951068192720413,
-0.00776991993188858,
-0.008656037040054798,
-0.021574893966317177,
0.021035639569163322,
-0.047315388917922974,
-0.06414476037025452,
0.0003265179693698883,
0.050615254789590836,
-0.024528900161385536,
-0.040216781198978424,
0.0553254596889019,
-0.042059317231178284,
-0.007674203719943762,
-0.03196006268262863,
-0.021735904738307,
-0.04399462416768074,
0.0483468659222126,
0.05236118659377098,
-0.02212640456855297,
0.001236717333085835,
-0.016381394118070602,
0.041642241179943085,
0.008345814421772957,
0.04911135509610176,
0.05389462038874626,
-0.014369355514645576,
0.061213959008455276,
-0.0068822032772004604,
0.01584354229271412,
-0.12757456302642822,
-0.06649746000766754,
0.06824380904436111,
0.03196302801370621,
0.01600748300552368,
0.010392313823103905,
0.01034068688750267,
0.03471338003873825,
-0.02663237228989601,
0.017746660858392715,
-0.08232727646827698,
-0.05530763417482376,
-0.11138900369405746,
0.013616388663649559,
0.07295776903629303,
0.021621065214276314,
-0.026760684326291084,
0.07926417887210846,
-0.028103947639465332,
0.1306980848312378,
0.009567962028086185,
0.007080665789544582,
-0.0679612010717392,
0.038709018379449844,
0.029480572789907455,
0.05845000222325325,
-0.010649042204022408,
0.004911594092845917,
-0.02342935837805271,
-0.09105326980352402,
-0.023294206708669662,
0.016536131501197815,
0.015133136883378029,
0.008436444215476513,
0.10298211127519608,
-0.03815409913659096,
-0.010283136740326881,
0.025893528014421463,
0.028590358793735504,
0.08569788932800293,
-0.07342172414064407,
-0.03840300813317299,
0.09553009271621704,
0.04512811824679375,
0.04939190298318863,
0.14075936377048492,
0.06950157135725021,
-0.015864863991737366,
0.010299543850123882,
0.12349722534418106,
0.03500119224190712,
-0.006708019413053989,
0.026097780093550682,
-0.06039372831583023,
-0.03550107032060623,
-0.049446951597929,
-0.049985382705926895,
0.024528872221708298,
-0.005240283906459808,
-0.10517775267362595,
-0.04704859480261803,
0.04053259640932083,
0.04601261764764786,
0.000011726536286005285,
-0.03347708284854889,
-0.04881986603140831,
0.0061389910988509655,
-0.034173961728811264,
0.060617927461862564,
-0.05838843435049057,
-0.013178667053580284,
0.048074379563331604,
0.04929768666625023,
0.015595365315675735,
0.09117641299962997,
-0.027870887890458107,
-0.0057197981514036655,
2.9517851500467924e-33,
0.08371566981077194,
0.07771327346563339,
0.0928177684545517,
0.03443599119782448,
-0.009190759621560574,
0.05839821696281433,
-0.06661764532327652,
-0.0015019617276266217,
0.006283936090767384,
-0.040779609233140945,
0.028646498918533325,
-0.024406181648373604,
0.006050670053809881,
0.02956920489668846,
-0.01657046750187874,
0.0582834854722023,
0.029334034770727158,
0.05943743884563446,
-0.11126811802387238,
0.005333013832569122,
0.048964325338602066,
-0.04912218078970909,
-0.03032705932855606,
-0.0189565010368824,
-0.007390109822154045,
0.022169796749949455,
0.049961965531110764,
-0.002205192344263196,
0.018588092178106308,
0.007606801111251116,
-0.10996941477060318,
-0.06497667729854584,
0.05882522836327553,
0.044329337775707245,
-0.008058290928602219,
-0.08086223900318146,
-0.06642645597457886,
-0.07672324031591415,
0.010613959282636642,
0.014851411804556847,
-0.03858612850308418,
-0.004975323099642992,
0.05683809518814087,
0.049250777810811996,
-0.05404731631278992,
-0.04217688739299774,
-0.013423671945929527,
0.03522975370287895,
-0.0646924301981926,
0.019594233483076096,
0.024999525398015976,
0.03355308622121811,
0.03301205858588219,
-0.006673458497971296,
0.04021207615733147,
0.046392928808927536,
0.0016774984542280436,
0.020831698551774025,
-0.07610128074884415,
-0.002276210580021143,
0.0061918399296700954,
0.05686038359999657,
0.007045224774628878,
0.03943644091486931,
0.02258945256471634,
0.06487022340297699,
-0.10760047286748886,
0.04642312601208687,
0.1394566297531128,
0.0005820659571327269,
-0.08590086549520493,
0.014222747646272182,
-0.0013534568715840578,
-0.04930209740996361,
0.02439558133482933,
-0.019093450158834457,
-0.06651424616575241,
-0.1086263358592987,
0.06622175127267838,
-0.009886321611702442,
-0.04496831074357033,
-0.12250778824090958,
0.012824763543903828,
-0.012339520268142223,
-0.03912409767508507,
-0.05432566627860069,
0.026580527424812317,
-0.03725776821374893,
0.04426717758178711,
-0.026074476540088654,
-0.010133006609976292,
-0.02150888741016388,
0.0359998494386673,
0.036397941410541534,
-0.010730915702879429,
-3.9112352104604593e-33,
-0.034356489777565,
0.03715105354785919,
0.06883985549211502,
0.05068616196513176,
0.02530616894364357,
-0.08633607625961304,
-0.021784091368317604,
-0.04876772314310074,
-0.013947593979537487,
-0.05011363327503204,
-0.006853514350950718,
-0.022315148264169693,
0.06269124895334244,
0.028970468789339066,
0.07231268286705017,
0.020006591454148293,
0.017650170251727104,
-0.008357064798474312,
0.025042010471224785,
-0.00027873748331330717,
0.02678675577044487,
0.024987509474158287,
-0.08702100068330765,
0.03287160024046898,
-0.032434940338134766,
-0.02686361037194729,
-0.03894465044140816,
0.01967490278184414,
-0.0294094979763031,
0.08084624260663986,
-0.06899961829185486,
-0.054928507655858994,
-0.07143115997314453,
-0.018403630703687668,
0.016907351091504097,
0.08432818949222565,
0.018184518441557884,
-0.046544868499040604,
-0.10345184803009033,
0.08998851478099823,
0.01685505174100399,
-0.014348632656037807,
-0.0976191833615303,
-0.009717417880892754,
-0.04401640221476555,
0.04033976420760155,
-0.0931612029671669,
0.058508001267910004,
0.016126004979014397,
0.0007183241541497409,
-0.06614794582128525,
0.04602472111582756,
-0.0015922351740300655,
-0.04442009702324867,
-0.05154584348201752,
-0.13077515363693237,
0.06404697149991989,
-0.02185133658349514,
0.04314984008669853,
-0.007274246774613857,
-0.11621831357479095,
-0.03988300636410713,
0.048946451395750046,
-0.02370222844183445,
-0.019955875352025032,
-0.053619470447301865,
-0.025149231776595116,
-0.025639908388257027,
0.03448432683944702,
-0.00972564984112978,
0.046725209802389145,
-0.011087486520409584,
-0.10010834038257599,
-0.05627598986029625,
0.07406488806009293,
0.07907494157552719,
-0.002691620495170355,
-0.05232507362961769,
-0.08456596732139587,
-0.051779743283987045,
0.029481086879968643,
0.04190006107091904,
0.06778726726770401,
0.05151912942528725,
-0.03984047845005989,
0.033281803131103516,
0.02026139758527279,
0.09006629139184952,
0.025783492252230644,
0.03850124031305313,
-0.034074872732162476,
0.03596828505396843,
-0.05030699819326401,
0.07247832417488098,
-0.07445178180932999,
-4.821588461823012e-8,
-0.09432899951934814,
0.05378745123744011,
-0.019620578736066818,
0.05174049735069275,
0.08310876041650772,
-0.06287866830825806,
-0.01081173773854971,
0.07856824994087219,
-0.02459753304719925,
0.030659860000014305,
0.0779896229505539,
0.027499690651893616,
-0.0007299887947738171,
-0.07609875500202179,
0.041087500751018524,
0.023259129375219345,
-0.023389097303152084,
-0.0008323901565745473,
-0.09297285228967667,
-0.041542887687683105,
0.08801564574241638,
-0.004670970607548952,
-0.048523519188165665,
0.058090683072805405,
0.0969020202755928,
-0.05864571034908295,
-0.01943630538880825,
0.0984669029712677,
-0.0011073033092543483,
0.024353599175810814,
-0.006722101476043463,
-0.01454889215528965,
-0.022931832820177078,
0.013231290504336357,
-0.0007615709328092635,
0.04434426873922348,
-0.004399753175675869,
-0.015375513583421707,
0.026878539472818375,
-0.06890718638896942,
-0.007391914259642363,
0.06561856716871262,
-0.10411710292100906,
0.06118027865886688,
0.06166859343647957,
-0.009939407929778099,
0.010205757804214954,
-0.10122708976268768,
0.0022315599489957094,
-0.078236423432827,
-0.10010409355163574,
-0.09854204207658768,
0.00326867145486176,
-0.0046586464159190655,
0.09255225956439972,
0.030909553170204163,
-0.03140178322792053,
0.020778099074959755,
-0.006070674397051334,
-0.022377967834472656,
0.06274893134832382,
0.06568581610918045,
-0.003642130410298705,
-0.02318158745765686
] | 0 | 0 | 0 | 0 |
|
Physics of Language Models: Part 2.2, How to Learn From Mistakes on Grade-School Math Problems | https://openreview.net/forum?id=zpDGwcmMV4 | [
"Tian Ye",
"Zicheng Xu",
"Yuanzhi Li",
"Zeyuan Allen-Zhu"
] | Poster | Language models have demonstrated remarkable performance in solving reasoning tasks; however, even the strongest models still occasionally make reasoning mistakes. Recently, there has been active research aimed at improving reasoning accuracy, particularly by using pretrained language models to "self-correct'' their mistakes via multi-round prompting. In this paper, we follow this line of work but focus on understanding the usefulness of incorporating ``error-correction'' data directly into the pretraining stage. This data consists of erroneous solution steps immediately followed by their corrections. Using a synthetic math dataset, we show promising results: this type of pretrain data can help language models achieve higher reasoning accuracy directly (i.e., through simple auto-regression, without multi-round prompting) compared to pretraining on the same amount of error-free data. We also delve into many details, such as (1) how this approach differs from beam search, (2) how such data can be prepared, (3) whether masking is needed on the erroneous tokens, (4) the amount of error required, (5) whether such data can be deferred to the fine-tuning stage, and many others. | pretraining, language model, error correction, error detection | null | 13,200 | 2408.16293 | [
-0.04150409996509552,
-0.00008126362808980048,
0.07428976893424988,
0.05789533630013466,
-0.004973303992301226,
0.005752494093030691,
0.017379991710186005,
0.03411247208714485,
-0.013740752823650837,
-0.017146233469247818,
-0.009376482106745243,
-0.03727122023701668,
0.03397517278790474,
0.09531031548976898,
0.027746057137846947,
-0.016127759590744972,
0.009678976610302925,
0.07919538766145706,
-0.07727095484733582,
-0.11165755987167358,
-0.03949432820081711,
-0.0012703286483883858,
0.029177973046898842,
0.03525535389780998,
0.056524503976106644,
-0.033466167747974396,
-0.0008441017707809806,
-0.09027546644210815,
-0.015067883767187595,
-0.0006095462595112622,
-0.029518969357013702,
0.02038416638970375,
0.019027607515454292,
0.0811077132821083,
-0.06555556505918503,
0.033223703503608704,
-0.031156178563833237,
0.024617711082100868,
0.0009402090217918158,
0.0018570935353636742,
-0.09537927061319351,
-0.07687439024448395,
0.06394244730472565,
0.022674620151519775,
0.0442722886800766,
-0.00940113514661789,
-0.04358813166618347,
0.030722636729478836,
0.02025035209953785,
-0.015474923886358738,
-0.11578533053398132,
-0.04263724386692047,
-0.029085181653499603,
0.018915396183729172,
-0.01077983807772398,
0.005635701585561037,
0.06959254294633865,
0.08866412937641144,
0.02753080613911152,
-0.01391688920557499,
-0.055538393557071686,
-0.11303701996803284,
-0.05391700193285942,
-0.012080627493560314,
-0.05388457328081131,
0.013895107433199883,
-0.009047475643455982,
0.029987245798110962,
-0.014394370838999748,
0.1045742854475975,
-0.025743691250681877,
0.112339548766613,
-0.009697839617729187,
0.01928826980292797,
0.003915689419955015,
0.06785783916711807,
0.009843377396464348,
0.029099447652697563,
0.07139226794242859,
-0.018338365480303764,
-0.045709315687417984,
-0.055020108819007874,
-0.016104551032185555,
-0.025743044912815094,
0.1482795625925064,
0.012918377295136452,
0.020456865429878235,
0.013320052064955235,
0.04380336403846741,
-0.021004658192396164,
-0.007779098581522703,
-0.05525750666856766,
0.03342655301094055,
0.08937094360589981,
0.04076918959617615,
0.11704995483160019,
0.022297268733382225,
-0.05248835310339928,
-0.018295520916581154,
0.06722595542669296,
0.059337470680475235,
0.12812276184558868,
-0.016041671857237816,
-0.03127769008278847,
-0.02548827789723873,
-0.021594040095806122,
0.0901828184723854,
0.04589200019836426,
0.05244487524032593,
-0.10894691199064255,
0.05087093636393547,
0.017123006284236908,
0.017946790903806686,
0.011288024485111237,
-0.049437183886766434,
-0.02127581275999546,
0.018292060121893883,
-0.042624544352293015,
-0.018530119210481644,
0.018155500292778015,
-0.023415209725499153,
0.04194628447294235,
0.07909995317459106,
0.019331442192196846,
0.07411127537488937,
-0.09943684190511703,
-0.08235783874988556,
5.429764894133599e-33,
0.04502780735492706,
0.08000026643276215,
0.027255747467279434,
0.02475629188120365,
-0.0008893607882782817,
-0.04637372866272926,
-0.026219794526696205,
0.006239316426217556,
0.034350547939538956,
-0.0066896239295601845,
0.022792533040046692,
-0.02353723719716072,
-0.08521925657987595,
0.0496310256421566,
0.0177485179156065,
0.03862452507019043,
-0.009259479120373726,
0.012446722015738487,
-0.01864042319357395,
-0.0012181939091533422,
0.04657995328307152,
-0.05520240217447281,
-0.003933577332645655,
-0.04257649555802345,
0.013872012495994568,
0.12955860793590546,
0.06620973348617554,
-0.05396999791264534,
-0.03715912997722626,
0.009120874106884003,
-0.09964535385370255,
-0.05983088165521622,
-0.00213120155967772,
0.09854153543710709,
-0.002398493466898799,
-0.06874663382768631,
0.06611315160989761,
-0.06737390905618668,
0.03244202211499214,
-0.0020882494281977415,
0.011252688243985176,
0.050502270460128784,
0.05247099697589874,
0.07215375453233719,
-0.027515390887856483,
-0.034953199326992035,
-0.015561585314571857,
0.0071656228974461555,
-0.012113813310861588,
0.014585085213184357,
-0.025352686643600464,
0.0477207750082016,
-0.0923628881573677,
-0.061423227190971375,
-0.015799876302480698,
0.04184696450829506,
0.02212873473763466,
0.02746233157813549,
0.013335474766790867,
0.0062728761695325375,
0.015962980687618256,
0.05823730304837227,
0.02365788258612156,
0.03194597363471985,
-0.0594966858625412,
0.010724854655563831,
-0.06839723140001297,
0.05229271948337555,
0.09972667694091797,
0.04833134263753891,
-0.08031076937913895,
-0.04386032745242119,
-0.03475920483469963,
0.034464411437511444,
0.04955677688121796,
-0.014583367854356766,
0.047083333134651184,
-0.09865811467170715,
0.07477210462093353,
-0.016839798539876938,
0.07336356490850449,
-0.018853137269616127,
-0.0428614616394043,
-0.03408081457018852,
-0.030472083017230034,
-0.06363785266876221,
0.05560974031686783,
-0.08877583593130112,
-0.005239998456090689,
-0.034731727093458176,
0.002777277259156108,
-0.017313167452812195,
0.031437039375305176,
-0.011270357295870781,
-0.03031468205153942,
-4.635374995397513e-33,
-0.013023076578974724,
0.02109365537762642,
-0.06420422345399857,
0.09422283619642258,
-0.06355742365121841,
-0.035812802612781525,
0.057749953120946884,
-0.008279557339847088,
0.048005420714616776,
-0.07045692950487137,
-0.048002637922763824,
-0.03280584141612053,
-0.029078936204314232,
0.028579264879226685,
-0.010250152088701725,
-0.026779120787978172,
-0.011270377784967422,
-0.013119673356413841,
-0.03508443757891655,
0.0815945491194725,
-0.007822109386324883,
0.07756409794092178,
-0.13627952337265015,
0.0027948396746069193,
-0.07107806950807571,
0.03344428539276123,
-0.05979844927787781,
0.05920011177659035,
-0.06756313145160675,
0.09257560968399048,
-0.004872781690210104,
-0.07479863613843918,
0.018743915483355522,
-0.026192624121904373,
-0.07210502028465271,
-0.015533131547272205,
0.0782434418797493,
-0.03899136558175087,
-0.024649351835250854,
0.03660060465335846,
0.08589117974042892,
-0.04840502515435219,
-0.08215302228927612,
0.002990271197631955,
-0.06505836546421051,
0.030343661084771156,
-0.05549883469939232,
0.037800177931785583,
-0.03623346984386444,
0.01463353168219328,
-0.027625035494565964,
-0.03349555283784866,
-0.03816334158182144,
-0.020594067871570587,
0.010159295052289963,
-0.04973718151450157,
-0.0016795616829767823,
-0.04559079930186272,
-0.0039651719853281975,
0.0641917884349823,
-0.13082177937030792,
-0.01631150394678116,
0.12308958172798157,
-0.05540555343031883,
-0.008998633362352848,
-0.042257823050022125,
-0.11407212913036346,
0.023707294836640358,
0.03388207405805588,
0.0014071050100028515,
0.030394315719604492,
-0.002850240794941783,
-0.007735759019851685,
0.03701627999544144,
0.03839477151632309,
-0.02121797949075699,
0.00896055344492197,
0.016235003247857094,
0.016855938360095024,
-0.026317259296774864,
0.03478800877928734,
0.022395912557840347,
0.027950115501880646,
0.07005037367343903,
-0.04906587302684784,
0.019453344866633415,
0.07772741466760635,
0.11185802519321442,
-0.007643414661288261,
0.02474125474691391,
-0.019343934953212738,
0.0736643522977829,
0.046926140785217285,
0.07613999396562576,
-0.09262139350175858,
-5.395138558128565e-8,
-0.08623025566339493,
0.06703335791826248,
0.034385789185762405,
0.05071776732802391,
0.018392985686659813,
-0.10916117578744888,
-0.06969483941793442,
0.07370592653751373,
-0.06039126589894295,
-0.05005524680018425,
0.014420523308217525,
0.03314211592078209,
-0.03610922768712044,
-0.04207022860646248,
-0.0561019703745842,
0.09756901860237122,
0.03273695707321167,
0.014330401085317135,
-0.05102244019508362,
-0.02836877852678299,
0.07098296284675598,
0.04244506359100342,
-0.030913883820176125,
-0.010159057565033436,
0.04683217406272888,
-0.018479254096746445,
-0.025123627856373787,
0.15843741595745087,
0.02993014082312584,
-0.013799927197396755,
-0.061845384538173676,
-0.023578684777021408,
0.009585287421941757,
-0.010497893206775188,
0.06090269982814789,
0.04081883653998375,
0.05747748911380768,
-0.039501871913671494,
0.012925994582474232,
-0.03824674338102341,
0.03773624449968338,
0.028964512050151825,
-0.08769477903842926,
0.019734308123588562,
0.026431869715452194,
0.014106781221926212,
-0.0523708201944828,
-0.08098669350147247,
0.006592450197786093,
-0.040397606790065765,
0.028359992429614067,
0.015871498733758926,
-0.08783749490976334,
0.01151146087795496,
0.04503330960869789,
0.0430426187813282,
-0.054569460451602936,
-0.03150029480457306,
-0.033230897039175034,
0.027763349935412407,
0.0002971758949570358,
0.07053230702877045,
-0.01400709804147482,
-0.001968814292922616
] | 0 | 1 | 0 | 0 |
|
Reasoning of Large Language Models over Knowledge Graphs with Super-Relations | https://openreview.net/forum?id=rTCJ29pkuA | [
"Song Wang",
"Junhong Lin",
"Xiaojie Guo",
"Julian Shun",
"Jundong Li",
"Yada Zhu"
] | Poster | While large language models (LLMs) have made significant progress in processing and reasoning over knowledge graphs, current methods suffer from a high non-retrieval rate. This limitation reduces the accuracy of answering questions based on these graphs. Our analysis reveals that the combination of greedy search and forward reasoning is a major contributor to this issue. To overcome these challenges, we introduce the concept of super-relations, which enables both forward and backward reasoning by summarizing and connecting various relational paths within the graph. This holistic approach not only expands the search space, but also significantly improves retrieval efficiency. In this paper, we propose the ReKnoS framework, which aims to Reason over Knowledge Graphs with Super-Relations. Our framework’s key advantages include the inclusion of multiple relation paths through super-relations, enhanced forward and backward reasoning capabilities, and increased efficiency in querying LLMs. These enhancements collectively lead to a substantial improvement in the successful retrieval rate and overall reasoning performance. We conduct extensive experiments on a variety of datasets to evaluate ReKnoS, and the results demonstrate the superior performance of ReKnoS over existing state-of-the-art baselines, with an average accuracy gain of 2.92% across nine real-world datasets. | Knowledge Graphs, Large Language Models, Question Answering | null | 13,191 | null | [
-0.033161018043756485,
-0.04255806654691696,
-0.0015344766434282064,
0.0375857874751091,
0.0284035112708807,
0.014460921287536621,
-0.030949082225561142,
0.07335109263658524,
0.005820278078317642,
0.005310333799570799,
-0.0403478667140007,
0.03787846863269806,
0.08371808379888535,
0.11911588907241821,
-0.0023659064900130033,
0.07852790504693985,
0.1280767023563385,
0.021076008677482605,
-0.06560312956571579,
-0.07365253567695618,
0.01731010340154171,
-0.02866378240287304,
0.025567611679434776,
0.002834061859175563,
-0.010577959939837456,
-0.04032457247376442,
0.0020967042073607445,
-0.014153513126075268,
0.002189901191741228,
0.06778595596551895,
0.025211341679096222,
0.0532197579741478,
-0.07189323753118515,
0.04131538048386574,
-0.023140376433730125,
0.0703226625919342,
-0.017038295045495033,
0.0209772028028965,
-0.01760416105389595,
-0.04175974801182747,
-0.02299979329109192,
0.013413871638476849,
0.06264954805374146,
0.02199845388531685,
0.10091479122638702,
0.08307862281799316,
-0.07002834230661392,
0.034895870834589005,
0.04767250642180443,
0.05049565061926842,
-0.10080640763044357,
-0.02251308225095272,
-0.06845814734697342,
0.07522430270910263,
0.08527461439371109,
0.039215195924043655,
-0.08907579630613327,
0.0146561274304986,
-0.06930813938379288,
-0.019112328067421913,
-0.01372530683875084,
-0.1320883333683014,
-0.05697305500507355,
-0.021950365975499153,
-0.05399639159440994,
-0.019237741827964783,
-0.02728176862001419,
0.06621652841567993,
-0.041650429368019104,
0.09540557116270065,
0.0397290363907814,
0.08830609172582626,
-0.047491807490587234,
0.07879816740751266,
-0.05731977894902229,
0.08911371976137161,
0.018823035061359406,
-0.015193808823823929,
-0.042513687163591385,
-0.08575282245874405,
0.02407906949520111,
0.015656914561986923,
0.0025642213877290487,
0.01483246125280857,
0.11806220561265945,
0.016147825866937637,
0.029587864875793457,
-0.0731964036822319,
0.02484552562236786,
0.06357365846633911,
0.0009351346525363624,
-0.10539045184850693,
0.09111198037862778,
0.0008503383141942322,
0.05930513143539429,
0.09951801598072052,
0.04667157307267189,
-0.011640734039247036,
-0.025915812700986862,
0.057126641273498535,
0.012087770737707615,
0.083611860871315,
0.019813435152173042,
-0.07666102796792984,
0.0042733740992844105,
0.020160648971796036,
-0.009880662895739079,
0.092170849442482,
-0.01977640762925148,
-0.08501917868852615,
-0.017449328675866127,
0.0610266849398613,
0.0221959687769413,
-0.004248947370797396,
-0.02779521606862545,
-0.04059496894478798,
0.030985645949840546,
-0.0208395104855299,
0.015436968766152859,
-0.048218827694654465,
-0.031927723437547684,
0.08632205426692963,
0.020198151469230652,
-0.0001535706833237782,
0.03976701572537422,
-0.06610459089279175,
-0.03971252962946892,
-7.230749064413538e-34,
0.10926440358161926,
0.06967480480670929,
0.058867447078228,
-0.012018399313092232,
0.018746955320239067,
0.021449584513902664,
-0.0011647045612335205,
0.08943354338407516,
-0.04611285403370857,
0.04024306684732437,
-0.048597317188978195,
0.018839111551642418,
-0.019786207005381584,
-0.054753225296735764,
-0.007857969962060452,
0.01531799603253603,
-0.006940983235836029,
0.020404016599059105,
-0.054074253886938095,
-0.03813277930021286,
0.04702576622366905,
0.005336985923349857,
0.008802182041108608,
0.04959283396601677,
0.022590668871998787,
0.023937327787280083,
0.024205094203352928,
-0.039598796516656876,
-0.01162748597562313,
-0.017245668917894363,
-0.06224183365702629,
-0.046285249292850494,
-0.017012685537338257,
0.03195292130112648,
-0.034660495817661285,
-0.05285908281803131,
-0.05985467880964279,
-0.09376030415296555,
0.01302466168999672,
-0.0330829918384552,
-0.03594006597995758,
0.05452443286776543,
-0.02247823029756546,
0.04688935726881027,
-0.04603152722120285,
-0.04869258776307106,
-0.03115888684988022,
0.010393858887255192,
0.03879522159695625,
-0.021073369309306145,
0.0716133639216423,
-0.0041253878735005856,
-0.07245153188705444,
0.019150014966726303,
-0.0008934422512538731,
-0.05686646327376366,
0.0009340952965430915,
0.08512349426746368,
-0.0053066471591591835,
0.06583131849765778,
-0.013535424135625362,
-0.02387293428182602,
0.00807962380349636,
0.06326208263635635,
0.029406683519482613,
0.014778516255319118,
-0.11442545801401138,
0.07212428748607635,
0.07190912216901779,
-0.006330796983093023,
0.006768890656530857,
0.022059936076402664,
0.07685596495866776,
-0.08437706530094147,
0.017303843051195145,
0.08356008678674698,
-0.044287774711847305,
-0.08855902403593063,
0.03251471370458603,
-0.024019138887524605,
-0.06562035530805588,
-0.02770869992673397,
-0.008335182443261147,
0.008832299150526524,
0.0015287610003724694,
-0.0005249973037280142,
0.029613660648465157,
-0.11067037284374237,
0.0496191680431366,
-0.0473988801240921,
-0.004405690357089043,
-0.036786701530218124,
-0.04522336274385452,
-0.029512017965316772,
0.034607358276844025,
1.292782605753443e-33,
0.01218623947352171,
0.004466227721422911,
0.030565084889531136,
0.009534133598208427,
0.017734220251441002,
-0.06078272685408592,
0.017522744834423065,
0.04304790124297142,
-0.07245834171772003,
-0.03873216733336449,
-0.0025616639759391546,
-0.08384475111961365,
0.07118517905473709,
-0.03701840713620186,
0.0501965694129467,
0.003517033765092492,
0.016830071806907654,
-0.021442990750074387,
-0.04589438438415527,
0.06595908850431442,
-0.06494470685720444,
0.14075469970703125,
-0.09158236533403397,
0.0759318396449089,
0.017448004335165024,
0.020560061559081078,
0.006745245773345232,
0.05271077901124954,
-0.03953470289707184,
0.07763199508190155,
-0.024738585576415062,
-0.038846418261528015,
-0.054071445018053055,
-0.07541488111019135,
-0.09362972527742386,
0.05235746130347252,
-0.008029208518564701,
-0.052329786121845245,
-0.0422336719930172,
-0.03606965020298958,
0.02991407737135887,
-0.04090295732021332,
-0.003781183622777462,
0.008688902482390404,
-0.06761692464351654,
0.05568477138876915,
-0.13745665550231934,
0.030342547222971916,
0.02021283283829689,
-0.023012574762105942,
-0.008273414336144924,
-0.07115630060434341,
0.032060734927654266,
-0.01667117327451706,
0.05171048268675804,
-0.08012045174837112,
-0.08610101044178009,
0.01676502823829651,
-0.014480390585958958,
-0.008240413852036,
0.011472361162304878,
0.014768330380320549,
0.0017921277321875095,
-0.013053534552454948,
0.015038993209600449,
-0.006480499170720577,
-0.028534214943647385,
0.006583023816347122,
-0.02435368299484253,
-0.028030985966324806,
0.010242696851491928,
0.012446494773030281,
-0.03722791746258736,
-0.0034258938394486904,
0.09770257025957108,
0.0177459754049778,
0.055104080587625504,
-0.054680127650499344,
-0.02548261173069477,
-0.060908954590559006,
-0.05886498838663101,
0.04948441684246063,
0.04132857173681259,
0.089055635035038,
0.062133125960826874,
0.013267292641103268,
-0.0007975154439918697,
0.11586354672908783,
0.04998995363712311,
0.03851473331451416,
-0.004552863072603941,
-0.019095400348305702,
-0.008175891824066639,
0.05038708075881004,
0.017615506425499916,
-3.646327684236894e-8,
-0.15133367478847504,
0.08299390971660614,
-0.05327349901199341,
0.02818339131772518,
0.047552440315485,
-0.0665450319647789,
-0.004088093061000109,
0.14338846504688263,
-0.057673994451761246,
0.004076547455042601,
0.0362481027841568,
-0.009883209131658077,
-0.037428420037031174,
-0.011815252713859081,
0.0009391066269017756,
0.02041548490524292,
0.028336720541119576,
-0.032403260469436646,
-0.022906593978405,
-0.004969351924955845,
0.005194804165512323,
-0.023043658584356308,
-0.04238307476043701,
0.06219736486673355,
0.044870514422655106,
-0.022565867751836777,
-0.02314820885658264,
0.08252827078104019,
0.057921528816223145,
-0.016163356602191925,
-0.056875407695770264,
-0.07812003791332245,
0.017013104632496834,
0.020741580054163933,
0.09025923162698746,
0.045716773718595505,
-0.024652618914842606,
0.005788709968328476,
-0.017763495445251465,
-0.04191344976425171,
0.08273741602897644,
0.0226738378405571,
-0.036328744143247604,
0.129916712641716,
-0.00461098225787282,
-0.04026713967323303,
-0.002397890668362379,
-0.009370516054332256,
-0.03513806313276291,
-0.021069886162877083,
-0.056865058839321136,
-0.099408358335495,
-0.01584428735077381,
-0.06286683678627014,
0.005050907842814922,
0.012414184398949146,
-0.07102417945861816,
0.040862075984478,
0.03420109674334526,
-0.05331162363290787,
0.06099352985620499,
0.01565352827310562,
-0.00666090939193964,
0.09877718240022659
] | 0 | 0 | 0 | 0 |
|
Physics of Language Models: Part 2.1, Grade-School Math and the Hidden Reasoning Process | https://openreview.net/forum?id=Tn5B6Udq3E | [
"Tian Ye",
"Zicheng Xu",
"Yuanzhi Li",
"Zeyuan Allen-Zhu"
] | Poster | Recent advances in language models have demonstrated their capability to solve mathematical reasoning problems, achieving near-perfect accuracy on grade-school level math benchmarks like GSM8K. In this paper, we formally study how language models solve these problems. We design a series of controlled experiments to address several fundamental questions: (1) Can language models truly develop reasoning skills, or do they simply memorize templates? (2) What is the model's hidden (mental) reasoning process? (3) Do models solve math questions using skills similar to or different from humans? (4) Do models trained on GSM8K-like datasets develop reasoning skills beyond those necessary for solving GSM8K problems? (5) What mental process causes models to make reasoning mistakes? (6) How large or deep must a model be to effectively solve GSM8K-level math questions?
Our study uncovers many hidden mechanisms by which language models solve mathematical questions, providing insights that extend beyond current understandings of LLMs. | linear probing, language model, grade math problems, logic following, reasoning | Probing reveals that LLMs develop some "level-2" reasoning skill beyond Humans on solving grade-school level math problems | 13,188 | 2407.20311 | [
-0.040482666343450546,
-0.010731496848165989,
0.054963137954473495,
0.022227779030799866,
-0.04526020959019661,
-0.010462014004588127,
-0.00032348590320907533,
0.03314239904284477,
0.015444163233041763,
0.021066807210445404,
-0.010614677332341671,
-0.02364797331392765,
0.06314659863710403,
0.04435522481799126,
0.05664173886179924,
-0.02913478948175907,
0.012500662356615067,
0.03880452737212181,
-0.13444800674915314,
-0.11201632767915726,
0.021370086818933487,
-0.020320327952504158,
0.014637894928455353,
-0.0021560306195169687,
0.018260929733514786,
0.010260299779474735,
0.01968122459948063,
-0.09585537016391754,
0.006379824131727219,
0.03139617666602135,
0.007202625274658203,
-0.009149010293185711,
-0.0262831449508667,
0.10342933982610703,
-0.05089275538921356,
0.04873369634151459,
-0.020342499017715454,
0.00032387449755333364,
0.005662696901708841,
-0.02782103605568409,
-0.09499074518680573,
-0.07676296681165695,
0.0497463196516037,
-0.04116865620017052,
0.10443225502967834,
-0.04934847354888916,
-0.011737238615751266,
0.04166045039892197,
-0.08840931206941605,
-0.026979872956871986,
-0.1703685224056244,
-0.06784572452306747,
-0.01793784834444523,
-0.0166232381016016,
-0.02286720834672451,
0.008158759213984013,
0.006337645463645458,
0.06585313379764557,
0.03184405714273453,
0.0009423824376426637,
-0.09195440262556076,
-0.10775808244943619,
-0.0504465326666832,
0.04299193248152733,
-0.030771993100643158,
-0.007661894895136356,
0.0027232859283685684,
0.015106018632650375,
0.012208481319248676,
0.054232899099588394,
0.0232379287481308,
0.10297535359859467,
-0.013971438631415367,
-0.005473803263157606,
0.013042783364653587,
0.031740088015794754,
0.018850654363632202,
0.03328818455338478,
0.057497598230838776,
-0.017498807981610298,
0.014247049577534199,
-0.06736205518245697,
-0.0004262841248419136,
0.03303435072302818,
0.10269521176815033,
-0.03086256980895996,
0.002825944684445858,
0.07954124361276627,
-0.01118782814592123,
-0.041993461549282074,
0.00043493154225870967,
-0.06670335680246353,
-0.015341395512223244,
0.03142784535884857,
0.03097325563430786,
0.10783939808607101,
0.02051415853202343,
-0.015071682631969452,
-0.053205836564302444,
0.02130412869155407,
0.059944819658994675,
0.08848094195127487,
0.04310207441449165,
0.004469994455575943,
0.024726277217268944,
-0.015179811045527458,
0.03416251391172409,
0.04402810335159302,
0.06900621950626373,
-0.09565237164497375,
-0.004349013790488243,
0.045122742652893066,
0.003586607286706567,
0.10242177546024323,
-0.04026542603969574,
-0.07041379809379578,
0.06957050412893295,
-0.05528558790683746,
-0.011263546533882618,
0.021326061338186264,
-0.02960878424346447,
0.06829388439655304,
0.012167908251285553,
-0.03497867286205292,
0.06533560901880264,
0.014418616890907288,
-0.10701417177915573,
4.024613477307192e-33,
0.04239840805530548,
0.01865818165242672,
0.03412802517414093,
0.046233102679252625,
0.050025541335344315,
-0.05306773632764816,
-0.005602508783340454,
0.006691317539662123,
0.10015588253736496,
0.05950779840350151,
-0.02382057160139084,
0.02569916471838951,
-0.07247137278318405,
0.06427761912345886,
0.056481268256902695,
0.05508839339017868,
-0.09813135862350464,
-0.01544308103621006,
-0.029927022755146027,
-0.014947300776839256,
0.07243836671113968,
-0.02521955408155918,
0.08964542299509048,
-0.04599868133664131,
-0.0029300269670784473,
0.11302923411130905,
0.05044938623905182,
-0.06829126924276352,
-0.03633652254939079,
0.0058571952395141125,
-0.0937887653708458,
-0.029491394758224487,
-0.09272602945566177,
0.0581846721470356,
-0.012395890429615974,
-0.002599900122731924,
0.04702313616871834,
-0.09259741753339767,
0.08051616698503494,
-0.06385525315999985,
-0.01760064996778965,
0.002258222782984376,
0.0022162063978612423,
0.06581133604049683,
-0.06001787632703781,
-0.0037986882962286472,
0.009042355231940746,
-0.043092019855976105,
-0.0034723905846476555,
0.015629181638360023,
-0.0002467046142555773,
0.011293510906398296,
-0.06336069107055664,
-0.06415422260761261,
0.03595942258834839,
0.05322235822677612,
0.007771814242005348,
0.0019340038998052478,
-0.04023541510105133,
0.051074862480163574,
0.024388741701841354,
0.038588881492614746,
0.05076893791556358,
0.05651293322443962,
0.006576803512871265,
0.0053745475597679615,
-0.094222791492939,
0.06753435730934143,
0.10492981225252151,
0.017816510051488876,
-0.031401876360177994,
-0.039273664355278015,
-0.020296962931752205,
0.018210554495453835,
0.03254547715187073,
-0.03403652459383011,
0.04314529895782471,
-0.06725457310676575,
0.037594303488731384,
0.06758423149585724,
0.0873408243060112,
-0.045857612043619156,
0.004691931884735823,
-0.065085768699646,
-0.05785118415951729,
-0.07626655697822571,
0.05657148361206055,
-0.07471737265586853,
-0.012648266740143299,
-0.11625516414642334,
-0.04044892638921738,
-0.05053938180208206,
0.03741633892059326,
-0.0054215858690440655,
-0.008697311393916607,
-5.697715157983693e-33,
-0.08644870668649673,
-0.03824077919125557,
-0.051280830055475235,
0.08471880108118057,
-0.07423845678567886,
-0.07155197113752365,
0.016905177384614944,
-0.006088580004870892,
0.0038748958613723516,
-0.08261735737323761,
-0.04011114686727524,
0.014727862551808357,
0.03718666359782219,
0.06284651160240173,
-0.008577919565141201,
-0.08681176602840424,
-0.02984306402504444,
-0.0033801558893173933,
0.022280648350715637,
0.015036899596452713,
-0.03608343005180359,
0.08428849279880524,
-0.10833392292261124,
0.034485165029764175,
0.000007050350632198388,
-0.020967787131667137,
-0.08337067067623138,
-0.004221232142299414,
-0.031494367867708206,
0.09333915263414383,
0.023861723020672798,
-0.053589750081300735,
0.02298717573285103,
-0.026697717607021332,
-0.04711882397532463,
-0.02260013297200203,
0.05172038450837135,
-0.040575891733169556,
0.03946271911263466,
0.042478181421756744,
0.04116249084472656,
-0.05951285734772682,
-0.014503665268421173,
0.0062554096803069115,
-0.01427557971328497,
0.08134329319000244,
-0.021389981731772423,
0.0009553037234582007,
-0.03913367912173271,
-0.06425502896308899,
0.0931755006313324,
-0.039073627442121506,
-0.015008965507149696,
-0.014076504856348038,
-0.03227700665593147,
0.017790788784623146,
-0.02859729714691639,
-0.03660416230559349,
0.01486059557646513,
0.03981321305036545,
-0.03328150883316994,
-0.002897966653108597,
0.07187116146087646,
-0.051043447107076645,
-0.0486043356359005,
-0.014775656163692474,
-0.053915396332740784,
0.023951077833771706,
0.006258344277739525,
-0.046162594109773636,
0.06530330330133438,
-0.02393590472638607,
-0.0046650199219584465,
0.08385904133319855,
0.03332575410604477,
0.013836449012160301,
-0.06824339181184769,
-0.04602132365107536,
0.0053775073029100895,
-0.1336134672164917,
0.04923374578356743,
-0.004197327420115471,
-0.009020178578794003,
0.02608640305697918,
0.0031332974322140217,
-0.040767524391412735,
0.05274437367916107,
0.06449013948440552,
0.011607026681303978,
-0.02639765292406082,
-0.0476447232067585,
0.051085058599710464,
-0.028047339990735054,
0.10740064084529877,
-0.11781627684831619,
-4.432939348930631e-8,
-0.041705455631017685,
0.05830554664134979,
0.005642512813210487,
0.024698814377188683,
0.02359093725681305,
-0.029207857325673103,
-0.06913845986127853,
0.024927474558353424,
0.0039132134988904,
-0.04213450476527214,
0.03735453635454178,
0.020123520866036415,
-0.06964725255966187,
0.04173882678151131,
-0.007620072457939386,
0.06775650382041931,
0.07090051472187042,
0.056526023894548416,
0.0065978458151221275,
-0.03782800957560539,
0.10435615479946136,
-0.01489969901740551,
0.012123269028961658,
0.05250338092446327,
0.04333816096186638,
0.004804269410669804,
-0.057978227734565735,
0.15669555962085724,
0.015008818358182907,
0.03908282145857811,
-0.03122095577418804,
-0.04088710620999336,
-0.03065299801528454,
0.0041777766309678555,
0.09896702319383621,
0.0627799853682518,
0.009488669224083424,
-0.053495071828365326,
0.002007502829656005,
-0.04227346181869507,
-0.003161222906783223,
0.0227703545242548,
-0.06413321942090988,
0.021938269957900047,
0.03943753242492676,
-0.003955451771616936,
-0.05050064995884895,
-0.08326633274555206,
-0.018878499045968056,
0.069084532558918,
-0.011134378612041473,
0.029941532760858536,
-0.10578704625368118,
-0.016213765367865562,
0.019865408539772034,
-0.03586079925298691,
-0.028936628252267838,
-0.027919169515371323,
-0.0729902908205986,
-0.026105044409632683,
0.049125753343105316,
0.0788268968462944,
-0.03647118806838989,
0.0015594218857586384
] | 0 | 1 | 1 | 0 |
|
Agent-Oriented Planning in Multi-Agent Systems | https://openreview.net/forum?id=EqcLAU6gyU | [
"Ao Li",
"Yuexiang Xie",
"Songze Li",
"Fugee Tsung",
"Bolin Ding",
"Yaliang Li"
] | Poster | Through the collaboration of multiple LLM-empowered agents possessing diverse expertise and tools, multi-agent systems achieve impressive progress in solving real-world problems. Given the user queries, the meta-agents, serving as the brain within multi-agent systems, are required to decompose the queries into multiple sub-tasks that can be allocated to suitable agents capable of solving them, so-called agent-oriented planning. In this study, we identify three critical design principles of agent-oriented planning, including solvability, completeness, and non-redundancy, to ensure that each sub-task can be effectively resolved, resulting in satisfactory responses to user queries. These principles further inspire us to propose AOP, a novel framework for agent-oriented planning in multi-agent systems, leveraging a fast task decomposition and allocation process followed by an effective and efficient evaluation via a reward model. According to the evaluation results, the meta-agent is also responsible for promptly making necessary adjustments to sub-tasks and scheduling. Besides, we integrate a feedback loop into AOP to further enhance the effectiveness and robustness of such a problem-solving process. Extensive experiments demonstrate the advancement of AOP in solving real-world problems compared to both single-agent systems and existing planning strategies for multi-agent systems. The source code is available at https://github.com/lalaliat/Agent-Oriented-Planning | Multi-Agent System; Planning | null | 13,187 | 2410.02189 | [
0.0025608516298234463,
0.0032994591165333986,
-0.061694346368312836,
0.03209371492266655,
-0.044192880392074585,
-0.022256014868617058,
-0.017056720331311226,
-0.015509199351072311,
-0.022768141701817513,
0.0023038210347294807,
-0.10323318839073181,
-0.04932938143610954,
0.05693695321679115,
0.05686737224459648,
0.04513710364699364,
0.02858744189143181,
0.061767105013132095,
-0.020898811519145966,
-0.047600891441106796,
-0.014158878475427628,
0.05715326964855194,
-0.0016930304700508714,
0.053990282118320465,
-0.046899907290935516,
-0.08727589249610901,
0.01756834052503109,
-0.019209541380405426,
-0.07168322801589966,
-0.031561944633722305,
0.0278947576880455,
0.08978544175624847,
0.0302469190210104,
-0.049985121935606,
0.007111291401088238,
-0.020455962046980858,
0.07222113013267517,
-0.13012902438640594,
0.04210465028882027,
-0.005991706158965826,
-0.0032429518178105354,
-0.02359832637012005,
-0.019641682505607605,
-0.07444571703672409,
-0.0205767210572958,
0.003000487806275487,
0.020197371020913124,
-0.1024903655052185,
-0.02551582269370556,
0.028436362743377686,
0.007480944972485304,
-0.09363047778606415,
-0.05775494873523712,
0.006738070398569107,
0.04026241973042488,
0.0060291471891105175,
0.06522607803344727,
-0.013710835948586464,
0.005418060347437859,
-0.01651090942323208,
-0.09316730499267578,
-0.03261740133166313,
-0.023231152445077896,
0.03177884221076965,
-0.013514072634279728,
0.004301407840102911,
0.014172310940921307,
0.009112314321100712,
-0.04837098717689514,
-0.03888700529932976,
-0.04199545085430145,
-0.028917627409100533,
0.015413505025207996,
-0.027454182505607605,
-0.005597093142569065,
0.021634675562381744,
0.0562506765127182,
-0.04447370395064354,
-0.0025017012376338243,
0.07433798909187317,
0.025677142664790154,
-0.047357745468616486,
0.05084235593676567,
-0.06953240931034088,
0.06595825403928757,
0.09578728675842285,
-0.05374637246131897,
0.016816865652799606,
0.0440029613673687,
0.0805848240852356,
-0.011443277820944786,
-0.027892373502254486,
-0.016255870461463928,
0.06389583647251129,
-0.006561029702425003,
0.027901308611035347,
0.10131723433732986,
-0.041567254811525345,
-0.04372946172952652,
-0.04452580586075783,
0.034380048513412476,
-0.031605154275894165,
0.04623229429125786,
-0.005516312085092068,
0.001012145192362368,
-0.009215768426656723,
0.008169949054718018,
0.046044643968343735,
0.019268706440925598,
-0.008175937458872795,
-0.10031019896268845,
-0.010208853520452976,
-0.020670972764492035,
0.12676946818828583,
0.0025880318135023117,
0.01596115343272686,
0.08269058167934418,
0.009492420591413975,
0.1135413646697998,
0.03193797171115875,
0.023770423606038094,
0.048886824399232864,
-0.0380820706486702,
0.10000976920127869,
0.034425999969244,
0.16023288667201996,
0.021813534200191498,
0.0033293673768639565,
-2.137438295969621e-33,
0.039518099278211594,
0.019583655521273613,
0.04318959638476372,
0.03051142208278179,
0.05894118919968605,
-0.0513090118765831,
0.06550869345664978,
0.07379060983657837,
-0.06925775855779648,
0.005425407085567713,
0.05127034708857536,
0.011623662896454334,
-0.04911671578884125,
0.046565230935811996,
0.059765782207250595,
-0.02965070679783821,
0.020814133808016777,
0.04887969419360161,
-0.04897438362240791,
-0.06492223590612411,
0.06339944899082184,
-0.022505708038806915,
0.08066295832395554,
-0.003684691619127989,
0.06825534254312515,
0.05804571509361267,
0.010089565068483353,
-0.027959687635302544,
0.04974421486258507,
-0.01661772094666958,
-0.05698030814528465,
0.047168031334877014,
-0.03970186412334442,
0.06429022550582886,
-0.0926574170589447,
0.038972869515419006,
-0.08235038071870804,
-0.06072685122489929,
0.04225128889083862,
0.06918395310640335,
-0.0041811405681073666,
0.03430229797959328,
0.0638498067855835,
-0.04882214963436127,
-0.06746389716863632,
-0.060821034014225006,
-0.02888241410255432,
0.02442026138305664,
0.011955452151596546,
0.0213759932667017,
0.045164406299591064,
0.015868594869971275,
0.025324417278170586,
-0.03931090980768204,
0.00966386403888464,
-0.009724722243845463,
-0.0462549664080143,
0.015114455483853817,
0.011544136330485344,
0.04885894060134888,
0.014287026599049568,
-0.06148015707731247,
-0.018612874671816826,
0.09139907360076904,
0.001709176693111658,
-0.022826889529824257,
-0.004308168776333332,
0.11229939013719559,
0.12629519402980804,
-0.019954586401581764,
0.0008783920202404261,
0.016420498490333557,
0.06864827126264572,
-0.02953324094414711,
-0.050068970769643784,
-0.007211082149296999,
0.0713951587677002,
-0.06298725306987762,
0.009315907955169678,
-0.032290659844875336,
-0.0003766886657103896,
0.037483587861061096,
-0.053349755704402924,
-0.01833140105009079,
0.03163231909275055,
-0.033763155341148376,
-0.022289330139756203,
-0.02235494554042816,
-0.0037264821585267782,
0.030409641563892365,
-0.0321739986538887,
0.05224744230508804,
-0.07111448794603348,
0.0033722571097314358,
-0.011828946880996227,
1.938001904028196e-33,
-0.024509049952030182,
-0.08961354196071625,
-0.1152488961815834,
-0.007504833396524191,
0.04207713156938553,
-0.05225416272878647,
-0.03896249085664749,
-0.17375628650188446,
-0.03957870602607727,
-0.030126944184303284,
-0.1282511204481125,
-0.013403122313320637,
0.055560071021318436,
-0.029391642659902573,
-0.009880940429866314,
-0.014352945610880852,
0.03183818608522415,
-0.06934935599565506,
0.03153947368264198,
0.0244778823107481,
-0.0445435494184494,
0.12085123360157013,
-0.06433925777673721,
-0.020743701606988907,
0.010497061535716057,
0.05552545562386513,
-0.008195774629712105,
-0.03508815914392471,
0.03861653432250023,
0.058395009487867355,
0.0004257843247614801,
-0.06584987044334412,
-0.03957662358880043,
-0.031241077929735184,
0.02793685346841812,
0.071271613240242,
-0.02381867729127407,
0.036901138722896576,
-0.07454971969127655,
0.12076446413993835,
0.04123356193304062,
-0.11049534380435944,
-0.01860901154577732,
-0.001624259864911437,
-0.013470480218529701,
0.04341035708785057,
-0.06447503715753555,
0.03790672495961189,
-0.03525715321302414,
0.012285453267395496,
-0.05164561793208122,
-0.006468401290476322,
-0.08150157332420349,
-0.06241075322031975,
-0.03886686637997627,
-0.010558078065514565,
0.018695887178182602,
-0.00512396777048707,
0.07221818715333939,
-0.045377906411886215,
-0.02692825347185135,
0.04907733201980591,
0.06711242347955704,
0.02863459102809429,
0.001229760586284101,
0.06713631004095078,
0.024973811581730843,
-0.024786731228232384,
0.01129587460309267,
-0.025824930518865585,
0.00789240375161171,
0.011365700513124466,
0.012050149030983448,
-0.0233286302536726,
0.13701143860816956,
0.06283904612064362,
-0.04997576028108597,
0.004702499136328697,
0.0075872489251196384,
-0.016413463279604912,
-0.051599644124507904,
0.08863609284162521,
-0.04777868092060089,
0.052947916090488434,
-0.11504024267196655,
-0.013333437032997608,
-0.03638690337538719,
0.14090164005756378,
-0.032396331429481506,
-0.01766103319823742,
0.00035879501956515014,
-0.013836881145834923,
0.01424485258758068,
0.049900662153959274,
-0.05708847567439079,
-4.152757071551605e-8,
0.023727530613541603,
0.0630117878317833,
0.017122086137533188,
-0.017678439617156982,
0.027065739035606384,
-0.07100297510623932,
-0.022844644263386726,
0.008388467133045197,
0.02935054525732994,
0.014040697365999222,
0.08349518477916718,
-0.023571735247969627,
0.05635399743914604,
0.01227463036775589,
0.04813866689801216,
-0.03558233380317688,
0.02219199948012829,
0.0025308115873485804,
-0.04470187425613403,
-0.11213985085487366,
-0.016811521723866463,
0.005034337751567364,
-0.010597649961709976,
0.011127790436148643,
0.03771790862083435,
0.0037414429243654013,
-0.040754400193691254,
0.0466151162981987,
0.07429571449756622,
0.07424122840166092,
0.006754451431334019,
-0.026621541008353233,
0.09351769834756851,
0.0760965496301651,
0.057668983936309814,
-0.047497592866420746,
0.0072121527045965195,
-0.04205876961350441,
-0.01340494304895401,
-0.09580658376216888,
0.037231411784887314,
0.06795995682477951,
0.019314894452691078,
0.022813845425844193,
0.024595821276307106,
0.007470083888620138,
-0.06184225529432297,
-0.055420324206352234,
0.013073054142296314,
-0.07968501001596451,
-0.07852856069803238,
-0.02952319197356701,
-0.02758515439927578,
0.09814612567424774,
0.04316965490579605,
0.008592579513788223,
0.04561503231525421,
-0.08459170162677765,
0.14030024409294128,
-0.025376785546541214,
-0.008123002015054226,
0.03692534938454628,
-0.07323643565177917,
-0.012269075959920883
] | https://github.com/lalaliat/agent-oriented-planning | 6 | 0 | 0 | 0 |
Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data | https://openreview.net/forum?id=zg3ec1TdAP | [
"Michael Wornow",
"Suhana Bedi",
"Miguel Angel Fuentes Hernandez",
"Ethan Steinberg",
"Jason Alan Fries",
"Christopher Re",
"Sanmi Koyejo",
"Nigam Shah"
] | Poster | Foundation Models (FMs) trained on Electronic Health Records (EHRs) have achieved state-of-the-art results on numerous clinical prediction tasks. However, prior EHR FMs typically have context windows of $<$1k tokens, which prevents them from modeling full patient EHRs which can exceed 10k's of events. For making clinical predictions, both model performance and robustness to the unique properties of EHR data are crucial. Recent advancements in subquadratic long-context architectures (e.g. Mamba) offer a promising solution. However, their application to EHR data has not been well-studied. We address this gap by presenting the first systematic evaluation of the effect of context length on modeling EHR data. We find that longer context models improve predictive performance -- our Mamba-based model surpasses the prior state-of-the-art on 9/14 tasks on the EHRSHOT prediction benchmark. Additionally, we measure robustness to three unique, previously underexplored properties of EHR data: (1) the prevalence of ``copy-forwarded" diagnoses which create artificial token repetition in EHR sequences; (2) the irregular time intervals between EHR events which can lead to a wide range of timespans within a context window; and (3) the natural increase in disease complexity over time which makes later tokens in the EHR harder to predict than earlier ones. Stratifying our EHRSHOT results, we find that higher levels of each property correlate negatively with model performance (e.g., a 14% higher Brier loss between the least and most irregular patients), but that longer context models are more robust to more extreme levels of these properties. Our work highlights the potential for using long-context architectures to model EHR data, and offers a case study on how to identify and quantify new challenges in modeling sequential data motivated by domains outside of natural language. We release all of our model checkpoints and code. | ehr, foundation model, long context, clinical prediction making, healthcare | An analysis of the impact of context length on foundation models trained on structured electronic health record (EHR) data for clinical prediction tasks. | 13,179 | null | [
-0.05142434686422348,
-0.009777409955859184,
0.03408372774720192,
0.03246023505926132,
0.07258260995149612,
-0.05436943843960762,
-0.03616056963801384,
0.03686165437102318,
0.053711555898189545,
-0.05871930345892906,
-0.041376374661922455,
-0.06210015341639519,
0.06278034299612045,
0.017472444102168083,
0.015121460892260075,
-0.041103851050138474,
0.049553148448467255,
-0.016380762681365013,
-0.08303789049386978,
0.019272807985544205,
-0.014571558684110641,
0.09802289307117462,
0.021769804880023003,
0.014793946407735348,
-0.008910872042179108,
-0.007658450398594141,
-0.05152108147740364,
-0.026490703225135803,
0.007503880653530359,
0.017070986330509186,
0.08950060606002808,
0.012758993543684483,
0.03523366153240204,
0.00886845588684082,
-0.07825618237257004,
-0.01549970917403698,
-0.03380274772644043,
0.05745837092399597,
-0.03355272114276886,
-0.046161506325006485,
0.01021178625524044,
-0.019630948081612587,
0.04909437894821167,
0.07250677049160004,
0.05878932774066925,
-0.010670388117432594,
-0.01094611082226038,
-0.03539854660630226,
0.004160142969340086,
0.060535307973623276,
-0.11072366684675217,
-0.014242502860724926,
0.026324234902858734,
0.050477880984544754,
-0.012877793051302433,
0.020176753401756287,
-0.0022842062171548605,
0.010372023098170757,
-0.07741014659404755,
0.049213409423828125,
-0.09317708760499954,
-0.0684882327914238,
-0.0511380210518837,
-0.00022599757357966155,
-0.06470215320587158,
0.04669933393597603,
-0.026585962623357773,
-0.0009508840157650411,
-0.0404670313000679,
0.022831300273537636,
-0.0716601088643074,
0.025684736669063568,
-0.08562841266393661,
0.12400747835636139,
0.007448054384440184,
0.02568867616355419,
0.0030605534557253122,
0.051958002150058746,
0.015156078152358532,
-0.046375662088394165,
0.024618936702609062,
0.023219408467411995,
0.0302434079349041,
-0.020520227029919624,
0.10174034535884857,
-0.007927771657705307,
0.06329431384801865,
0.011688973754644394,
-0.06606622785329819,
-0.014921946451067924,
0.028356444090604782,
-0.08491060137748718,
0.04500859975814819,
-0.007437689229846001,
0.03727167099714279,
0.005968102719634771,
-0.017911136150360107,
-0.004311354365199804,
-0.016353648155927658,
0.052546802908182144,
-0.026467105373740196,
0.07920953631401062,
0.022309748455882072,
0.001362760434858501,
0.013603036291897297,
-0.02926187589764595,
0.005600878037512302,
0.014046763069927692,
0.016079822555184364,
0.00905232410877943,
0.06032731384038925,
0.0858738049864769,
0.02090730331838131,
-0.04679931327700615,
0.07998606562614441,
-0.025438247248530388,
-0.029601886868476868,
-0.048532042652368546,
0.05229497328400612,
0.12810522317886353,
0.018895627930760384,
-0.036914367228746414,
0.0028665701393038034,
-0.016289807856082916,
-0.0043630702421069145,
0.003684011986479163,
-0.08277520537376404,
2.3161494368091025e-33,
0.012357604689896107,
0.019240714609622955,
0.08612167090177536,
-0.0017084430437535048,
0.05774124711751938,
-0.035868238657712936,
-0.06797274947166443,
0.01896578259766102,
0.04012942314147949,
-0.025565676391124725,
-0.027944164350628853,
0.095524862408638,
-0.04726647585630417,
0.07122261077165604,
0.034333620220422745,
0.006639411672949791,
-0.08150748908519745,
0.04921022802591324,
-0.07716745883226395,
0.029328837990760803,
0.018157141283154488,
-0.14078763127326965,
-0.052115924656391144,
0.001247311825864017,
0.03804803639650345,
0.050516560673713684,
0.05956520885229111,
0.0052671246230602264,
0.020955687388777733,
-0.0038414502050727606,
-0.17922718822956085,
0.05338388308882713,
0.023775042966008186,
0.011555276811122894,
0.026808975264430046,
-0.012110000476241112,
0.032029904425144196,
-0.021479329094290733,
-0.0037180513609200716,
-0.0010897063184529543,
-0.02211453765630722,
0.01781691424548626,
0.015311257913708687,
0.018454324454069138,
-0.05770053714513779,
-0.15058408677577972,
-0.05000041052699089,
-0.006790780927985907,
-0.053217802196741104,
0.005576652009040117,
0.047349099069833755,
-0.038658272475004196,
-0.0003804758016485721,
-0.057488586753606796,
-0.08424986153841019,
0.011975320056080818,
-0.0473550409078598,
-0.010235711932182312,
0.001040185452438891,
0.06892865896224976,
0.03742778301239014,
-0.007176354061812162,
0.028270088136196136,
0.022328300401568413,
-0.052072782069444656,
0.0524468757212162,
-0.001171936048194766,
-0.024473026394844055,
0.007663085125386715,
-0.038885585963726044,
0.021398339420557022,
0.05941714718937874,
0.016717102378606796,
-0.021978721022605896,
0.0630268007516861,
-0.046129729598760605,
-0.003988479729741812,
-0.020084630697965622,
-0.055500682443380356,
0.03862781822681427,
0.06065376475453377,
0.028605446219444275,
-0.04951542243361473,
0.019661910831928253,
0.006849715951830149,
-0.07296691089868546,
-0.024324271827936172,
-0.04037419706583023,
-0.20785243809223175,
-0.06778845191001892,
-0.05395929887890816,
0.023116905242204666,
0.04090525582432747,
0.042301177978515625,
-0.01729460433125496,
-3.263110604429252e-33,
0.02231491170823574,
-0.045795511454343796,
-0.0191456638276577,
0.009589060209691525,
0.063635915517807,
-0.08475609123706818,
0.02774147130548954,
0.12413736432790756,
-0.04645776376128197,
-0.08693137764930725,
0.052087243646383286,
-0.05290236324071884,
0.08827459067106247,
-0.0446048341691494,
0.0485437773168087,
0.03180134668946266,
-0.06065719574689865,
0.004464849829673767,
-0.006234862841665745,
0.036606770008802414,
0.08357197046279907,
-0.007932756096124649,
-0.07868270576000214,
0.06807215511798859,
-0.01859941892325878,
0.025451460853219032,
-0.07280299812555313,
0.03780071437358856,
0.024643490090966225,
-0.12345616519451141,
-0.12436755746603012,
-0.05664658173918724,
-0.01556423306465149,
-0.0540640689432621,
-0.05617481842637062,
0.009423011913895607,
0.03401445224881172,
-0.0444505512714386,
-0.014742260798811913,
0.002720883348956704,
0.10666905343532562,
0.030513526871800423,
-0.08529271930456161,
0.0031949947588145733,
0.043592579662799835,
0.021192599087953568,
-0.02573855035007,
-0.02020934969186783,
0.0645521879196167,
-0.011069010943174362,
-0.010210543870925903,
0.03461224213242531,
-0.0841035395860672,
0.0550445131957531,
-0.027411868795752525,
-0.10457118600606918,
-0.05416486784815788,
-0.06853228062391281,
-0.036832354962825775,
-0.02736610919237137,
-0.04835493490099907,
0.012049076147377491,
-0.032669611275196075,
0.057453520596027374,
0.03962074592709541,
-0.045208562165498734,
0.07076061517000198,
-0.030212614685297012,
-0.057389479130506516,
0.0006792043568566442,
-0.06351400166749954,
-0.010297445580363274,
0.0037728867027908564,
-0.006198420654982328,
0.05730564519762993,
-0.004905750043690205,
-0.04205026477575302,
-0.11570976674556732,
-0.03800942748785019,
-0.007269371300935745,
-0.030915137380361557,
-0.062984399497509,
-0.0006054549594409764,
-0.020366841927170753,
0.035806890577077866,
0.09706895798444748,
0.06997150182723999,
0.021096527576446533,
0.050623394548892975,
0.04561571404337883,
-0.04486098513007164,
0.010841683484613895,
-0.05854957550764084,
0.06893543899059296,
-0.08138634264469147,
-4.563559130588146e-8,
0.013831205666065216,
0.05255912244319916,
0.005967827048152685,
-0.018054910004138947,
-0.006402747239917517,
0.0094691701233387,
-0.021838292479515076,
0.12033360451459885,
0.005003712140023708,
0.012059718370437622,
0.0919979065656662,
-0.027646370232105255,
-0.05439312383532524,
-0.056606657803058624,
0.04555642604827881,
0.07467147707939148,
0.07365284115076065,
0.07313276082277298,
0.014560767449438572,
0.0060379114001989365,
0.08011846244335175,
0.035521116107702255,
0.03499208763241768,
-0.011850387789309025,
0.09273119270801544,
-0.03850811347365379,
-0.034969884902238846,
0.08742935210466385,
0.030751680955290794,
-0.05261887609958649,
-0.06044066697359085,
0.021357988938689232,
-0.0049704937264323235,
0.012032495811581612,
-0.08122391253709793,
-0.025054853409528732,
0.07931637018918991,
-0.07652539759874344,
-0.02331830933690071,
-0.021149279549717903,
0.038065098226070404,
0.03490043058991432,
-0.09099613130092621,
0.022433094680309296,
0.07145020365715027,
0.01585393399000168,
0.058502793312072754,
-0.019308455288410187,
0.020525116473436356,
-0.02899421751499176,
0.03873298689723015,
-0.015335244126617908,
0.07843418419361115,
0.04088315740227699,
0.051955245435237885,
0.04908192157745361,
0.0071014645509421825,
0.015322095714509487,
0.06241887807846069,
-0.004265887662768364,
0.01578228548169136,
-0.013840670697391033,
-0.056121837347745895,
0.10992981493473053
] | 0 | 0 | 0 | 0 |
|
Plastic Learning with Deep Fourier Features | https://openreview.net/forum?id=NIkfix2eDQ | [
"Alex Lewandowski",
"Dale Schuurmans",
"Marlos C. Machado"
] | Poster | Deep neural networks can struggle to learn continually in the face of non-stationarity, a
phenomenon known as loss of plasticity.
In this paper, we identify underlying principles that lead to plastic algorithms.
We provide theoretical results showing that linear function approximation, as well as a special case of deep linear networks, do not suffer from loss of plasticity.
We then propose deep Fourier features, which are the concatenation of a sine and cosine in every layer, and we show that this combination provides a dynamic balance between the trainability obtained through linearity and the effectiveness obtained through the nonlinearity of neural networks.
Deep networks composed entirely of deep Fourier features are highly trainable and sustain their trainability over the course of learning.
Our empirical results show that continual learning performance can be improved by replacing ReLU activations with deep Fourier features combined with regularization.
These results hold for different continual learning scenarios (e.g., label noise, class incremental learning, pixel permutations)
on all major supervised learning datasets used for continual learning research, such as CIFAR10, CIFAR100, and tiny-ImageNet. | Fourier, plasticity, neural networks, continual learning | null | 13,172 | 2410.20634 | [
-0.12058078497648239,
-0.03442494198679924,
0.06444715708494186,
0.06547036021947861,
0.029238762333989143,
0.06039092689752579,
0.0037190262228250504,
-0.08728018403053284,
-0.019228698685765266,
-0.07852025330066681,
0.027143385261297226,
-0.01837182231247425,
-0.04838467761874199,
0.0660109594464302,
-0.045752525329589844,
-0.028116442263126373,
0.015349146910011768,
0.1227569505572319,
-0.13546039164066315,
-0.04381784051656723,
0.04174492508172989,
-0.0014701925683766603,
-0.021141845732927322,
0.018361609429121017,
0.061621688306331635,
0.06013147905468941,
0.014833073131740093,
-0.05341872200369835,
0.0730716660618782,
-0.05841757729649544,
0.011255578137934208,
-0.004329602234065533,
-0.06045011058449745,
0.028153447434306145,
-0.03662877529859543,
0.05085430666804314,
-0.04580402001738548,
0.029355952516198158,
0.04557909071445465,
0.020323453471064568,
0.024005431681871414,
0.028413869440555573,
0.008351917378604412,
0.004536829888820648,
0.057608745992183685,
-0.021714355796575546,
0.0516849122941494,
-0.11747004836797714,
0.05401685833930969,
0.006598937790840864,
-0.005006920546293259,
-0.00033827044535428286,
-0.020816422998905182,
0.08247639983892441,
0.09986318647861481,
-0.02106437087059021,
0.07640302181243896,
0.03738407418131828,
-0.037304334342479706,
0.032612208276987076,
0.04927222430706024,
-0.07584170252084732,
0.0447726845741272,
-0.03851849213242531,
-0.011678086593747139,
0.028458360582590103,
-0.009440546855330467,
-0.010563182644546032,
0.028657080605626106,
0.031151290982961655,
-0.005254777614027262,
0.15437470376491547,
-0.11346951127052307,
0.005715031176805496,
0.0601230263710022,
0.03487081825733185,
-0.02901686355471611,
-0.0017444165423512459,
0.05331495404243469,
0.027695998549461365,
0.017621980980038643,
0.005519001744687557,
0.03355519846081734,
-0.08955559134483337,
0.10954757034778595,
0.009005489759147167,
-0.002481713192537427,
0.018275700509548187,
-0.07316970080137253,
-0.012867573648691177,
-0.035029035061597824,
0.048729058355093,
-0.037904102355241776,
-0.06061220169067383,
-0.004327623639255762,
-0.009809678420424461,
0.0032164768781512976,
-0.07045940309762955,
-0.0643094927072525,
0.06961291283369064,
-0.013740641996264458,
-0.020286230370402336,
-0.00857542548328638,
0.0712362751364708,
0.010600169189274311,
-0.026619818061590195,
0.04279157146811485,
0.02571364864706993,
0.07688890397548676,
-0.035696931183338165,
0.018784992396831512,
0.008890459313988686,
-0.005286285188049078,
-0.0016651564510539174,
0.026439081877470016,
-0.013072741217911243,
0.0028724533040076494,
-0.027026306837797165,
0.046533189713954926,
0.09605487436056137,
-0.02189101278781891,
0.00707191601395607,
-0.06203994154930115,
0.02960919216275215,
-0.007638813927769661,
-0.04113520681858063,
-0.0678633525967598,
3.095631312111877e-33,
0.003246854990720749,
0.057573407888412476,
0.015427318401634693,
-0.040553927421569824,
-0.0023216200061142445,
-0.07608415186405182,
-0.05050993338227272,
0.08081556856632233,
0.03863854706287384,
-0.006191709544509649,
-0.04360067844390869,
0.022764060646295547,
-0.07692268490791321,
0.09468495845794678,
0.08614281564950943,
-0.06564807146787643,
-0.013080988079309464,
-0.014027848839759827,
0.06822811812162399,
-0.09341584146022797,
0.01723862998187542,
-0.023641018196940422,
0.03137137368321419,
-0.06122302636504173,
0.005444698967039585,
-0.06012804061174393,
0.05862659960985184,
0.019887376576662064,
-0.053467318415641785,
0.015950143337249756,
-0.09061963856220245,
0.06311722099781036,
0.024002166464924812,
0.025222081691026688,
0.07394400238990784,
-0.048092372715473175,
-0.006901968736201525,
0.023468751460313797,
0.028152309358119965,
-0.015494040213525295,
0.002417835406959057,
0.03066389448940754,
-0.050453174859285355,
-0.024384982883930206,
-0.05000819265842438,
-0.033172354102134705,
0.059923071414232254,
0.0014559883857145905,
-0.060764409601688385,
-0.051302582025527954,
-0.02183469384908676,
0.014145554974675179,
-0.09737559407949448,
-0.033553361892700195,
0.018376275897026062,
-0.018163321539759636,
-0.013386509381234646,
0.0003709601878654212,
-0.018950138241052628,
0.023738615214824677,
0.04013267159461975,
0.03854909539222717,
0.011506500653922558,
0.08825989067554474,
-0.04232107102870941,
0.08849258720874786,
0.006272857543081045,
0.06386743485927582,
0.037563078105449677,
-0.0009359109681099653,
0.004307172726839781,
-0.00615623127669096,
-0.02078067697584629,
-0.12447452545166016,
0.08198507875204086,
-0.023572999984025955,
0.04010092094540596,
-0.08316028863191605,
-0.054689180105924606,
0.03136738762259483,
0.025832954794168472,
0.01737024635076523,
-0.01020484883338213,
-0.03181638568639755,
-0.07152269035577774,
-0.05130184069275856,
0.027378523722290993,
-0.08533737063407898,
0.06435147672891617,
-0.007614294998347759,
-0.06517129391431808,
-0.018560130149126053,
0.037441011518239975,
-0.0009298644727095962,
-0.010374466888606548,
-3.3335392464328154e-33,
-0.06962008774280548,
0.0677502229809761,
-0.032661329954862595,
0.013693446293473244,
-0.05008068308234215,
0.011011183261871338,
-0.028700100257992744,
0.05994362011551857,
-0.01793544925749302,
-0.02758258394896984,
0.04906119778752327,
-0.03828122094273567,
-0.052141714841127396,
-0.002209180500358343,
-0.04752159118652344,
-0.0019208816811442375,
-0.10854946821928024,
-0.032032500952482224,
0.0017666926141828299,
0.029697787016630173,
0.007629445753991604,
0.028905263170599937,
-0.08277872949838638,
0.04790177941322327,
-0.04457258805632591,
0.0541083998978138,
-0.1087719053030014,
0.049768947064876556,
0.03412351757287979,
0.002801298862323165,
-0.0618416890501976,
-0.03354828059673309,
-0.09154631197452545,
-0.03928244113922119,
0.0035815942101180553,
0.067552350461483,
0.06089591234922409,
0.047031741589307785,
-0.017043132334947586,
0.01151666697114706,
0.09372494369745255,
-0.051363080739974976,
0.02967580035328865,
-0.01168966293334961,
0.0034208197612315416,
-0.08951083570718765,
-0.008241712115705013,
-0.03261222690343857,
0.007498422171920538,
0.08582176268100739,
0.009655830450356007,
0.0038482178933918476,
-0.1443522721529007,
0.02257470227777958,
-0.05158571898937225,
0.02434757724404335,
-0.02561628818511963,
0.04492480680346489,
0.06669427454471588,
0.02287057787179947,
-0.09356650710105896,
-0.0475732795894146,
-0.03905284404754639,
0.014666423201560974,
0.002541434019804001,
-0.021116331219673157,
-0.02677840366959572,
0.009437103755772114,
0.033420730382204056,
0.03683933615684509,
0.04847665876150131,
0.026326483115553856,
0.033627524971961975,
0.03455832973122597,
-0.10393068939447403,
-0.024370694532990456,
-0.0051160636357963085,
-0.08871258050203323,
-0.0706828162074089,
0.05182710662484169,
0.0656701847910881,
0.009789327159523964,
-0.06501947343349457,
0.04136992618441582,
0.0567350834608078,
0.0673535019159317,
-0.015378007665276527,
0.01185932569205761,
0.008378302678465843,
-0.0198165625333786,
0.012552081607282162,
0.035640984773635864,
-0.07812110334634781,
0.009089764207601547,
0.030090559273958206,
-4.0647737620247426e-8,
-0.11113979667425156,
0.061430055648088455,
0.010837862268090248,
-0.0418059304356575,
0.08533233404159546,
-0.08784210681915283,
0.07760151475667953,
0.05864911153912544,
-0.026849154382944107,
-0.04097893461585045,
0.11189102381467819,
0.04535644128918648,
-0.03493602201342583,
-0.050194140523672104,
-0.03318427503108978,
0.12090242654085159,
-0.003366834484040737,
-0.026916516944766045,
0.09096301347017288,
-0.05812959745526314,
-0.01477566547691822,
-0.0017365304520353675,
0.0734657421708107,
-0.04001583904027939,
-0.053108081221580505,
-0.10645369440317154,
0.039284490048885345,
0.13585545122623444,
0.044143591076135635,
0.046749722212553024,
-0.005990253761410713,
-0.016862789168953896,
0.02768787555396557,
0.007538714911788702,
0.024816090241074562,
0.048819053918123245,
0.03939631208777428,
-0.0446905680000782,
-0.07376639544963837,
-0.05363951623439789,
-0.0425773486495018,
0.09773494303226471,
-0.025121863931417465,
0.022247396409511566,
0.03073696419596672,
-0.040816884487867355,
0.07455769926309586,
-0.06873270869255066,
0.021073993295431137,
0.012115687131881714,
0.018533041700720787,
0.025116978213191032,
0.019455358386039734,
0.07346730679273605,
0.09071001410484314,
0.004908070899546146,
-0.018783239647746086,
-0.04836941510438919,
-0.06653852015733719,
0.09017857164144516,
-0.046017926186323166,
0.009138369001448154,
0.016610372811555862,
-0.00489877350628376
] | 0 | 0 | 0 | 0 |
|
STAFF: Speculative Coreset Selection for Task-Specific Fine-tuning | https://openreview.net/forum?id=FAfxvdv1Dy | [
"Xiaoyu Zhang",
"Juan Zhai",
"Shiqing Ma",
"Chao Shen",
"Tianlin Li",
"Weipeng Jiang",
"Yang Liu"
] | Poster | Task-specific fine-tuning is essential for the deployment of large language models (LLMs), but it requires significant computational resources and time. Existing solutions have proposed coreset selection methods to improve data efficiency and reduce model training overhead, but they still have limitations: ❶ Overlooking valuable samples at high pruning rates, which degrades the coreset’s performance.
❷ Requiring high time overhead during coreset selection to fine-tune and evaluate the target LLM. In this paper, we introduce STAFF, a speculative coreset selection method. STAFF leverages a small model from the same family as the target LLM to efficiently estimate data scores and then verifies the scores on the target LLM to accurately identify and allocate more selection budget to important regions while maintaining coverage of easy regions. We evaluate STAFF on three LLMs and three downstream tasks and show that STAFF improves the performance of SOTA methods by up to 54.3% and reduces selection overhead by up to 70.5% at different pruning rates. Furthermore, we observe that the coreset selected by STAFF at low pruning rates (i.e., 20%) can even obtain better fine-tuning performance than the full dataset. | Task-specific fine-tuning, coreset selection, speculative execution | We propose STAFF, a speculative coreset selection method for task-specific fine-tuning that outperforms SOTA methods with much lower selection overhead. | 13,170 | null | [
0.028172269463539124,
-0.04186480864882469,
0.0008640842279419303,
0.028966886922717094,
0.0664668157696724,
0.0020056080538779497,
-0.04601169750094414,
0.017175210639834404,
-0.018476879224181175,
-0.007406043820083141,
-0.03568745031952858,
-0.04789680615067482,
-0.005462908651679754,
-0.002285944065079093,
0.035085082054138184,
-0.04121273756027222,
0.11680872738361359,
-0.029879653826355934,
-0.049572452902793884,
-0.06242997571825981,
-0.0008308115648105741,
-0.021742207929491997,
0.005236522294580936,
0.0011010559974238276,
0.09829743951559067,
-0.0838303193449974,
-0.02780485525727272,
-0.005701873917132616,
-0.026050396263599396,
0.017431538552045822,
0.016827529296278954,
0.05125756189227104,
0.0159144327044487,
0.04180087149143219,
-0.08964169770479202,
0.07010718435049057,
-0.021939652040600777,
0.0004212931962683797,
0.034047480672597885,
0.0016413283301517367,
-0.03538317233324051,
-0.0234832800924778,
-0.01065164152532816,
-0.0019109892891719937,
0.05026533082127571,
-0.007722583133727312,
-0.0068566082045435905,
-0.041587069630622864,
-0.015429927967488766,
-0.001479113008826971,
-0.07534202188253403,
-0.05275276303291321,
0.023612577468156815,
0.029759161174297333,
-0.060714345425367355,
0.032178521156311035,
0.08245425671339035,
-0.031012002378702164,
-0.03697583079338074,
-0.05485824868083,
-0.08143731951713562,
-0.14179009199142456,
-0.06735794246196747,
-0.029911499470472336,
0.01014937274158001,
-0.05521809309720993,
0.03842999041080475,
0.015692397952079773,
-0.01722477190196514,
-0.02784629538655281,
0.04057236760854721,
0.06872609257698059,
0.00994851440191269,
0.06166206672787666,
0.06170910224318504,
0.09519799053668976,
0.052101004868745804,
0.010713160037994385,
0.10036692768335342,
-0.10876349359750748,
-0.009978351183235645,
0.005387917160987854,
-0.00022704486036673188,
-0.007917829789221287,
0.06235978752374649,
0.016540786251425743,
0.02938242442905903,
-0.0035737017169594765,
0.04925181716680527,
0.016310255974531174,
0.020747428759932518,
-0.03223143890500069,
0.06585712730884552,
0.018031327053904533,
-0.04691910743713379,
0.04703094810247421,
-0.004427206236869097,
-0.021055256947875023,
-0.06829552352428436,
0.047164708375930786,
-0.01717335544526577,
0.04032561928033829,
0.05359082296490669,
-0.07186862081289291,
-0.08606340736150742,
-0.05118738114833832,
0.0346786268055439,
0.06326142698526382,
-0.01909623108804226,
-0.11229898035526276,
0.003954190295189619,
0.050013020634651184,
-0.007952063344419003,
0.04608095437288284,
0.02039664052426815,
-0.004415284376591444,
0.05437770485877991,
-0.03254130482673645,
-0.008655582554638386,
0.13352268934249878,
0.014168579131364822,
0.05677983909845352,
0.026437396183609962,
0.0029046947602182627,
0.023795578628778458,
0.08162449300289154,
-0.055643852800130844,
3.823938959206565e-33,
0.06301876902580261,
0.07243606448173523,
-0.006417096126824617,
-0.027709338814020157,
-0.00003521927646943368,
-0.08474767208099365,
0.03871568664908409,
0.025956934317946434,
-0.012984723784029484,
0.002317763166502118,
-0.04944545775651932,
-0.01284298487007618,
-0.0019286739407107234,
0.05359215289354324,
0.07790088653564453,
-0.048824504017829895,
0.016156304627656937,
0.13830098509788513,
-0.09699856489896774,
0.017521124333143234,
0.049536723643541336,
-0.018499867990612984,
0.014029303565621376,
-0.1349927932024002,
0.08785713464021683,
-0.0007966813864186406,
0.05147881805896759,
-0.028315963223576546,
-0.07838035374879837,
0.028337061405181885,
-0.0733160749077797,
-0.024449311196804047,
0.007527367677539587,
0.04469137638807297,
0.017577705904841423,
-0.008118881843984127,
-0.05210648104548454,
-0.03845387324690819,
0.08557761460542679,
0.028089288622140884,
-0.005540995392948389,
0.035420212894678116,
0.08476037532091141,
-0.0008481350378133357,
-0.04659703001379967,
-0.029146548360586166,
-0.04019150137901306,
0.022596556693315506,
-0.025785639882087708,
0.031120948493480682,
0.0669131726026535,
-0.007550633046776056,
0.019383516162633896,
0.029579192399978638,
-0.039311498403549194,
0.02676193229854107,
0.048748929053545,
-0.024236716330051422,
0.040185216814279556,
0.042728107422590256,
-0.007413468323647976,
0.004775102250277996,
-0.054345425218343735,
0.11274725943803787,
0.0758867934346199,
-0.03286579251289368,
-0.03870378062129021,
0.05235182121396065,
0.07821361720561981,
-0.04671647399663925,
-0.0702502653002739,
-0.01406504400074482,
0.05798109620809555,
-0.009923703968524933,
0.005957688204944134,
-0.032589107751846313,
0.12572388350963593,
-0.006220956798642874,
-0.010886814445257187,
-0.07193050533533096,
0.04194814711809158,
0.04822876676917076,
-0.0246288999915123,
-0.1140771433711052,
0.01264986116439104,
-0.019114641472697258,
0.03261404484510422,
-0.03814045712351799,
-0.014709355309605598,
-0.012939076870679855,
-0.030162224546074867,
-0.002336789621040225,
-0.06908264756202698,
-0.04892517253756523,
-0.043151237070560455,
-2.897757592973077e-33,
-0.07803277671337128,
-0.019443731755018234,
-0.051099520176649094,
0.06856177002191544,
0.026973828673362732,
-0.02918788231909275,
0.05437523499131203,
-0.02488907054066658,
-0.1096416562795639,
-0.09376239031553268,
-0.060502488166093826,
-0.0546654649078846,
0.08312233537435532,
-0.015394194051623344,
0.007192757911980152,
0.021329596638679504,
-0.026366351172327995,
-0.016514314338564873,
0.04714922606945038,
0.026199569925665855,
0.01436051819473505,
0.04173547402024269,
-0.0641012042760849,
0.031757671386003494,
-0.00980073306709528,
-0.02093859575688839,
-0.10033857822418213,
-0.006534392945468426,
-0.00003884279794874601,
-0.044356342405080795,
-0.057018980383872986,
-0.024995801970362663,
-0.02253669500350952,
-0.05168010666966438,
-0.01878739520907402,
0.02571532502770424,
-0.011557504534721375,
0.01972893625497818,
0.01846284419298172,
0.19523747265338898,
0.010022786445915699,
0.006205055862665176,
-0.03421451896429062,
0.028642866760492325,
-0.0579552985727787,
0.05452965199947357,
-0.06350488215684891,
-0.037155065685510635,
-0.06970862299203873,
-0.00030857292586006224,
-0.040093839168548584,
-0.010671586729586124,
-0.03238065913319588,
0.03488359600305557,
-0.0007793069235049188,
-0.0937146469950676,
0.017172278836369514,
-0.10054252296686172,
-0.009937340393662453,
-0.02340596355497837,
-0.053972501307725906,
0.04074252396821976,
0.06370509415864944,
-0.05039519816637039,
0.052418243139982224,
-0.06480511277914047,
-0.026458168402314186,
-0.019741220399737358,
-0.04850529506802559,
-0.012008841149508953,
-0.021286653354763985,
-0.047506652772426605,
0.033600714057683945,
-0.00927977915853262,
-0.06949537247419357,
0.04468109458684921,
-0.03435675799846649,
-0.017021268606185913,
-0.00035736089921556413,
0.028897179290652275,
-0.016694124788045883,
0.02291174978017807,
-0.01665978692471981,
-0.008678341284394264,
0.06055901199579239,
0.10347869247198105,
0.10748830437660217,
0.06384388357400894,
-0.013084282167255878,
-0.0632234662771225,
-0.05453173816204071,
-0.003715856233611703,
0.11531868577003479,
0.01430356502532959,
-0.03019443340599537,
-5.004662995133913e-8,
0.03497910127043724,
0.019837185740470886,
-0.02625497803092003,
0.08295643329620361,
0.039587195962667465,
-0.05950349196791649,
-0.0795636847615242,
0.0974264070391655,
0.03346148133277893,
-0.0018477244302630424,
0.03373177722096443,
-0.04355696588754654,
-0.027178267017006874,
-0.01976938359439373,
0.052511729300022125,
0.014062739908695221,
0.002536170883104205,
0.07664622366428375,
-0.07150139659643173,
-0.03430837020277977,
0.034281354397535324,
0.08052580803632736,
-0.03680980205535889,
-0.009109128266572952,
0.12007423490285873,
0.018264686688780785,
-0.035956259816884995,
0.1165291890501976,
0.10767977684736252,
0.04151587188243866,
-0.0026627809274941683,
-0.019194092601537704,
-0.05305195599794388,
-0.007293523754924536,
0.08926701545715332,
0.05027160421013832,
-0.0036603561602532864,
0.035881318151950836,
0.0038443091325461864,
0.05238686874508858,
0.06135121360421181,
-0.006269339006394148,
-0.07762005180120468,
0.05671697109937668,
0.04914315789937973,
-0.060235727578401566,
-0.07904519885778427,
-0.05917206406593323,
0.03808281570672989,
-0.07115265727043152,
0.01927879825234413,
-0.06406722217798233,
-0.05671964958310127,
0.051362086087465286,
0.022819144651293755,
0.07442115992307663,
-0.01265906821936369,
-0.03452742099761963,
-0.007069196552038193,
0.02493342198431492,
0.04840821400284767,
-0.025561334565281868,
-0.11377653479576111,
0.0674719586968422
] | 0 | 0 | 0 | 0 |
|
Time-to-Event Pretraining for 3D Medical Imaging | https://openreview.net/forum?id=zcTLpIfj9u | [
"Zepeng Frazier Huo",
"Jason Alan Fries",
"Alejandro Lozano",
"Jeya Maria Jose Valanarasu",
"Ethan Steinberg",
"Louis Blankemeier",
"Akshay S Chaudhari",
"Curtis Langlotz",
"Nigam Shah"
] | Poster | With the rise of medical foundation models and the growing availability of imaging data, scalable pretraining techniques offer a promising way to identify imaging biomarkers predictive of future disease risk. While current self-supervised methods for 3D medical imaging models capture local structural features like organ morphology, they fail to link pixel biomarkers with long-term health outcomes due to a missing context problem. Current approaches lack the temporal context necessary to identify biomarkers correlated with disease progression, as they rely on supervision derived only from images and concurrent text descriptions. To address this, we introduce time-to-event pretraining, a pretraining framework for 3D medical imaging models that leverages large-scale temporal supervision from paired, longitudinal electronic health records (EHRs). Using a dataset of 18,945 CT scans (4.2 million 2D images) and time-to-event distributions across thousands of EHR-derived tasks, our method improves outcome prediction, achieving an average AUROC increase of 23.7% and a 29.4% gain in Harrell’s C-index across 8 benchmark tasks. Importantly, these gains are achieved without sacrificing diagnostic classification performance. This study lays the foundation for integrating longitudinal EHR and 3D imaging data to advance clinical risk prediction. | Multimodal learning, medical imaging, Electronic Health Records | null | 13,164 | 2411.09361 | [
0.029021453112363815,
-0.03320914879441261,
0.039954252541065216,
0.006173125933855772,
0.07121651619672775,
0.019584933295845985,
-0.036261264234781265,
-0.011001896113157272,
0.030670862644910812,
-0.08714298903942108,
-0.050067149102687836,
-0.07717879116535187,
-0.014697808772325516,
0.12770499289035797,
-0.03449764847755432,
-0.05895571783185005,
-0.0041366335935890675,
0.00484450114890933,
-0.07689114660024643,
0.03267315402626991,
-0.05615543946623802,
0.0184896532446146,
0.008315946906805038,
0.0383375883102417,
0.0927487462759018,
-0.04407544806599617,
0.03508657589554787,
-0.11499021202325821,
0.007728717289865017,
0.026397105306386948,
0.06088853254914284,
-0.06320793926715851,
-0.00379680166952312,
0.02395656332373619,
-0.06781197339296341,
-0.0155918775126338,
-0.03925585374236107,
0.03903069719672203,
-0.07331658899784088,
-0.047806210815906525,
0.04248733073472977,
0.006587836425751448,
0.00558226415887475,
0.08819261193275452,
0.05117415264248848,
0.03020872361958027,
-0.009253004565834999,
-0.09001666307449341,
0.03681536763906479,
0.024540308862924576,
-0.07393638044595718,
0.008952825330197811,
-0.028260216116905212,
0.011454354971647263,
0.00641203299164772,
0.13008970022201538,
0.053192295134067535,
-0.06878776848316193,
-0.04806874319911003,
0.005302290432155132,
-0.023784758523106575,
-0.07471805065870285,
-0.01955827698111534,
-0.0018630451522767544,
0.0301884226500988,
0.017818357795476913,
0.035854991525411606,
0.010571328923106194,
-0.0393805094063282,
0.04754617065191269,
-0.06333219259977341,
0.04841659590601921,
0.012454304844141006,
0.07723528891801834,
0.0031046424992382526,
0.035863183438777924,
0.019355883821845055,
0.024719292297959328,
0.030726000666618347,
-0.07634290307760239,
0.004337773192673922,
0.07616385072469711,
0.07889749854803085,
-0.011679708026349545,
0.10049895942211151,
0.02699708752334118,
0.0014037605142220855,
0.07955522835254669,
-0.09910643845796585,
0.03476307541131973,
-0.018887102603912354,
-0.04553968831896782,
-0.015649069100618362,
0.008980149403214455,
-0.007248726207762957,
-0.0002543221053201705,
0.03488920256495476,
-0.07704021781682968,
0.07734362781047821,
0.034199271351099014,
0.002639115322381258,
-0.025939621031284332,
-0.0006077004945836961,
0.07911202311515808,
0.02864563651382923,
-0.057606734335422516,
0.023903165012598038,
-0.006855932530015707,
0.0093439482152462,
0.026703940704464912,
0.11650560051202774,
0.012659689411520958,
0.020316261798143387,
-0.05953318253159523,
0.03885205090045929,
0.02729802392423153,
-0.05735697224736214,
-0.021149948239326477,
0.09178536385297775,
0.05020987242460251,
-0.04224594682455063,
-0.07630830258131027,
-0.005210611037909985,
-0.02177238091826439,
0.013256564736366272,
-0.04482915997505188,
-0.04989500716328621,
-6.308404080114301e-35,
0.06769099086523056,
0.049233999103307724,
0.032534416764974594,
0.011917386204004288,
-0.012795284390449524,
-0.010810715146362782,
-0.0620662085711956,
-0.012394487857818604,
-0.011732451617717743,
-0.015512055717408657,
-0.037269968539476395,
-0.010970031842589378,
-0.04306163638830185,
0.08969566226005554,
0.01024194248020649,
0.017602942883968353,
-0.04943835362792015,
0.10186873376369476,
-0.004810718819499016,
0.014759918674826622,
-0.009875926189124584,
-0.08571723848581314,
-0.06469940394163132,
-0.027268702164292336,
0.03597039356827736,
-0.004825533367693424,
-0.014018463902175426,
-0.039637189358472824,
-0.007700427435338497,
-0.0010811827378347516,
-0.10456599295139313,
0.04968029633164406,
0.04556197300553322,
0.0008726456435397267,
0.04420233517885208,
-0.018940996378660202,
0.010128572583198547,
-0.0034312617499381304,
-0.011786126531660557,
0.05653291568160057,
0.0492018386721611,
0.020358426496386528,
0.010756840929389,
-0.06652036309242249,
-0.0631459653377533,
-0.07811912894248962,
0.010984373278915882,
0.042074523866176605,
-0.10710860043764114,
-0.049906522035598755,
0.07178576290607452,
-0.01741730608046055,
-0.02511834166944027,
-0.0632447898387909,
-0.0690823644399643,
0.009504102170467377,
-0.03954207897186279,
-0.02823912352323532,
0.027132084593176842,
0.04436497762799263,
0.08281885087490082,
0.034250035881996155,
-0.03759780526161194,
0.08570826053619385,
-0.05154596269130707,
-0.006543691270053387,
0.04315382242202759,
0.027561215683817863,
0.06962631642818451,
0.06460433453321457,
-0.03673822060227394,
0.029940079897642136,
-0.0011799645144492388,
-0.1090361624956131,
0.09919553995132446,
0.06325988471508026,
0.015987329185009003,
-0.03400076925754547,
-0.016163598746061325,
0.04240874573588371,
-0.022287826985120773,
-0.02182917855679989,
-0.07334841787815094,
-0.02896488830447197,
-0.0017245891503989697,
-0.0511513277888298,
0.03335794061422348,
-0.026971470564603806,
-0.08501778542995453,
-0.028206508606672287,
-0.04512804374098778,
-0.006531839724630117,
0.005487821064889431,
0.05226979777216911,
-0.024397671222686768,
-8.750098930902865e-34,
0.0765080600976944,
0.010639225132763386,
-0.011425328440964222,
-0.045474786311388016,
0.08988785743713379,
-0.05492589995265007,
-0.02065671980381012,
0.06030980870127678,
0.03510313108563423,
-0.06526162475347519,
0.07437500357627869,
-0.0678074061870575,
0.023991620168089867,
-0.04971893131732941,
-0.00564512750133872,
0.005726184695959091,
-0.0756574422121048,
-0.011557292193174362,
-0.01436472125351429,
0.02811080403625965,
0.059565234929323196,
0.01479126513004303,
-0.07638533413410187,
0.042995885014534,
-0.0971616730093956,
0.05112157016992569,
0.04236745461821556,
0.01897878386080265,
-0.03580765053629875,
-0.028710197657346725,
-0.10602077096700668,
-0.011257576756179333,
0.04260352626442909,
-0.07320134341716766,
-0.07520367205142975,
0.09764198213815689,
0.0680534765124321,
-0.09305542707443237,
-0.027983546257019043,
0.016146842390298843,
0.1137133240699768,
-0.01090127881616354,
-0.11713048070669174,
-0.0008272834238596261,
-0.042379848659038544,
-0.003395459149032831,
-0.03662290424108505,
0.03031807392835617,
0.10453144460916519,
0.1171271800994873,
-0.06495343148708344,
0.019627079367637634,
-0.06725934892892838,
0.028840001672506332,
-0.07149561494588852,
-0.1183813065290451,
-0.049787383526563644,
-0.07307378202676773,
0.002200346440076828,
0.014997678808867931,
-0.11126746982336044,
0.000023579856133437715,
-0.013777715153992176,
-0.0028061282355338335,
0.0031020422466099262,
-0.041439782828092575,
0.008484743535518646,
-0.039200324565172195,
-0.023239070549607277,
0.07621683180332184,
-0.05866250768303871,
0.03151000663638115,
-0.024766206741333008,
-0.032552581280469894,
0.03826897591352463,
-0.06361449509859085,
-0.017224546521902084,
0.0003711522731464356,
0.006271440535783768,
-0.06081879884004593,
0.03797675296664238,
-0.048929937183856964,
0.04014541208744049,
0.08473803848028183,
0.06167236715555191,
0.14535118639469147,
0.04960675537586212,
0.005703364964574575,
0.026668813079595566,
0.0008493622881360352,
-0.024045925587415695,
0.023231929168105125,
0.0057912259362638,
0.040394511073827744,
-0.06462674587965012,
-3.963916128668643e-8,
-0.03521536663174629,
0.050937894731760025,
0.015704400837421417,
-0.016409529373049736,
-0.021184761077165604,
-0.015823399648070335,
-0.03174830600619316,
0.13352636992931366,
0.019357681274414062,
0.009672782383859158,
0.06973929703235626,
-0.06102215498685837,
-0.07324158400297165,
-0.11505136638879776,
0.04326813668012619,
0.057873472571372986,
-0.018743202090263367,
0.02470002882182598,
-0.01677541807293892,
-0.04512500390410423,
-0.020344845950603485,
0.013640238903462887,
0.0759642943739891,
-0.0709957703948021,
0.019609713926911354,
-0.013899530284106731,
0.004919042810797691,
0.06146949902176857,
0.07169982045888901,
0.005873130634427071,
-0.03439638763666153,
-0.008353669196367264,
0.012277388945221901,
0.04354797676205635,
-0.015410776250064373,
-0.026232805103063583,
0.06064969301223755,
-0.07782003283500671,
-0.03233013674616814,
0.0162266343832016,
0.01986461877822876,
0.08144056797027588,
-0.05095291510224342,
0.012396229431033134,
0.03990458697080612,
0.002318315440788865,
0.01988808438181877,
0.0001992750767385587,
0.03520538657903671,
-0.060506097972393036,
-0.038263119757175446,
-0.03452256694436073,
0.06019344925880432,
0.10495460033416748,
0.09460614621639252,
0.09268258512020111,
0.05564405024051666,
0.0018420314881950617,
0.06552555412054062,
0.07073108851909637,
0.02402700111269951,
-0.01775841787457466,
-0.03360743820667267,
-0.018667159602046013
] | 0 | 0 | 0 | 0 |
|
ExACT: Teaching AI Agents to Explore with Reflective-MCTS and Exploratory Learning | https://openreview.net/forum?id=GBIUbwW9D8 | [
"Xiao Yu",
"Baolin Peng",
"Vineeth Vajipey",
"Hao Cheng",
"Michel Galley",
"Jianfeng Gao",
"Zhou Yu"
] | Poster | Autonomous agents have demonstrated significant potential in automating complex multistep decision-making tasks. However, even state-of-the-art vision-language models (VLMs), such as GPT-4o, still fall short of human-level performance, particularly in intricate web environments and long-horizon planning tasks. To address these limitations, we introduce Reflective Monte Carlo Tree Search (R-MCTS), a novel test-time algorithm designed to enhance the ability of AI agents, e.g., powered by GPT-4o, to explore decision space on the fly.
R-MCTS extends traditional MCTS by 1) incorporating contrastive reflection, allowing agents to learn from past interactions and dynamically improve their search efficiency; and 2) using multi-agent debate to provide reliable state evaluation. Moreover, we improve the agent's performance by fine-tuning GPT-4o through self-learning, using R-MCTS generated tree traversals without any human-provided labels. On the challenging VisualWebArena benchmark, our GPT-4o-based R-MCTS agent achieves a 6% to 30% relative improvement across various tasks compared to the previous state-of-the-art. Additionally, we show that the knowledge gained from test-time search can be effectively transferred back to GPT-4o via fine-tuning. The fine-tuned GPT-4o matches 97\% of R-MCTS's performance while reducing compute usage by a factor of four at test time. Furthermore, qualitative results reveal that the fine-tuned GPT-4o model demonstrates the ability to explore the environment, evaluate a state, and backtrack to viable ones when it detects that the current state cannot lead to success. Moreover, our work demonstrates the compute scaling properties in both training - data collection with R-MCTS - and testing time. These results suggest a promising research direction to enhance VLMs' reasoning and planning capabilities for agentic applications via test-time search and self-learning. | AI agent, tree search, self-improvement | null | 13,162 | 2410.02052 | [
-0.01691902056336403,
-0.08916417509317398,
0.002016076585277915,
0.060608264058828354,
0.12235090136528015,
-0.05667325481772423,
-0.033861760050058365,
0.00690285861492157,
0.007299617398530245,
-0.01983986794948578,
-0.10134775936603546,
-0.08289068937301636,
-0.0074196322821080685,
0.06128866970539093,
0.02179850824177265,
0.011955388821661472,
0.11646678298711777,
0.06918975710868835,
-0.011722138151526451,
-0.05580919608473778,
0.05927400290966034,
-0.023137910291552544,
0.047266244888305664,
-0.08367006480693817,
-0.01132652722299099,
0.01961442083120346,
0.030573807656764984,
-0.10581109672784805,
0.005884990096092224,
-0.004160167183727026,
0.07544505596160889,
-0.0485222265124321,
0.007921750657260418,
0.0030324957333505154,
-0.056518279016017914,
0.07431179285049438,
-0.094056136906147,
-0.005835050251334906,
0.03500106558203697,
-0.05880128964781761,
-0.027758609503507614,
0.04411492496728897,
0.047838546335697174,
-0.04709918051958084,
0.06386704742908478,
-0.02103017456829548,
-0.04045813903212547,
-0.05044424161314964,
0.07589966803789139,
-0.026103319600224495,
-0.12598392367362976,
-0.09559943526983261,
-0.003135633422061801,
-0.03007381036877632,
-0.010533859953284264,
0.12376988679170609,
0.047428715974092484,
-0.0028614026959985495,
0.03889176622033119,
-0.01733420230448246,
0.04333767667412758,
-0.06086549535393715,
0.019253766164183617,
-0.0004787665675394237,
-0.01142066065222025,
0.046750422567129135,
0.022441009059548378,
0.023399367928504944,
0.02168254181742668,
-0.026732871308922768,
-0.012629717588424683,
0.09345381706953049,
0.019986968487501144,
-0.003843204816803336,
0.05717828497290611,
0.049694471061229706,
-0.01664070226252079,
-0.04521210119128227,
0.09004402905702591,
-0.05296825245022774,
-0.024978231638669968,
0.012872062623500824,
0.006375241558998823,
0.02967325784265995,
0.10479677468538284,
0.029743127524852753,
0.027296757325530052,
0.07996820658445358,
0.08476279675960541,
0.04816191643476486,
0.038820780813694,
-0.07324739545583725,
-0.005702534224838018,
0.008165333420038223,
-0.010519496165215969,
-0.015937689691781998,
0.005052432883530855,
-0.07699116319417953,
-0.04956655576825142,
0.03657311201095581,
-0.03438562527298927,
0.0637381374835968,
-0.016922904178500175,
-0.01892482303082943,
-0.06741709262132645,
0.008238162845373154,
0.050203729420900345,
0.01680818386375904,
0.02516344003379345,
-0.037368252873420715,
-0.004782403353601694,
-0.00677899643778801,
0.023578893393278122,
0.06673145294189453,
-0.023423070088028908,
0.03162126615643501,
0.02384284883737564,
-0.024855012074112892,
-0.004064531996846199,
0.07140225172042847,
0.01536622829735279,
0.027144376188516617,
0.02366652898490429,
-0.007504256442189217,
0.10273835062980652,
-0.017900682985782623,
0.020282145589590073,
-1.8094292706711314e-33,
0.08876560628414154,
0.05333920940756798,
0.06308500468730927,
0.01879914104938507,
0.0897575318813324,
0.02700917236506939,
0.04041900113224983,
0.018744777888059616,
-0.03288979083299637,
-0.01799350418150425,
0.03339444473385811,
-0.011928361840546131,
-0.0854654610157013,
0.047611135989427567,
0.01961962692439556,
0.022396134212613106,
-0.06682120263576508,
0.06361280381679535,
-0.0937863439321518,
-0.05360030010342598,
0.02064635045826435,
0.04576525092124939,
0.058093875646591187,
-0.08781296014785767,
0.06073525547981262,
0.020778339356184006,
-0.0022576539777219296,
-0.012050916440784931,
0.04540276527404785,
0.011361401528120041,
-0.12586364150047302,
0.030616004019975662,
-0.07953107357025146,
0.12731638550758362,
0.013876065611839294,
0.03507658839225769,
-0.06980594247579575,
-0.08102307468652725,
-0.02914370596408844,
0.023911433294415474,
-0.05168578028678894,
-0.01236609648913145,
0.07449061423540115,
-0.011780019849538803,
-0.06862488389015198,
-0.08472218364477158,
-0.03206966444849968,
0.02605327218770981,
-0.038599591702222824,
0.017337294295430183,
0.009853031486272812,
0.039514657109975815,
-0.04445485398173332,
-0.08252018690109253,
-0.01586036942899227,
0.03858156502246857,
0.03421292454004288,
0.06123688817024231,
0.02417331002652645,
0.013453721068799496,
0.04057534784078598,
0.02087930031120777,
0.00022054566943552345,
0.1307520866394043,
0.0075592948123812675,
0.09737929701805115,
-0.06999747455120087,
0.06085445359349251,
0.07287424802780151,
0.035006146878004074,
0.018613841384649277,
0.010931611992418766,
0.03488176316022873,
-0.06107712909579277,
0.06665973365306854,
-0.053280964493751526,
0.046814993023872375,
-0.07595887035131454,
0.001861201599240303,
-0.007202198263257742,
-0.0365152433514595,
-0.02617630921304226,
-0.05477902293205261,
-0.04701734334230423,
0.046639248728752136,
-0.06887514144182205,
0.008588609285652637,
-0.07525075972080231,
0.04608896002173424,
-0.04937109723687172,
-0.08079897612333298,
-0.03326468542218208,
-0.06910223513841629,
0.0008106418536044657,
-0.02505965530872345,
3.3864313953497816e-34,
-0.01176745630800724,
-0.05312441289424896,
0.015807470306754112,
0.09015554189682007,
-0.03696964681148529,
-0.10940932482481003,
-0.01906179077923298,
-0.030068831518292427,
0.008375615812838078,
-0.05122626572847366,
-0.04782285541296005,
-0.014183021150529385,
-0.020997267216444016,
0.024875782430171967,
0.011181303299963474,
-0.011307450011372566,
0.03557950258255005,
-0.03676885738968849,
-0.006405494175851345,
0.010714596137404442,
0.038224250078201294,
0.103068046271801,
-0.07575565576553345,
0.01581007055938244,
-0.04172534868121147,
0.0059182350523769855,
0.07384411245584488,
-0.01750757545232773,
0.05048017576336861,
0.014629911631345749,
0.008297395892441273,
0.02272154577076435,
-0.01939203590154648,
0.014220204204320908,
-0.033678703010082245,
0.11246227473020554,
0.02538176067173481,
-0.005571991205215454,
-0.05634959042072296,
0.1692975014448166,
0.07049844413995743,
-0.07307374477386475,
-0.08661101013422012,
-0.023822888731956482,
-0.03821682184934616,
0.05438142642378807,
-0.04177001118659973,
0.056199558079242706,
-0.00438761105760932,
0.018550174310803413,
-0.017601042985916138,
-0.02389759011566639,
-0.07425207644701004,
0.004945769906044006,
-0.0787791982293129,
-0.05157944932579994,
-0.016521429643034935,
0.08036968111991882,
0.07427670061588287,
0.03342776745557785,
-0.09011006355285645,
0.021000945940613747,
0.04884450137615204,
-0.028726359829306602,
-0.049323100596666336,
-0.07472413033246994,
-0.09664800018072128,
-0.024206126108765602,
0.051353223621845245,
-0.03336707502603531,
-0.0027574992273002863,
-0.031041905283927917,
-0.03627363592386246,
-0.006153604481369257,
0.03877101466059685,
0.036305081099271774,
-0.019796865060925484,
0.007020920515060425,
-0.01184526365250349,
-0.05174606665968895,
0.013612777926027775,
0.025305375456809998,
0.018927499651908875,
0.06033429875969887,
0.043227456510066986,
0.05963568761944771,
-0.025548599660396576,
0.10569947957992554,
0.051830973476171494,
-0.0070790620520710945,
-0.03274308517575264,
0.03987685218453407,
0.03011537343263626,
0.05678940564393997,
-0.12059569358825684,
-4.24148396405144e-8,
-0.04589921608567238,
0.05559185892343521,
0.0691971555352211,
0.04041203483939171,
0.012997351586818695,
-0.052858248353004456,
-0.01388625893741846,
0.026428889483213425,
-0.056933045387268066,
0.018056899309158325,
0.0653291642665863,
-0.03024720586836338,
-0.008489507250487804,
-0.052223365753889084,
0.018709098920226097,
-0.004813327454030514,
-0.042887985706329346,
-0.056465987116098404,
-0.004972401075065136,
0.0031202679965645075,
0.011166744865477085,
-0.08019432425498962,
-0.01650974154472351,
0.022146960720419884,
0.02415492571890354,
0.005919076967984438,
-0.0785556361079216,
0.018416864797472954,
0.007743386551737785,
0.054114606231451035,
0.08806662261486053,
0.06785623729228973,
0.02477887086570263,
0.020184800028800964,
0.0062912930734455585,
0.012566064484417439,
-0.041256267577409744,
-0.03580315783619881,
0.03858834132552147,
-0.0036566013004630804,
-0.022763509303331375,
0.06771311908960342,
-0.009377827867865562,
0.004237926099449396,
-0.0032523248810321093,
0.02214384637773037,
-0.08432912826538086,
-0.10054842382669449,
0.0068572270683944225,
-0.05360144004225731,
-0.047413021326065063,
-0.07260898500680923,
-0.06019081920385361,
0.07301465421915054,
0.09328844398260117,
0.03516208380460739,
0.058761194348335266,
-0.07864908874034882,
0.023861128836870193,
0.0743795856833458,
0.03246752917766571,
-0.030817486345767975,
-0.0575823076069355,
0.0032511604949831963
] | 0 | 0 | 1 | 0 |
|
TOMATO: Assessing Visual Temporal Reasoning Capabilities in Multimodal Foundation Models | https://openreview.net/forum?id=fCi4o83Mfs | [
"Ziyao Shangguan",
"Chuhan Li",
"Yuxuan Ding",
"Yanan Zheng",
"Yilun Zhao",
"Tesca Fitzgerald",
"Arman Cohan"
] | Poster | Existing benchmarks often highlight the remarkable performance achieved by state-of-the-art Multimodal Foundation Models (MFMs) in leveraging temporal context for video understanding.
However, *how well do the models truly perform visual temporal reasoning?*
Our study of existing benchmarks shows that this capability of MFMs is likely overestimated as many questions can be solved by using a single, few, or out-of-order frames.
To systematically examine current visual temporal reasoning tasks, we propose three principles with corresponding metrics:
(1) *Multi-Frame Gain*,
(2) *Frame Order Sensitivity*,
and (3) *Frame Information Disparity*.
Following these principles, we introduce **TOMATO**, **T**emp**O**ral Reasoning **M**ultimod**A**l Evalua**T**i**O**n, a novel benchmark crafted to rigorously assess MFMs' temporal reasoning capabilities in video understanding.
TOMATO comprises 1,484 carefully curated, *human-annotated* questions spanning *six* tasks (i.e. *action count, direction, rotation, shape & trend, velocity & frequency, and visual cues*), applied to 1,417 videos, including 805 self-recorded and -generated videos, that encompass human-centric, real-world, and simulated scenarios.
Our comprehensive evaluation reveals a human-model performance gap of 57.3% with the best-performing model.
Moreover, our in-depth analysis uncovers more fundamental limitations beyond this gap in current MFMs. While they can accurately recognize events in isolated frames, they fail to interpret these frames as a continuous sequence.
We believe TOMATO will serve as a crucial testbed for evaluating the next-generation MFMs and as a call to the community to develop AI systems capable of comprehending the human world dynamics through the video modality. | visual temporal reasoning, video understanding, benchmark, vision-language benchmark, video-language models, evaluation | null | 13,155 | 2410.23266 | [
0.024161674082279205,
-0.10214051604270935,
0.0031338355038315058,
-0.04115324467420578,
0.19187502562999725,
-0.03981330245733261,
-0.005853800103068352,
0.015228456817567348,
0.06321467459201813,
-0.050925396382808685,
-0.07108237594366074,
-0.0936068743467331,
-0.023942476138472557,
0.13045816123485565,
-0.031889479607343674,
0.01453394815325737,
0.030360626056790352,
0.06183463707566261,
-0.1025664433836937,
-0.05037529021501541,
-0.0009597507887519896,
-0.07427085191011429,
0.03493427112698555,
-0.05820610374212265,
0.058894600719213486,
0.009845560416579247,
-0.047992341220378876,
-0.036310210824012756,
0.0069811358116567135,
-0.03912258893251419,
-0.0021004220470786095,
0.01319956686347723,
0.019279100000858307,
0.0333237387239933,
-0.036608994007110596,
0.04350264370441437,
-0.025333259254693985,
-0.023614129051566124,
-0.057296622544527054,
-0.053987693041563034,
-0.04825977608561516,
0.02753196656703949,
0.08205920457839966,
-0.017818903550505638,
0.045946184545755386,
0.02880827523767948,
0.004512642975896597,
-0.052562348544597626,
0.013296046294271946,
-0.043978191912174225,
-0.11563199758529663,
0.0018526118947193027,
-0.02692439593374729,
-0.0364249087870121,
-0.04325892776250839,
0.08445458114147186,
0.04821845516562462,
-0.00958940852433443,
0.021868882700800896,
-0.007314974442124367,
-0.012275618501007557,
-0.11077819019556046,
-0.0025178566575050354,
0.0467369444668293,
0.03738056868314743,
-0.030542945489287376,
0.022764205932617188,
-0.060537684708833694,
0.020563559606671333,
0.021901879459619522,
-0.06275497376918793,
0.0393439345061779,
-0.00942953396588564,
0.01593279466032982,
-0.04339395463466644,
0.060043636709451675,
0.043019309639930725,
-0.0385393388569355,
0.02292763441801071,
-0.14067819714546204,
0.06783288717269897,
-0.0009094845154322684,
-0.038282185792922974,
-0.023862868547439575,
0.16390429437160492,
0.0037483167834579945,
-0.025488274171948433,
0.10203853249549866,
-0.04435346648097038,
0.08593355864286423,
-0.03793365880846977,
-0.07001259922981262,
0.02248704433441162,
0.02340446226298809,
0.05371479317545891,
0.020289432257413864,
0.03899126499891281,
-0.09975763410329819,
-0.05297023802995682,
0.08298692107200623,
0.016116857528686523,
0.03223112225532532,
-0.02614997886121273,
-0.010412243194878101,
0.0776672214269638,
-0.021172190085053444,
0.047507189214229584,
0.02841227315366268,
-0.019388487562537193,
-0.05465148761868477,
0.02116820402443409,
0.034118086099624634,
0.06253618001937866,
0.02238018810749054,
0.010019778273999691,
-0.007377966307103634,
0.03352554887533188,
-0.00046051733079366386,
0.08213412016630173,
-0.045314498245716095,
0.03063797578215599,
-0.0026936272624880075,
-0.04722311347723007,
-0.0051496815867722034,
0.04320872575044632,
-0.03381435200572014,
0.03045034594833851,
-4.92169153195974e-34,
0.011728190816938877,
0.0067702909000217915,
0.07042291760444641,
0.05552450940012932,
0.06708644330501556,
0.011207880452275276,
0.0068296710960567,
0.03349067270755768,
-0.04423537105321884,
0.02039857767522335,
-0.013471103273332119,
-0.06947821378707886,
-0.053963590413331985,
0.07703284174203873,
-0.02974785678088665,
-0.044722408056259155,
-0.04197470098733902,
0.0253083948045969,
-0.027437033131718636,
0.04150504246354103,
0.015838591381907463,
-0.037611402571201324,
-0.045628588646650314,
-0.026721537113189697,
-0.013411886058747768,
-0.0285195205360651,
0.050897445529699326,
-0.02539846859872341,
-0.020604290068149567,
-0.017898958176374435,
-0.13147318363189697,
-0.021605823189020157,
-0.056548770517110825,
0.038528233766555786,
0.05238860845565796,
-0.06711956858634949,
-0.007974139414727688,
-0.06309609115123749,
0.05497436597943306,
-0.05115482956171036,
-0.06797492504119873,
0.03197318688035011,
-0.03433375060558319,
0.022661857306957245,
-0.07302214950323105,
-0.019452963024377823,
-0.007728251162916422,
0.14632061123847961,
-0.08265203982591629,
0.04204973950982094,
0.04018079489469528,
0.0507061630487442,
0.012294924817979336,
-0.13917715847492218,
-0.01848398707807064,
-0.015891827642917633,
0.06584349274635315,
0.00461988802999258,
0.0340270958840847,
0.04681885242462158,
0.0022842432372272015,
0.04545111954212189,
-0.031813476234674454,
0.045910175889730453,
-0.03484122082591057,
0.05866522341966629,
-0.07094498723745346,
0.05267689377069473,
0.09336678683757782,
0.050051383674144745,
-0.03767741098999977,
0.02936781942844391,
0.09356804937124252,
-0.06760693341493607,
0.026629609987139702,
0.05090431496500969,
0.029593270272016525,
-0.0012390757910907269,
0.05355114862322807,
0.052150849252939224,
-0.010205686092376709,
0.03931121155619621,
-0.008148166351020336,
-0.05805196985602379,
0.0006073705153539777,
-0.04547753930091858,
0.019307106733322144,
-0.09564660489559174,
-0.024140534922480583,
-0.07526643574237823,
0.025560295209288597,
-0.045761652290821075,
0.05048635974526405,
-0.0446748286485672,
-0.010157839395105839,
5.442672924131582e-34,
0.07378564029932022,
0.04382995516061783,
-0.029372667893767357,
0.013216235674917698,
0.028017306700348854,
-0.0745152086019516,
-0.051496703177690506,
-0.022055961191654205,
0.008449782617390156,
-0.0660485252737999,
0.005547540262341499,
-0.026622222736477852,
-0.060048460960388184,
-0.04373425617814064,
0.026936916634440422,
0.014965678565204144,
0.019717425107955933,
-0.04412878677248955,
-0.0355231799185276,
0.04307975620031357,
-0.016322243958711624,
0.08937903493642807,
-0.0857781171798706,
-0.02200796641409397,
-0.10648129880428314,
0.0738537460565567,
-0.001054941094480455,
0.017275501042604446,
-0.061733927577733994,
-0.030746787786483765,
0.001977700274437666,
-0.09604263305664062,
-0.04371540993452072,
-0.02272358350455761,
-0.03489388898015022,
0.008329996839165688,
0.04712195694446564,
-0.08889865130186081,
-0.06613076478242874,
0.11471183598041534,
0.052128881216049194,
0.03159596771001816,
-0.07145123183727264,
0.02646745555102825,
-0.011323656886816025,
0.06635969132184982,
-0.056846100836992264,
-0.006875294726341963,
-0.04534464702010155,
0.04046306014060974,
-0.01085931621491909,
0.02204957790672779,
-0.04201126471161842,
-0.062379688024520874,
0.01036945078521967,
-0.06702978163957596,
-0.02421753481030464,
0.010953420773148537,
0.0543709434568882,
0.035517483949661255,
-0.016468560323119164,
0.0026302712503820658,
-0.060173019766807556,
-0.003844214603304863,
0.019475936889648438,
0.059922315180301666,
-0.0419822633266449,
-0.06112821772694588,
-0.014935951679944992,
0.011507930234074593,
0.031603146344423294,
-0.0377439484000206,
-0.0534994900226593,
0.0014147384790703654,
0.08393222838640213,
0.043519388884305954,
0.013202151283621788,
0.005189511924982071,
0.0032354879658669233,
-0.06028018146753311,
-0.027080640196800232,
-0.08670838922262192,
0.066756971180439,
0.09084506332874298,
-0.00269430223852396,
0.090849868953228,
0.006865075323730707,
0.06417063623666763,
0.08847174793481827,
0.03147909417748451,
-0.04492030292749405,
0.043559398502111435,
0.04200588911771774,
0.10709546506404877,
0.05145297199487686,
-3.7577208900074766e-8,
-0.08216715604066849,
0.0681091696023941,
-0.01172831654548645,
-0.028956225141882896,
-0.000288660783553496,
-0.06784377992153168,
-0.01770026423037052,
0.04683127626776695,
0.06442669779062271,
-0.025847313925623894,
0.09032322466373444,
0.00898880697786808,
-0.020390911027789116,
0.005136653780937195,
-0.011330125853419304,
0.04381587356328964,
0.023765360936522484,
-0.01253405399620533,
-0.004192928317934275,
0.03303639590740204,
0.06133986636996269,
0.02546997368335724,
-0.02965870127081871,
-0.008786609396338463,
-0.0040393853560090065,
-0.02967514656484127,
-0.014954022131860256,
0.15840446949005127,
0.0964418277144432,
0.02292219176888466,
0.02082092873752117,
0.022384585812687874,
-0.012368389405310154,
-0.05725384131073952,
0.016427358612418175,
-0.0021561707835644484,
0.0059763407334685326,
-0.051090922206640244,
-0.012150241993367672,
0.013765953481197357,
-0.023484645411372185,
0.060006726533174515,
-0.051210079342126846,
0.04128801077604294,
-0.012876039370894432,
0.023781081661581993,
-0.08846230059862137,
-0.06499125808477402,
-0.04612968862056732,
-0.008749268017709255,
-0.00913734920322895,
-0.017019696533679962,
-0.027840210124850273,
0.097647525370121,
0.10355813801288605,
-0.06064238026738167,
0.05195023491978645,
-0.004571922123432159,
0.005374129395931959,
0.05401979759335518,
0.012231207452714443,
0.05557628720998764,
-0.0049280826933681965,
-0.004248655866831541
] | https://github.com/yale-nlp/TOMATO | 25 | 0 | 1 | 0 |
Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers | https://openreview.net/forum?id=yzloNYH3QN | [
"Shijie Chen",
"Bernal Jimenez Gutierrez",
"Yu Su"
] | Poster | Information retrieval (IR) systems have played a vital role in modern digital life and have cemented their continued usefulness in this new era of generative AI via retrieval-augmented generation. With strong language processing capabilities and remarkable versatility, large language models (LLMs) have become popular choices for zero-shot re-ranking in IR systems. So far, LLM-based re-ranking methods rely on strong generative capabilities, which restricts their use to either specialized or powerful proprietary models. Given these restrictions, we ask: is autoregressive generation necessary and optimal for LLMs to perform re-ranking? We hypothesize that there are abundant signals relevant to re-ranking within LLMs that might not be used to their full potential via generation. To more directly leverage such signals, we propose in-context re-ranking (ICR), a novel method that leverages the change in attention pattern caused by the search query for accurate and efficient re-ranking. We assume that more relevant documents should receive more attention weights when an LLM is processing the query tokens, and leverage such signals for re-ranking. To mitigate the intrinsic biases in LLMs, we propose a calibration method using a content-free query. Due to the absence of generation, ICR only requires two ($O(1)$) forward passes to re-rank $N$ documents, making it substantially more efficient than generative re-ranking methods that require at least $O(N)$ forward passes. Our novel design also enables ICR to be applied to any LLM without specialized training while guaranteeing a well-formed ranking. Extensive experiments with two popular open-weight LLMs on standard single-hop and multi-hop information retrieval benchmarks show that ICR outperforms RankGPT while cutting the latency by more than 60% in practice. Through detailed analyses, we show that ICR's performance is specially strong on tasks that require more complex re-ranking signals, such as handling contextualization and contradiction between the query and passages, as well as information integration across multiple passages. Our findings call for further exploration on novel ways of utilizing open-weight LLMs beyond text generation. | Large Language Model, Information Retrieval | We propose an efficient LLM-based re-ranking method that outperforms RankGPT while only requiring two forward passes without specialized training. | 13,153 | 2410.02642 | [
-0.06763888895511627,
-0.08576849848031998,
0.04066389426589012,
0.06980524212121964,
0.036152102053165436,
0.08933773636817932,
-0.003735230304300785,
0.041421424597501755,
0.09466937929391861,
-0.021235579624772072,
-0.05170561745762825,
-0.01775088720023632,
0.05877204239368439,
0.009403819218277931,
-0.045293353497982025,
-0.006131904665380716,
0.14139141142368317,
0.032227568328380585,
-0.0706515982747078,
-0.06752817332744598,
0.006153285037726164,
0.008681935258209705,
0.11096391826868057,
-0.07534261047840118,
0.028792571276426315,
-0.07387737184762955,
-0.06822826713323593,
-0.028862344101071358,
0.038047630339860916,
-0.01917000114917755,
0.024163268506526947,
0.10953262448310852,
-0.054079215973615646,
0.026742301881313324,
-0.07289331406354904,
0.050556380301713943,
-0.07076243311166763,
0.01702100969851017,
0.00902572926133871,
-0.0652390867471695,
0.021216318011283875,
0.0015649873530492187,
-0.006563705392181873,
0.0006300758686847985,
0.09008586406707764,
-0.01489975955337286,
-0.013381052762269974,
0.0011110816849395633,
0.019350042566657066,
0.02646138146519661,
-0.11459063738584518,
-0.007567724213004112,
-0.048508524894714355,
0.09054400026798248,
-0.012553985230624676,
0.03134856000542641,
0.014962789602577686,
0.04556575044989586,
0.006225325167179108,
-0.06739293038845062,
-0.06904982775449753,
-0.10561525821685791,
-0.005253891926258802,
-0.05205411836504936,
0.015843404456973076,
-0.07138540595769882,
0.027885984629392624,
-0.006194652058184147,
-0.04240195080637932,
0.039221782237291336,
0.046412743628025055,
0.05288054421544075,
-0.01622607931494713,
0.07640247792005539,
-0.011435145512223244,
0.05387397110462189,
-0.03103114850819111,
-0.06067773699760437,
0.04739253968000412,
-0.0034829461947083473,
0.04779895395040512,
-0.023782946169376373,
0.0496663935482502,
-0.012486523948609829,
0.10581853240728378,
0.004158644005656242,
0.06784000992774963,
-0.06579253077507019,
0.06541905552148819,
0.0283682718873024,
-0.07673804461956024,
-0.07773695141077042,
0.04888239502906799,
0.014135053381323814,
0.053803879767656326,
0.008225596509873867,
0.008484062738716602,
-0.05640827491879463,
0.0021155700087547302,
0.05022283270955086,
0.039039451628923416,
0.0723671168088913,
-0.043759703636169434,
-0.04945460334420204,
-0.005399012006819248,
0.017564935609698296,
0.004389818292111158,
0.12124791741371155,
0.014460291713476181,
-0.09462017565965652,
0.013809564523398876,
0.041686687618494034,
0.06277945637702942,
-0.047068919986486435,
0.004354910459369421,
-0.022973496466875076,
0.016911787912249565,
-0.026677431538701057,
0.016098134219646454,
0.006670478731393814,
-0.003329358296468854,
0.03461220860481262,
-0.02785446308553219,
-0.03046136163175106,
0.008264528587460518,
-0.010030017234385014,
0.0005384684191085398,
3.066195050939616e-34,
0.05817274749279022,
0.09319767355918884,
0.026755230501294136,
0.054072409868240356,
0.020346011966466904,
-0.00944498460739851,
0.03615950047969818,
0.10183388739824295,
-0.041851241141557693,
-0.07323331385850906,
-0.03092906065285206,
0.04438536614179611,
-0.026174984872341156,
0.04423539340496063,
0.024820072576403618,
-0.006400402169674635,
-0.033376771956682205,
0.0367470383644104,
-0.04150199145078659,
-0.017269771546125412,
-0.04361226409673691,
0.03454555571079254,
0.03796587511897087,
-0.0852711871266365,
0.03260505199432373,
0.00747197400778532,
0.02221955731511116,
-0.10388798266649246,
-0.10520407557487488,
-0.02094184421002865,
-0.09162986278533936,
-0.030896969139575958,
0.015382872894406319,
0.005057604983448982,
-0.01704566553235054,
-0.05827147141098976,
-0.06652560830116272,
-0.01757972314953804,
0.025767844170331955,
0.0127927390858531,
-0.005311342887580395,
0.06710005551576614,
-0.015676245093345642,
-0.04153776168823242,
-0.08213068544864655,
-0.09677407145500183,
-0.04907277971506119,
0.06759773939847946,
-0.035926107317209244,
-0.046863723546266556,
0.11592793464660645,
0.07533366233110428,
-0.02254483290016651,
0.00680631585419178,
-0.01703401282429695,
-0.03542613610625267,
-0.034634336829185486,
0.04457712918519974,
-0.028293265029788017,
0.012187017127871513,
0.08250193297863007,
0.034879278391599655,
0.022164223715662956,
0.10147079080343246,
0.05107956752181053,
0.010949506424367428,
0.03932536765933037,
0.05184372887015343,
0.07589592784643173,
0.02532411366701126,
0.020278336480259895,
0.03004078008234501,
-0.026666993275284767,
-0.1060132384300232,
0.05427518114447594,
0.03195779398083687,
0.03680718317627907,
-0.08051123470067978,
0.040507931262254715,
0.00916372798383236,
0.04313639551401138,
-0.03862881660461426,
-0.0057145655155181885,
-0.03141693025827408,
-0.020419156178832054,
-0.05466870591044426,
0.034410227090120316,
-0.05648487061262131,
0.00010359544830862433,
-0.05593327060341835,
0.03857385367155075,
-0.03471057116985321,
-0.07175149023532867,
0.007335908245295286,
-0.025995686650276184,
2.6498785737390368e-34,
0.008627350442111492,
-0.011726155877113342,
-0.013925345614552498,
0.0012391062919050455,
-0.0488954596221447,
-0.023260917514562607,
-0.0024755813647061586,
0.015022732317447662,
-0.09143079072237015,
-0.0624239519238472,
-0.04435482621192932,
-0.053606245666742325,
0.08676516264677048,
-0.009334374219179153,
0.039380189031362534,
-0.002264084992930293,
-0.01452331431210041,
-0.017144860699772835,
-0.10929862409830093,
0.051652129739522934,
0.04245753586292267,
0.07860508561134338,
-0.09039640426635742,
0.07912552356719971,
0.00770660862326622,
0.0557958148419857,
-0.0009045019396580756,
0.03865595534443855,
-0.04062328115105629,
-0.06416930258274078,
0.005805693566799164,
-0.022131895646452904,
-0.03459837660193443,
-0.04657665267586708,
-0.06036827340722084,
0.08750877529382706,
0.080522820353508,
-0.028239596635103226,
-0.08748514950275421,
0.16068977117538452,
0.09901310503482819,
0.00900342408567667,
-0.04603036120533943,
0.020148152485489845,
-0.04937798157334328,
-0.0205399002879858,
-0.11829423159360886,
0.015774251893162727,
0.038382045924663544,
0.0015018130652606487,
-0.025061756372451782,
0.034322526305913925,
-0.07889622449874878,
0.009659189730882645,
-0.05545898154377937,
-0.06805726885795593,
-0.08028864860534668,
-0.017817992717027664,
0.030353015288710594,
0.025723354890942574,
0.005775592755526304,
-0.016790034249424934,
0.032810453325510025,
-0.08659449219703674,
-0.008450459688901901,
-0.01590908318758011,
-0.000943968421779573,
-0.03028256632387638,
-0.053956929594278336,
-0.060337524861097336,
0.0674133151769638,
-0.015132997184991837,
0.051871370524168015,
-0.013190307654440403,
-0.014874322339892387,
0.06383364647626877,
0.05122116580605507,
-0.030818844214081764,
0.00750910397619009,
-0.11179763823747635,
-0.030617741867899895,
0.031147364526987076,
-0.019460931420326233,
0.06756675988435745,
0.04358072578907013,
0.0430402047932148,
0.05887020751833916,
0.038036659359931946,
0.041613686829805374,
0.0013145594857633114,
0.004691847134381533,
-0.06636741757392883,
0.013554959557950497,
0.019068479537963867,
-0.025222172960639,
-3.8463571883085024e-8,
-0.08367522805929184,
0.05000195652246475,
-0.051791705191135406,
0.14680646359920502,
0.006158456671983004,
-0.005772022996097803,
-0.06140550225973129,
0.10645348578691483,
-0.01282487716525793,
-0.040560342371463776,
0.022832904011011124,
-0.03979812562465668,
-0.014397382736206055,
-0.031770963221788406,
0.05238720402121544,
0.0469401590526104,
-0.012751278467476368,
0.024125676602125168,
-0.03596636652946472,
-0.07618581503629684,
-0.005545963067561388,
0.05656203255057335,
0.04296301305294037,
-0.0034724916331470013,
0.014803287573158741,
-0.022829117253422737,
0.051101211458444595,
0.027625788003206253,
0.10204938799142838,
-0.0036250660195946693,
-0.024838807061314583,
-0.0016808892833068967,
0.01285751722753048,
-0.00892728567123413,
0.009390213526785374,
0.04312964528799057,
-0.011461753398180008,
-0.02499762363731861,
-0.05150342732667923,
-0.06359104067087173,
0.04077262431383133,
0.05330878123641014,
-0.030843647196888924,
0.07840067893266678,
0.02198418602347374,
-0.0325738824903965,
0.006575172767043114,
-0.08033278584480286,
0.08319763094186783,
-0.03808290511369705,
0.000801978399977088,
-0.1391986906528473,
0.03160495683550835,
0.011278795078396797,
0.06260057538747787,
0.0451878160238266,
0.025310466066002846,
0.016792265698313713,
0.06503259390592575,
0.025744760408997536,
0.061527319252491,
0.017215630039572716,
-0.054423678666353226,
0.03215063363313675
] | 0 | 0 | 0 | 0 |
|
Inspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct | https://openreview.net/forum?id=wWnsoLhHwt | [
"Christopher Ackerman",
"Nina Panickssery"
] | Poster | It has been reported that LLMs can recognize their own writing. As this has potential implications for AI safety, yet is relatively understudied, we investigate the phenomenon, seeking to establish: whether it robustly occurs at the behavioral level, how the observed behavior is achieved, and whether it can be controlled. First, we find that the Llama3-8b–Instruct chat model - but not the base Llama3-8b model - can reliably distinguish its own outputs from those of humans, and present evidence that the chat model is likely using its experience with its own outputs, acquired during post-training, to succeed at the writing recognition task. Second, we identify a vector in the residual stream of the model that is differentially activated when the model makes a correct self-written-text recognition judgment, show that the vector activates in response to information relevant to self-authorship, present evidence that the vector is related to the concept of ``self'' in the model, and demonstrate that the vector is causally related to the model’s ability to perceive and assert self-authorship. Finally, we show that the vector can be used to control both the model’s behavior and its perception, steering the model to claim or disclaim authorship by applying the vector to the model’s output as it generates it, and steering the model to believe or disbelieve it wrote arbitrary texts by applying the vector to them as the model reads them. | LLM, Interpretability, AI, Activation Steering, Representation Engineering, Control | null | 13,146 | 2410.02064 | [
-0.059107810258865356,
-0.053813908249139786,
0.001623895950615406,
0.006872196216136217,
0.0005228116642683744,
0.032615602016448975,
0.07364454120397568,
-0.015318538062274456,
0.11117095500230789,
0.022241177037358284,
-0.07704910635948181,
-0.04062565788626671,
0.037716351449489594,
0.021045906469225883,
-0.004669641610234976,
-0.061215125024318695,
0.0876728743314743,
-0.06667161732912064,
-0.09236221015453339,
-0.01799566112458706,
0.07257742434740067,
0.03988594189286232,
0.07547555863857269,
0.008031426928937435,
-0.019715402275323868,
0.0191932525485754,
-0.008264688774943352,
-0.06259351968765259,
0.010575996711850166,
-0.03208748623728752,
0.021364348009228706,
0.0028756919782608747,
0.09697222709655762,
0.045329537242650986,
-0.07628601044416428,
-0.012571551837027073,
-0.018997790291905403,
0.05188271775841713,
-0.033886346966028214,
-0.06326930224895477,
-0.055920954793691635,
-0.06406442075967789,
0.023263219743967056,
-0.024324672296643257,
0.08695263415575027,
-0.018349867314100266,
-0.011969491839408875,
0.007989457808434963,
-0.0328952893614769,
0.04811182990670204,
-0.1003156304359436,
-0.01697440817952156,
-0.01659228466451168,
0.06234973296523094,
-0.0231377761811018,
0.023201925680041313,
0.025815512984991074,
0.085905522108078,
-0.015374568291008472,
0.05913393571972847,
0.01843036152422428,
-0.01798069477081299,
0.02648693509399891,
0.014423469081521034,
0.07723553478717804,
0.059319768100976944,
-0.05330197513103485,
0.030107907950878143,
-0.046640489250421524,
-0.009769405238330364,
0.06985671818256378,
0.020866086706519127,
-0.028247002512216568,
0.02999005652964115,
0.044186513870954514,
-0.006302937865257263,
0.03713299706578255,
-0.015225433744490147,
0.045661214739084244,
-0.0422804094851017,
-0.012190169654786587,
0.03184404969215393,
0.046531494706869125,
0.059802114963531494,
-0.050340522080659866,
0.0368199348449707,
-0.046912193298339844,
0.03274747356772423,
0.03829435259103775,
-0.006529534235596657,
-0.03950987383723259,
-0.08599437773227692,
-0.0381295345723629,
-0.011614319868385792,
0.03926539421081543,
0.00003504613050608896,
0.011004691012203693,
0.02485332079231739,
0.02035311609506607,
0.03605237603187561,
0.04387018084526062,
0.03872968629002571,
-0.009694505482912064,
0.04921133443713188,
0.07370951771736145,
0.03226384148001671,
0.06090712547302246,
-0.043585825711488724,
0.02203330770134926,
-0.05086921900510788,
-0.010416326113045216,
-0.003299220697954297,
-0.029708679765462875,
0.05036041885614395,
0.06322232633829117,
-0.011991716921329498,
-0.006446570623666048,
0.003664926392957568,
0.010169602930545807,
0.02405533753335476,
0.020223017781972885,
-0.040773507207632065,
0.07108582556247711,
0.02152896113693714,
0.020744089037179947,
-0.05456487461924553,
-0.05726982280611992,
-1.3418384645716533e-33,
0.06540202349424362,
0.04736492037773132,
0.007021841127425432,
0.03473931550979614,
0.01384167093783617,
-0.03780588507652283,
-0.04960646107792854,
0.0067675309255719185,
-0.0168786458671093,
0.011366133578121662,
-0.0324053093791008,
-0.0097942054271698,
0.028725851327180862,
0.0803685262799263,
0.054531048983335495,
0.08415537327528,
-0.04971098527312279,
-0.014965887181460857,
0.02363104373216629,
-0.05608462914824486,
0.11937927454710007,
-0.026221703737974167,
0.06978549808263779,
-0.028845705091953278,
-0.0003967759548686445,
0.0554991215467453,
0.09884866327047348,
-0.08865496516227722,
-0.05125446245074272,
0.028671447187662125,
-0.04453077167272568,
0.04213973879814148,
-0.00404104869812727,
-0.03758833184838295,
0.007475551217794418,
-0.02352289855480194,
-0.013230628333985806,
-0.013268058188259602,
0.02041931077837944,
0.03820333629846573,
0.008336491882801056,
-0.034916363656520844,
0.013947679661214352,
-0.057092323899269104,
-0.08812651038169861,
-0.006158828269690275,
-0.013195410370826721,
0.005637919995933771,
-0.06807198375463486,
-0.011219190433621407,
-0.012343328446149826,
0.08370529860258102,
0.010152776725590229,
-0.03265247121453285,
-0.05245256423950195,
0.027839642018079758,
-0.0023132439237087965,
0.038828905671834946,
0.004270520061254501,
-0.008770961314439774,
-0.017619546502828598,
0.06285186111927032,
-0.00982602871954441,
0.09270156919956207,
0.04689349979162216,
-0.030011162161827087,
-0.024998126551508904,
0.004013486672192812,
0.093319371342659,
-0.045591261237859726,
-0.03578684478998184,
-0.026311786845326424,
-0.11008009314537048,
-0.03786651790142059,
-0.05126851797103882,
-0.03587108850479126,
-0.003910345025360584,
-0.05401630327105522,
0.01861489564180374,
0.03252313286066055,
0.034498851746320724,
-0.030528835952281952,
0.006150726694613695,
-0.07063762843608856,
-0.07374799251556396,
-0.08720183372497559,
0.013800378888845444,
-0.10734976828098297,
-0.0009513776749372482,
0.05067887529730797,
0.02343817614018917,
-0.0038332343101501465,
0.005877017974853516,
0.03432708606123924,
-0.05660664662718773,
2.237240108159585e-34,
-0.06025616452097893,
-0.002435410162433982,
-0.1148485317826271,
0.00008879640517989174,
-0.045885197818279266,
-0.08936316519975662,
0.0725376307964325,
0.06497860699892044,
-0.0504605695605278,
-0.03197144716978073,
-0.07856184989213943,
-0.0006915779667906463,
0.016081033274531364,
0.012835076078772545,
0.034149155020713806,
-0.06564927101135254,
-0.009482691995799541,
0.0130084827542305,
-0.0006100375903770328,
-0.07958522439002991,
0.015098555944859982,
0.046768419444561005,
-0.10074304789304733,
0.021502133458852768,
-0.025284869596362114,
0.0065865968354046345,
-0.007288487162441015,
0.04970359429717064,
-0.003610533894971013,
0.018485907465219498,
0.010468972846865654,
0.06843604147434235,
-0.04152176156640053,
0.022979089990258217,
-0.018379371613264084,
0.030553337186574936,
-0.0026323534548282623,
-0.03493572026491165,
-0.00899432972073555,
0.020397698506712914,
0.0651133805513382,
0.0456206388771534,
-0.10145724564790726,
-0.039804838597774506,
-0.024475622922182083,
-0.09624595195055008,
-0.07254413515329361,
-0.007462360896170139,
0.05433857440948486,
0.04437139630317688,
0.01812579110264778,
-0.07677996158599854,
0.03346652537584305,
-0.0352805070579052,
-0.08556618541479111,
-0.05813271552324295,
0.05579861253499985,
-0.029535813257098198,
0.03581435978412628,
0.0407395213842392,
-0.018844401463866234,
-0.024622080847620964,
0.07626499235630035,
-0.014825265854597092,
0.04015149548649788,
-0.06136445701122284,
-0.00995629746466875,
0.06724143773317337,
0.09031461179256439,
-0.04977571964263916,
0.1405499279499054,
-0.05908098816871643,
-0.03067639283835888,
0.0008008684380911291,
0.015480462461709976,
-0.027612963691353798,
-0.09141264855861664,
-0.051644179970026016,
-0.07619193941354752,
-0.06291346997022629,
-0.060652609914541245,
-0.02594771422445774,
0.06116027012467384,
0.12501372396945953,
0.011138985864818096,
0.031168919056653976,
0.022604918107390404,
0.07523506134748459,
-0.03681639954447746,
0.042899612337350845,
0.005670660175383091,
0.0458206832408905,
-0.10290083289146423,
0.08268792182207108,
-0.07836493849754333,
-4.503372963426955e-8,
-0.11722082644701004,
-0.07672727853059769,
0.004098803736269474,
0.1081877052783966,
0.06290309876203537,
0.028294267132878304,
-0.005235715303570032,
-0.031198330223560333,
0.011265231296420097,
-0.059953104704618454,
0.06494668126106262,
-0.07001298666000366,
-0.06191713735461235,
-0.06046750023961067,
0.06514456868171692,
0.00998207088559866,
0.05062665790319443,
0.021475166082382202,
-0.02846432663500309,
-0.06815263628959656,
0.10873126238584518,
-0.02177783101797104,
-0.04514048621058464,
-0.046207379549741745,
-0.020556189119815826,
-0.02885306626558304,
-0.09601502865552902,
0.07816187292337418,
-0.09195858240127563,
0.09977652132511139,
-0.010567573830485344,
0.04023563861846924,
-0.02040909416973591,
-0.00923911202698946,
0.07907704263925552,
0.05098745971918106,
0.0947040542960167,
-0.06927022337913513,
0.011994966305792332,
0.0073089501820504665,
0.06333667039871216,
0.0802282765507698,
-0.02511799894273281,
0.03625970706343651,
0.014321085065603256,
-0.03018971160054207,
-0.019079871475696564,
-0.1629413217306137,
-0.00580050190910697,
-0.032037317752838135,
0.02700994722545147,
-0.04199822619557381,
-0.008736363612115383,
0.1364334374666214,
0.04915286973118782,
-0.023505231365561485,
0.03336801379919052,
0.03363658860325813,
0.020408794283866882,
0.08739444613456726,
0.08583509922027588,
0.11185288429260254,
0.006325290072709322,
-0.06963037699460983
] | 0 | 0 | 0 | 0 |
|
Towards Learning High-Precision Least Squares Algorithms with Sequence Models | https://openreview.net/forum?id=snocoXIQXz | [
"Jerry Weihong Liu",
"Jessica Grogan",
"Owen M Dugan",
"Ashish Rao",
"Simran Arora",
"Atri Rudra",
"Christopher Re"
] | Poster | This paper investigates whether sequence models can learn to perform numerical algorithms, e.g. gradient descent, on the fundamental problem of least squares. Our goal is to inherit two properties of standard algorithms from numerical analysis: (1) machine precision, i.e. we want to obtain solutions that are accurate to near floating point error, and (2) numerical generality, i.e. we want them to apply broadly across problem instances. We find that prior approaches using Transformers fail to meet these criteria, and identify limitations present in existing architectures and training procedures. First, we show that softmax Transformers struggle to perform high-precision multiplications, which prevents them from precisely learning numerical algorithms. Second, we identify an alternate class of architectures, comprised entirely of polynomials, that can efficiently represent high-precision gradient descent iterates. Finally, we investigate precision bottlenecks during training and address them via a high-precision training recipe that reduces stochastic gradient noise. Our recipe enables us to train two polynomial architectures, gated convolutions and linear attention, to perform gradient descent iterates on least squares problems. For the first time, we demonstrate the ability to train to near machine precision. Applied iteratively, our models obtain $100,000\times$ lower MSE than standard Transformers trained end-to-end and they incur a $10,000\times$ smaller generalization gap on out-of-distribution problems. We make progress towards end-to-end learning of numerical algorithms for least squares. | high precision, least squares, algorithm learning, Transformers, gated convolutions, linear regression, in-context learning | This work uncovers limitations of Transformers for high-precision numerical tasks, and makes progress towards training gated convolutional models to machine precision as an alternative. | 13,145 | 2503.12295 | [
-0.12684009969234467,
-0.058749713003635406,
-0.03611248731613159,
0.004469596315175295,
-0.007624308578670025,
-0.019509967416524887,
-0.06834997981786728,
-0.015155864879488945,
-0.022063791751861572,
-0.10358203202486038,
-0.043737225234508514,
-0.01188377384096384,
-0.02783278189599514,
0.044502563774585724,
-0.08100960403680801,
0.017936842516064644,
0.026419559493660927,
0.1746169626712799,
-0.04414815083146095,
-0.0608065091073513,
0.014438290148973465,
0.0131165636703372,
-0.04005686566233635,
0.020263444632291794,
0.06648137420415878,
-0.025444982573390007,
-0.017861858010292053,
-0.029314853250980377,
0.0022038782481104136,
0.018814196810126305,
0.014025264419615269,
-0.07267259806394577,
-0.001970555167645216,
0.03291663900017738,
-0.0202370285987854,
0.09762488305568695,
0.008207876235246658,
-0.019789904356002808,
0.04513843357563019,
-0.056093767285346985,
0.059587255120277405,
-0.0320383757352829,
0.05571135878562927,
0.01442320179194212,
0.11944352090358734,
-0.04773760214447975,
0.05079624801874161,
-0.05446654558181763,
0.03763158991932869,
-0.023944394662976265,
-0.029798032715916634,
-0.011943518184125423,
-0.0017044610576704144,
0.0843159556388855,
-0.08821658045053482,
-0.0343070887029171,
0.06300088763237,
0.038245994597673416,
-0.006609851494431496,
0.01565372385084629,
-0.03919769078493118,
-0.10309877246618271,
-0.017555896192789078,
-0.01845085248351097,
-0.06753075122833252,
-0.03355639800429344,
0.05030292645096779,
0.017557503655552864,
-0.02068566158413887,
0.06197012960910797,
0.0006137709715403616,
0.1017947569489479,
-0.09299745410680771,
0.060925718396902084,
0.016720853745937347,
-0.0081752585247159,
0.08774247765541077,
0.004453599918633699,
0.026226086542010307,
-0.04479892551898956,
0.03249165788292885,
-0.022102398797869682,
0.024308882653713226,
-0.0325222909450531,
0.1007460430264473,
-0.0421377532184124,
-0.02984454110264778,
0.036699678748846054,
0.006532559171319008,
0.0021911696530878544,
-0.04407469183206558,
-0.06231384351849556,
-0.0037761135026812553,
-0.0033456850796937943,
0.029181770980358124,
0.07822853326797485,
0.035559069365262985,
-0.05073381960391998,
-0.07282279431819916,
0.09469214826822281,
-0.01193002238869667,
0.02382720448076725,
0.0029118142556399107,
-0.05199825391173363,
0.009794509038329124,
-0.031152432784438133,
0.04572129249572754,
0.007942643947899342,
-0.004228116944432259,
-0.1447085738182068,
0.03523361682891846,
-0.00376906618475914,
0.026555566117167473,
0.0447925366461277,
0.08153259009122849,
-0.04466190189123154,
-0.02276313304901123,
-0.05414009094238281,
-0.002879928331822157,
0.05803752318024635,
-0.0757368877530098,
0.05863986164331436,
-0.013766399584710598,
0.10711826384067535,
0.07478686422109604,
-0.028303006663918495,
-0.10292944312095642,
8.585803391566295e-34,
0.0319693349301815,
0.046933356672525406,
0.0771007239818573,
-0.016401179134845734,
0.04202140495181084,
-0.031165068969130516,
0.04670533537864685,
0.06186942011117935,
-0.04579846188426018,
0.09798312187194824,
-0.026732442900538445,
-0.0038538677617907524,
-0.07496046274900436,
0.022423963993787766,
0.047024961560964584,
-0.008342993445694447,
-0.030310133472085,
0.004205819685012102,
0.024737317115068436,
-0.017798855900764465,
0.016306212171912193,
-0.06882389634847641,
-0.028091443702578545,
-0.02840445749461651,
-0.013879122212529182,
0.05169251188635826,
0.06580980122089386,
-0.018649639561772346,
-0.0008858498185873032,
0.031192395836114883,
-0.0939723551273346,
0.0019557590130716562,
0.01777692511677742,
0.007843871600925922,
0.07444840669631958,
0.006233395077288151,
0.031632453203201294,
-0.035759277641773224,
0.021304545924067497,
-0.03325970470905304,
-0.01721019670367241,
0.028471196070313454,
-0.020938673987984657,
0.0002971403009723872,
-0.057497017085552216,
-0.018429763615131378,
0.030760522931814194,
-0.0017202121671289206,
0.026993324980139732,
-0.07535946369171143,
0.004232258070260286,
0.02416267804801464,
-0.11791961640119553,
-0.039891306310892105,
0.0029420473147183657,
0.00519578717648983,
-0.0020296277943998575,
0.06923699378967285,
0.010005669668316841,
0.058033768087625504,
-0.010364595800638199,
0.015328571200370789,
-0.05029618740081787,
0.03325202688574791,
-0.04027264565229416,
0.00933818519115448,
0.0026547929737716913,
0.08579288423061371,
-0.01626446098089218,
0.020476387813687325,
-0.047333378344774246,
-0.05150068923830986,
-0.007351535372436047,
-0.06821268051862717,
0.11327818036079407,
0.026251589879393578,
0.12676773965358734,
-0.10730625689029694,
-0.03190511465072632,
-0.016595520079135895,
0.0027169175446033478,
0.07828335464000702,
0.01632024347782135,
-0.020414521917700768,
-0.09517195075750351,
-0.029249751940369606,
0.020166434347629547,
-0.044057413935661316,
-0.03563756123185158,
-0.05025765299797058,
-0.11930396407842636,
-0.01525998953729868,
-0.002601632149890065,
0.011576171964406967,
-0.07249445468187332,
-2.4438033478626027e-33,
-0.05137589946389198,
0.018831806257367134,
-0.03433792665600777,
0.046478986740112305,
-0.010719380341470242,
-0.04978340119123459,
-0.04513641074299812,
-0.05679703876376152,
-0.026303546503186226,
-0.0017960441764444113,
-0.02397809736430645,
-0.028148755431175232,
0.004082652274519205,
-0.003776339814066887,
0.0550684928894043,
0.008489293046295643,
-0.05366041883826256,
0.03401166945695877,
-0.02449874021112919,
-0.00178141996730119,
0.016991544514894485,
0.13063247501850128,
-0.042112965136766434,
0.009257609024643898,
-0.07666251808404922,
0.035911381244659424,
-0.06377305835485458,
0.07358314841985703,
0.02618297003209591,
-0.04345450922846794,
-0.060251496732234955,
-0.0368938073515892,
-0.0060969567857682705,
-0.030538039281964302,
-0.04779546707868576,
0.0494571216404438,
0.025433633476495743,
0.006941919215023518,
0.008873308077454567,
0.03286219760775566,
0.10275745391845703,
-0.08304711431264877,
-0.026354268193244934,
0.005255660507827997,
0.005621303338557482,
-0.006272794213145971,
-0.0689486414194107,
-0.034843601286411285,
0.03124123625457287,
0.03902909904718399,
-0.022625694051384926,
0.03380914032459259,
-0.07077983021736145,
0.05127020925283432,
-0.058957211673259735,
-0.009693773463368416,
-0.06775122880935669,
0.04415375739336014,
0.10494096577167511,
-0.00912063755095005,
-0.08771656453609467,
-0.012884731404483318,
-0.024484889581799507,
-0.012068750336766243,
-0.011619781143963337,
-0.002537196036428213,
-0.002221114933490753,
-0.006980448961257935,
0.045566484332084656,
0.09353452175855637,
-0.009051000699400902,
-0.010438067838549614,
0.03613525256514549,
0.06127304583787918,
-0.1200665608048439,
-0.05488749220967293,
-0.000632009468972683,
-0.07167696952819824,
-0.03911852464079857,
-0.009585360996425152,
0.08885479718446732,
0.07322774827480316,
0.04340176284313202,
-0.010860683396458626,
0.047394875437021255,
0.02549515664577484,
0.13620471954345703,
0.05029957741498947,
0.039843615144491196,
-0.03269382193684578,
-0.03288950398564339,
0.07505979388952255,
0.010741104371845722,
-0.02477666176855564,
-0.07390177249908447,
-4.2607794625837414e-8,
-0.011288237757980824,
0.038558557629585266,
0.029882054775953293,
-0.010079276748001575,
0.05956946313381195,
-0.04049025475978851,
-0.037025436758995056,
0.11659856140613556,
-0.02893909066915512,
-0.06752417236566544,
0.032810866832733154,
-0.026566464453935623,
-0.014405228197574615,
-0.049214962869882584,
-0.0034790036734193563,
0.11058688163757324,
-0.000039392689359374344,
-0.10031415522098541,
0.008713610470294952,
-0.05480355769395828,
-0.034381452947854996,
0.07706791162490845,
-0.00003403723530936986,
0.01863246038556099,
0.02280850149691105,
-0.0854516327381134,
-0.029536817222833633,
0.12234221398830414,
0.11245355010032654,
0.08927594125270844,
-0.013945425860583782,
0.027859359979629517,
0.038086261600255966,
-0.051873426884412766,
0.003812342183664441,
0.05557655170559883,
0.06514399498701096,
0.018303554505109787,
0.005134400445967913,
-0.08708404749631882,
-0.025917964056134224,
0.07048285752534866,
-0.06376161426305771,
0.013024846091866493,
0.023029455915093422,
-0.012630305252969265,
-0.011691934429109097,
-0.08626337349414825,
0.029541119933128357,
0.018425121903419495,
0.07803663611412048,
0.016214167699217796,
0.013323282822966576,
0.041431646794080734,
0.0024435888044536114,
0.0026093064807355404,
-0.04302847385406494,
-0.11552050709724426,
-0.01227373443543911,
0.07315067946910858,
-0.0002568829513620585,
0.018683981150388718,
-0.0004288765776436776,
0.030215833336114883
] | https://github.com/HazyResearch/precision-ls | 5 | 0 | 0 | 0 |
Learning Efficient Positional Encodings with Graph Neural Networks | https://openreview.net/forum?id=AWg2tkbydO | [
"Charilaos Kanatsoulis",
"Evelyn Choi",
"Stefanie Jegelka",
"Jure Leskovec",
"Alejandro Ribeiro"
] | Poster | Positional encodings (PEs) are essential for effective graph representation learning because they provide position awareness in inherently position-agnostic transformer architectures and increase the expressive capacity of Graph Neural Networks (GNNs). However, designing powerful and efficient PEs for graphs poses significant challenges due to the absence of canonical node ordering and the scale of the graph. In this work, we identify four key properties that graph PEs should satisfy: stability, expressive power, scalability, and genericness. We find that existing eigenvector-based PE methods often fall short of jointly satisfying these criteria. To address this gap, we introduce PEARL, a novel framework of learnable PEs for graphs. Our primary insight is that message-passing GNNs function as nonlinear mappings of eigenvectors, enabling the design of GNN architectures for generating powerful and efficient PEs. A crucial challenge lies in initializing node features in a manner that is both expressive and permutation equivariant. We tackle this by initializing GNNs with random node inputs or standard basis vectors, thereby unlocking the expressive power of message-passing operations, while employing statistical pooling functions to maintain permutation equivariance. Our analysis demonstrates that PEARL approximates equivariant functions of eigenvectors with linear complexity, while rigorously establishing its stability and high expressive power. Experimental evaluations show that PEARL outperforms lightweight versions of eigenvector-based PEs and achieves comparable performance to full eigenvector-based PEs, but with one or two orders of magnitude lower complexity. Our code is available at https://github.com/ehejin/Pearl-PE. | positional encodings, graph neural networks, graph transformers | null | 13,142 | 2502.01122 | [
-0.05020107328891754,
-0.052613794803619385,
0.029029782861471176,
-0.020340148359537125,
-0.006683630868792534,
0.05666932091116905,
-0.02311885543167591,
-0.08365564048290253,
0.00043261778773739934,
0.0031208298169076443,
0.057301197201013565,
0.04468367248773575,
0.03357226774096489,
0.07628142833709717,
0.01660129800438881,
0.06087426468729973,
0.056973740458488464,
0.062218938022851944,
-0.03198274224996567,
0.0695313960313797,
0.0648079514503479,
0.0025134470779448748,
-0.0028782389126718044,
0.0072079855017364025,
0.059232261031866074,
0.0025262890849262476,
-0.00047570274909958243,
-0.04195534065365791,
0.06604472547769547,
-0.04711296036839485,
0.07520968466997147,
-0.03117307275533676,
-0.039148494601249695,
0.02774803154170513,
-0.08269735425710678,
0.05789325758814812,
-0.03025851771235466,
0.0011536253150552511,
0.026581736281514168,
0.007353659253567457,
0.052805475890636444,
0.023227451369166374,
0.004566117189824581,
0.05893769860267639,
0.10933321714401245,
0.04447641968727112,
0.06910167634487152,
-0.009364907629787922,
-0.02229764312505722,
-0.008300085552036762,
-0.02397971972823143,
-0.03143606334924698,
-0.038564663380384445,
0.03254615515470505,
0.05892981216311455,
0.027858711779117584,
-0.0019317242549732327,
-0.04740883409976959,
0.019924819469451904,
-0.000496705761179328,
0.01710359752178192,
-0.048637159168720245,
0.047191180288791656,
-0.01766933500766754,
-0.08929453790187836,
-0.05443090200424194,
0.04263623431324959,
0.04941021651029587,
-0.05070369690656662,
0.03401925787329674,
0.014115850441157818,
0.013847870752215385,
-0.0952933207154274,
0.09728517383337021,
0.04489533230662346,
0.09188440442085266,
0.06791799515485764,
-0.03315482661128044,
-0.025311868637800217,
0.024009132757782936,
-0.013054676353931427,
-0.003003697143867612,
0.04537777602672577,
0.033088572323322296,
0.09008462727069855,
-0.011641768738627434,
-0.07734467089176178,
0.027163362130522728,
-0.054991234093904495,
-0.006713006645441055,
-0.04462968930602074,
-0.02652014046907425,
0.03609933704137802,
-0.019976602867245674,
0.04852147772908211,
0.01651132106781006,
0.05820639058947563,
-0.05869597569108009,
-0.008617713116109371,
0.07252814620733261,
-0.04181797057390213,
0.031593699008226395,
0.03456785902380943,
-0.024887697771191597,
-0.02614494413137436,
0.010640125721693039,
-0.0808984562754631,
0.06436387449502945,
-0.03663823381066322,
-0.11487629264593124,
-0.0021126968786120415,
0.03657189756631851,
-0.029947461560368538,
0.02486529015004635,
-0.023700395599007607,
-0.03218870609998703,
-0.02787436731159687,
-0.015373713336884975,
0.09925081580877304,
0.04635695740580559,
-0.014265311881899834,
0.029480991885066032,
-0.10726434737443924,
0.08207875490188599,
-0.04603346437215805,
0.014073412865400314,
-0.07736577838659286,
2.745055734013477e-33,
0.06270664185285568,
0.057601623237133026,
0.06758144497871399,
0.07063142210245132,
0.06776320189237595,
0.040885079652071,
-0.01538289338350296,
-0.022869622334837914,
0.005278029479086399,
0.0015683072851970792,
-0.04221532493829727,
0.005655387416481972,
0.02099769562482834,
0.08895190060138702,
-0.014345376752316952,
-0.01024317741394043,
-0.004606346599757671,
-0.02124013379216194,
0.00677500432357192,
-0.12129542976617813,
0.09049780666828156,
-0.04117472097277641,
0.014312632381916046,
-0.027444690465927124,
-0.05086569860577583,
0.035297542810440063,
0.01598740555346012,
-0.032011035829782486,
-0.06187388300895691,
-0.030994638800621033,
-0.11122837662696838,
0.023811817169189453,
0.00834080670028925,
0.025506889447569847,
-0.000058520687161944807,
-0.0945015400648117,
-0.014065679162740707,
-0.07488280534744263,
0.02364996448159218,
-0.06111150234937668,
-0.008306991308927536,
0.02630108781158924,
-0.025708254426717758,
-0.02104746177792549,
-0.02116590179502964,
-0.010203703306615353,
-0.018515458330512047,
-0.0034080154728144407,
-0.021495696157217026,
-0.009378448128700256,
-0.014879449270665646,
-0.0321912057697773,
-0.0671163946390152,
0.09674036502838135,
0.05651414021849632,
0.01428207941353321,
0.09254197776317596,
0.023095628246665,
0.0244982298463583,
-0.0031244412530213594,
-0.06572026014328003,
0.044062692672014236,
-0.004120394587516785,
-0.022082272917032242,
0.02532166987657547,
-0.03263590857386589,
-0.02202686108648777,
0.026068871840834618,
0.057434454560279846,
-0.07690615206956863,
0.01293124444782734,
0.07237426936626434,
-0.028494814410805702,
-0.051645245403051376,
-0.011445127427577972,
-0.01652284897863865,
-0.05117751285433769,
-0.07652803510427475,
-0.07751688361167908,
0.08414659649133682,
-0.030361061915755272,
-0.008042833767831326,
0.011515962891280651,
0.03092023730278015,
-0.030314503237605095,
-0.07948827743530273,
0.033020179718732834,
-0.04835900291800499,
0.032302819192409515,
-0.0416683554649353,
-0.03833938017487526,
-0.04885135591030121,
0.013615136966109276,
-0.016342122107744217,
-0.002793043153360486,
-3.160871248582514e-33,
-0.019581444561481476,
0.08625981211662292,
0.0372360534965992,
0.047738123685121536,
0.0690155029296875,
-0.12470421195030212,
0.016020022332668304,
-0.014813653193414211,
-0.06234687939286232,
0.00012179294571978971,
-0.10719748586416245,
-0.05935432389378548,
0.004270673729479313,
-0.06536437571048737,
0.0651838406920433,
0.07622732222080231,
-0.0862794741988182,
-0.013698714785277843,
-0.06623842567205429,
0.014483505859971046,
-0.04477909207344055,
0.06284008175134659,
-0.07420692592859268,
0.058169927448034286,
0.11095357686281204,
0.03329341858625412,
0.004927212372422218,
0.007760187145322561,
0.07190024107694626,
-0.017239093780517578,
-0.08783212304115295,
-0.04566476121544838,
-0.03688058629631996,
-0.010053694248199463,
0.015438005328178406,
0.032379087060689926,
-0.041490305215120316,
-0.01845168136060238,
0.014424581080675125,
0.10627803206443787,
0.046435315161943436,
-0.04277992993593216,
-0.06910169869661331,
0.04100184142589569,
0.016736023128032684,
0.0513937771320343,
0.005366831552237272,
0.04530099406838417,
-0.013679672963917255,
0.056180328130722046,
-0.029754921793937683,
0.08002971857786179,
-0.06703106313943863,
-0.006133697926998138,
-0.0024998588487505913,
-0.06908740103244781,
0.047369036823511124,
0.08470051735639572,
0.05946213752031326,
-0.034249670803546906,
-0.003219753969460726,
-0.06152527034282684,
0.01615455001592636,
-0.053197793662548065,
-0.05359054356813431,
-0.029796261340379715,
0.0025761432480067015,
0.005201100837439299,
0.028116760775446892,
0.029640594497323036,
-0.00021719290816690773,
-0.07752865552902222,
-0.027076231315732002,
-0.05561848357319832,
0.05262675881385803,
-0.04834766313433647,
-0.050031717866659164,
-0.0007649147883057594,
-0.03647422417998314,
-0.03798271343111992,
-0.022535035386681557,
-0.0001951178564922884,
-0.005193090531975031,
0.09723086655139923,
0.1079278290271759,
0.07702012360095978,
0.021416405215859413,
0.022753937169909477,
0.03322627767920494,
0.000005838006472913548,
0.020794769749045372,
0.09100920706987381,
-0.10997886210680008,
0.09775529056787491,
0.024070153012871742,
-4.279310061860997e-8,
-0.2056327760219574,
-0.0038717836141586304,
-0.04175649210810661,
0.020790237933397293,
0.06443897634744644,
-0.07353782653808594,
0.10769698768854141,
0.11847497522830963,
-0.0008556116954423487,
0.050075847655534744,
0.09095647931098938,
-0.04617743566632271,
-0.006495460867881775,
-0.10101199150085449,
-0.0066214012913405895,
0.028779156506061554,
0.019834809005260468,
-0.05416936054825783,
0.027309132739901543,
-0.02958422526717186,
-0.05092734843492508,
0.024543119594454765,
0.0005448099691420794,
0.024061208590865135,
-0.04227828606963158,
-0.14928747713565826,
-0.07251597195863724,
0.0434272326529026,
0.056004781275987625,
0.028524450957775116,
-0.024696655571460724,
-0.003995970822870731,
0.06089767813682556,
0.011658884584903717,
-0.007027811836451292,
0.08192531019449234,
0.05469507724046707,
-0.049432411789894104,
-0.006759643089026213,
-0.05133059620857239,
-0.014480877667665482,
0.08773670345544815,
-0.011846824549138546,
0.02676953375339508,
-0.032502103596925735,
0.006893011275678873,
0.03770381584763527,
-0.010218888521194458,
0.042880475521087646,
-0.03094394877552986,
-0.0071992757730185986,
-0.05244312807917595,
-0.052004680037498474,
-0.02360321395099163,
-0.0022347630001604557,
0.05946042761206627,
-0.017228228971362114,
-0.0892486423254013,
0.03250874578952789,
0.07334247976541519,
-0.09273623675107956,
0.02965223416686058,
-0.018469059839844704,
-0.0544525645673275
] | 0 | 0 | 0 | 0 |
|
Scale-Aware Contrastive Reverse Distillation for Unsupervised Medical Anomaly Detection | https://openreview.net/forum?id=HNOo4UNPBF | [
"Chunlei Li",
"Yilei Shi",
"Jingliang Hu",
"Xiao Xiang Zhu",
"Lichao Mou"
] | Poster | Unsupervised anomaly detection using deep learning has garnered significant research attention due to its broad applicability, particularly in medical imaging where labeled anomalous data are scarce. While earlier approaches leverage generative models like autoencoders and generative adversarial networks (GANs), they often fall short due to overgeneralization. Recent methods explore various strategies, including memory banks, normalizing flows, self-supervised learning, and knowledge distillation, to enhance discrimination. Among these, knowledge distillation, particularly reverse distillation, has shown promise. Following this paradigm, we propose a novel scale-aware contrastive reverse distillation model that addresses two key limitations of existing reverse distillation methods: insufficient feature discriminability and inability to handle anomaly scale variations. Specifically, we introduce a contrastive student-teacher learning approach to derive more discriminative representations by generating and exploring out-of-normal distributions. Further, we design a scale adaptation mechanism to softly weight contrastive distillation losses at different scales to account for the scale variation issue. Extensive experiments on benchmark datasets demonstrate state-of-the-art performance, validating the efficacy of the proposed method. The code will be made publicly available. | unsupervised anomaly detection, medical images, contrastive reverse distillation, student-teacher | null | 13,123 | 2503.13828 | [
-0.03552740439772606,
-0.08214949071407318,
0.08235257118940353,
0.06596878916025162,
0.021293053403496742,
-0.043208200484514236,
0.012429208494722843,
-0.038936082273721695,
0.0020574978552758694,
-0.07792599499225616,
-0.024702172726392746,
-0.04489894211292267,
-0.004902901127934456,
0.06180229038000107,
-0.15015380084514618,
-0.01593564823269844,
0.05273823067545891,
0.12389999628067017,
-0.12677377462387085,
-0.004421497229486704,
-0.027514342218637466,
0.07869217544794083,
-0.03953809663653374,
0.040295928716659546,
0.06398779898881912,
0.005202264524996281,
0.043112192302942276,
-0.06220097839832306,
0.03540967032313347,
-0.0635865181684494,
0.10163429379463196,
0.0022839384619146585,
-0.08332324773073196,
0.03300892934203148,
-0.05606909841299057,
0.05070026218891144,
-0.009171283803880215,
0.04191482812166214,
-0.0016304159071296453,
-0.042542714625597,
0.09441708028316498,
0.006436167284846306,
-0.0005508529138751328,
0.041655033826828,
0.035734813660383224,
-0.047261543571949005,
-0.07560029625892639,
-0.0198083333671093,
0.07723218947649002,
-0.009547056630253792,
0.038226377218961716,
-0.027312325313687325,
-0.05436248704791069,
0.07857074588537216,
0.011661493219435215,
-0.027343204244971275,
0.03691916540265083,
-0.02034347876906395,
-0.06366800516843796,
0.032427478581666946,
-0.01012738049030304,
-0.06047708913683891,
-0.028613509610295296,
-0.048093367367982864,
-0.003750571981072426,
-0.001140074455179274,
0.009939203038811684,
0.040318943560123444,
0.027478264644742012,
0.02206762135028839,
0.04076680168509483,
0.0629647746682167,
-0.00200242199935019,
0.02932082489132881,
0.08010957390069962,
0.04508659616112709,
-0.015918947756290436,
-0.0023914813064038754,
0.03248142823576927,
-0.049874939024448395,
0.0629906877875328,
0.018733780831098557,
0.1219455823302269,
0.014403066597878933,
0.1477213203907013,
0.046969782561063766,
-0.023581942543387413,
0.04818647727370262,
0.020552504807710648,
0.03541778400540352,
-0.07676762342453003,
0.024731911718845367,
-0.0612126886844635,
-0.01227634772658348,
0.11480728536844254,
-0.02009792998433113,
0.010910534299910069,
-0.037643399089574814,
0.04547615349292755,
0.041664011776447296,
-0.01443750225007534,
0.01869969442486763,
0.014598878100514412,
0.033580221235752106,
0.05772050842642784,
-0.026685714721679688,
0.07625465840101242,
0.0011444285046309233,
0.05895669758319855,
-0.08355243504047394,
0.07234017550945282,
0.049832459539175034,
0.07894720137119293,
-0.040712952613830566,
0.023488294333219528,
-0.06934607028961182,
0.024753108620643616,
0.03314865380525589,
-0.07749148458242416,
-0.0020570443011820316,
-0.05692894756793976,
-0.013516562059521675,
-0.06538303196430206,
-0.020965397357940674,
-0.002189488150179386,
0.0000081955304267467,
-0.0498930998146534,
8.84826648452095e-34,
0.020284399390220642,
0.03944991156458855,
0.014469488523900509,
0.02249336801469326,
0.04465377330780029,
-0.011924550868570805,
-0.10379233211278915,
0.012639740481972694,
-0.014765152707695961,
0.03837556391954422,
-0.018753090873360634,
0.035761334002017975,
-0.0954691544175148,
0.03585079684853554,
0.016332121565937996,
-0.011680946685373783,
0.030326224863529205,
-0.021300271153450012,
0.0092548206448555,
-0.04988180100917816,
-0.002640664577484131,
-0.04336458072066307,
-0.10133571922779083,
-0.0216498002409935,
-0.07541769742965698,
-0.022862199693918228,
0.006834212690591812,
0.014365331269800663,
0.03485879302024841,
-0.005624768789857626,
-0.09860726445913315,
0.05161671340465546,
0.02717486396431923,
-0.0029723686166107655,
-0.013248027302324772,
-0.06249101087450981,
0.001725615351460874,
-0.011949911713600159,
0.023285234346985817,
-0.024739129468798637,
0.024464601650834084,
0.06395281851291656,
-0.05661264806985855,
-0.042293284088373184,
-0.028979012742638588,
-0.01788213662803173,
0.08500757813453674,
-0.027618056163191795,
-0.07354339212179184,
-0.13563300669193268,
-0.01882115565240383,
0.05076155066490173,
-0.00839726347476244,
-0.08159106224775314,
-0.01653822511434555,
-0.06206038221716881,
0.022608807310461998,
0.04483834281563759,
0.0272707287222147,
0.020989783108234406,
0.018754154443740845,
0.05212779715657234,
-0.026119645684957504,
0.12879124283790588,
-0.022090239450335503,
-0.0011851141462102532,
-0.029564792290329933,
0.046038784086704254,
0.019281428307294846,
0.03954881429672241,
-0.037973422557115555,
0.02286963351070881,
-0.037745747715234756,
-0.05447402596473694,
0.01190190203487873,
0.023969819769263268,
0.05711005628108978,
-0.03848625719547272,
0.00105280673597008,
-0.0030071076471358538,
-0.038949042558670044,
-0.0664660856127739,
-0.008824007585644722,
0.02074201963841915,
-0.09739352762699127,
0.03594915196299553,
0.0010690789204090834,
-0.013410782441496849,
-0.03238782659173012,
-0.017813336104154587,
-0.05475284904241562,
-0.01686098799109459,
0.037073176354169846,
0.0012594553409144282,
0.008498533628880978,
-1.648542297609962e-33,
-0.025343121960759163,
0.07691161334514618,
-0.09365620464086533,
0.0031547248363494873,
0.05278721824288368,
-0.022983361035585403,
-0.0015421831049025059,
0.1243845745921135,
-0.0006590877892449498,
-0.06308325380086899,
0.025277959182858467,
0.03878355398774147,
-0.012843298725783825,
-0.06212398037314415,
-0.0188238974660635,
0.004405721090734005,
-0.05600643903017044,
0.06596078723669052,
-0.03885345906019211,
0.021235594525933266,
0.009286907501518726,
0.11962899565696716,
-0.03982115909457207,
0.05668998882174492,
-0.013750892132520676,
0.04407934844493866,
-0.05312907695770264,
0.04121888056397438,
-0.03620437905192375,
-0.03337116166949272,
-0.026827579364180565,
0.017738724127411842,
0.09637894481420517,
-0.018832651898264885,
-0.10681891441345215,
0.07233666628599167,
0.04716014489531517,
-0.061502452939748764,
-0.05139932408928871,
0.056943971663713455,
0.041577503085136414,
0.0012217684416100383,
-0.04782913997769356,
0.028591327369213104,
-0.01308777928352356,
-0.021739289164543152,
0.02008071541786194,
-0.03276332840323448,
0.06434502452611923,
0.052041687071323395,
0.0014879490481689572,
0.04079151898622513,
-0.053570833057165146,
0.06711143255233765,
-0.0731637105345726,
-0.060960378497838974,
0.030308496206998825,
-0.03609995171427727,
0.003121358575299382,
0.11056394875049591,
-0.04482191428542137,
-0.09871736913919449,
0.008317753672599792,
-0.06378777325153351,
-0.10173928737640381,
-0.012038035318255424,
-0.00382422492839396,
0.06211937218904495,
0.010683136060833931,
0.02085288241505623,
-0.010843170806765556,
0.016366519033908844,
0.037852246314287186,
0.0041609215550124645,
0.003981976304203272,
-0.0801537036895752,
0.0006508696242235601,
-0.046945709735155106,
-0.02091248705983162,
-0.00704182730987668,
0.06094517931342125,
-0.07465959340333939,
0.041170209646224976,
0.07762058079242706,
0.09348417073488235,
0.01062505692243576,
0.014967401511967182,
-0.03943304717540741,
0.0068144043907523155,
0.058204855769872665,
-0.009475266560912132,
-0.03455530107021332,
-0.056844547390937805,
-0.05553774908185005,
0.029040157794952393,
-3.700493067526622e-8,
-0.13719573616981506,
0.01589754968881607,
0.041851092129945755,
-0.015384810045361519,
-0.0007601085817441344,
-0.052171848714351654,
-0.08413677662611008,
0.10245176404714584,
-0.0037647360004484653,
0.04570847377181053,
0.0232015922665596,
-0.061163295060396194,
-0.05456927791237831,
-0.10150069743394852,
0.034884415566921234,
0.0739022046327591,
0.04854227975010872,
0.06763503700494766,
-0.021350057795643806,
-0.012026197277009487,
-0.011533941142261028,
-0.008809790946543217,
0.03343641757965088,
-0.08143281191587448,
0.03161608427762985,
-0.08917644619941711,
-0.02934039942920208,
0.03560764715075493,
0.05730979144573212,
-0.028023846447467804,
0.0020775252487510443,
-0.006772011984139681,
0.07416728138923645,
0.04189577326178551,
-0.06679246574640274,
0.10611361265182495,
-0.018236832693219185,
-0.029566459357738495,
-0.019310589879751205,
-0.06618182361125946,
0.025514716282486916,
0.0471501462161541,
-0.021281488239765167,
0.05769599229097366,
-0.10232489556074142,
-0.005913039669394493,
0.019050130620598793,
-0.002451198175549507,
0.06149708479642868,
0.03402498736977577,
0.08965425193309784,
-0.03492332249879837,
0.11991279572248459,
0.055037666112184525,
0.08198226243257523,
0.012464717961847782,
-0.019841890782117844,
0.053931910544633865,
-0.006695943884551525,
0.007886041887104511,
0.06701218336820602,
0.03623926267027855,
-0.04674472659826279,
-0.058990225195884705
] | https://github.com/medaitech/scrd4ad | 2 | 0 | 0 | 0 |
Scalable Influence and Fact Tracing for Large Language Model Pretraining | https://openreview.net/forum?id=gLa96FlWwn | [
"Tyler A. Chang",
"Dheeraj Rajagopal",
"Tolga Bolukbasi",
"Lucas Dixon",
"Ian Tenney"
] | Poster | Training data attribution (TDA) methods aim to attribute model outputs back to specific training examples, and the application of these methods to large language model (LLM) outputs could significantly advance model transparency and data curation. However, it has been challenging to date to apply these methods to the full scale of LLM pretraining. In this paper, we refine existing gradient-based methods to work effectively at scale, allowing us to retrieve influential examples for an 8B-parameter language model from a pretraining corpus of over 160B tokens with no need for subsampling or pre-filtering. Our method combines several techniques, including optimizer state correction, a task-specific Hessian approximation, and normalized encodings, which we find to be critical for performance at scale. In quantitative evaluations on a fact tracing task, our method performs best at identifying examples that influence model predictions, but classical, model-agnostic retrieval methods such as BM25 still perform better at finding passages which explicitly contain relevant facts. These results demonstrate a misalignment between factual *attribution* and causal *influence*. With increasing model size and training tokens, we find that influence more closely aligns with factual attribution. Finally, we examine different types of examples identified as influential by our method, finding that while many directly entail a particular fact, others support the same output by reinforcing priors on relation types, common entities, and names. We release our prompt set and model outputs, along with a web-based visualization tool to explore influential examples for factual predictions, commonsense reasoning, arithmetic, and open-ended generation for an 8B-parameter LLM. | training data attribution, LLM pretraining, influence functions, factual knowledge | We present a gradient-based influence method that scales to LLM pretraining, and we demonstrate its ability to retrieve examples that influence LLM factual predictions. | 13,120 | 2410.17413 | [
-0.02194981276988983,
-0.06838828325271606,
0.0027010717894881964,
0.09644105285406113,
0.09129508584737778,
0.07885835319757462,
-0.012313260696828365,
-0.027565550059080124,
-0.016129981726408005,
-0.020730426535010338,
-0.03068383038043976,
-0.034975793212652206,
0.059972893446683884,
0.01037729624658823,
-0.05420678108930588,
0.01362320501357317,
0.115550197660923,
0.06932000070810318,
-0.13413593173027039,
-0.11793076246976852,
0.055751387029886246,
0.04182353615760803,
0.01130031980574131,
0.032085858285427094,
0.07443716377019882,
-0.10170074552297592,
-0.026801206171512604,
-0.048045892268419266,
0.0024803548585623503,
-0.008177575655281544,
-0.013373063877224922,
0.012311093509197235,
-0.04363677278161049,
0.004549486096948385,
-0.0469759926199913,
0.032177623361349106,
-0.02029126137495041,
0.049856506288051605,
0.05783435329794884,
-0.011106004007160664,
0.012009203433990479,
-0.03100655786693096,
0.05259326472878456,
0.007798343896865845,
0.09568268805742264,
-0.03599897399544716,
0.051315922290086746,
-0.0007414643769152462,
-0.033508434891700745,
0.021370625123381615,
-0.08246131986379623,
-0.023286018520593643,
0.028428319841623306,
0.008237562142312527,
-0.03800160065293312,
-0.057014577090740204,
0.0520462729036808,
0.04851390793919563,
-0.06739169359207153,
-0.05293882265686989,
-0.1531122624874115,
-0.11589251458644867,
-0.0727885365486145,
-0.012657341547310352,
-0.014536957256495953,
-0.010550517588853836,
0.031136158853769302,
0.030745504423975945,
-0.06225792318582535,
0.07146433740854263,
0.025334343314170837,
0.07148205488920212,
0.011913229711353779,
0.036767639219760895,
0.014437990263104439,
0.07143653929233551,
0.02074122615158558,
0.09466361999511719,
0.024377144873142242,
-0.07116314023733139,
0.028498193249106407,
0.02385980822145939,
0.06419040262699127,
0.03249160572886467,
0.08933403342962265,
-0.013752734288573265,
0.01411362923681736,
-0.013036753050982952,
0.0003476968267932534,
0.05835907533764839,
0.007899065501987934,
-0.042263634502887726,
0.0922616496682167,
-0.004552717786282301,
0.006143032573163509,
0.005955621600151062,
-0.023686286062002182,
-0.07468851655721664,
0.05948236212134361,
0.06857037544250488,
0.018804600462317467,
0.1264384686946869,
0.010442202910780907,
-0.0363180935382843,
-0.015822207555174828,
-0.004922519903630018,
0.04586956277489662,
0.07705343514680862,
0.024074897170066833,
-0.03620623052120209,
0.029924213886260986,
0.050630949437618256,
-0.03056543879210949,
0.004481661599129438,
0.1161869689822197,
-0.01747889071702957,
0.04710589721798897,
-0.07578275352716446,
-0.0367765910923481,
0.042988553643226624,
-0.03898013383150101,
0.06518770754337311,
0.047880206257104874,
0.009584205225110054,
-0.00877390243113041,
-0.0071726529859006405,
-0.07677565515041351,
1.8159209382235474e-33,
0.06709906458854675,
0.05659010633826256,
0.016565967351198196,
0.05598125606775284,
-0.016492415219545364,
-0.009430318139493465,
0.0017671816749498248,
0.035432375967502594,
-0.026741161942481995,
0.018105633556842804,
-0.0314478799700737,
0.04049886763095856,
-0.09289226680994034,
0.014778368175029755,
0.019286703318357468,
0.020160725340247154,
-0.08824872970581055,
0.03088010661303997,
-0.025937756523489952,
-0.027455337345600128,
0.055773042142391205,
-0.06753382086753845,
-0.004661160055547953,
-0.0684240311384201,
0.03936988487839699,
-0.027666490525007248,
0.0681840032339096,
-0.03465601056814194,
-0.10960736125707626,
0.03504703566431999,
-0.07802706211805344,
-0.0799289420247078,
0.06432384997606277,
-0.03441697731614113,
-0.010431838221848011,
-0.07397741824388504,
-0.06257878988981247,
-0.046093396842479706,
0.05519820377230644,
0.00027423916617408395,
-0.014953839592635632,
0.03859556093811989,
-0.018352512270212173,
-0.018729889765381813,
-0.09246809780597687,
-0.04679357260465622,
-0.04006492346525192,
-0.0041274442337453365,
-0.008955519646406174,
0.005133764818310738,
0.08113915473222733,
0.03207090124487877,
0.009454682469367981,
-0.06359029561281204,
-0.0033077914267778397,
0.01998746581375599,
0.02925720252096653,
0.030025217682123184,
-0.005828340537846088,
-0.016981054097414017,
0.06816276907920837,
0.037973277270793915,
-0.034199897199869156,
0.07060842216014862,
-0.001224457984790206,
0.04084198176860809,
-0.03460821509361267,
0.07214014232158661,
0.016671422868967056,
0.018174821510910988,
-0.04198979213833809,
0.014869713224470615,
-0.008958879858255386,
-0.054695770144462585,
0.05475814640522003,
-0.013296919874846935,
-0.010051162913441658,
-0.08889918029308319,
0.021983928978443146,
-0.011949297972023487,
0.0029548148158937693,
-0.029464319348335266,
0.013113598339259624,
-0.05974277853965759,
-0.09872248768806458,
-0.052400555461645126,
0.06445333361625671,
-0.0523974634706974,
0.0019617697689682245,
-0.03726368770003319,
0.018258409574627876,
-0.022556375712156296,
-0.04185893386602402,
-0.027696529403328896,
-0.04520132765173912,
-9.323950477390343e-34,
-0.07622656226158142,
0.014132662676274776,
0.03962916508316994,
0.021926511079072952,
-0.005524287931621075,
-0.07350198179483414,
-0.0032091999892145395,
0.06211736053228378,
0.03204651176929474,
-0.026149675250053406,
-0.011352723464369774,
-0.08193279802799225,
0.03200608491897583,
-0.004997289273887873,
0.10197395086288452,
0.005525784566998482,
0.011024347506463528,
0.05198366567492485,
-0.01038700994104147,
0.042582396417856216,
-0.05832874774932861,
0.03365067392587662,
-0.05497536063194275,
0.020728031173348427,
-0.023704852908849716,
-0.0002929523470811546,
0.011180700734257698,
0.048083264380693436,
-0.0034386967308819294,
0.002627098001539707,
-0.09138412773609161,
0.019578013569116592,
0.028158346191048622,
-0.05725419148802757,
-0.1269099861383438,
0.057010769844055176,
0.0548359714448452,
-0.04657010734081268,
-0.02215699665248394,
0.07548226416110992,
0.04733767360448837,
-0.03302761912345886,
-0.10104051232337952,
-0.0270210150629282,
-0.07702858000993729,
-0.021048685535788536,
-0.14017915725708008,
-0.0002532015205360949,
0.030186563730239868,
-0.007801123429089785,
-0.04298632591962814,
0.00734845781698823,
0.00429513631388545,
0.09537729620933533,
-0.06281174719333649,
-0.08862293511629105,
0.006883236113935709,
-0.043252069503068924,
0.030022231861948967,
-0.007343890145421028,
-0.06388461589813232,
0.05886797606945038,
0.009849068708717823,
-0.07516366243362427,
0.006043106783181429,
-0.07646533101797104,
0.004636612255126238,
0.014338614419102669,
0.019015464931726456,
-0.0029564588330686092,
-0.012883043847978115,
-0.03160935267806053,
-0.004328009206801653,
0.03960965573787689,
-0.04272891953587532,
-0.03563372790813446,
0.04284684360027313,
-0.05290582403540611,
-0.03617604821920395,
-0.04774654656648636,
0.05008652061223984,
0.026074862107634544,
0.012793743051588535,
0.03363052010536194,
0.10477979481220245,
0.07062622159719467,
0.0685560554265976,
0.08595582842826843,
0.012870312668383121,
0.06582578271627426,
-0.056330204010009766,
-0.0028805951587855816,
-0.023942625150084496,
0.054294388741254807,
-0.0564703606069088,
-4.44316228254138e-8,
-0.08003158867359161,
0.04900215193629265,
-0.01212629396468401,
0.12260954082012177,
-0.006253409199416637,
-0.012308876030147076,
-0.05869739502668381,
0.08550258725881577,
-0.056599605828523636,
-0.025128256529569626,
0.01626654528081417,
0.03595750778913498,
-0.06043732911348343,
-0.07854779809713364,
0.01269694697111845,
0.05665721744298935,
-0.04856029152870178,
0.013533110730350018,
-0.031368885189294815,
-0.008666593581438065,
0.04243528097867966,
0.04651637002825737,
0.03130189701914787,
0.011759419925510883,
0.10251358896493912,
-0.061356451362371445,
-0.05972716212272644,
0.1125793531537056,
0.13025259971618652,
-0.039389580488204956,
0.005823396146297455,
0.006292757112532854,
-0.035384126007556915,
-0.03972688689827919,
0.004973961506038904,
0.10497614741325378,
0.06852703541517258,
-0.02074357122182846,
-0.0018180845072492957,
0.06567643582820892,
0.03798284754157066,
0.059481434524059296,
-0.039619576185941696,
0.07351672649383545,
0.05483592674136162,
-0.0061263591051101685,
-0.07618829607963562,
-0.013519958592951298,
0.020253542810678482,
-0.05095931887626648,
0.006206543650478125,
-0.0572255402803421,
-0.028797771781682968,
0.09588142484426498,
0.011106587015092373,
-0.008732501417398453,
-0.03368072956800461,
0.04402844235301018,
0.006183539051562548,
0.03202689066529274,
0.10100913792848587,
0.005243596155196428,
-0.0022287494502961636,
0.04944300651550293
] | https://github.com/pair-code/pretraining-tda | 16 | 0 | 0 | 0 |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.