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
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning | https://openreview.net/forum?id=5U1rlpX68A | [
"Yichen Wu",
"Hongming Piao",
"Long-Kai Huang",
"Renzhen Wang",
"Wanhua Li",
"Hanspeter Pfister",
"Deyu Meng",
"Kede Ma",
"Ying Wei"
] | Oral | Continual Learning (CL) with foundation models has recently emerged as a promising paradigm to exploit abundant knowledge acquired during pre-training for tackling sequential tasks. However, existing prompt-based and Low-Rank Adaptation-based (LoRA-based) methods often require expanding a prompt/LoRA pool or retaining samples of previous tasks, which poses significant scalability challenges as the number of tasks grows.
To address these limitations, we propose Scalable Decoupled LoRA (SD-LoRA) for class incremental learning, which continually separates the learning of the magnitude and direction of LoRA components without rehearsal. Our empirical and theoretical analysis reveals that SD-LoRA tends to follow a low-loss trajectory and converges to an overlapping low-loss region for all learned tasks, resulting in an excellent stability-plasticity trade-off. Building upon these insights, we introduce two variants of SD-LoRA with further improved parameter efficiency. All parameters of SD-LoRAs can be end-to-end optimized for CL objectives. Meanwhile, they support efficient inference by allowing direct evaluation with the finally trained model, obviating the need for component selection. Extensive experiments across multiple CL benchmarks and foundation models consistently validate the effectiveness of SD-LoRA. The code is available at https://github.com/WuYichen-97/SD-Lora-CL. | Continual learning; Low-rank adaptation | null | 6,765 | null | [
-0.003789611393585801,
-0.08700986206531525,
0.04297790676355362,
0.04664197564125061,
0.11223362386226654,
0.02296304889023304,
-0.027607696130871773,
-0.06761135905981064,
-0.00475012231618166,
-0.03892058506608009,
-0.06597495079040527,
0.015147421509027481,
-0.026112811639904976,
0.012121287174522877,
-0.03323198854923248,
0.011628963984549046,
0.050731584429740906,
0.0666959211230278,
-0.03940844163298607,
0.012385227717459202,
-0.012547070160508156,
0.016836846247315407,
0.09864826500415802,
0.07534430921077728,
0.028588825836777687,
0.030017608776688576,
-0.009721532464027405,
-0.05997198447585106,
0.05636431649327278,
-0.094321109354496,
0.03155910223722458,
0.0793500766158104,
-0.02621505968272686,
-0.02859867550432682,
-0.08879027515649796,
-0.005652557127177715,
-0.0054871877655386925,
0.06640256196260452,
-0.02175772748887539,
0.01860443502664566,
0.03214164823293686,
0.04517576843500137,
0.004390287678688765,
-0.013741150498390198,
0.02721427194774151,
-0.027769479900598526,
0.003994368948042393,
-0.14461694657802582,
0.03557954728603363,
0.035053811967372894,
0.025573361665010452,
-0.0547441765666008,
-0.02388186939060688,
0.04246683046221733,
-0.013857487589120865,
0.02639741078019142,
0.028344707563519478,
0.016027487814426422,
-0.012023347429931164,
0.003620370989665389,
0.004123888909816742,
-0.0609465017914772,
0.05415060743689537,
-0.02908659353852272,
-0.040121663361787796,
-0.04190893843770027,
0.038290835916996,
0.05227605253458023,
-0.027030479162931442,
0.024856239557266235,
-0.05426817014813423,
0.042687345296144485,
-0.005107993260025978,
0.035587262362241745,
0.0775623619556427,
0.0073586818762123585,
0.0558246485888958,
0.028930336236953735,
0.09371083229780197,
-0.016334474086761475,
-0.019676093012094498,
0.048019181936979294,
-0.014018452726304531,
-0.050370510667562485,
0.1137666255235672,
-0.018071701750159264,
0.048401568084955215,
0.0048254248686134815,
0.03878239542245865,
-0.0704846903681755,
0.004658398684114218,
0.020505690947175026,
0.01774926669895649,
-0.006809095852077007,
-0.02663041092455387,
0.04539529234170914,
-0.033572033047676086,
-0.0949450433254242,
-0.01448991522192955,
0.09812992066144943,
-0.024289026856422424,
0.04831793159246445,
-0.024586385115981102,
-0.025394482538104057,
-0.05852251499891281,
-0.07385782152414322,
-0.0005832816823385656,
0.07874802500009537,
0.05446434020996094,
-0.09348349273204803,
0.06354334205389023,
0.015223686583340168,
0.01985875889658928,
0.07549925893545151,
0.003338434034958482,
0.014655875973403454,
0.029882948845624924,
-0.047296565026044846,
0.0002792824525386095,
0.0424383170902729,
0.03762920945882797,
-0.013892380520701408,
0.0016755691030994058,
0.006795053835958242,
-0.015318574383854866,
-0.02249648980796337,
-0.0301104336977005,
3.6187218901239796e-33,
0.08977939188480377,
-0.005280756391584873,
0.0438641719520092,
-0.005116824526339769,
0.011510957963764668,
-0.04276329651474953,
0.10329598188400269,
0.07171940058469772,
-0.06948641687631607,
-0.038015998899936676,
-0.04501606523990631,
0.0035176239907741547,
-0.07571278512477875,
0.0912085473537445,
0.08155390620231628,
-0.04974501579999924,
-0.03306117281317711,
0.04296424239873886,
-0.037415359169244766,
-0.11112579703330994,
0.07118649780750275,
-0.035967033356428146,
-0.0062636882066726685,
-0.09466397762298584,
0.07494285702705383,
0.031375546008348465,
0.07514245808124542,
0.012221543118357658,
-0.05756993591785431,
0.0178813599050045,
-0.0529288686811924,
-0.04357225447893143,
0.014405245892703533,
0.0330219492316246,
-0.008781851269304752,
-0.05971113592386246,
-0.013772721402347088,
0.025580283254384995,
0.049410946667194366,
-0.10461144894361496,
0.002984344493597746,
0.004941856488585472,
-0.007783185690641403,
-0.01609080657362938,
-0.07520057260990143,
-0.11507662385702133,
0.0376557931303978,
0.03292972221970558,
-0.0969969853758812,
-0.04275792837142944,
-0.033625081181526184,
-0.02373914048075676,
-0.08797820657491684,
-0.04789048433303833,
-0.08161611109972,
0.0015981222968548536,
0.0077049206010997295,
0.03347517177462578,
-0.016393274068832397,
-0.010772145353257656,
0.02316603995859623,
0.045097433030605316,
-0.07138802856206894,
0.14084935188293457,
0.01903636008501053,
0.03542999178171158,
-0.01677463948726654,
0.041148241609334946,
0.05915048345923424,
-0.03592860698699951,
0.050592485815286636,
-0.013879593461751938,
-0.039120182394981384,
-0.08931244909763336,
0.08856787532567978,
-0.07816727459430695,
0.0228588730096817,
-0.08388808369636536,
0.02429528534412384,
0.04281105101108551,
0.04139423742890358,
0.024280959740281105,
-0.01698339730501175,
-0.0706152394413948,
-0.021649062633514404,
-0.012408625334501266,
0.06852135807275772,
-0.07411809265613556,
-0.01639041304588318,
0.021949488669633865,
-0.07188910990953445,
-0.037817906588315964,
0.031317900866270065,
-0.07564164698123932,
-0.004720797762274742,
-2.8959511887636492e-33,
0.03171258047223091,
0.008946621790528297,
-0.0366307757794857,
0.03626507148146629,
-0.002082552295178175,
0.028732702136039734,
-0.03512265905737877,
0.00014660692249890417,
-0.13559499382972717,
-0.1108788475394249,
-0.040215399116277695,
-0.05611570179462433,
-0.06030094251036644,
0.00024902107543312013,
0.023444276303052902,
0.03606773540377617,
-0.09488697350025177,
-0.018944451585412025,
0.00723497848957777,
0.0013434902066364884,
-0.009470327757298946,
0.05716169252991676,
-0.050193868577480316,
-0.00048517793766222894,
-0.00006670701259281486,
0.02244524657726288,
-0.008694440126419067,
0.1325293481349945,
0.004810674581676722,
-0.010888643562793732,
-0.0823054313659668,
-0.03511099889874458,
-0.03209338337182999,
-0.0064321355894207954,
-0.06543496996164322,
0.07111577689647675,
0.024096505716443062,
-0.0391988568007946,
-0.07929468899965286,
0.12738224864006042,
0.032998017966747284,
-0.0930955782532692,
-0.018049469217658043,
0.024923592805862427,
0.0236368328332901,
-0.05928066372871399,
-0.06430032849311829,
0.054812993854284286,
0.012440506368875504,
0.0644344761967659,
0.002609060611575842,
0.04473443701863289,
-0.08581463992595673,
-0.009598763659596443,
-0.036434587091207504,
-0.023898882791399956,
-0.03626072779297829,
0.00109101680573076,
-0.00803703349083662,
-0.043912384659051895,
0.027406776323914528,
-0.027030929923057556,
0.050067730247974396,
-0.018997780978679657,
0.04522457718849182,
0.023486724123358727,
0.008071399293839931,
0.04435301572084427,
-0.026720190420746803,
0.0036980127915740013,
0.03806747496128082,
0.0609334371984005,
0.005993885453790426,
-0.07057725638151169,
-0.005057359114289284,
-0.015974577516317368,
-0.03652148321270943,
-0.04291153326630592,
0.0034403654281049967,
0.006943671498447657,
-0.03694596886634827,
-0.0018357146764174104,
-0.04016854986548424,
0.07305854558944702,
0.05348489060997963,
0.09806361794471741,
0.01803228072822094,
0.14446555078029633,
0.08274350315332413,
0.02713059075176716,
0.0032904336694628,
-0.0025644926354289055,
0.020866379141807556,
-0.004746688529849052,
-0.008857029490172863,
-4.724793001287253e-8,
-0.11562075465917587,
0.030563702806830406,
-0.06490815430879593,
0.05945158004760742,
0.05495312064886093,
-0.023306312039494514,
-0.0016131214797496796,
0.07403984665870667,
-0.010136888362467289,
-0.025316162034869194,
0.06733617931604385,
0.03120645321905613,
0.05442076548933983,
-0.04092622175812721,
0.042676229029893875,
0.0935363844037056,
0.07588202506303787,
-0.0060275946743786335,
-0.013273142278194427,
-0.046290840953588486,
0.04970342293381691,
0.008211232721805573,
0.012160802260041237,
0.01626787707209587,
0.014334803447127342,
-0.10604359954595566,
-0.027678588405251503,
0.1573021411895752,
0.11552057415246964,
-0.019222740083932877,
0.007267214357852936,
-0.008239737711846828,
0.023931223899126053,
-0.009136018343269825,
0.04529226943850517,
0.080487921833992,
0.07170712202787399,
-0.0007206980371847749,
-0.01746608503162861,
-0.0052515659481287,
-0.041733432561159134,
0.053500402718782425,
-0.0628097802400589,
0.03667832165956497,
-0.04317576810717583,
-0.010469737462699413,
-0.06275927275419235,
-0.07612709701061249,
0.003138807602226734,
-0.05165877938270569,
-0.005370308645069599,
-0.05847744643688202,
-0.032376401126384735,
0.08775573968887329,
0.04604847729206085,
-0.008658144623041153,
-0.00804544985294342,
-0.031942836940288544,
-0.012531661428511143,
0.07792039960622787,
0.029785314574837685,
-0.014009973034262657,
-0.07044732570648193,
0.10624941438436508
] | 0 | 0 | 0 | 0 |
|
Improving Probabilistic Diffusion Models With Optimal Diagonal Covariance Matching | https://openreview.net/forum?id=fV0t65OBUu | [
"Zijing Ou",
"Mingtian Zhang",
"Andi Zhang",
"Tim Z. Xiao",
"Yingzhen Li",
"David Barber"
] | Oral | The probabilistic diffusion model has become highly effective across various domains. Typically, sampling from a diffusion model involves using a denoising distribution characterized by a Gaussian with a learned mean and either fixed or learned covariances. In this paper, we leverage the recently proposed covariance moment matching technique and introduce a novel method for learning the diagonal covariances. Unlike traditional data-driven covariance approximation approaches, our method involves directly regressing the optimal analytic covariance using a new, unbiased objective named Optimal Covariance Matching (OCM). This approach can significantly reduce the approximation error in covariance prediction. We demonstrate how our method can substantially enhance the sampling efficiency, recall rate and likelihood of both diffusion models and latent diffusion models. | Diffusion Model, Generative Model, Probalistic Modelling | We introduce Optimal Covariance Matching (OCM), a novel method that improves sampling efficiency and accuracy in diffusion models by directly regressing optimal analytic covariances. | 6,659 | 2406.10808 | [
-0.009653352200984955,
-0.12877650558948517,
0.0046144877560436726,
0.0026092284824699163,
0.06125537306070328,
-0.033946357667446136,
0.022002000361680984,
-0.04230542108416557,
0.04692288860678673,
-0.010695543140172958,
0.017057672142982483,
-0.0015461177099496126,
-0.006382443476468325,
0.01684723235666752,
-0.01224503293633461,
-0.01340433582663536,
0.09107182174921036,
0.054620757699012756,
-0.00015500681183766574,
-0.019118009135127068,
-0.039287760853767395,
0.0447915643453598,
-0.038361161947250366,
-0.02519945241510868,
0.04539908841252327,
-0.011292378418147564,
0.11089657992124557,
-0.025766579434275627,
-0.09254106879234314,
0.01408339012414217,
0.06140628829598427,
-0.023669453337788582,
-0.023889001458883286,
-0.024933936074376106,
-0.013595019467175007,
-0.016022570431232452,
-0.05293036997318268,
-0.02705514430999756,
0.08007147163152695,
0.059478629380464554,
0.003406099509447813,
0.016033656895160675,
0.009166961535811424,
0.09048835933208466,
0.03603693097829819,
-0.006231891456991434,
0.016175465658307076,
-0.0699712261557579,
0.03279067575931549,
0.050222061574459076,
0.03539106994867325,
-0.038087520748376846,
-0.1310489922761917,
0.01686512865126133,
0.022585468366742134,
-0.013419638387858868,
-0.016607491299510002,
-0.00490014161914587,
0.07795264571905136,
0.0448446050286293,
-0.02916646935045719,
-0.03465999290347099,
-0.03458757698535919,
0.051033344119787216,
-0.07554513961076736,
-0.054589636623859406,
-0.026717370375990868,
0.05896180495619774,
-0.02891220897436142,
0.0732782855629921,
-0.010429645888507366,
0.04978059232234955,
-0.1338834911584854,
0.005753664765506983,
-0.09349019080400467,
0.029708167538046837,
-0.0024111077655106783,
-0.06943463534116745,
0.018496908247470856,
-0.023526590317487717,
-0.02605394832789898,
0.10442963242530823,
0.0982302576303482,
-0.044572245329618454,
0.11433745175600052,
-0.06216892972588539,
0.014886506833136082,
0.03948037326335907,
0.00807538628578186,
-0.09390042722225189,
-0.0016912277787923813,
0.048892125487327576,
-0.0477570965886116,
-0.004908863455057144,
-0.01692608930170536,
0.029731549322605133,
0.04627872258424759,
-0.011366192251443863,
0.057184237986803055,
0.07350634783506393,
-0.030578911304473877,
0.1063946858048439,
-0.03251934051513672,
0.05977458879351616,
0.014070271514356136,
-0.11450115591287613,
0.07749943435192108,
0.07661604136228561,
0.028857765719294548,
-0.051713913679122925,
-0.049402669072151184,
-0.002396738389506936,
-0.010796940885484219,
0.0067315539345145226,
0.003841559635475278,
-0.06264878809452057,
0.07386225461959839,
0.022152172401547432,
0.013722983188927174,
0.09072604030370712,
-0.0031940704211592674,
-0.00026393000734969974,
-0.02072696015238762,
-0.02071286179125309,
-0.00579924788326025,
0.05966322496533394,
-0.03480016067624092,
5.367657282765919e-33,
0.00033137446735054255,
-0.007025136146694422,
0.04763083532452583,
-0.0213210079818964,
-0.044526033103466034,
0.007630014326423407,
-0.036546170711517334,
-0.07720153033733368,
0.04827136546373367,
-0.04954852908849716,
-0.020602894946932793,
0.05497266352176666,
-0.08727417141199112,
0.010761016048491001,
-0.06765075027942657,
0.04232117533683777,
-0.09764306992292404,
0.07504347711801529,
0.032404180616140366,
0.011206001043319702,
0.05545886978507042,
-0.009076017886400223,
-0.007962530478835106,
-0.0008706600056029856,
-0.009920910932123661,
0.007977945730090141,
0.046334683895111084,
-0.037429969757795334,
0.04451247677206993,
0.03752204775810242,
-0.02733776345849037,
-0.020714741200208664,
-0.042039550840854645,
0.0014201975427567959,
0.02933741733431816,
-0.01760222762823105,
-0.015816615894436836,
-0.005177420563995838,
-0.030497029423713684,
-0.08033781498670578,
-0.06481585651636124,
0.01670142263174057,
0.008229534141719341,
-0.05170969292521477,
-0.14675305783748627,
-0.011670168489217758,
-0.01418593991547823,
-0.027304187417030334,
0.07502451539039612,
-0.04679486155509949,
-0.023952357470989227,
-0.06141814962029457,
-0.10589862614870071,
0.011220959946513176,
-0.11116398870944977,
0.046101320534944534,
0.02246367558836937,
-0.03915325924754143,
-0.0009713434847071767,
-0.043328557163476944,
0.014105216600000858,
-0.0018321349052712321,
0.03907141089439392,
-0.028685808181762695,
0.09959360212087631,
0.017360754311084747,
0.0011656292481347919,
0.03221002593636513,
0.05190330743789673,
0.047273147851228714,
0.052306123077869415,
-0.0015657164622098207,
-0.07157652080059052,
-0.07002013176679611,
0.013179084286093712,
-0.01535759400576353,
0.01930881291627884,
-0.01422872580587864,
-0.013757899403572083,
-0.003516529221087694,
-0.03221140801906586,
-0.040638841688632965,
-0.015939073637127876,
-0.1323598325252533,
0.027891969308257103,
0.014852557331323624,
-0.06141261011362076,
-0.0887560322880745,
-0.08560680598020554,
-0.03754308819770813,
-0.016913793981075287,
0.036540623754262924,
-0.07203194499015808,
0.04219335317611694,
-0.019262243062257767,
-4.5907906987227315e-33,
-0.0009261596715077758,
0.01193607784807682,
0.14920347929000854,
0.03901677206158638,
0.004747624509036541,
-0.058153510093688965,
-0.00030177299049682915,
-0.018580611795186996,
0.10700719803571701,
-0.005709420889616013,
-0.038617972284555435,
-0.04743380844593048,
-0.02730032056570053,
-0.052666373550891876,
0.01047536451369524,
0.015875522047281265,
0.01728643663227558,
0.04304536059498787,
-0.010081930086016655,
-0.017724817618727684,
-0.0456320084631443,
-0.031366247683763504,
-0.028550773859024048,
-0.006589612923562527,
0.0034863564651459455,
0.033622466027736664,
-0.00683374609798193,
-0.03254169598221779,
0.0014220996527001262,
-0.016809159889817238,
-0.08510499447584152,
-0.018033331260085106,
-0.025156239047646523,
-0.0029263803735375404,
-0.07479429244995117,
0.13122855126857758,
-0.032278891652822495,
0.08123160153627396,
-0.022470170632004738,
0.00637038005515933,
0.06070893630385399,
-0.01494626421481371,
-0.03650358319282532,
-0.04452970251441002,
0.05488356575369835,
-0.021536828950047493,
0.01427619718015194,
0.08918685466051102,
0.07113948464393616,
-0.010709719732403755,
0.013368197716772556,
0.04434255510568619,
-0.1440170407295227,
0.05165252834558487,
-0.04362469166517258,
-0.003986023366451263,
-0.0333273783326149,
-0.05196244269609451,
0.044623974710702896,
-0.007785620167851448,
0.03845362737774849,
0.08138913661241531,
-0.006137019954621792,
-0.006042846944183111,
-0.05399183928966522,
0.044754840433597565,
0.007712635677307844,
0.02126847207546234,
0.03281175345182419,
0.04572084918618202,
-0.022816769778728485,
-0.01226200070232153,
-0.03874226659536362,
-0.06627622991800308,
0.01107348594814539,
-0.10246805101633072,
0.054172009229660034,
-0.01195222046226263,
-0.07326436042785645,
0.029158178716897964,
0.053952500224113464,
0.02707027830183506,
0.032554902136325836,
-0.011318661272525787,
0.04263181984424591,
0.047206033021211624,
-0.027346136048436165,
-0.07791923731565475,
0.023968560621142387,
-0.000514285929966718,
-0.033179741352796555,
-0.020876266062259674,
0.1087910383939743,
0.0016262904973700643,
-0.1707148402929306,
-4.161665856372565e-8,
-0.09338708221912384,
0.08900652825832367,
0.06849059462547302,
0.037925440818071365,
0.0579984188079834,
0.03488142415881157,
0.016466597095131874,
0.051736000925302505,
0.010475192219018936,
-0.022789238020777702,
0.06109657138586044,
-0.03959745913743973,
-0.011415049433708191,
0.0354665108025074,
-0.022728925570845604,
0.06520451605319977,
0.056521158665418625,
0.05499504879117012,
0.04674272984266281,
-0.03458103537559509,
-0.04335913807153702,
0.002621800173074007,
0.06939730793237686,
0.016090547665953636,
0.08099322021007538,
0.04174487665295601,
-0.05363192781805992,
0.08564000576734543,
0.09653738141059875,
0.06384474784135818,
-0.04956826567649841,
0.018690166994929314,
0.025726472958922386,
-0.052542950958013535,
0.0455738864839077,
0.06409834325313568,
0.05174705386161804,
-0.05062698945403099,
-0.03554442152380943,
-0.07152929902076721,
-0.0295499786734581,
0.022904664278030396,
0.01927560567855835,
0.04488386958837509,
0.050682224333286285,
0.053993530571460724,
0.09216757863759995,
0.010337160900235176,
0.04749032109975815,
0.04977141693234444,
0.08730092644691467,
-0.11436144262552261,
0.03533968701958656,
0.06405968964099884,
0.052939243614673615,
-0.04109374061226845,
0.052374351769685745,
-0.03823093697428703,
0.059718769043684006,
0.002663264749571681,
-0.008047569543123245,
0.02012980356812477,
0.015030154958367348,
0.06800591945648193
] | 0 | 0 | 0 | 0 |
|
PathGen-1.6M: 1.6 Million Pathology Image-text Pairs Generation through Multi-agent Collaboration | https://openreview.net/forum?id=rFpZnn11gj | [
"Yuxuan Sun",
"Yunlong Zhang",
"Yixuan Si",
"Chenglu Zhu",
"Kai Zhang",
"Zhongyi Shui",
"Jingxiong Li",
"Xuan Gong",
"XINHENG LYU",
"Tao Lin",
"Lin Yang"
] | Oral | Vision Language Models (VLMs) like CLIP have attracted substantial attention in pathology, serving as backbones for applications such as zero-shot image classification and Whole Slide Image (WSI) analysis. Additionally, they can function as vision encoders when combined with large language models (LLMs) to support broader capabilities. Current efforts to train pathology VLMs rely on pathology image-text pairs from platforms like PubMed, YouTube, and Twitter, which provide limited, unscalable data with generally suboptimal image quality. In this work, we leverage large-scale WSI datasets like TCGA to extract numerous high-quality image patches. We then train a large multimodal model (LMM) to generate captions for extracted images, creating PathGen-1.6M, a dataset containing 1.6 million high-quality image-caption pairs. Our approach involves multiple agent models collaborating to extract representative WSI patches, generating and refining captions to obtain high-quality image-text pairs. Extensive experiments show that integrating these generated pairs with existing datasets to train a pathology-specific CLIP model, PathGen-CLIP, significantly enhances its ability to analyze pathological images, with substantial improvements across nine pathology-related zero-shot image classification tasks and three whole-slide image tasks. Furthermore, we construct 200K instruction-tuning data based on PathGen-1.6M and integrate PathGen-CLIP with the Vicuna LLM to create more powerful multimodal models through instruction tuning. Overall, we provide a scalable pathway for high-quality data generation in pathology, paving the way for next-generation general pathology models. Our dataset, code, and model are open-access at https://github.com/PathFoundation/PathGen-1.6M. | Image-text pairs generation, Vision-language models, Multi-agent collaboration | We present PathGen-1.6M, an open-source large-scale pathology dataset with 1.6M high-quality image-caption pairs, enabling the creation of powerful multimodal models for pathology analysis. | 6,633 | null | [
0.025998001918196678,
-0.06695249676704407,
-0.008269855752587318,
-0.019589368253946304,
0.06575708091259003,
-0.02956812083721161,
-0.05905779451131821,
-0.032380469143390656,
0.07661963254213333,
-0.08644405752420425,
-0.009907565079629421,
-0.06780733913183212,
0.03603994473814964,
0.08760049939155579,
-0.004069858230650425,
0.00048093366785906255,
0.032685551792383194,
0.07989892363548279,
-0.12462583184242249,
-0.01662312261760235,
0.03294866904616356,
0.0345248319208622,
0.08667626231908798,
-0.05695590004324913,
0.011853822506964207,
-0.047155413776636124,
-0.0076614925637841225,
-0.06029485911130905,
0.04884934797883034,
-0.0069306110963225365,
0.014961852692067623,
-0.005869537591934204,
-0.026854513213038445,
0.10252943634986877,
-0.02546866238117218,
0.09041235595941544,
-0.03896165266633034,
0.0558948777616024,
0.00397910550236702,
-0.01947721093893051,
0.018817611038684845,
0.009707693941891193,
0.02121390588581562,
0.015329061076045036,
0.0854148119688034,
-0.058715399354696274,
-0.0840214192867279,
-0.058453306555747986,
0.023861011490225792,
0.03577036038041115,
-0.07246584445238113,
-0.05675099790096283,
-0.008404274471104145,
0.08310411870479584,
-0.055208779871463776,
-0.036009058356285095,
0.03242054209113121,
-0.018299942836165428,
-0.011022822931408882,
-0.039057258516550064,
-0.026619447395205498,
-0.10307369381189346,
0.011301719583570957,
0.01208923663944006,
0.02972915582358837,
-0.05592694878578186,
0.04783105105161667,
0.01072005182504654,
-0.0422140397131443,
-0.013192174956202507,
-0.1255849003791809,
0.01491621509194374,
0.01742560602724552,
0.06390362232923508,
-0.0794791728258133,
0.046099890023469925,
0.045578956604003906,
-0.001223149592988193,
0.06300070136785507,
-0.07775886356830597,
0.049130652099847794,
0.001791023532859981,
0.11391666531562805,
0.01370539516210556,
0.08903005719184875,
0.0034800663124769926,
0.010444201529026031,
0.04998021945357323,
-0.05619623512029648,
0.013465339317917824,
-0.07599962502717972,
0.007293683476746082,
0.0783274695277214,
-0.016130004078149796,
0.052683088928461075,
0.06953837722539902,
-0.04550113528966904,
-0.09598955512046814,
0.01819564774632454,
0.06978771090507507,
-0.053910017013549805,
-0.007233167067170143,
0.007839735597372055,
-0.015877429395914078,
0.026369936764240265,
0.011331327259540558,
0.07215440273284912,
0.029219500720500946,
0.05166631191968918,
-0.04615410417318344,
0.04554183408617973,
0.032512959092855453,
-0.010298489592969418,
-0.02901213802397251,
0.10202577710151672,
-0.015623807907104492,
-0.015049162320792675,
0.01151827722787857,
0.02344593219459057,
-0.007196193560957909,
-0.03509048372507095,
-0.004823387134820223,
-0.017473939806222916,
-0.054363101720809937,
0.05368883162736893,
0.008525828830897808,
0.01093707699328661,
1.584462610797785e-33,
0.09545446187257767,
0.06854381412267685,
0.09893294423818588,
0.008048256859183311,
0.1110585629940033,
-0.03705362603068352,
-0.033772170543670654,
0.008121787570416927,
-0.0976773276925087,
-0.02810407243669033,
0.00032836274476721883,
-0.019009103998541832,
-0.05565030500292778,
0.0927080512046814,
0.023509299382567406,
-0.029998088255524635,
-0.027158018201589584,
0.06856518238782883,
-0.018039798364043236,
0.07880339026451111,
0.004629424773156643,
-0.01324465498328209,
0.05148680880665779,
-0.00847002025693655,
0.003411827376112342,
-0.0157624464482069,
0.023096919059753418,
-0.04447147250175476,
0.02104257233440876,
-0.007899776101112366,
-0.13562652468681335,
0.05881451070308685,
0.09472336620092392,
0.04792698845267296,
0.0320529080927372,
-0.05167648196220398,
-0.048925142735242844,
-0.011331760324537754,
0.04823000729084015,
0.01747804507613182,
0.038735631853342056,
0.06279117614030838,
-0.029889175668358803,
-0.0681021437048912,
-0.04856450483202934,
-0.035073187202215195,
0.011447766795754433,
0.04224696010351181,
-0.054011955857276917,
-0.06239329278469086,
0.047938812524080276,
0.061439674347639084,
-0.027956439182162285,
-0.06453026086091995,
-0.026887450367212296,
0.004791163839399815,
0.06144776567816734,
-0.009211582131683826,
0.038983117789030075,
-0.03006797283887863,
0.043013930320739746,
0.04723828285932541,
-0.005823565647006035,
0.09167172759771347,
0.048084408044815063,
-0.03509771078824997,
0.050502073019742966,
0.024937406182289124,
-0.011748281307518482,
0.01671484485268593,
-0.04867032170295715,
0.06662959605455399,
-0.05140465870499611,
-0.05517406761646271,
0.038363564759492874,
-0.001771825016476214,
0.04583490267395973,
-0.08787377923727036,
-0.010802320204675198,
0.08123841136693954,
-0.05101735517382622,
0.00385829946026206,
0.007380941417068243,
-0.13093259930610657,
-0.052882272750139236,
-0.023219803348183632,
-0.0705423504114151,
-0.08481179922819138,
-0.015446344390511513,
-0.020174596458673477,
-0.06129272282123566,
0.03539194166660309,
-0.11622702330350876,
0.015097049064934254,
0.04262765869498253,
-1.0156148131539712e-33,
0.02188178524374962,
0.08019409328699112,
-0.008194660767912865,
0.06085562705993652,
-0.025043053552508354,
-0.05808325111865997,
-0.037984125316143036,
-0.0029527219012379646,
0.08191709965467453,
-0.005132075399160385,
0.04480162262916565,
-0.01838552951812744,
-0.017678119242191315,
-0.11987246572971344,
-0.018019527196884155,
-0.0757124051451683,
0.06552363187074661,
0.007108236197382212,
-0.04099424183368683,
-0.022931179031729698,
0.04746295511722565,
0.11336520314216614,
-0.03696410730481148,
0.03390451893210411,
-0.026769768446683884,
0.08466296643018723,
0.028126826509833336,
0.038441985845565796,
-0.002106285421177745,
-0.018916940316557884,
-0.003341318340972066,
0.01637846603989601,
0.004217278212308884,
-0.060968756675720215,
-0.08170866966247559,
0.06294582784175873,
0.041368529200553894,
-0.05665721744298935,
0.01501739677041769,
0.06278381496667862,
0.11015799641609192,
-0.09308397024869919,
-0.11906187981367111,
0.01569461263716221,
-0.04683144390583038,
-0.07176552712917328,
-0.0577089823782444,
0.028249744325876236,
0.03130599856376648,
0.045306749641895294,
-0.03595021739602089,
0.003490840084850788,
-0.1078183725476265,
0.021247131749987602,
-0.02081187441945076,
-0.13053524494171143,
-0.0800902247428894,
-0.02657797746360302,
0.03155858442187309,
-0.06854143738746643,
-0.04862821847200394,
0.016476597636938095,
-0.06938591599464417,
0.020807284861803055,
-0.018916644155979156,
-0.03424853831529617,
-0.011406626552343369,
0.0020385656971484423,
-0.01306151133030653,
0.032227348536252975,
0.027491310611367226,
-0.04617566242814064,
0.007415683474391699,
0.045674677938222885,
0.021784039214253426,
0.007065846584737301,
-0.015506277792155743,
0.018377285450696945,
0.031156150624155998,
-0.049633968621492386,
0.02620081976056099,
-0.01318548433482647,
0.04232894256711006,
0.07887960225343704,
0.1041298359632492,
0.09356408566236496,
0.0019837983418256044,
0.06449811905622482,
0.028888002038002014,
0.04380760341882706,
-0.052651531994342804,
-0.018506644293665886,
0.010312512516975403,
0.027612583711743355,
-0.08853816241025925,
-3.7745888192830535e-8,
-0.0017768493853509426,
-0.012130492366850376,
-0.07737461477518082,
-0.0287084449082613,
-0.05352352187037468,
-0.05157748982310295,
-0.06602388620376587,
0.10526439547538757,
0.04834185540676117,
0.029236426576972008,
0.02163098193705082,
0.0290669072419405,
-0.05793401971459389,
-0.0421665795147419,
0.01733543910086155,
0.06068647280335426,
-0.06022210791707039,
0.05186118558049202,
-0.003698062850162387,
-0.0388568751513958,
-0.04101455211639404,
-0.013381819240748882,
0.06597558408975601,
-0.01278746873140335,
0.061744604259729385,
-0.025052398443222046,
-0.061140026897192,
0.03522836044430733,
-0.006896292790770531,
-0.0007081710500642657,
0.05102800950407982,
-0.02287081442773342,
-0.03832004591822624,
-0.005331206135451794,
0.01952420547604561,
0.03482953459024429,
0.0021576001308858395,
-0.04010112211108208,
0.01582980528473854,
-0.005662804003804922,
0.0530509352684021,
0.055549826472997665,
0.01097029633820057,
-0.011453908868134022,
-0.04348495975136757,
0.00011798654304584488,
-0.025962121784687042,
-0.02775331400334835,
0.06202487275004387,
-0.08848389983177185,
-0.048326265066862106,
-0.032460767775774,
-0.031001802533864975,
0.084593765437603,
0.03098326548933983,
-0.029523517936468124,
0.0848802924156189,
-0.020776113495230675,
0.13852755725383759,
0.0347696915268898,
0.029445357620716095,
-0.021889586001634598,
0.016875866800546646,
-0.0461047887802124
] | 0 | 0 | 0 | 0 |
|
Training on the Test Task Confounds Evaluation and Emergence | https://openreview.net/forum?id=jOmk0uS1hl | [
"Ricardo Dominguez-Olmedo",
"Florian E. Dorner",
"Moritz Hardt"
] | Oral | We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of techniques to include task-relevant data in the pretraining stage of a language model. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data before evaluation. Lastly, we show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models with broad implications for benchmarking and the study of emergent capabilities. | language models, benchmarking, emergence | null | 6,619 | 2407.07890 | [
-0.06552910059690475,
-0.022980734705924988,
-0.0017970274202525616,
0.08165887743234634,
0.1035735011100769,
0.024032412096858025,
-0.013296441175043583,
0.004680626094341278,
0.012862378731369972,
-0.021428851410746574,
-0.022259682416915894,
-0.08510586619377136,
0.0356457456946373,
0.03175096958875656,
-0.021597525104880333,
-0.08793970942497253,
0.05806644633412361,
-0.011765030212700367,
-0.06600553542375565,
-0.04739583283662796,
0.0341339148581028,
0.04408607259392738,
0.07292342185974121,
0.044322721660137177,
0.01423391792923212,
-0.08139507472515106,
-0.05953310802578926,
-0.04107089713215828,
0.05808473750948906,
0.029246537014842033,
0.021351784467697144,
0.020957928150892258,
-0.006387930829077959,
0.08338744938373566,
-0.004291204269975424,
0.0680467039346695,
0.024130912497639656,
0.030021853744983673,
0.07254444807767868,
-0.015990683808922768,
0.0002535755920689553,
-0.02017863467335701,
0.010959181934595108,
-0.007930057123303413,
0.08502928167581558,
-0.07103527337312698,
-0.029991399496793747,
0.004571686964482069,
-0.09475143253803253,
-0.031507883220911026,
-0.12066338211297989,
-0.05160738155245781,
0.04774324223399162,
0.007214306388050318,
-0.04479072988033295,
-0.0001950719451997429,
0.08288612961769104,
0.008451204746961594,
-0.06276890635490417,
-0.027581844478845596,
-0.08827479928731918,
-0.12896908819675446,
-0.05573229864239693,
-0.01431497372686863,
-0.003819146426394582,
0.003137324471026659,
0.0007086773402988911,
0.027996445074677467,
-0.041965220123529434,
0.013910573907196522,
-0.03423178195953369,
0.0904914140701294,
0.019484400749206543,
0.07283525913953781,
0.06891463696956635,
0.07528563588857651,
0.0014402936212718487,
0.020421402528882027,
0.017909996211528778,
-0.09753292053937912,
-0.0044016847386956215,
-0.004384943284094334,
0.01066073402762413,
0.0023674415424466133,
0.08449547737836838,
0.06083846837282181,
0.0779372975230217,
-0.036865174770355225,
0.0044405413791537285,
-0.012365791015326977,
-0.03664460405707359,
-0.04787643626332283,
0.022814039140939713,
0.050157468765974045,
0.04914655163884163,
0.032765988260507584,
0.04967794194817543,
-0.08510220050811768,
0.02954396791756153,
0.050301291048526764,
0.02696995623409748,
0.07385240495204926,
0.03661493584513664,
-0.027900276705622673,
-0.04368460550904274,
-0.03935316950082779,
0.016930149868130684,
-0.028542902320623398,
0.017487993463873863,
-0.06443855911493301,
0.014647520147264004,
0.03426487743854523,
0.015052207745611668,
0.02775704115629196,
0.009031324647367,
-0.017496416345238686,
-0.04239746555685997,
0.002782465424388647,
-0.03238998353481293,
0.09225749969482422,
0.0518256276845932,
0.04955712705850601,
0.005448258947581053,
-0.039490051567554474,
-0.008095945231616497,
-0.004679698497056961,
-0.02546171471476555,
4.089216993789454e-34,
0.07736868411302567,
0.04641985148191452,
0.008546652272343636,
0.09618394076824188,
0.034392714500427246,
-0.06744357198476791,
0.010933135636150837,
0.043162669986486435,
-0.0256928913295269,
-0.02532116137444973,
-0.0637892633676529,
0.023467596620321274,
-0.05062974989414215,
-0.0012660946231335402,
0.031911615282297134,
0.0647561252117157,
-0.07488104701042175,
0.07963903248310089,
-0.023953720927238464,
0.04247424006462097,
0.09642673283815384,
0.007372106425464153,
0.022874943912029266,
-0.13478238880634308,
0.001628414262086153,
0.009351693093776703,
0.06660587340593338,
-0.06920371204614639,
-0.1256520301103592,
0.03211823105812073,
-0.09546861052513123,
-0.02125328779220581,
0.016599131748080254,
0.09075979143381119,
-0.01626008003950119,
0.0018303223187103868,
0.006583763286471367,
-0.0204079020768404,
0.05721179023385048,
-0.01600678078830242,
-0.03058529458940029,
0.02124321646988392,
0.04564782604575157,
-0.0023538388777524233,
-0.009016010910272598,
-0.09844711422920227,
-0.03360392525792122,
-0.08086280524730682,
0.0016116313636302948,
-0.05295330658555031,
0.04885514825582504,
0.02706918865442276,
0.025656849145889282,
-0.048922762274742126,
-0.00488812942057848,
0.055972933769226074,
0.0032572918571531773,
-0.01014757715165615,
0.017168626189231873,
0.04474491626024246,
0.018839016556739807,
0.030558062717318535,
-0.03005564957857132,
0.0639517679810524,
-0.00036304222885519266,
-0.03697628527879715,
-0.005580487195402384,
0.004250429105013609,
0.08780939877033234,
-0.030984969809651375,
-0.027966175228357315,
-0.05768909305334091,
-0.0778050348162651,
-0.035933900624513626,
0.05834631621837616,
-0.013339728116989136,
0.07289423048496246,
-0.08527039736509323,
0.005233631934970617,
-0.030038028955459595,
0.07460937649011612,
-0.07661916315555573,
0.05595315247774124,
-0.08456023782491684,
-0.048442620784044266,
-0.08952181041240692,
-0.005913045257329941,
-0.06198490783572197,
0.009464593604207039,
-0.028547126799821854,
-0.05570583790540695,
0.010828127153217793,
-0.022573841735720634,
-0.06827709078788757,
0.07372332364320755,
-1.9167842309958538e-33,
-0.03637668117880821,
-0.0076967752538621426,
-0.08176054060459137,
0.06012056767940521,
-0.007130597252398729,
-0.05358113348484039,
0.044371090829372406,
0.007212827447801828,
-0.13953913748264313,
-0.05061526969075203,
-0.0713416039943695,
-0.06339417397975922,
0.014276280999183655,
0.018471864983439445,
-0.015914490446448326,
-0.02432718500494957,
0.024017302319407463,
0.031420499086380005,
0.07452960312366486,
0.025975437834858894,
0.05347846820950508,
0.12412871420383453,
-0.046806417405605316,
0.004208572208881378,
-0.027972877025604248,
-0.022811466827988625,
-0.0214366652071476,
0.05545210465788841,
0.019407063722610474,
0.0019868372473865747,
-0.024758784100413322,
0.0355830155313015,
-0.007194710895419121,
0.03859380632638931,
0.017894327640533447,
0.03425810486078262,
0.10515430569648743,
-0.07597382366657257,
-0.037974126636981964,
0.12956461310386658,
0.03689006343483925,
0.014314018189907074,
-0.09711699187755585,
-0.0053823404014110565,
-0.03053586184978485,
-0.004122006241232157,
-0.039661843329668045,
-0.023923825472593307,
0.010686497204005718,
0.018597763031721115,
0.007272045593708754,
0.031148934736847878,
-0.0568842887878418,
0.005846291780471802,
-0.077751524746418,
-0.09016839414834976,
-0.011772600933909416,
-0.08039414882659912,
-0.03358697518706322,
0.0717773288488388,
-0.04047602042555809,
0.07923611998558044,
-0.024695610627532005,
0.025902634486556053,
0.008826634846627712,
-0.06741098314523697,
-0.02014634758234024,
0.044663138687610626,
0.055636417120695114,
-0.010672359727323055,
0.0025007110089063644,
-0.05173978582024574,
-0.00738054933026433,
0.007285988423973322,
0.01479926984757185,
0.02759774774312973,
-0.06339287012815475,
-0.03955134376883507,
-0.03148028999567032,
-0.06551273912191391,
0.0024395242799073458,
-0.06298089772462845,
-0.03100130334496498,
0.010465751402080059,
0.026899617165327072,
0.06728820502758026,
0.055570945143699646,
0.11635962873697281,
-0.045833270996809006,
0.05532713979482651,
-0.00981059018522501,
-0.06524014472961426,
-0.06797721982002258,
0.008883832022547722,
-0.05604270473122597,
-4.465069380898967e-8,
-0.02714516967535019,
0.03448709473013878,
0.03752053901553154,
0.10334628075361252,
-0.031186500564217567,
-0.05212580785155296,
-0.09875556826591492,
0.08296385407447815,
-0.00729057053104043,
-0.07866628468036652,
-0.001884414697997272,
0.04279821366071701,
-0.03001084364950657,
0.01128421165049076,
0.00557243125513196,
0.030231038108468056,
-0.07009442895650864,
0.0607193186879158,
-0.027014028280973434,
0.04303620755672455,
0.09233269095420837,
0.08693571388721466,
0.01397347915917635,
-0.04391540214419365,
0.013695389032363892,
-0.05403652414679527,
-0.04335641860961914,
0.06883671134710312,
0.03832066059112549,
-0.032206661999225616,
0.033201511949300766,
-0.0036257540341466665,
-0.05106961354613304,
-0.041323527693748474,
0.03882088512182236,
0.1223207488656044,
-0.026007816195487976,
-0.025560131296515465,
0.02351089008152485,
0.038933608680963516,
0.04353472962975502,
0.09073546528816223,
-0.026480836793780327,
0.029516467824578285,
0.011172477155923843,
-0.0640948936343193,
-0.09377685189247131,
-0.050853874534368515,
-0.006597171537578106,
-0.10888871550559998,
-0.02227609232068062,
-0.03390005975961685,
-0.06030281260609627,
0.028775611892342567,
0.07344619929790497,
0.027667490765452385,
-0.0036803807597607374,
-0.04291597008705139,
-0.04567908123135567,
0.08205679804086685,
0.06977705657482147,
0.07229695469141006,
0.07511832565069199,
0.031072983518242836
] | https://github.com/socialfoundations/training-on-the-test-task | 9 | 0 | 0 | 0 |
Subgraph Federated Learning for Local Generalization | https://openreview.net/forum?id=cH65nS5sOz | [
"Sungwon Kim",
"Yoonho Lee",
"Yunhak Oh",
"Namkyeong Lee",
"Sukwon Yun",
"Junseok Lee",
"Sein Kim",
"Carl Yang",
"Chanyoung Park"
] | Oral | Federated Learning (FL) on graphs enables collaborative model training to enhance performance without compromising the privacy of each client. However, existing methods often overlook the mutable nature of graph data, which frequently introduces new nodes and leads to shifts in label distribution. Since they focus solely on performing well on each client's local data, they are prone to overfitting to their local distributions (i.e., local overfitting), which hinders their ability to generalize to unseen data with diverse label distributions. In contrast, our proposed method, FedLoG, effectively tackles this issue by mitigating local overfitting. Our model generates global synthetic data by condensing the reliable information from each class representation and its structural information across clients. Using these synthetic data as a training set, we alleviate the local overfitting problem by adaptively generalizing the absent knowledge within each local dataset. This enhances the generalization capabilities of local models, enabling them to handle unseen data effectively. Our model outperforms baselines in our proposed experimental settings, which are designed to measure generalization power to unseen data in practical scenarios.
Our code is available at https://github.com/sung-won-kim/FedLoG | Graph Neural Networks, Graph Federated Learning | null | 6,521 | 2503.03995 | [
-0.04772266373038292,
-0.0642642080783844,
0.03660338371992111,
0.07053810358047485,
0.10599236190319061,
-0.02198667824268341,
-0.00812617689371109,
-0.04934016242623329,
-0.005346956197172403,
-0.06993553042411804,
-0.027873221784830093,
-0.035659193992614746,
0.04696471616625786,
0.03410404920578003,
0.00837633479386568,
0.01617124117910862,
0.12284751236438751,
0.031921759247779846,
-0.06575265526771545,
-0.020261211320757866,
-0.027761327102780342,
0.023451510816812515,
0.0421433262526989,
0.06262151151895523,
0.048353973776102066,
-0.039406441152095795,
-0.0084298774600029,
-0.04721858352422714,
0.00556514598429203,
-0.043746236711740494,
0.03125951066613197,
0.077010378241539,
-0.06843572109937668,
0.02009192854166031,
0.022772012278437614,
0.054554738104343414,
-0.08466875553131104,
0.10288375616073608,
0.06306562572717667,
0.044176459312438965,
0.031166763976216316,
0.006065089255571365,
0.040034472942352295,
0.03963766247034073,
0.013705747202038765,
0.0011010391172021627,
0.016586538404226303,
-0.02363668568432331,
0.000080045749200508,
0.0772813931107521,
-0.03043697215616703,
-0.05123484879732132,
-0.053784552961587906,
0.01817149668931961,
0.0047501567751169205,
-0.03805479034781456,
0.030019225552678108,
-0.08656420558691025,
-0.025612369179725647,
0.06023961678147316,
-0.047581061720848083,
-0.07613047957420349,
-0.036054827272892,
-0.008364916779100895,
-0.04422139376401901,
-0.009496355429291725,
0.05023493990302086,
0.09210193157196045,
0.011616754345595837,
0.04777584597468376,
0.02758149243891239,
0.09165322035551071,
-0.05444617196917534,
0.025526057928800583,
0.05142880603671074,
0.056090835481882095,
-0.02316524088382721,
0.015282339416444302,
0.04331548511981964,
-0.10299863666296005,
-0.04208209365606308,
0.06271927058696747,
0.028904348611831665,
-0.039372533559799194,
0.0998632162809372,
-0.01741345413029194,
0.040225762873888016,
0.004379194229841232,
-0.034852996468544006,
0.036589302122592926,
-0.11188722401857376,
0.0012450264766812325,
0.003966816235333681,
0.00862674880772829,
-0.029724406078457832,
0.06019168347120285,
0.04634230211377144,
-0.04903046041727066,
0.028471097350120544,
0.1040945053100586,
-0.06751132011413574,
0.10310231149196625,
0.051603250205516815,
-0.01006819773465395,
-0.016311759129166603,
0.019062962383031845,
-0.012532460503280163,
0.044703949242830276,
0.04364921152591705,
-0.0360381118953228,
0.03592229634523392,
0.08082626014947891,
-0.07497961819171906,
0.0256339143961668,
0.009314559400081635,
0.04017456993460655,
0.046365175396203995,
-0.05121653154492378,
-0.07453620433807373,
0.05630908161401749,
-0.05795611813664436,
0.029744304716587067,
0.018004434183239937,
-0.0005093299550935626,
-0.048092443495988846,
-0.009126276709139347,
-0.07911554723978043,
2.64968273454348e-33,
0.12203603237867355,
0.023280860856175423,
0.018385356292128563,
-0.00988775584846735,
0.08690036833286285,
-0.018239285796880722,
-0.06271534413099289,
0.024370873346924782,
-0.09137160331010818,
0.0612298846244812,
0.012307269498705864,
0.041634030640125275,
0.0320410393178463,
0.04284080117940903,
0.03024139069020748,
0.011924260295927525,
-0.0028188906144350767,
0.05335237830877304,
0.0052053052932024,
-0.030893290415406227,
0.05619967356324196,
-0.00937732309103012,
0.01072918251156807,
-0.03341112658381462,
0.02464834228157997,
0.021589767187833786,
0.001627729507163167,
0.05507664754986763,
0.024119677022099495,
0.011259093880653381,
-0.042991701513528824,
-0.024726269766688347,
0.07769982516765594,
0.10304774343967438,
0.010945881716907024,
-0.0546552799642086,
-0.009718548506498337,
-0.0922185555100441,
-0.008918805047869682,
0.017464371398091316,
0.014655648730695248,
0.014695653691887856,
0.007367773912847042,
-0.04847168177366257,
-0.07717608660459518,
-0.010870110243558884,
-0.09983272105455399,
-0.021822074428200722,
-0.030648652464151382,
-0.055403877049684525,
0.004135945346206427,
-0.045995958149433136,
-0.029688922688364983,
-0.02363504096865654,
-0.06543067842721939,
-0.014171425253152847,
0.03208232298493385,
-0.0003368578909430653,
-0.0001595990324858576,
-0.021118149161338806,
-0.01444188877940178,
-0.01545900758355856,
0.01726725697517395,
0.038779016584157944,
-0.05005033314228058,
0.012961715459823608,
-0.005595291033387184,
0.0062591303139925,
0.09808703511953354,
-0.05872249975800514,
0.005746441427618265,
0.08524376899003983,
-0.07192501425743103,
0.022207297384738922,
0.04476870968937874,
-0.006288511212915182,
0.03819524496793747,
-0.04272427409887314,
0.01608116552233696,
0.0613330714404583,
0.005539682228118181,
-0.035275451838970184,
-0.05149507522583008,
0.041869621723890305,
-0.10124261677265167,
-0.07026887685060501,
0.035307954996824265,
-0.027524124830961227,
-0.028253044933080673,
0.024319421499967575,
-0.11257027834653854,
-0.05324650555849075,
-0.046369630843400955,
-0.08797287195920944,
-0.02444191835820675,
-2.1994630717197666e-33,
-0.05563428997993469,
0.06691376119852066,
0.029758701100945473,
0.0024786489084362984,
0.011793856509029865,
-0.07866955548524857,
0.0026836921460926533,
0.027241794392466545,
-0.03963238373398781,
0.011789808049798012,
-0.07929840683937073,
-0.049669940024614334,
0.029698988422751427,
-0.04512424021959305,
-0.014147014357149601,
0.033578649163246155,
0.0038859383203089237,
0.01832604594528675,
-0.02654421329498291,
0.019345354288816452,
-0.0008021970861591399,
0.10031528770923615,
-0.08553396165370941,
0.06841087341308594,
0.12002454698085785,
0.036366477608680725,
-0.12310463190078735,
0.10270243883132935,
0.0014231739332899451,
-0.023366887122392654,
-0.013745573349297047,
-0.045570164918899536,
-0.02139696665108204,
-0.12829327583312988,
-0.07252564281225204,
0.06062323972582817,
0.042961981147527695,
0.04582832753658295,
-0.0501214899122715,
0.04618639871478081,
-0.030358413234353065,
-0.05257858708500862,
-0.1858629584312439,
0.028614535927772522,
0.019913114607334137,
-0.028568854555487633,
-0.06162390112876892,
0.007670850958675146,
-0.022513631731271744,
0.02301628887653351,
-0.056845538318157196,
0.04102548584342003,
0.02784011699259281,
0.009714421816170216,
-0.06676899641752243,
-0.09800837934017181,
0.06870338320732117,
0.020709048956632614,
0.03949078917503357,
-0.027619503438472748,
-0.026735655963420868,
-0.05281674489378929,
0.010579348541796207,
0.01687193289399147,
-0.005657005123794079,
-0.09370037913322449,
-0.023605193942785263,
0.04389197379350662,
0.0033667203970253468,
0.03488083556294441,
0.058665890246629715,
-0.009603271260857582,
-0.07964435964822769,
0.03424448519945145,
0.07392886281013489,
0.04357152432203293,
0.01873987913131714,
-0.036617476493120193,
-0.047585371881723404,
0.05040740594267845,
0.059186581522226334,
-0.018306026235222816,
0.016939500346779823,
0.09028047323226929,
0.1236494779586792,
0.02334219217300415,
0.08940143138170242,
0.0933888778090477,
0.01446507964283228,
-0.026262464001774788,
-0.03932357206940651,
0.04378398880362511,
-0.13401296734809875,
0.0746450424194336,
-0.06533680856227875,
-3.9728831779939355e-8,
-0.12572187185287476,
0.05272553116083145,
-0.03970722854137421,
0.029604095965623856,
0.015877731144428253,
0.018521858379244804,
0.017386168241500854,
0.11867958307266235,
-0.04252873733639717,
0.012383732944726944,
0.035506896674633026,
-0.019188346341252327,
0.00925395730882883,
-0.05591863766312599,
-0.005773182958364487,
0.02204054594039917,
0.02621767483651638,
0.03588365390896797,
-0.028759721666574478,
0.07381244003772736,
-0.03796369209885597,
0.026468098163604736,
0.014508994296193123,
-0.0076004150323569775,
0.03168212249875069,
-0.11160068958997726,
0.04761480540037155,
0.06417971104383469,
0.03956542909145355,
0.03283476084470749,
-0.060874149203300476,
0.006071559153497219,
0.010796220041811466,
0.014507369138300419,
-0.03378242999315262,
0.11554791778326035,
0.022894971072673798,
-0.04254600778222084,
-0.036218345165252686,
0.014517546631395817,
0.003373742802068591,
0.0037700419779866934,
-0.08925852179527283,
0.011372331529855728,
0.03426756709814072,
-0.04026250168681145,
-0.015697471797466278,
-0.04151197150349617,
0.07796140760183334,
-0.006455527152866125,
-0.012432546354830265,
-0.03952893987298012,
-0.036249689757823944,
0.015725748613476753,
0.04381818696856499,
0.009118580259382725,
-0.06727335602045059,
-0.005757860839366913,
0.0772053450345993,
0.03513462841510773,
-0.03835311532020569,
-0.033823784440755844,
-0.034151364117860794,
0.007021026220172644
] | https://github.com/sung-won-kim/fedlog | 7 | 0 | 0 | 0 |
A Probabilistic Perspective on Unlearning and Alignment for Large Language Models | https://openreview.net/forum?id=51WraMid8K | [
"Yan Scholten",
"Stephan Günnemann",
"Leo Schwinn"
] | Oral | Comprehensive evaluation of Large Language Models (LLMs) is an open research problem. Existing evaluations rely on deterministic point estimates generated via greedy decoding. However, we find that deterministic evaluations fail to capture the whole output distribution of a model, yielding inaccurate estimations of model capabilities. This is particularly problematic in critical contexts such as unlearning and alignment, where precise model evaluations are crucial. To remedy this, we introduce the first formal probabilistic evaluation framework for LLMs. Namely, we propose novel metrics with high probability guarantees concerning the output distribution of a model. Our metrics are application-independent and allow practitioners to make more reliable estimates about model capabilities before deployment. Our experimental analysis reveals that deterministic evaluations falsely indicate successful unlearning and alignment, whereas our probabilistic evaluations better capture model capabilities. We show how to overcome challenges associated with probabilistic outputs in a case study on unlearning by introducing (1) a novel loss based on entropy optimization, and (2) adaptive temperature scaling. We demonstrate that our approach significantly enhances unlearning in probabilistic settings on recent benchmarks. Overall, our proposed shift from point estimates to probabilistic evaluations of output distributions represents an important step toward comprehensive evaluations of LLMs. | Machine Unlearning, Alignment, Large Language Models | We demonstrate that existing deterministic evaluations in large language models are insufficient and propose a novel probabilistic evaluation framework that considers the whole output distribution of a model. | 6,509 | 2410.03523 | [
-0.0783272311091423,
-0.10545025765895844,
-0.02609238028526306,
0.05203323811292648,
0.0699981227517128,
0.0006386461318470538,
-0.0239811223000288,
-0.04407316818833351,
0.1079852357506752,
-0.024179456755518913,
-0.020369166508316994,
-0.11308100819587708,
0.08420682698488235,
0.003421981818974018,
-0.010042566806077957,
-0.024992380291223526,
0.08445044606924057,
0.04703043773770332,
-0.09533639252185822,
-0.0810701996088028,
0.0061181699857115746,
0.07452617585659027,
0.022399043664336205,
0.019420308992266655,
0.04709237441420555,
-0.08353446424007416,
-0.0009059796575456858,
0.025810081511735916,
-0.016812965273857117,
0.008972960524260998,
-0.0013017577584832907,
0.005015198606997728,
0.01916249468922615,
0.02443557046353817,
0.04650215432047844,
0.05908357724547386,
-0.08823652565479279,
-0.05215880647301674,
0.02994128316640854,
-0.027237869799137115,
-0.01690170355141163,
0.01382364984601736,
0.02109570801258087,
0.004245284479111433,
0.06287337839603424,
-0.05552923306822777,
-0.015989530831575394,
-0.021001027897000313,
-0.06542780250310898,
-0.000024347715225303546,
-0.07006892561912537,
-0.006143408827483654,
-0.00428665429353714,
0.016531381756067276,
-0.02956622652709484,
0.043993085622787476,
-0.0028739580884575844,
0.0055695208720862865,
-0.03707491233944893,
-0.05802487954497337,
-0.08034857362508774,
-0.08595927059650421,
-0.06659585982561111,
0.031205089762806892,
0.0029485293198376894,
-0.02254125103354454,
0.03825082257390022,
0.07118579745292664,
-0.058926139026880264,
0.08979235589504242,
-0.10147328674793243,
0.06834077835083008,
0.01771707274019718,
0.08964061737060547,
0.013293439522385597,
0.0392036959528923,
0.04884767532348633,
-0.018999971449375153,
0.015436910092830658,
-0.060305122286081314,
0.008351316675543785,
-0.008988804183900356,
0.07846815139055252,
0.02449515275657177,
0.0864652544260025,
-0.007561839185655117,
0.055274736136198044,
0.021997913718223572,
0.13179916143417358,
0.003713648533448577,
-0.13325005769729614,
-0.09087829291820526,
0.022290950641036034,
0.03874616697430611,
0.01720951683819294,
0.0895528495311737,
0.013894989155232906,
-0.06732740998268127,
0.01651226542890072,
0.07311142981052399,
0.005395038984715939,
0.07537667453289032,
0.06690020859241486,
-0.1252613067626953,
-0.006635558325797319,
-0.04880819842219353,
-0.02208779938519001,
0.03765048086643219,
0.010050848126411438,
-0.06351288408041,
0.058223284780979156,
0.03605365380644798,
0.026751436293125153,
0.05916615203022957,
0.05317133292555809,
-0.036027539521455765,
-0.007669289596378803,
-0.05306638404726982,
-0.052254825830459595,
0.04006200656294823,
-0.06260747462511063,
0.06928670406341553,
0.03051559440791607,
0.02194545976817608,
-0.06407056748867035,
0.033372361212968826,
0.024328023195266724,
3.5746191787910506e-33,
0.05001845210790634,
0.061872873455286026,
0.043001074343919754,
0.0670371800661087,
0.06379804760217667,
0.014036042615771294,
-0.01042194478213787,
0.020331352949142456,
-0.06478723883628845,
0.0005319655756466091,
-0.03424802049994469,
0.03757587820291519,
-0.017698241397738457,
-0.002218262292444706,
0.04151132330298424,
0.04197833687067032,
-0.029580870643258095,
0.1303766965866089,
-0.0624438151717186,
0.03603266924619675,
0.051305580884218216,
0.001329062506556511,
-0.029218215495347977,
-0.09729799628257751,
0.016497017815709114,
0.020370332524180412,
-0.0019401988247409463,
-0.013036662712693214,
-0.10142466425895691,
0.007258784491568804,
-0.07316653430461884,
-0.05674249678850174,
0.02832985483109951,
-0.023769382387399673,
0.04273790493607521,
-0.09030142426490784,
-0.0643896833062172,
-0.03516154736280441,
-0.041492924094200134,
-0.051368411630392075,
-0.0056648715399205685,
0.05707283318042755,
0.013305864296853542,
-0.0010938318446278572,
-0.061032798141241074,
-0.06168089061975479,
-0.013488506898283958,
-0.03302932530641556,
0.010330443270504475,
-0.1330205351114273,
0.12736248970031738,
0.023181600496172905,
0.0017200240399688482,
0.016788607463240623,
-0.02828763797879219,
-0.02169341780245304,
0.040827684104442596,
-0.009478136897087097,
0.009092990309000015,
0.05826341360807419,
-0.06498818099498749,
0.049752991646528244,
-0.010670223273336887,
-0.012746776454150677,
0.09979411959648132,
0.01703626848757267,
-0.012886489741504192,
0.08804759383201599,
0.05116725340485573,
-0.019385840743780136,
0.013564796186983585,
-0.04821585863828659,
-0.039768993854522705,
-0.026725169271230698,
0.03200861066579819,
0.006579487584531307,
0.0600268729031086,
-0.09995129704475403,
0.0655997022986412,
0.04586441442370415,
-0.011634187772870064,
0.013913851231336594,
0.04057445004582405,
-0.10328643023967743,
-0.04035409539937973,
-0.11452572047710419,
0.02499539963901043,
-0.0518869049847126,
-0.0365561842918396,
0.021906372159719467,
-0.01524337474256754,
-0.02587793581187725,
-0.06331301480531693,
-0.012093422003090382,
0.022086406126618385,
-5.83564839979315e-33,
-0.06478673964738846,
-0.036999598145484924,
0.04248538985848427,
0.06753157079219818,
-0.05398030951619148,
-0.09450779855251312,
-0.05315686762332916,
0.039463289082050323,
-0.02658490277826786,
-0.0937853530049324,
-0.035292353481054306,
-0.056655969470739365,
0.06493784487247467,
-0.02339407615363598,
0.021534355357289314,
0.0018791973125189543,
-0.006419481709599495,
0.0252385213971138,
0.03396601229906082,
0.07241210341453552,
0.04786291345953941,
0.0740056186914444,
-0.0947219729423523,
0.06009345129132271,
0.02735764905810356,
0.01062305923551321,
-0.004001684486865997,
0.04409637674689293,
0.013777162879705429,
-0.017481422051787376,
-0.027941133826971054,
-0.02310599572956562,
-0.10003896057605743,
0.012525851838290691,
-0.09036560356616974,
0.01874047890305519,
0.04129970446228981,
-0.015671666711568832,
0.018973808735609055,
0.10774759948253632,
0.03334102779626846,
-0.035503827035427094,
-0.11507067829370499,
-0.027899624779820442,
-0.011002590879797935,
-0.0538688600063324,
-0.03742162510752678,
-0.02349543757736683,
0.035785213112831116,
-0.016460424289107323,
-0.010198049247264862,
0.058121390640735626,
-0.02435605600476265,
0.06490422785282135,
-0.06439701467752457,
-0.09572622179985046,
-0.11320950835943222,
-0.061234667897224426,
0.03862902149558067,
0.023338057100772858,
-0.037825796753168106,
-0.005573522299528122,
0.016951274126768112,
-0.018402772024273872,
0.01319881808012724,
-0.03931600600481033,
-0.007071256637573242,
0.030125819146633148,
0.02118508331477642,
0.0005081974668428302,
0.0948345959186554,
-0.025740455836057663,
-0.07877334952354431,
0.03249837085604668,
-0.023831674829125404,
-0.0190679132938385,
0.0674515962600708,
-0.05000987648963928,
0.011593062430620193,
-0.08799301087856293,
0.035773053765296936,
0.052576176822185516,
0.04032167047262192,
0.024343030527234077,
0.04248301684856415,
0.013095594942569733,
0.08680409938097,
0.023187104612588882,
0.027978111058473587,
0.01641882210969925,
-0.007279347162693739,
-0.014448469504714012,
0.0009106111247092485,
0.07180079072713852,
-0.08095783740282059,
-4.6859973679147515e-8,
-0.09744267165660858,
0.06074044480919838,
0.01809118501842022,
0.0812915712594986,
-0.035572201013565063,
-0.009945637546479702,
-0.06823565065860748,
0.007027107290923595,
0.03502162918448448,
-0.06346770375967026,
0.043112244457006454,
0.030822178348898888,
-0.04050430655479431,
-0.0092387106269598,
-0.04617885500192642,
0.0760035291314125,
-0.04133286699652672,
0.07764938473701477,
-0.012930784374475479,
-0.008777035400271416,
0.019281750544905663,
0.05875289440155029,
-0.011155975982546806,
-0.0007968018762767315,
0.002735280664637685,
-0.07881690561771393,
-0.04807191714644432,
0.0729394406080246,
0.052989162504673004,
-0.04804900661110878,
-0.05866847187280655,
-0.013596568256616592,
-0.005548081826418638,
-0.020706770941615105,
0.05350485444068909,
0.09285707026720047,
0.006333495490252972,
-0.01805964484810829,
0.04135102033615112,
-0.06019458547234535,
0.0323883481323719,
0.07949204742908478,
-0.07946327328681946,
0.044804222881793976,
-0.007218099199235439,
-0.04470095783472061,
-0.03294483199715614,
-0.019218022003769875,
0.020899256691336632,
-0.034840378910303116,
0.023453399538993835,
-0.049555785953998566,
-0.012208936735987663,
0.022767474874854088,
0.006224439945071936,
0.05739673972129822,
-0.03559385985136032,
0.0011054426431655884,
0.03896743059158325,
0.09422390162944794,
-0.004879648331552744,
-0.016007954254746437,
-0.037270475178956985,
-0.0009149211691692472
] | https://github.com/yascho/probabilistic-unlearning | 6 | 0 | 0 | 0 |
MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering | https://openreview.net/forum?id=6s5uXNWGIh | [
"Jun Shern Chan",
"Neil Chowdhury",
"Oliver Jaffe",
"James Aung",
"Dane Sherburn",
"Evan Mays",
"Giulio Starace",
"Kevin Liu",
"Leon Maksin",
"Tejal Patwardhan",
"Aleksander Madry",
"Lilian Weng"
] | Oral | We introduce MLE-bench, a benchmark for measuring how well AI agents perform at machine learning engineering. To this end, we curate 75 ML engineering-related competitions from Kaggle, creating a diverse set of challenging tasks that test real-world ML engineering skills such as training models, preparing datasets, and running experiments. We establish human baselines for each competition using Kaggle's publicly available leaderboards. We use open-source agent scaffolds to evaluate several frontier language models on our benchmark, finding that the best-performing setup — OpenAI's o1-preview with AIDE scaffolding — achieves at least the level of a Kaggle bronze medal in 16.9% of competitions. In addition to our main results, we investigate various forms of resource-scaling for AI agents and the impact of contamination from pre-training. We open-source our benchmark code https://github.com/openai/mle-bench to facilitate future research in understanding the ML engineering capabilities of AI agents. | benchmark, evals, evaluations, dataset, tasks, data science, engineering, agents, language agents, scaffold, coding, swe, mle | We introduce MLE-bench, a benchmark for measuring how well AI agents perform on machine learning engineering problems. | 6,441 | null | [
-0.09763632714748383,
-0.0868692696094513,
0.006305431015789509,
0.05343106389045715,
0.09301812946796417,
0.0038456874899566174,
-0.005664135795086622,
-0.02951459400355816,
-0.05649920925498009,
-0.038159605115652084,
-0.10218476504087448,
-0.06355263292789459,
0.023933663964271545,
0.015159279108047485,
-0.04128356650471687,
0.014361150562763214,
0.05281839147210121,
0.04321255162358284,
-0.07339853048324585,
-0.06190825626254082,
0.016813145950436592,
-0.012772136367857456,
0.02857677824795246,
-0.02457943744957447,
0.0453917421400547,
-0.05918728932738304,
-0.04023556783795357,
-0.036680061370134354,
0.028925243765115738,
-0.02047686092555523,
0.017166176810860634,
-0.07748949527740479,
0.0049772984348237514,
0.04125116392970085,
-0.004179974086582661,
0.03923974558711052,
-0.02303665690124035,
-0.019120564684271812,
0.011632471345365047,
0.006665811873972416,
0.010031585581600666,
-0.058806609362363815,
0.022166017442941666,
-0.03413860872387886,
0.09619103372097015,
0.031633149832487106,
0.0015166737139225006,
-0.11832588165998459,
0.006036167498677969,
-0.003323298878967762,
-0.11335401237010956,
-0.10767628997564316,
0.03984697908163071,
-0.028939807787537575,
-0.07079999148845673,
0.02753104828298092,
0.0541364885866642,
-0.029769524931907654,
0.02474617399275303,
-0.010919974185526371,
-0.02774370089173317,
-0.07373407483100891,
-0.013068278320133686,
0.030422912910580635,
-0.025270456448197365,
-0.04517753794789314,
-0.04557189345359802,
0.07889223843812943,
0.026634065434336662,
-0.02253485657274723,
0.028294818475842476,
0.0615311935544014,
0.025921007618308067,
0.023931195959448814,
0.07237973809242249,
0.07822585850954056,
0.010432668961584568,
0.018182851374149323,
0.09146721661090851,
-0.02213514782488346,
-0.021366095170378685,
-0.04094889014959335,
0.014527449384331703,
0.022040849551558495,
0.05767138674855232,
-0.0012119370512664318,
0.05089277774095535,
0.003552684560418129,
0.053165704011917114,
0.042746223509311676,
-0.02584836632013321,
0.026003221049904823,
0.04746146500110626,
0.03580474853515625,
0.020905114710330963,
0.0741218850016594,
-0.008654575794935226,
-0.1030307412147522,
-0.04405920207500458,
0.14671045541763306,
-0.07669506967067719,
0.09143531322479248,
-0.008391594514250755,
0.014377481304109097,
-0.02494628168642521,
-0.03694068267941475,
0.0686258003115654,
-0.00443859351798892,
0.04256737604737282,
-0.05667920410633087,
0.08505211025476456,
-0.000822768488433212,
0.06620237231254578,
-0.00003265896157245152,
0.03748316690325737,
0.009033247828483582,
-0.043084073811769485,
0.05980662256479263,
-0.00765585619956255,
0.03696626424789429,
-0.004133102484047413,
0.029744001105427742,
0.018209772184491158,
0.023556994274258614,
0.05320902168750763,
0.04688655957579613,
-0.02714562602341175,
2.6316492277516755e-34,
0.06284255534410477,
0.035904400050640106,
0.07494619488716125,
0.01227116771042347,
0.035032209008932114,
-0.12476756423711777,
0.03015296161174774,
-0.03302493691444397,
-0.05487120896577835,
0.017806531861424446,
-0.05630633607506752,
0.07798337191343307,
-0.0553617849946022,
0.06956521421670914,
0.04544701427221298,
0.028967617079615593,
-0.0346902497112751,
0.02284379117190838,
-0.04518989846110344,
-0.01801692508161068,
0.07940076291561127,
-0.05209115520119667,
0.0042185052298009396,
-0.013326267711818218,
0.036778610199689865,
0.01523883268237114,
0.040388431400060654,
-0.07256417721509933,
-0.004661381244659424,
0.040384095162153244,
-0.09087014943361282,
0.02245500683784485,
-0.00812121294438839,
0.06449680030345917,
-0.03916241601109505,
0.009605764411389828,
-0.06974788010120392,
-0.050767261534929276,
0.030218038707971573,
0.0954781323671341,
-0.05232977122068405,
0.015942012891173363,
0.061818916350603104,
-0.06380369514226913,
-0.07337155938148499,
-0.08132479339838028,
0.017800036817789078,
-0.006894329562783241,
0.036771081387996674,
-0.06657931208610535,
-0.018686803057789803,
0.008940815925598145,
0.05302387848496437,
0.002152440370991826,
0.0008230119128711522,
0.01501378696411848,
0.05810500681400299,
0.049074407666921616,
0.031732287257909775,
0.06271985173225403,
-0.0061101666651666164,
0.031436894088983536,
0.006082381587475538,
0.08386671543121338,
-0.0498090460896492,
-0.005757256411015987,
0.0105679240077734,
0.047635965049266815,
0.035677406936883926,
-0.023427249863743782,
-0.06763982772827148,
0.014322643168270588,
-0.007978175766766071,
-0.09378700703382492,
0.0473775789141655,
-0.02804657071828842,
0.06429595500230789,
-0.0767119750380516,
0.04226052388548851,
-0.023417800664901733,
0.018586060032248497,
0.03428355231881142,
-0.004649714566767216,
-0.05530234053730965,
-0.017625020816922188,
-0.08208956569433212,
-0.03367435559630394,
-0.06478946655988693,
0.008994430303573608,
0.0073398114182055,
-0.12830810248851776,
-0.07972857356071472,
-0.03300251066684723,
0.010547406040132046,
-0.03690874204039574,
3.115244389437599e-34,
-0.016873806715011597,
-0.014745119959115982,
-0.023495705798268318,
0.09063396602869034,
0.004624736029654741,
-0.05232606083154678,
-0.011138616129755974,
0.004488344769924879,
-0.0237332321703434,
0.04899133741855621,
-0.009251882322132587,
-0.04811558499932289,
0.06739214062690735,
0.006673541851341724,
-0.007647991180419922,
-0.09296365827322006,
-0.0502888448536396,
0.01776275411248207,
0.0005669856909662485,
-0.027065442875027657,
0.05323461443185806,
0.11335477977991104,
-0.00414638314396143,
-0.040954288095235825,
0.021680884063243866,
0.01635708101093769,
-0.029438558965921402,
0.058966659009456635,
-0.0003517711302265525,
-0.006803145632147789,
0.015310434624552727,
-0.023204097524285316,
-0.03628974407911301,
-0.008333764038980007,
-0.019191965460777283,
0.08630601316690445,
0.090641550719738,
-0.04548263177275658,
0.008444047532975674,
0.10672786831855774,
0.10519156605005264,
-0.051746051758527756,
-0.07583930343389511,
-0.06197243183851242,
-0.038194864988327026,
0.04589904472231865,
-0.09490837156772614,
0.006484364625066519,
-0.012692030519247055,
-0.06910619139671326,
-0.008676148019731045,
0.07368221879005432,
-0.07756020873785019,
-0.057221587747335434,
-0.05991624295711517,
-0.04677810147404671,
-0.03510359674692154,
-0.007026484701782465,
-0.002885602880269289,
0.04635496437549591,
-0.054288361221551895,
-0.017824430018663406,
-0.01660960167646408,
0.05834538862109184,
-0.027408620342612267,
-0.06749481707811356,
0.023457126691937447,
0.12058823555707932,
-0.06855178624391556,
-0.04874887689948082,
-0.0171989593654871,
-0.03748944029211998,
0.011970622465014458,
0.043390244245529175,
-0.029341235756874084,
0.05807275325059891,
-0.009421220980584621,
0.03894869610667229,
0.039777759462594986,
-0.017586346715688705,
0.02465328387916088,
-0.07364730536937714,
0.04811617359519005,
0.025359202176332474,
0.11188967525959015,
0.11003419011831284,
0.07203233242034912,
0.07023188471794128,
0.0390070341527462,
0.034217070788145065,
-0.05084457993507385,
0.001201158040203154,
0.005878190975636244,
0.04874013736844063,
-0.047294098883867264,
-3.667868853085565e-8,
-0.038818977773189545,
0.05775803327560425,
0.06210164353251457,
0.049185384064912796,
-0.03993714973330498,
0.013774005696177483,
-0.04839611053466797,
0.06406931579113007,
-0.05917789414525032,
0.006794915068894625,
0.08044464141130447,
-0.02094368264079094,
-0.028170641511678696,
-0.059379033744335175,
-0.011299165897071362,
0.09515555202960968,
-0.09570144861936569,
0.06363428384065628,
-0.03407276049256325,
-0.07727733999490738,
0.06200944632291794,
0.0572623535990715,
0.07158375531435013,
-0.061672039330005646,
-0.0029742398764938116,
-0.12324723601341248,
-0.05765927582979202,
0.07842958718538284,
-0.02295227162539959,
-0.017822783440351486,
-0.0278604868799448,
0.0024090660735964775,
0.04626287519931793,
-0.055671390146017075,
0.08112679421901703,
0.1255168467760086,
-0.05436078459024429,
-0.03992409631609917,
0.015443420968949795,
0.05930139124393463,
-0.04977750405669212,
0.04847021400928497,
-0.03989188000559807,
-0.05409860238432884,
0.04016658291220665,
0.05348388850688934,
-0.06971095502376556,
-0.037833329290151596,
0.030980758368968964,
-0.06470181047916412,
-0.009789208881556988,
-0.04801098629832268,
-0.024834226816892624,
0.07323955744504929,
0.06884292513132095,
0.0705503597855568,
-0.004448773339390755,
-0.05410945042967796,
0.022630730643868446,
0.06771412491798401,
0.03147440403699875,
-0.01854686439037323,
-0.037550006061792374,
0.014303320087492466
] | 0 | 0 | 0 | 0 |
|
Learning Randomized Algorithms with Transformers | https://openreview.net/forum?id=UV5p3JZMjC | [
"Johannes Von Oswald",
"Seijin Kobayashi",
"Yassir Akram",
"Angelika Steger"
] | Oral | Randomization is a powerful tool that endows algorithms with remarkable properties. For instance, randomized algorithms excel in adversarial settings, often surpassing the worst-case performance of deterministic algorithms with large margins. Furthermore, their success probability can be amplified by simple strategies such as repetition and majority voting. In this paper, we enhance deep neural networks, in particular transformer models, with randomization. We demonstrate for the first time that randomized algorithms can be instilled in transformers through learning, in a purely data- and objective-driven manner. First, we analyze known adversarial objectives for which randomized algorithms offer a distinct advantage over deterministic ones. We then show that common optimization techniques, such as gradient descent or evolutionary strategies, can effectively learn transformer parameters that make use of the randomness provided to the model. To illustrate the broad applicability of randomization in empowering neural networks, we study three conceptual tasks: associative recall, graph coloring, and agents that explore grid worlds. In addition to demonstrating increased robustness against oblivious adversaries through learned randomization, our experiments reveal remarkable performance improvements due to the inherently random nature of the neural networks' computation and predictions. | Randomized algorithms, Learning under adversarial losses, Adversarial robustness, In-context learning algorithms | null | 6,351 | 2408.10818 | [
-0.14939145743846893,
-0.03474781662225723,
-0.016507625579833984,
0.04333571344614029,
0.05804181098937988,
-0.01182557549327612,
0.010515715926885605,
-0.11545619368553162,
-0.006316841579973698,
-0.0005050686886534095,
-0.06225041300058365,
-0.014211652800440788,
0.02239459939301014,
0.05586472153663635,
-0.0434485487639904,
0.0054363226518034935,
0.09054984152317047,
0.1401040107011795,
-0.12028983235359192,
-0.04397554323077202,
0.03609190136194229,
0.018084311857819557,
0.0676693320274353,
-0.04148838296532631,
0.029139619320631027,
-0.0664297491312027,
0.01959903910756111,
-0.01995113305747509,
-0.006508977152407169,
-0.03057125210762024,
0.08295278251171112,
-0.07214075326919556,
-0.07466068863868713,
0.0328337699174881,
-0.06372050940990448,
0.05402129143476486,
-0.09777668863534927,
0.0023612352088093758,
0.07712963968515396,
-0.02823355793952942,
0.03981848061084747,
-0.03947988897562027,
-0.04290695860981941,
0.006198585499078035,
0.018990173935890198,
0.009940001182258129,
0.012540314346551895,
-0.04041861742734909,
0.013803968206048012,
0.007121434900909662,
-0.023642845451831818,
-0.0650540366768837,
-0.05584239214658737,
0.04419936612248421,
-0.005484537221491337,
0.01210935041308403,
0.05561835318803787,
-0.009985196404159069,
-0.0007511267904192209,
-0.0373433418571949,
-0.008429999463260174,
-0.11004450917243958,
-0.03420322760939598,
0.03180694207549095,
-0.07455334067344666,
-0.045401014387607574,
0.06747858226299286,
0.03881583362817764,
-0.03170734643936157,
0.015168684534728527,
0.057177748531103134,
0.10329499840736389,
-0.025759637355804443,
0.03172195702791214,
0.12505392730236053,
0.0765940323472023,
0.0271682757884264,
-0.07171127945184708,
-0.031080525368452072,
-0.029302341863512993,
0.04678092524409294,
0.024000216275453568,
0.0208754800260067,
0.10484322160482407,
0.09948456287384033,
-0.01979026384651661,
-0.034527551382780075,
0.013109881430864334,
0.05418434739112854,
-0.001235782983712852,
-0.08792354166507721,
0.025933800265192986,
-0.026181094348430634,
-0.030978519469499588,
0.05205914005637169,
0.026976438239216805,
0.05245427042245865,
-0.11550392210483551,
-0.05726446956396103,
0.10177549719810486,
-0.036258235573768616,
-0.059415556490421295,
0.042630311101675034,
-0.0344550684094429,
0.04867422580718994,
-0.06152401119470596,
0.02055317349731922,
0.019596200436353683,
0.02573602832853794,
-0.04555000364780426,
0.09713201969861984,
0.04813681170344353,
0.03484858199954033,
0.02390531450510025,
-0.001069774036295712,
-0.03090760111808777,
-0.038717228919267654,
0.022735122591257095,
-0.08685366809368134,
0.0905701220035553,
-0.02205299399793148,
0.06624957919120789,
-0.011673374101519585,
0.02276114746928215,
0.014362883754074574,
0.044596876949071884,
-0.03891326114535332,
4.85295036719808e-34,
0.04511507973074913,
-0.035307977348566055,
0.04596346616744995,
0.035863157361745834,
0.08711874485015869,
0.023702697828412056,
0.009837075136601925,
0.0700145810842514,
0.009505919180810452,
0.04537881910800934,
-0.05808040872216225,
0.06369563937187195,
-0.05312735214829445,
0.08789736777544022,
0.04012191668152809,
-0.005222204141318798,
-0.008762674406170845,
0.035467904061079025,
-0.02765541896224022,
-0.05255843698978424,
0.08111345022916794,
-0.01684364303946495,
0.06977415829896927,
-0.08675355464220047,
-0.043377507477998734,
-0.035210493952035904,
0.015151224099099636,
0.020618600770831108,
0.05223561450839043,
0.02226104959845543,
-0.08788998425006866,
0.06905575841665268,
-0.014387651346623898,
0.03761831298470497,
0.0037726559676229954,
0.007088711019605398,
-0.028213180601596832,
-0.053089506924152374,
-0.044289495795965195,
-0.018082084134221077,
-0.028741413727402687,
0.027450306341052055,
-0.0027817687951028347,
-0.017046364024281502,
-0.04079556092619896,
-0.01982514001429081,
0.048295482993125916,
-0.00719175161793828,
-0.013409982435405254,
-0.05964737385511398,
0.031682197004556656,
0.021748607978224754,
-0.05198626220226288,
0.02072799950838089,
0.07710358500480652,
-0.006836223416030407,
0.05123352259397507,
0.019876837730407715,
-0.016848774626851082,
0.012272711843252182,
-0.071195587515831,
0.032565195113420486,
0.007434236351400614,
0.0476314052939415,
-0.010765600018203259,
-0.0018482826417312026,
-0.002482594223693013,
0.03975691646337509,
0.0218538548797369,
-0.04390421882271767,
0.017080800607800484,
-0.0015621853526681662,
-0.03798558562994003,
-0.043143533170223236,
-0.010507975704967976,
0.019748395308852196,
0.09957468509674072,
-0.014942503534257412,
0.008183560334146023,
-0.03991730138659477,
-0.007736446335911751,
0.022203722968697548,
-0.03129321336746216,
0.013927585445344448,
-0.07937942445278168,
0.006793934386223555,
0.06663239002227783,
-0.06841062009334564,
-0.02296060137450695,
-0.004384222440421581,
-0.0867827758193016,
-0.02670598030090332,
-0.0055405329912900925,
-0.02017807960510254,
-0.05491571128368378,
-3.031496707672998e-33,
-0.12219706177711487,
0.011429287493228912,
-0.04194439575076103,
0.0610244981944561,
-0.008966636843979359,
-0.050071801990270615,
-0.06530378758907318,
-0.03267672285437584,
-0.04676065966486931,
-0.01456893514841795,
-0.04319566488265991,
0.06360099464654922,
0.05845901370048523,
-0.04071483761072159,
0.027199331670999527,
-0.054693661630153656,
0.022503698244690895,
0.040502503514289856,
-0.019806964322924614,
0.005155864171683788,
0.0344458669424057,
0.14057649672031403,
-0.053138986229896545,
-0.027101576328277588,
-0.009609230794012547,
-0.0051184422336518764,
-0.0736372321844101,
0.0621219165623188,
0.09174177050590515,
0.060882747173309326,
-0.0685477927327156,
0.014248307794332504,
-0.07031780481338501,
-0.021058320999145508,
-0.008652973920106888,
0.08144069463014603,
0.05256178230047226,
-0.0030903806909918785,
-0.04814065247774124,
0.07235336303710938,
0.00014808107516728342,
-0.08678415417671204,
-0.08594264835119247,
0.03798806667327881,
-0.005630680825561285,
-0.08557049185037613,
-0.026469295844435692,
-0.011410159058868885,
0.015152845531702042,
0.03464370593428612,
-0.004150861874222755,
0.018597854301333427,
-0.055501945316791534,
0.023209450766444206,
-0.028704402968287468,
-0.07510937005281448,
-0.00079687888501212,
0.008872691541910172,
0.12265665084123611,
0.050062235444784164,
-0.08617012947797775,
-0.011532047763466835,
-0.03272196277976036,
-0.0276175644248724,
-0.08996273577213287,
0.004873798228800297,
-0.02638770267367363,
-0.03589882701635361,
0.07878507673740387,
0.07956015318632126,
0.03728008642792702,
0.02199431322515011,
-0.02732785977423191,
-0.036696456372737885,
0.007164441514760256,
-0.015280052088201046,
-0.0013574338518083096,
-0.009822389110922813,
-0.040417738258838654,
-0.004701171535998583,
0.052079834043979645,
0.0026360570918768644,
0.022940371185541153,
-0.029701227322220802,
0.027942756190896034,
-0.0007279157871380448,
0.06006631627678871,
0.02197016030550003,
0.09915943443775177,
0.029499730095267296,
0.03515211120247841,
0.02193213254213333,
-0.022282490506768227,
0.007950916886329651,
-0.06308194249868393,
-4.15637941841851e-8,
-0.07898934185504913,
0.09641597419977188,
0.08144629746675491,
0.013739253394305706,
0.03728289529681206,
-0.04884311184287071,
0.02617308311164379,
0.06320304423570633,
-0.03188164532184601,
-0.04102328047156334,
0.15327656269073486,
0.008039657026529312,
0.01303221844136715,
-0.021022427827119827,
0.02128194086253643,
0.08411437273025513,
-0.033593934029340744,
0.00320812058635056,
0.010038446635007858,
-0.010245881974697113,
-0.06735304743051529,
0.0017548176692798734,
-0.02799944579601288,
-0.05555739998817444,
0.013005942106246948,
-0.1190759614109993,
-0.037984151393175125,
0.010151091031730175,
0.011640728451311588,
0.08561387658119202,
-0.08951679617166519,
-0.026901720091700554,
0.03912145644426346,
0.018847201019525528,
-0.05123897269368172,
0.1344982236623764,
-0.003546907100826502,
-0.06338923424482346,
0.02385130152106285,
-0.059723325073719025,
0.00019209404126740992,
0.06844067573547363,
-0.009975981898605824,
0.016597598791122437,
-0.08458971232175827,
-0.013426749035716057,
0.022579528391361237,
-0.03248719871044159,
0.04151797667145729,
0.005691309459507465,
-0.0180670078843832,
-0.038725294172763824,
-0.007853775285184383,
0.013205611146986485,
0.0662628561258316,
0.03828561678528786,
-0.028998173773288727,
-0.08331766724586487,
0.00034483367926441133,
0.13261955976486206,
-0.0048170192167162895,
0.04190410301089287,
-0.1023387610912323,
-0.07601331174373627
] | 0 | 0 | 0 | 0 |
|
Data Scaling Laws in Imitation Learning for Robotic Manipulation | https://openreview.net/forum?id=pISLZG7ktL | [
"Fanqi Lin",
"Yingdong Hu",
"Pingyue Sheng",
"Chuan Wen",
"Jiacheng You",
"Yang Gao"
] | Oral | Data scaling has revolutionized fields like natural language processing and computer vision, providing models with remarkable generalization capabilities. In this paper, we investigate whether similar data scaling laws exist in robotics, particularly in robotic manipulation, and whether appropriate data scaling can yield single-task robot policies that can be deployed zero-shot for any object within the same category in any environment. To this end, we conduct a comprehensive empirical study on data scaling in imitation learning. By collecting data across numerous environments and objects, we study how a policy’s generalization performance changes with the number of training environments, objects, and demonstrations. Throughout our research, we collect over 40,000 demonstrations and execute more than 15,000 real-world robot rollouts under a rigorous evaluation protocol. Our findings reveal several intriguing results: the generalization performance of the policy follows a roughly power-law relationship with the number of environments and objects. The diversity of environments and objects is far more important than the absolute number of demonstrations; once the number of demonstrations per environment or object reaches a certain threshold, additional demonstrations have minimal effect. Based on these insights, we propose an efficient data collection strategy. With four data collectors working for one afternoon, we collect sufficient data to enable the policies for two tasks to achieve approximately 90\% success rates in novel environments with unseen objects. | Data Scaling Laws, Imitation Learning, Robotic Manipulation | null | 6,331 | 2410.18647 | [
0.00957921240478754,
-0.05230950564146042,
0.03883332386612892,
-0.009070315398275852,
0.057800550013780594,
-0.01820823736488819,
-0.015825821086764336,
-0.03587031364440918,
0.025601230561733246,
0.05056198313832283,
0.020370541140437126,
-0.037217218428850174,
0.032205842435359955,
0.053031016141176224,
0.05268585681915283,
0.03135561943054199,
0.11896177381277084,
0.053606633096933365,
-0.1270689070224762,
0.0062568034045398235,
-0.02644653245806694,
-0.01071520708501339,
0.10735426098108292,
0.004783277399837971,
0.003572397166863084,
0.04150617867708206,
-0.011872689239680767,
-0.10103241354227066,
-0.004713698290288448,
-0.06101679056882858,
-0.02505888044834137,
-0.03706337511539459,
-0.002837158739566803,
-0.008830663748085499,
-0.015075246803462505,
0.05194023251533508,
-0.044913213700056076,
-0.009111016057431698,
-0.026773257181048393,
-0.013031382113695145,
-0.004149488639086485,
0.00925174169242382,
0.04347241297364235,
0.007358433213084936,
0.037831466645002365,
0.0715884268283844,
0.027389492839574814,
-0.06780684739351273,
0.07755391299724579,
0.01068231649696827,
-0.05240810662508011,
-0.022814834490418434,
0.020565353333950043,
0.001967349322512746,
-0.003469474846497178,
0.0040488410741090775,
0.06721864640712738,
-0.04435265436768532,
0.11258042603731155,
-0.012329235672950745,
0.014820534735918045,
-0.08029516041278839,
-0.0009728566510602832,
-0.00954305287450552,
-0.05680094286799431,
-0.057822905480861664,
0.003344608936458826,
-0.0337006151676178,
0.012394966557621956,
0.03938983380794525,
-0.004206367768347263,
0.04161325842142105,
0.03239433839917183,
0.029290050268173218,
0.018546294420957565,
-0.053504087030887604,
-0.06281645596027374,
-0.08340166509151459,
-0.007761780638247728,
-0.09209681302309036,
-0.011708520352840424,
-0.07505370676517487,
-0.023025011643767357,
0.06740979105234146,
0.0516112819314003,
-0.04879013076424599,
0.04410528019070625,
0.04175957664847374,
0.057047054171562195,
0.0282928254455328,
-0.03566158562898636,
-0.013529895804822445,
-0.0021274075843393803,
-0.03430936112999916,
-0.025985417887568474,
-0.05309353396296501,
-0.03626447543501854,
-0.14415602385997772,
0.03534107282757759,
0.0615006759762764,
-0.007795024663209915,
0.06563031673431396,
0.045313067734241486,
0.02429325319826603,
-0.02538207732141018,
-0.036767031997442245,
-0.001576099544763565,
-0.006515734363347292,
0.06433684378862381,
-0.0023591523058712482,
0.033307578414678574,
-0.010700209066271782,
0.0070672957226634026,
0.029466619715094566,
0.024591829627752304,
-0.038867805153131485,
-0.010071360506117344,
-0.0010783803882077336,
-0.11085399240255356,
0.10758619010448456,
0.054585766047239304,
-0.04720741882920265,
0.0370618961751461,
0.031415797770023346,
0.06156540289521217,
-0.006380611564964056,
0.0008604920003563166,
-1.70529987143684e-34,
0.12453017383813858,
0.005905899219214916,
0.04338622838258743,
0.04255014285445213,
0.10167254507541656,
0.008501552976667881,
0.045875582844018936,
0.008120965212583542,
-0.040589287877082825,
0.008427267894148827,
-0.011716472916305065,
0.03927406668663025,
-0.04918137192726135,
0.0839390978217125,
0.06472780555486679,
0.046605564653873444,
0.014885696582496166,
0.0486006923019886,
-0.036162134259939194,
0.03947443142533302,
0.09170866757631302,
-0.06960457563400269,
-0.04440038651227951,
-0.026414459571242332,
0.00930595863610506,
-0.042469002306461334,
-0.033370960503816605,
0.018092596903443336,
0.0016137725906446576,
-0.0267986711114645,
-0.04794655367732048,
-0.0045873806811869144,
-0.023128913715481758,
0.1148044764995575,
0.04812854155898094,
0.006676771678030491,
-0.0042733363807201385,
-0.0914648175239563,
0.022759994491934776,
0.0011945292353630066,
-0.003572791116312146,
0.010675473138689995,
0.009703313000500202,
-0.0623277984559536,
-0.05773734301328659,
-0.04740491136908531,
0.0501534678041935,
0.043732546269893646,
-0.05869999900460243,
-0.070279560983181,
0.08566389232873917,
0.049031682312488556,
-0.03398190066218376,
-0.09067355841398239,
0.015096951276063919,
-0.04048539325594902,
-0.021313274279236794,
0.008292518556118011,
0.019185921177268028,
0.023186279460787773,
-0.000429618579801172,
0.011617235839366913,
-0.037732064723968506,
0.09360954165458679,
-0.02431274950504303,
0.06198681518435478,
0.0009151693084277213,
0.04894094169139862,
0.08667022734880447,
0.011778906919062138,
-0.01052829623222351,
0.02540973387658596,
-0.055008500814437866,
-0.0875331461429596,
0.08983200788497925,
0.011387960985302925,
0.030951762571930885,
-0.14723730087280273,
0.014106008224189281,
0.07765739411115646,
-0.08434563130140305,
-0.03128901496529579,
0.023589061573147774,
-0.02166306972503662,
-0.014972051605582237,
-0.03715914860367775,
0.013380772434175014,
-0.0839271992444992,
-0.03147532790899277,
-0.0025279601104557514,
-0.11801302433013916,
-0.0626557394862175,
-0.14807429909706116,
-0.004876124206930399,
-0.01785690151154995,
-1.728845275831425e-33,
-0.015115465968847275,
-0.04584766924381256,
0.038980938494205475,
0.09767431020736694,
-0.05105297267436981,
-0.023822156712412834,
-0.03931781277060509,
0.007424087729305029,
-0.04541442543268204,
-0.08120611310005188,
-0.09648153185844421,
-0.004048742353916168,
0.04964885488152504,
-0.05444721877574921,
0.005825418047606945,
-0.014509245753288269,
-0.002520211273804307,
0.04467209056019783,
-0.004537769127637148,
-0.024268386885523796,
0.02597217820584774,
0.11383949220180511,
-0.010511385276913643,
0.0390288308262825,
-0.0324682891368866,
0.0015021051513031125,
-0.0805901288986206,
0.0191927719861269,
0.035990502685308456,
0.030328428372740746,
-0.03649239242076874,
-0.06081122159957886,
0.008332569152116776,
0.013209710828959942,
-0.051154348999261856,
0.03278874233365059,
0.0621088445186615,
0.03059528023004532,
-0.04600545018911362,
0.07480160892009735,
0.030600661411881447,
-0.06889290362596512,
-0.031076127663254738,
0.016713425517082214,
0.010883787646889687,
-0.12192417681217194,
-0.07037542015314102,
0.05660794675350189,
-0.01718215085566044,
-0.010840756818652153,
-0.0605972521007061,
0.06916899234056473,
-0.057853709906339645,
-0.03223826736211777,
-0.018371501937508583,
-0.03634393960237503,
-0.03687814250588417,
-0.015822023153305054,
0.056097082793712616,
0.011869985610246658,
-0.056808166205883026,
-0.011860300786793232,
-0.012524361722171307,
0.10752309113740921,
-0.03719208762049675,
0.021640855818986893,
-0.018029285594820976,
0.05524907633662224,
-0.0001441441272618249,
-0.004510493949055672,
0.017517274245619774,
-0.0207145344465971,
-0.011305952444672585,
-0.018005046993494034,
-0.01703743264079094,
-0.025787999853491783,
0.03993396461009979,
-0.03540613502264023,
0.03325948491692543,
0.0547749325633049,
0.030464500188827515,
0.03819232061505318,
0.0781836062669754,
0.07883161306381226,
0.039942096918821335,
0.11336866021156311,
-0.009421788156032562,
0.06435828655958176,
0.022011568769812584,
0.06875183433294296,
-0.04409797489643097,
-0.004643789026886225,
-0.019084148108959198,
0.035023339092731476,
-0.030909426510334015,
-4.023596744673341e-8,
-0.11601950228214264,
0.10489384084939957,
0.1202886626124382,
0.03357747942209244,
0.020625445991754532,
0.02837875671684742,
-0.02060757577419281,
0.017729656770825386,
-0.059686753898859024,
-0.036352213472127914,
0.09305761754512787,
0.023416882380843163,
0.0009330286411568522,
-0.020272493362426758,
0.022208787500858307,
0.047989845275878906,
0.01670299656689167,
-0.05251069366931915,
-0.011337601579725742,
0.06636016815900803,
-0.015046162530779839,
-0.02104761265218258,
0.030627675354480743,
-0.061422791332006454,
-0.02163536101579666,
-0.060381051152944565,
-0.05914175510406494,
0.0651937872171402,
0.024402745068073273,
-0.01995040848851204,
-0.010195549577474594,
-0.05652117729187012,
0.030707955360412598,
-0.003276368835940957,
0.0588097907602787,
-0.04327400401234627,
-0.026400242000818253,
-0.02429109811782837,
0.0179554745554924,
-0.094745934009552,
0.016117151826620102,
0.003772580996155739,
0.004781613126397133,
0.02500235103070736,
-0.0014173107920214534,
0.032766617834568024,
-0.10416552424430847,
-0.14502108097076416,
0.019843576475977898,
-0.03130434453487396,
0.038206666707992554,
-0.05585560202598572,
-0.016032427549362183,
0.11847251653671265,
0.14608898758888245,
0.0691981166601181,
-0.013536627404391766,
-0.03844692185521126,
-0.010441998019814491,
0.11148080974817276,
0.02646826207637787,
0.00015026902838144451,
-0.12923432886600494,
-0.03290238231420517
] | https://github.com/Fanqi-Lin/Data-Scaling-Laws | 164 | 0 | 2 | 0 |
Amortized Control of Continuous State Space Feynman-Kac Model for Irregular Time Series | https://openreview.net/forum?id=8zJRon6k5v | [
"Byoungwoo Park",
"Hyungi Lee",
"Juho Lee"
] | Oral | Many real-world datasets, such as healthcare, climate, and economics, are often collected as irregular time series, which poses challenges for accurate modeling. In this paper, we propose the Amortized Control of continuous State Space Model (ACSSM) for continuous dynamical modeling of time series for irregular and discrete observations. We first present a multi-marginal Doob's $h$-transform to construct a continuous dynamical system conditioned on these irregular observations. Following this, we introduce a variational inference algorithm with a tight evidence lower bound (ELBO), leveraging stochastic optimal control (SOC) theory to approximate the intractable Doob's $h$-transform and simulate the conditioned dynamics. To improve efficiency and scalability during both training and inference, ACSSM leverages auxiliary variable to flexibly parameterize the latent dynamics and amortized control. Additionally, it incorporates a simulation-free latent dynamics framework and a transformer-based data assimilation scheme, facilitating parallel inference of the latent states and ELBO computation. Through empirical evaluations across a variety of real-world datasets, ACSSM demonstrates superior performance in tasks such as classification, regression, interpolation, and extrapolation, while maintaining computational efficiency. | stochastic optimal control, variational inference, state space model, irregular time series | We propose a multi-marginal Doob's $h$-transform for irregular time series and variational inference with stochastic optimal control to approximate it. | 6,305 | 2410.05602 | [
-0.10111209750175476,
-0.10954537242650986,
0.0107766417786479,
0.049126312136650085,
-0.0017677765572443604,
0.04285941645503044,
-0.043666135519742966,
0.0023054201155900955,
0.09606486558914185,
0.01269396860152483,
-0.02862129919230938,
0.015205706469714642,
-0.01661689579486847,
-0.002269667573273182,
-0.02765653282403946,
-0.007977214641869068,
0.04818848893046379,
0.03167571872472763,
-0.061028748750686646,
-0.1037953644990921,
0.028203733265399933,
0.07343175262212753,
-0.018900757655501366,
-0.003265355248004198,
0.013306739740073681,
0.02371770143508911,
0.00688069686293602,
-0.02359584905207157,
-0.04243723303079605,
0.007843976840376854,
0.06653913110494614,
0.005597732029855251,
-0.027779042720794678,
-0.10673360526561737,
0.0013852108968421817,
-0.01669228821992874,
-0.0769323781132698,
-0.04370560497045517,
-0.14115723967552185,
0.021975746378302574,
0.07247663289308548,
-0.01628836989402771,
-0.03840935602784157,
0.0461251437664032,
0.010797610506415367,
-0.059449438005685806,
0.023892587050795555,
-0.04900633916258812,
0.054430000483989716,
0.05513238534331322,
0.0814795047044754,
0.043104369193315506,
-0.0031869395170360804,
0.011321804486215115,
0.028409967198967934,
-0.010099470615386963,
0.0749015212059021,
0.017677713185548782,
0.023613959550857544,
-0.05339525640010834,
-0.02206519804894924,
-0.031187988817691803,
0.010401980020105839,
-0.013915767893195152,
-0.04448515921831131,
0.0345199890434742,
-0.0018011641222983599,
0.08550897240638733,
-0.0035907959099859,
0.021835187450051308,
-0.07752083241939545,
0.023450110107660294,
-0.05471384525299072,
-0.004304361995309591,
0.0006074415287002921,
-0.13007614016532898,
0.10945767164230347,
-0.030732668936252594,
0.13009727001190186,
-0.04549356549978256,
-0.0035423184745013714,
0.027411971241235733,
0.04594472050666809,
-0.02873038686811924,
0.03999669849872589,
0.027674736455082893,
-0.033134300261735916,
0.08792386204004288,
0.07448242604732513,
-0.004062339663505554,
0.06762086600065231,
0.02443120628595352,
-0.011868477798998356,
-0.03652619943022728,
0.0024235527962446213,
-0.0062896185554564,
0.005612687207758427,
-0.004887705203145742,
0.02153739519417286,
0.07761970907449722,
-0.006132213864475489,
0.0739409551024437,
0.08063597977161407,
0.06096741929650307,
0.04111453890800476,
-0.14832927286624908,
-0.008122851140797138,
0.06245419755578041,
-0.000875097350217402,
-0.020832622423768044,
0.040657103061676025,
0.023894328624010086,
0.0645020455121994,
0.05916652828454971,
0.022361887618899345,
-0.05035297945141792,
0.04208661988377571,
-0.0791870653629303,
0.0013983247336000204,
0.028721028938889503,
-0.02416146546602249,
-0.037077274173498154,
-0.005741769913583994,
0.06897681951522827,
0.09892092645168304,
0.019197724759578705,
-0.03300845995545387,
1.014799773125569e-33,
0.04434617981314659,
-0.05298155918717384,
0.07173383980989456,
-0.0462694875895977,
-0.03992384672164917,
-0.006004283204674721,
0.031877726316452026,
0.01045996230095625,
-0.026837855577468872,
0.039486855268478394,
-0.005037669092416763,
-0.009504197165369987,
-0.07358402013778687,
0.021902253851294518,
0.04266594350337982,
-0.05246349424123764,
-0.11942194402217865,
0.015277210623025894,
0.051711101084947586,
0.005993369966745377,
0.08206052333116531,
-0.006717534735798836,
-0.0456571988761425,
0.00021984426712151617,
-0.05375286936759949,
-0.04178629070520401,
0.03666013479232788,
0.0632934644818306,
-0.02063417248427868,
0.039052814245224,
-0.093699149787426,
0.04580729082226753,
-0.07979714870452881,
-0.004322768654674292,
-0.0013095574686303735,
-0.01807444728910923,
-0.013312898576259613,
-0.007597392424941063,
-0.012907657772302628,
-0.0630144476890564,
0.0859493613243103,
-0.01688285730779171,
0.013537650927901268,
0.004677778575569391,
-0.07372087240219116,
-0.08891277015209198,
0.07015150785446167,
0.011544082313776016,
0.040680695325136185,
-0.03921550139784813,
-0.03075914829969406,
-0.02564837783575058,
0.03178468719124794,
-0.0806649923324585,
-0.029827166348695755,
0.018158338963985443,
0.013326120562851429,
-0.04091610386967659,
-0.01417434960603714,
0.023756589740514755,
-0.05572652444243431,
0.01117800548672676,
0.0682632252573967,
0.04651603102684021,
0.060785435140132904,
0.042380623519420624,
-0.02404826693236828,
0.07912392169237137,
0.006011048797518015,
-0.024313008412718773,
0.03665757551789284,
-0.006290535908192396,
-0.010917378589510918,
-0.09947848320007324,
0.07228930294513702,
-0.055588241666555405,
0.1160731390118599,
0.016199588775634766,
-0.10766799002885818,
0.024331364780664444,
0.013289541006088257,
-0.04770714044570923,
0.002237859182059765,
0.05219084396958351,
0.008183247409760952,
-0.09084463864564896,
-0.019571760669350624,
0.02379542402923107,
-0.06745927780866623,
-0.09524796903133392,
-0.05533595010638237,
-0.0500175841152668,
0.05773843452334404,
0.04378516599535942,
-0.03529881313443184,
-1.325185474875844e-33,
0.02458661049604416,
0.01763611100614071,
0.04855997860431671,
0.047839611768722534,
0.02968316711485386,
-0.043009888380765915,
-0.009118657559156418,
0.006878075189888477,
0.041805755347013474,
-0.08684606850147247,
-0.03371788561344147,
-0.04070739075541496,
0.02273711748421192,
-0.0163040179759264,
0.05433899909257889,
-0.004683809354901314,
-0.02778821624815464,
-0.016676513478159904,
-0.013885678723454475,
0.016065688803792,
-0.04709363728761673,
-0.0343591570854187,
0.01837073639035225,
-0.02685687504708767,
-0.027193263173103333,
0.016081394627690315,
-0.038249120116233826,
0.11608685553073883,
-0.014802874065935612,
0.009842931292951107,
-0.09081708639860153,
-0.0712541788816452,
-0.05797908455133438,
0.02745928429067135,
-0.14106829464435577,
0.03201470151543617,
0.0389375165104866,
0.026557505130767822,
0.005508623085916042,
0.03059041127562523,
0.03193536773324013,
-0.05008484423160553,
-0.03440948203206062,
0.0072155361995100975,
0.05135006830096245,
-0.0014529478503391147,
-0.023411447182297707,
0.010660971514880657,
-0.01193614024668932,
0.012700929306447506,
-0.05068466067314148,
0.0689648762345314,
-0.049940209835767746,
0.07397253066301346,
-0.06475184857845306,
-0.026029547676444054,
0.03996463492512703,
-0.041039519011974335,
0.005190384108573198,
0.006221704185009003,
-0.06262004375457764,
-0.03842110186815262,
0.08604074269533157,
-0.01145960297435522,
-0.02567238360643387,
-0.04420434683561325,
0.04964541271328926,
-0.027420630678534508,
0.047229524701833725,
-0.030753768980503082,
0.024321366101503372,
-0.00596333434805274,
-0.0021289405412971973,
0.05672479793429375,
0.11941598355770111,
-0.08339279890060425,
0.037456996738910675,
-0.025442739948630333,
0.005492794327437878,
-0.03738474100828171,
-0.030587777495384216,
0.007127879187464714,
0.037622030824422836,
0.0034945502411574125,
-0.050592269748449326,
0.05236919969320297,
-0.03635421395301819,
-0.012028551660478115,
0.054036155343055725,
0.015080713666975498,
-0.054580867290496826,
-0.02086227387189865,
0.037424392998218536,
0.037376102060079575,
-0.08656951040029526,
-4.5788148383962834e-8,
-0.04625990614295006,
0.09021524339914322,
0.1463319957256317,
0.02981286682188511,
0.07070469111204147,
-0.057098522782325745,
0.07368729263544083,
-0.04952504113316536,
0.07196927815675735,
-0.007017107214778662,
0.07964253425598145,
0.021990526467561722,
0.07127386331558228,
-0.11851389706134796,
-0.05507199838757515,
0.030336894094944,
0.06954135000705719,
-0.013728533871471882,
0.012948717921972275,
0.00525798462331295,
0.029329946264624596,
0.028202632442116737,
0.012394841760396957,
0.00604165717959404,
0.1222120001912117,
-0.0032166880555450916,
-0.10275588929653168,
0.11934607475996017,
0.10458556562662125,
0.02983497455716133,
-0.010106726549565792,
0.04273808002471924,
0.023084234446287155,
-0.00955817848443985,
-0.06702055782079697,
-0.007343754172325134,
0.0690159872174263,
-0.003339623101055622,
-0.028786351904273033,
-0.013889080844819546,
-0.06040457263588905,
0.06122767925262451,
-0.06084800139069557,
-0.0040766410529613495,
0.003565064165741205,
0.024297958239912987,
0.0005981116555631161,
-0.06559748202562332,
0.09692800790071487,
0.02706695720553398,
0.05805307626724243,
0.008659102022647858,
-0.006440937053412199,
0.052626632153987885,
0.03738247603178024,
-0.007229165639728308,
0.04618228226900101,
-0.0715053379535675,
0.022251877933740616,
0.01809244230389595,
-0.030652178451418877,
0.00827047973871231,
-0.15645717084407806,
-0.008849908597767353
] | https://github.com/bw-park/ACSSM | 7 | 0 | 0 | 0 |
Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates | https://openreview.net/forum?id=syThiTmWWm | [
"Xiaosen Zheng",
"Tianyu Pang",
"Chao Du",
"Qian Liu",
"Jing Jiang",
"Min Lin"
] | Oral | Automatic LLM benchmarks, such as AlpacaEval 2.0, Arena-Hard-Auto, and MT-Bench, have become popular for evaluating language models due to their cost-effectiveness and scalability compared to human evaluation. Achieving high win rates on these benchmarks can significantly boost the promotional impact of newly released language models. This promotional benefit may motivate tricks, such as manipulating model output length or style to game win rates, even though several mechanisms have been developed to control length and disentangle style to reduce gameability. Nonetheless, we show that even a **"null model"** that always outputs a **constant** response (*irrelevant to input instructions*) can cheat automatic benchmarks and achieve top-ranked win rates: an $86.5\\%$ LC win rate on AlpacaEval 2.0; an $83.0$ score on Arena-Hard-Auto; and a $9.55$ score on MT-Bench. Moreover, the crafted cheating outputs are **transferable** because we assume that the instructions of these benchmarks (e.g., $805$ samples of AlpacaEval 2.0) are *private* and cannot be accessed. While our experiments are primarily proof-of-concept, an adversary could use LLMs to generate more imperceptible cheating responses, unethically benefiting from high win rates and promotional impact. Our findings call for the development of anti-cheating mechanisms for reliable automatic benchmarks. The code is available at https://github.com/sail-sg/Cheating-LLM-Benchmarks. | Large Language Models, Cheating, Automatic LLM Benchmarks | We show that null models that always return the same cheating responses can achieve high win rates on automatic LLM benchmarks. | 6,258 | 2410.07137 | [
-0.038747578859329224,
-0.037917591631412506,
-0.029334155842661858,
-0.018462618812918663,
0.016729576513171196,
0.023770535364747047,
-0.023218154907226562,
-0.00014555863162968308,
0.034807849675416946,
0.0510152205824852,
-0.06378713250160217,
-0.10623263567686081,
0.053954292088747025,
-0.046610768884420395,
-0.013636806048452854,
-0.037117850035429,
0.12587179243564606,
0.041845474392175674,
-0.07217278331518173,
-0.05784197524189949,
0.0563729889690876,
-0.058949217200279236,
0.003555887145921588,
-0.01700434274971485,
0.07677488029003143,
-0.014377781189978123,
-0.09301465004682541,
0.0037661390379071236,
0.013897251337766647,
-0.07826995104551315,
-0.08388911187648773,
0.010540444403886795,
0.01942949742078781,
-0.03027951531112194,
-0.017616089433431625,
-0.00765724154189229,
-0.0671129897236824,
-0.004857643507421017,
-0.028314542025327682,
-0.028297705575823784,
0.0005993918748572469,
0.0023493433836847544,
0.019650176167488098,
0.031045639887452126,
-0.015562653541564941,
-0.009030493907630444,
-0.017156245186924934,
0.044104769825935364,
-0.09779069572687149,
-0.02201364003121853,
-0.07719985395669937,
0.006254070438444614,
0.020960628986358643,
-0.009795686230063438,
-0.007961027324199677,
-0.01484418660402298,
0.0904010683298111,
0.06998326629400253,
-0.05216647684574127,
0.016441626474261284,
-0.07424511760473251,
-0.09978150576353073,
-0.04300267994403839,
0.005995211191475391,
0.024917926639318466,
0.0024486344773322344,
0.011566471308469772,
0.01536715216934681,
-0.05027803033590317,
0.05785740539431572,
-0.027407310903072357,
-0.04974216967821121,
-0.08182163536548615,
0.008460156619548798,
0.08077866584062576,
0.04199640452861786,
-0.007931296713650227,
0.00026978302048519254,
0.03398927301168442,
-0.033011917024850845,
-0.033972471952438354,
-0.07811732590198517,
-0.021793657913804054,
-0.03139184042811394,
0.06710780411958694,
-0.011669682338833809,
0.010315920226275921,
0.017935404554009438,
0.15867578983306885,
-0.0031954627484083176,
-0.031860288232564926,
-0.004550866782665253,
0.005458409432321787,
0.017323767766356468,
0.023469945415854454,
0.07087083905935287,
0.04403536766767502,
0.0026010842993855476,
-0.0417536124587059,
0.08180100470781326,
0.0443904809653759,
0.06674893200397491,
0.026009079068899155,
-0.08550916612148285,
0.0767044723033905,
-0.007087256293743849,
0.03376360610127449,
0.06913977861404419,
-0.027675608173012733,
-0.03575318679213524,
0.01769436150789261,
0.11557113379240036,
0.0128514114767313,
0.015358245931565762,
-0.01007536705583334,
-0.010658083483576775,
-0.04830488562583923,
-0.04503030329942703,
-0.0013531934237107635,
0.05460908263921738,
0.03746814653277397,
-0.014666522853076458,
0.04296892136335373,
0.060797590762376785,
0.03625244274735451,
-0.029548514634370804,
-0.011396788991987705,
1.7340253488277126e-33,
0.018191510811448097,
0.0662614032626152,
0.01675855740904808,
-0.02105221338570118,
-0.010857164859771729,
-0.03834669291973114,
0.04151188209652901,
0.06612895429134369,
0.017979159951210022,
0.01043332926928997,
-0.06467951089143753,
0.024740708991885185,
0.0028019926976412535,
0.11251987516880035,
0.1228804960846901,
0.08347506076097488,
0.011430573649704456,
0.02696920931339264,
-0.019440999254584312,
-0.02294493094086647,
0.11385763436555862,
0.0113116055727005,
0.021085066720843315,
-0.1377890557050705,
0.02125503309071064,
0.04807719588279724,
-0.03940992057323456,
-0.07211525738239288,
-0.06777667999267578,
0.0259733647108078,
-0.07244045287370682,
-0.09111417084932327,
0.038663215935230255,
0.007479318883270025,
0.045914679765701294,
0.0053060464560985565,
-0.005094597116112709,
-0.06267323344945908,
0.02551873028278351,
-0.004701611120253801,
-0.06970304995775223,
-0.0030599378515034914,
-0.06482262909412384,
-0.028267009183764458,
-0.024861404672265053,
-0.051025424152612686,
0.0253140851855278,
-0.040752701461315155,
-0.07239025086164474,
0.0035928848665207624,
0.08120090514421463,
0.0562155544757843,
0.02180604077875614,
0.020974580198526382,
-0.021040067076683044,
0.0011309880064800382,
0.04887453094124794,
-0.030020901933312416,
-0.03492814302444458,
0.14363791048526764,
0.011318503879010677,
0.01789553090929985,
0.0066285328939557076,
0.025226157158613205,
-0.047836847603321075,
0.05451760068535805,
-0.04139457270503044,
-0.032402683049440384,
0.05049111694097519,
-0.05401793494820595,
-0.013932516798377037,
-0.03641797974705696,
-0.04754456877708435,
0.024885687977075577,
0.039552852511405945,
-0.04391198977828026,
0.06376390159130096,
-0.09714797884225845,
0.08405882865190506,
-0.09537750482559204,
0.0664021372795105,
0.012596039101481438,
-0.05677244812250137,
-0.07195088267326355,
-0.02345326356589794,
-0.016224052757024765,
0.029629649594426155,
-0.048155877739191055,
-0.017506614327430725,
0.006165859289467335,
-0.02718949504196644,
-0.025421468541026115,
-0.04478812590241432,
-0.010254832915961742,
0.056724559515714645,
-2.9477594492372107e-33,
-0.05732915177941322,
-0.00924067385494709,
-0.02287420444190502,
0.13123762607574463,
-0.02713290974497795,
-0.057657405734062195,
0.005441057495772839,
-0.00618613138794899,
-0.00407404126599431,
0.02267109788954258,
-0.06969133019447327,
0.0022634032648056746,
0.012234959751367569,
0.0345301553606987,
0.06674958020448685,
-0.10212834179401398,
0.02813156507909298,
0.010410673916339874,
0.028536271303892136,
0.01404193602502346,
0.10280138999223709,
0.10520868748426437,
-0.06314349919557571,
-0.004221805837005377,
-0.059499673545360565,
0.019099757075309753,
-0.11606167256832123,
0.02836023084819317,
0.018901003524661064,
0.05858045443892479,
0.02111184224486351,
0.04040535166859627,
-0.0379195362329483,
-0.026462502777576447,
-0.02379569225013256,
0.029052715748548508,
0.08575735241174698,
0.06786655634641647,
-0.023528708145022392,
0.11734862625598907,
0.045796383172273636,
-0.030128512531518936,
-0.08769813925027847,
0.008734147995710373,
0.022706247866153717,
0.02604062482714653,
-0.1165711060166359,
-0.06996612250804901,
0.028386466205120087,
-0.044724240899086,
0.0664038434624672,
-0.014712653122842312,
-0.00787714496254921,
0.061139583587646484,
-0.11870714277029037,
-0.053294792771339417,
-0.07068049907684326,
-0.023564444854855537,
0.036783576011657715,
0.03622349724173546,
-0.04778647422790527,
0.022383825853466988,
-0.09710035473108292,
-0.05403056740760803,
0.07591360807418823,
0.002830695593729615,
-0.01911773532629013,
-0.02109496295452118,
-0.007792316377162933,
-0.02653435431420803,
0.0249622855335474,
0.004600119311362505,
-0.04788043722510338,
0.007510565221309662,
-0.00835737306624651,
0.09333616495132446,
-0.03511311486363411,
-0.02415580302476883,
0.038730744272470474,
-0.055523745715618134,
-0.020001282915472984,
0.011088116094470024,
0.09329479932785034,
-0.0028953207656741142,
-0.01789913885295391,
0.02347569540143013,
0.030345916748046875,
0.02696223556995392,
-0.051160164177417755,
-0.01792849786579609,
-0.024945341050624847,
0.048290327191352844,
0.0012823812430724502,
0.08848728984594345,
-0.03948912024497986,
-5.446233686257074e-8,
-0.005326895974576473,
0.010028812102973461,
0.024427572265267372,
0.11380585283041,
0.0005883618723601103,
0.026377791538834572,
-0.033958226442337036,
0.012511326931416988,
0.033016253262758255,
-0.010305969975888729,
0.05062336102128029,
-0.009345918893814087,
-0.057298582047224045,
-0.010779104195535183,
0.01604466512799263,
0.08229599893093109,
-0.08877596259117126,
0.037992823868989944,
-0.07971973717212677,
-0.026788175106048584,
0.048554714769124985,
0.10271517187356949,
-0.06257164478302002,
-0.06385156512260437,
-0.009647039696574211,
-0.04515465721487999,
0.032431505620479584,
0.1234850063920021,
0.0890703946352005,
0.034203920513391495,
0.026819782331585884,
0.0013064208906143904,
0.03931132331490517,
0.010712839663028717,
0.03548534959554672,
0.06337455660104752,
-0.0195816308259964,
0.013265415094792843,
0.011112933978438377,
0.018095921725034714,
-0.03165491297841072,
0.0140923410654068,
-0.1414228081703186,
-0.005304246675223112,
0.06448602676391602,
-0.027692211791872978,
-0.11400463432073593,
-0.12037460505962372,
-0.02156161330640316,
-0.07940356433391571,
0.035773538053035736,
-0.02558644860982895,
-0.07718013226985931,
0.02681637741625309,
0.018225599080324173,
-0.008643052540719509,
-0.015897370874881744,
0.00042564349132589996,
0.0005506803863681853,
0.05447292700409889,
0.06480752676725388,
-0.03027886152267456,
-0.025818796828389168,
-0.017137110233306885
] | https://github.com/sail-sg/Cheating-LLM-Benchmarks | 77 | 0 | 0 | 0 |
On the Hölder Stability of Multiset and Graph Neural Networks | https://openreview.net/forum?id=P7KIGdgW8S | [
"Yair Davidson",
"Nadav Dym"
] | Oral | Extensive research efforts have been put into characterizing and constructing maximally separating multiset and graph neural networks.
However, recent empirical evidence suggests the notion of separation itself doesn't capture several interesting phenomena. On the one hand, the quality of this separation may be very weak, to the extent that the embeddings of "separable" objects might even be considered identical when using fixed finite precision. On the other hand, architectures which aren't capable of separation in theory, somehow achieve separation when taking the network to be wide enough.
In this work, we address both of these issues, by proposing a novel pair-wise separation quality analysis framework which is based on an adaptation of Lipschitz and Hölder stability to parametric functions. The proposed framework, which we name Hölder in expectation, allows for separation quality analysis, without restricting the analysis to embeddings that can separate all the input space simultaneously. We prove that common sum-based models are lower-Hölder in expectation, with an exponent
that decays rapidly with the network's depth . Our analysis leads to adversarial examples of graphs which can be separated by three 1-WL iterations, but cannot be separated in practice by standard maximally powerful Message Passing Neural Networks (MPNNs). To remedy this, we propose two novel MPNNs with improved separation quality, one of which is lower Lipschitz in expectation. We show these MPNNs can easily classify our adversarial examples, and compare favorably with standard MPNNs on standard graph learning tasks. | graph neural networks, message passing neural networks, multiset neural networks, neural network stability, expressive power, WL tests | null | 5,998 | null | [
-0.09509112685918808,
-0.0471348911523819,
-0.025060242041945457,
-0.028499197214841843,
0.03989922255277634,
-0.03512914106249809,
-0.012623591348528862,
-0.07031811028718948,
0.009821532294154167,
-0.03461599722504616,
-0.010236775502562523,
0.03542213514447212,
0.0538376048207283,
0.020110903307795525,
-0.029770437628030777,
0.03901907056570053,
0.047080740332603455,
0.05692673102021217,
-0.11705933511257172,
0.02885645627975464,
0.01135606225579977,
0.003989634569734335,
-0.03105422481894493,
0.013762962073087692,
0.0835600271821022,
0.00021737579663749784,
-0.0602700300514698,
-0.016631076112389565,
0.05479016155004501,
0.03286642208695412,
0.04582243040204048,
-0.10547713190317154,
-0.0075716194696724415,
0.0056041451171040535,
-0.05920998752117157,
0.061403509229421616,
-0.027633432298898697,
-0.010135217569768429,
0.028452467173337936,
0.02768261544406414,
-0.03022286295890808,
0.10537648946046829,
0.010916492901742458,
-0.005898874718695879,
0.08689837902784348,
0.07407285273075104,
-0.08455066382884979,
-0.01240348070859909,
-0.005981213413178921,
-0.0011740365298464894,
-0.025632046163082123,
0.044438742101192474,
-0.06243877112865448,
0.04075491055846214,
0.0187684353441,
-0.030289841815829277,
-0.013564346358180046,
0.023001808673143387,
-0.0860813558101654,
0.025460660457611084,
-0.002534382976591587,
-0.02107962965965271,
0.02890569344162941,
-0.006999787408858538,
0.04949973151087761,
0.012831220403313637,
0.034914012998342514,
0.07125677168369293,
-0.008664839901030064,
0.14817893505096436,
0.003816821612417698,
0.056129321455955505,
-0.124881312251091,
0.020009083673357964,
-0.04099135473370552,
0.12647463381290436,
0.07896993309259415,
0.016772840172052383,
-0.01126742921769619,
-0.04369509592652321,
0.05893378704786301,
-0.05890079587697983,
0.025994913652539253,
-0.01695728302001953,
0.051654964685440063,
-0.011334989219903946,
-0.030629485845565796,
-0.009487594477832317,
0.025359002873301506,
0.0479096956551075,
-0.11574956029653549,
0.09428196400403976,
0.018196340650320053,
0.006428161170333624,
0.06913842260837555,
-0.04251110926270485,
0.03285043314099312,
0.004725201521068811,
-0.027884557843208313,
0.16443388164043427,
-0.00956802535802126,
0.0016251979395747185,
0.0250773373991251,
-0.04940151795744896,
0.028560571372509003,
0.006872877478599548,
0.03675103560090065,
-0.0029189593624323606,
0.01720726676285267,
-0.07593692094087601,
0.022659122943878174,
0.041409555822610855,
-0.04919050261378288,
0.06691481918096542,
0.022955048829317093,
-0.07257745414972305,
0.037015680223703384,
0.01244125422090292,
0.10887559503316879,
0.0234608743339777,
-0.07405119389295578,
-0.03215762972831726,
-0.0003169738338328898,
-0.03798431530594826,
0.04381466284394264,
-0.02198847383260727,
0.03511444851756096,
1.682777284880715e-33,
-0.011507507413625717,
0.0007999618537724018,
0.00834685005247593,
0.00033259097835980356,
0.10353536903858185,
0.07044389843940735,
-0.03352333977818489,
0.013980554416775703,
-0.02211374044418335,
0.06350673735141754,
-0.06739525496959686,
0.00006888795178383589,
-0.04400010406970978,
0.03151937946677208,
0.07346388697624207,
0.03136802464723587,
0.03405070677399635,
-0.007974796928465366,
0.00837822537869215,
-0.10302772372961044,
0.09330815821886063,
-0.0832672044634819,
0.001530561363324523,
0.025098759680986404,
-0.02867087349295616,
-0.02293899469077587,
0.017003042623400688,
0.026936590671539307,
-0.04670848324894905,
-0.0023149321787059307,
-0.11656398326158524,
0.05552377179265022,
0.019181707873940468,
0.03202293440699577,
0.05623387545347214,
-0.024709459394216537,
-0.059583213180303574,
-0.02638917602598667,
0.010403059422969818,
-0.05234580859541893,
0.0038967113941907883,
-0.007638437207788229,
-0.01714462600648403,
-0.026970945298671722,
-0.061555299907922745,
-0.05994517356157303,
-0.059850435703992844,
0.034209877252578735,
-0.016068408265709877,
-0.10769692063331604,
-0.041519373655319214,
0.0009171036654151976,
0.026605932042002678,
-0.04562748968601227,
-0.050511930137872696,
-0.07750100642442703,
0.06378362327814102,
0.10424051433801651,
0.020610617473721504,
0.022036870941519737,
-0.08110535889863968,
0.05082171782851219,
0.008277337066829205,
0.022966131567955017,
0.020752279087901115,
0.11548088490962982,
-0.11129393428564072,
0.04134301096200943,
0.07786067575216293,
-0.0696457251906395,
-0.015874318778514862,
0.07776359468698502,
-0.0713251382112503,
-0.0836174413561821,
-0.029458751901984215,
-0.0228683240711689,
-0.0007708063931204379,
-0.06020199880003929,
-0.012734541669487953,
0.06735284626483917,
-0.003598859068006277,
-0.03194944187998772,
0.04216601327061653,
-0.02443377673625946,
-0.08667489886283875,
-0.02258503995835781,
0.06468439102172852,
0.012541016563773155,
0.06508477032184601,
-0.012382116168737411,
-0.08715657889842987,
0.031125346198678017,
-0.0374627560377121,
-0.03186634182929993,
0.026297228410840034,
-2.5230529729431336e-33,
-0.04894230142235756,
0.10727991163730621,
0.027242431417107582,
0.059839364141225815,
-0.0004877839528489858,
-0.017405059188604355,
-0.045378267765045166,
-0.005281626246869564,
-0.00412846589460969,
0.013766282238066196,
-0.01687724143266678,
-0.06794272363185883,
0.034814443439245224,
-0.13605627417564392,
0.02705340087413788,
-0.032399170100688934,
-0.07810092717409134,
-0.03980501368641853,
0.018275123089551926,
-0.004541107919067144,
0.05034345015883446,
-0.03748965635895729,
-0.020883483812212944,
0.0951104462146759,
0.022341055795550346,
0.02298186533153057,
-0.1107465997338295,
-0.0011464741546660662,
-0.026358671486377716,
0.03622090443968773,
-0.03395524621009827,
-0.09177112579345703,
-0.02213534712791443,
-0.06882155686616898,
-0.014340627007186413,
-0.021976033225655556,
0.013993462547659874,
0.02793445810675621,
-0.0028800189029425383,
0.08325450867414474,
-0.03499748185276985,
-0.019030580297112465,
-0.10450603067874908,
0.00766880763694644,
0.04715641215443611,
0.045129597187042236,
-0.0029809705447405577,
-0.010020137764513493,
-0.07229918241500854,
-0.003302267286926508,
-0.005872211419045925,
0.03491406887769699,
-0.042922377586364746,
0.041301459074020386,
-0.019697492942214012,
-0.04967426508665085,
0.013243140652775764,
0.03547076880931854,
0.08546280860900879,
-0.0404401458799839,
-0.03424282371997833,
-0.036517780274152756,
-0.119843028485775,
0.048795539885759354,
0.042680419981479645,
-0.027226794511079788,
0.04695574566721916,
0.022503186017274857,
0.061249278485774994,
0.022950153797864914,
0.013376077637076378,
-0.017244527116417885,
0.022900305688381195,
-0.003586574923247099,
-0.016055162996053696,
-0.008722408674657345,
0.04351722076535225,
-0.06093668192625046,
-0.0010730944341048598,
0.02228333242237568,
0.03914957866072655,
-0.006080956198275089,
0.10346993803977966,
-0.014501946046948433,
0.023182816803455353,
0.032737307250499725,
-0.0024237483739852905,
-0.011100280098617077,
-0.026073334738612175,
0.009526336565613747,
0.02627643011510372,
0.07807372510433197,
-0.046552665531635284,
-0.012673183344304562,
-0.02387230657041073,
-4.65342075983699e-8,
-0.0532071553170681,
0.037348970770835876,
-0.0033902483992278576,
-0.011504877358675003,
0.02416364476084709,
-0.1014765128493309,
0.057846326380968094,
0.09299857169389725,
0.04312949255108833,
0.03576043248176575,
0.048303183168172836,
-0.009853135794401169,
-0.0698159709572792,
-0.08796718716621399,
-0.04438520222902298,
-0.027251768857240677,
-0.05128798261284828,
0.05437391623854637,
0.02951233647763729,
0.01626397855579853,
0.0046718004159629345,
-0.05039513111114502,
0.08135035634040833,
-0.02741045132279396,
0.06573591381311417,
-0.08356839418411255,
-0.07528530806303024,
0.008361541666090488,
0.0008291457197628915,
0.04052232950925827,
-0.004643810447305441,
0.03026527166366577,
0.13452304899692535,
-0.021295856684446335,
0.002058684825897217,
0.16418100893497467,
-0.015136889182031155,
0.06376012414693832,
-0.04754985496401787,
-0.02127939648926258,
-0.0606175996363163,
0.10880506783723831,
-0.03875244781374931,
0.0028418987058103085,
0.02016516588628292,
-0.06008465215563774,
0.08719713985919952,
-0.01452531386166811,
0.045744672417640686,
-0.011311201378703117,
0.030014408752322197,
0.01692177541553974,
-0.017328571528196335,
-0.028470007702708244,
0.01276844646781683,
-0.037964217364788055,
-0.024423331022262573,
-0.10033036023378372,
0.03437722474336624,
0.02077828161418438,
-0.007572326809167862,
-0.034361861646175385,
-0.03623088821768761,
-0.05954442173242569
] | 0 | 0 | 0 | 0 |
|
On Conformal Isometry of Grid Cells: Learning Distance-Preserving Position Embedding | https://openreview.net/forum?id=Xo0Q1N7CGk | [
"Dehong Xu",
"Ruiqi Gao",
"Wenhao Zhang",
"Xue-Xin Wei",
"Ying Nian Wu"
] | Oral | This paper investigates the conformal isometry hypothesis as a potential explanation for the hexagonal periodic patterns in grid cell response maps. We posit that grid cell activities form a high-dimensional vector in neural space, encoding the agent's position in 2D physical space. As the agent moves, this vector rotates within a 2D manifold in the neural space, driven by a recurrent neural network. The conformal hypothesis proposes that this neural manifold is a conformal isometric embedding of 2D physical space, where local physical distance is preserved by the embedding up to a scaling factor (or unit of metric). Such distance-preserving position embedding is indispensable for path planning in navigation, especially planning local straight path segments. We conduct numerical experiments to show that this hypothesis leads to the hexagonal grid firing patterns by learning maximally distance-preserving position embedding, agnostic to the choice of the recurrent neural network. Furthermore, we present a theoretical explanation of why hexagon periodic patterns emerge by minimizing our loss function by showing that hexagon flat torus is maximally distance preserving. | grid cells, conformal isometry, distance-preserving, position embedding, representation learning | We investigate the conformal isometry hypothesis that leads to the emergence of hexagon periodic patterns in grid cells, showing that learning a maximally distance-preserving position embedding naturally leads to these patterns. | 5,957 | 2405.16865 | [
-0.012160265818238258,
-0.07754725962877274,
-0.021062705665826797,
-0.048932790756225586,
-0.07882820814847946,
0.05492495745420456,
-0.07367288321256638,
-0.03809693828225136,
0.07750459015369415,
0.020525509491562843,
0.011146326549351215,
0.05598616600036621,
-0.012682903558015823,
0.03136654198169708,
-0.03427519276738167,
-0.04517735168337822,
0.019898900762200356,
0.10219266265630722,
-0.051058921962976456,
-0.022185303270816803,
0.04416031017899513,
-0.06095627322793007,
0.029000407084822655,
-0.027271663770079613,
-0.031181706115603447,
0.053654491901397705,
0.04136865958571434,
0.009644118137657642,
0.031191756948828697,
-0.018466489389538765,
0.09270845353603363,
0.03704637661576271,
-0.06525639444589615,
-0.04047965630888939,
-0.06608029454946518,
0.09471353143453598,
-0.017057130113244057,
0.025989947840571404,
-0.029346058145165443,
-0.009820766746997833,
0.08815551549196243,
0.07227484881877899,
0.02375635877251625,
0.03730713948607445,
0.03281150013208389,
0.005140659864991903,
0.019799256697297096,
-0.06317023187875748,
-0.00048628399963490665,
-0.013006126508116722,
-0.08034970611333847,
-0.06567413359880447,
0.03175602853298187,
-0.026003079488873482,
0.023207837715744972,
-0.017198456451296806,
0.038995299488306046,
0.04683564230799675,
0.06115500256419182,
-0.09213226288557053,
0.09505950659513474,
-0.09493044018745422,
-0.0015415434027090669,
0.03559684753417969,
-0.06777957081794739,
-0.03702547028660774,
-0.006700852885842323,
-0.008085436187684536,
-0.02456420101225376,
0.03899059072136879,
-0.021042179316282272,
0.04167194664478302,
-0.01689877174794674,
-0.016334231942892075,
0.07370856404304504,
0.008804059587419033,
-0.044619638472795486,
-0.028314920142292976,
0.02856866642832756,
-0.016060737892985344,
-0.009414107538759708,
0.056656401604413986,
0.061948198825120926,
0.11109177023172379,
0.019593428820371628,
0.03623536229133606,
0.0060166059993207455,
0.02848159894347191,
0.046706877648830414,
-0.02399151213467121,
0.04915335774421692,
0.03700142353773117,
-0.007306348066776991,
-0.07544374465942383,
-0.035743869841098785,
-0.010291154496371746,
-0.060097258538007736,
0.0073105525225400925,
0.015988007187843323,
0.002420059870928526,
0.05919262021780014,
-0.01309569925069809,
-0.02422601729631424,
0.09211676567792892,
0.10819761455059052,
0.04253237694501877,
0.041620079427957535,
-0.025321291759610176,
0.033891431987285614,
0.02045118436217308,
-0.030127454549074173,
-0.04622328281402588,
-0.04572967812418938,
0.06434313207864761,
0.015964142978191376,
-0.08752428740262985,
0.04689812287688255,
0.008441317826509476,
0.0005477133090607822,
0.08421070873737335,
-0.01841604709625244,
-0.04923398792743683,
-0.05712786689400673,
0.026723092421889305,
0.04476616904139519,
-0.025459321215748787,
-0.04136781021952629,
2.7151869276543114e-34,
-0.03304881602525711,
0.0882570818066597,
0.016240468248724937,
-0.029581798240542412,
0.09744269400835037,
-0.03278368338942528,
0.004880142398178577,
0.03372865915298462,
0.11346834152936935,
0.04749958589673042,
-0.12180634588003159,
0.06715380400419235,
-0.016043992713093758,
0.08777575194835663,
0.03270535543560982,
-0.05126045644283295,
0.03630100190639496,
-0.055781491100788116,
-0.07701313495635986,
-0.05179501697421074,
0.023883868008852005,
-0.02622307650744915,
0.04999084025621414,
-0.11312543600797653,
-0.028997110202908516,
-0.040898654609918594,
-0.03668783977627754,
-0.003509462345391512,
-0.06650662422180176,
-0.007825694046914577,
-0.05907170847058296,
0.010642671957612038,
-0.08094632625579834,
0.04100785776972771,
0.020159127190709114,
0.01698712445795536,
-0.013194669969379902,
-0.00022860485478304327,
0.01302067469805479,
0.003347499296069145,
0.03856377303600311,
-0.02836574614048004,
-0.024876758456230164,
0.005852801259607077,
0.043565765023231506,
-0.062448371201753616,
0.07502515614032745,
0.059598784893751144,
-0.041552186012268066,
-0.08509334921836853,
-0.010044518858194351,
0.07250592112541199,
-0.017041338607668877,
-0.16957303881645203,
-0.01107004377990961,
-0.00022192997857928276,
-0.001095513696782291,
-0.020400401204824448,
0.0264748502522707,
0.08670902997255325,
0.014112645760178566,
0.010119847021996975,
0.05896386131644249,
0.09280329197645187,
0.014544106088578701,
-0.08019185811281204,
-0.0173452440649271,
0.0257431510835886,
0.07038922607898712,
0.006808432284742594,
0.07087638974189758,
0.05345460772514343,
-0.05069341883063316,
-0.02655436098575592,
0.04895191639661789,
-0.027238372713327408,
-0.023705000057816505,
-0.07036843150854111,
-0.049383603036403656,
-0.0018206631066277623,
-0.08008188754320145,
-0.048541631549596786,
-0.12163941562175751,
-0.04115630313754082,
-0.019828010350465775,
-0.04626400023698807,
0.06321825832128525,
-0.036304496228694916,
-0.03932258114218712,
-0.009367885068058968,
-0.0005538510158658028,
-0.10368813574314117,
-0.021148543804883957,
0.023464161902666092,
0.010164921171963215,
-3.3193947433145616e-33,
-0.03966245800256729,
0.0007309913635253906,
-0.001837987801991403,
-0.017714358866214752,
-0.04646215960383415,
0.029432548210024834,
-0.03460688889026642,
0.0701291561126709,
-0.05961837247014046,
0.05600247532129288,
-0.0431390106678009,
0.07469702512025833,
0.06398586928844452,
-0.008185653015971184,
-0.0014749903930351138,
-0.018456069752573967,
0.05553776025772095,
-0.008177471347153187,
-0.015995685011148453,
-0.07920042425394058,
-0.0028399371076375246,
0.02126551792025566,
-0.07324789464473724,
0.05675749108195305,
-0.009899025782942772,
0.0478251688182354,
0.05885279178619385,
0.011117618530988693,
-0.0065046558156609535,
0.11679413169622421,
-0.07369941473007202,
-0.02535942755639553,
-0.003048777347430587,
0.012219812721014023,
0.008964551612734795,
0.04755130037665367,
-0.008327726274728775,
0.009872683323919773,
-0.03312026336789131,
-0.05021604523062706,
-0.015384454280138016,
-0.04938892275094986,
0.03150973841547966,
0.002101735444739461,
0.050068747252225876,
-0.042023368179798126,
-0.013826148584485054,
0.06640596687793732,
-0.08307072520256042,
0.04607497528195381,
-0.04119228571653366,
0.07539120316505432,
-0.08088207989931107,
-0.013736488297581673,
0.0038750520907342434,
0.08237617462873459,
-0.12258561700582504,
0.00008803548553260043,
0.09520915150642395,
-0.04198707267642021,
-0.05166226997971535,
-0.020117653533816338,
-0.013047176413238049,
0.05632218345999718,
0.014156103134155273,
0.05926842242479324,
-0.024980107322335243,
0.0006976405857130885,
0.021214136853814125,
0.04285353422164917,
0.002439960604533553,
0.04439215734601021,
-0.02215600572526455,
-0.022149039432406425,
-0.02811557613313198,
-0.02712070383131504,
0.028611136600375175,
-0.0571976974606514,
-0.07373998314142227,
0.01102283876389265,
0.01052857469767332,
-0.031037498265504837,
-0.019913479685783386,
0.018609318882226944,
-0.01326022483408451,
-0.01252945140004158,
0.012419031001627445,
0.00746558653190732,
0.03568354994058609,
0.07523418217897415,
0.15409372746944427,
-0.02267436496913433,
-0.07959666103124619,
0.013069004751741886,
-0.015368085354566574,
-4.131452158162574e-8,
-0.11158674210309982,
0.04815279319882393,
0.12021519988775253,
-0.03048064187169075,
-0.011350814253091812,
-0.06684904545545578,
0.09451915323734283,
-0.049863606691360474,
-0.04570654034614563,
0.012340930290520191,
0.08038779348134995,
-0.014493423514068127,
-0.046246081590652466,
0.013452555052936077,
-0.013091996312141418,
0.08020707964897156,
0.013998540118336678,
-0.024492360651493073,
0.029700083658099174,
0.03407059982419014,
-0.07868823409080505,
0.00499781034886837,
0.01573389396071434,
0.030346570536494255,
0.0197727270424366,
-0.07942632585763931,
-0.14989951252937317,
-0.013444200158119202,
0.07035701721906662,
-0.03534071519970894,
0.033045265823602676,
0.03466617316007614,
0.09827909618616104,
0.02297281287610531,
-0.09362974762916565,
0.01148893591016531,
0.024826155975461006,
-0.04634964093565941,
-0.0007605021819472313,
-0.051231175661087036,
-0.02075856365263462,
0.033478349447250366,
0.07478628307580948,
-0.021054046228528023,
-0.010514948517084122,
0.04040832072496414,
0.08995156735181808,
-0.027554789558053017,
0.03865111619234085,
-0.053392622619867325,
-0.04554947838187218,
-0.004766879137605429,
-0.020913945510983467,
0.0020281109027564526,
0.05544762685894966,
0.03255346417427063,
-0.05279729515314102,
-0.13714107871055603,
0.023004520684480667,
0.015786731615662575,
-0.07061276584863663,
0.05810729041695595,
-0.07252403348684311,
-0.06349801272153854
] | 0 | 0 | 0 | 0 |
|
Combatting Dimensional Collapse in LLM Pre-Training Data via Submodular File Selection | https://openreview.net/forum?id=f4gF6AIHRy | [
"Ziqing Fan",
"Siyuan Du",
"Shengchao Hu",
"Pingjie Wang",
"Li Shen",
"Ya Zhang",
"Dacheng Tao",
"Yanfeng Wang"
] | Oral | Selecting high-quality pre-training data for large language models (LLMs) is crucial for enhancing their overall performance under limited computation budget, improving both training and sample efficiency. Recent advancements in file selection primarily rely on using an existing or trained proxy model to assess the similarity of samples to a target domain, such as high quality sources BookCorpus and Wikipedia. However, upon revisiting these methods, the domain-similarity selection criteria demonstrates a diversity dilemma, i.e. dimensional collapse in the feature space, improving performance on the domain-related tasks but causing severe degradation on generic performance.To prevent collapse and enhance diversity, we propose a DiverSified File selection algorithm (DiSF), which selects the most decorrelated text files in the feature space. We approach this with a classical greedy algorithm to achieve more uniform eigenvalues in the feature covariance matrix of the selected texts, analyzing its approximation to the optimal solution under a formulation of $\gamma$-weakly submodular optimization problem. Empirically, we establish a benchmark and conduct extensive experiments on the TinyLlama architecture with models from 120M to 1.1B parameters. Evaluating across nine tasks from the Harness framework, DiSF demonstrates a significant improvement on overall performance. Specifically, DiSF saves 98.5\% of 590M training files in SlimPajama, outperforming the full-data pre-training within a 50B training budget, and achieving about 1.5x training efficiency and 5x data efficiency. Source code
is available at: https://github.com/MediaBrain-SJTU/DiSF.git. | file selection, large language model, pre-training, submodular optimization | null | 5,918 | null | [
-0.03265996277332306,
-0.06114211678504944,
0.06435337662696838,
0.04634328931570053,
0.08448476344347,
0.016013801097869873,
-0.038751427084207535,
0.014552833512425423,
0.03462684527039528,
-0.032777708023786545,
-0.011237070895731449,
0.04454141482710838,
-0.00440253596752882,
0.04047925025224686,
-0.03828361630439758,
0.05702657997608185,
0.06356729567050934,
0.03486412763595581,
-0.087900809943676,
-0.08688376843929291,
0.039194948971271515,
0.0895942747592926,
0.02478691376745701,
0.011919256299734116,
0.07040362805128098,
-0.008904922753572464,
0.01781282387673855,
0.01798401027917862,
0.0007541004451923072,
0.001718628336675465,
0.016623064875602722,
0.07615906745195389,
-0.014616108499467373,
-0.07503212243318558,
-0.00827072560787201,
0.06776054203510284,
-0.046994373202323914,
-0.005633666645735502,
0.05959447845816612,
0.026790671050548553,
0.03346412628889084,
0.030861608684062958,
-0.012893879786133766,
0.027734210714697838,
0.02614171989262104,
0.005444501526653767,
-0.02027137391269207,
-0.04880501702427864,
0.0009240815415978432,
0.006607623305171728,
-0.04977833852171898,
0.04506590589880943,
-0.04702221229672432,
0.09821197390556335,
-0.005466376431286335,
-0.10465125739574432,
-0.00930539146065712,
-0.0004451751592569053,
-0.03345189243555069,
0.022652694955468178,
-0.03939047455787659,
-0.12645936012268066,
-0.05476151779294014,
-0.040183283388614655,
0.07332000881433487,
0.019596906378865242,
0.05356065556406975,
0.04541146755218506,
-0.030111096799373627,
-0.03146832436323166,
-0.04446931928396225,
0.1252729892730713,
-0.0550173744559288,
0.17795558273792267,
-0.01738852635025978,
0.02577948570251465,
0.001665433868765831,
0.006745941936969757,
0.0679916962981224,
-0.06431588530540466,
0.025035766884684563,
0.024768909439444542,
0.0533703938126564,
-0.06405042111873627,
0.07486157864332199,
0.02297046035528183,
0.030744854360818863,
-0.017656216397881508,
0.07270509749650955,
-0.002520452719181776,
-0.047770995646715164,
-0.008935186080634594,
0.08464136719703674,
0.030906599014997482,
-0.07663468271493912,
0.03204894810914993,
0.017185432836413383,
0.0165774654597044,
0.04193517938256264,
0.07252935320138931,
0.054726067930459976,
0.05125436186790466,
0.06391320377588272,
0.027467437088489532,
-0.003177366917952895,
-0.06669150292873383,
0.08679711073637009,
0.04341718927025795,
0.047678373754024506,
-0.08686253428459167,
0.05769093707203865,
0.027614543214440346,
-0.02382083423435688,
0.060556989163160324,
0.010195557028055191,
0.0360996313393116,
0.08052346110343933,
0.042141471058130264,
-0.0002553416707087308,
0.11157842725515366,
-0.044730495661497116,
-0.024598723277449608,
-0.010528471320867538,
-0.04855562746524811,
-0.0023398087359964848,
0.0015539595624431968,
-0.09731478989124298,
3.381461212379963e-33,
0.07183126360177994,
0.04460505396127701,
0.009027590043842793,
0.050725214183330536,
0.02417239546775818,
-0.07942983508110046,
0.010060327127575874,
0.010117138735949993,
-0.07613840699195862,
0.004467248450964689,
-0.08604910224676132,
0.01207574550062418,
-0.048361457884311676,
0.01470034196972847,
0.03720898926258087,
-0.01912647858262062,
-0.06677170097827911,
0.09368551522493362,
-0.032788004726171494,
-0.06384950131177902,
0.05538668856024742,
0.05206257849931717,
0.03299875929951668,
-0.07086575776338577,
-0.001275978167541325,
-0.03289243206381798,
0.06138767674565315,
-0.10828817635774612,
-0.10069140791893005,
0.012708516791462898,
-0.12041985243558884,
-0.03034006804227829,
0.04288743808865547,
0.06135805323719978,
0.02017921581864357,
0.007975694723427296,
-0.04587977007031441,
-0.05409717932343483,
0.03603440523147583,
-0.01370815560221672,
-0.01909538172185421,
0.04774155095219612,
0.010649135336279869,
-0.06106606498360634,
-0.06450872123241425,
-0.024177679792046547,
0.034352343529462814,
-0.03891048580408096,
-0.03293083608150482,
-0.07036761939525604,
0.07102807611227036,
0.0058151958510279655,
-0.04050799459218979,
0.03777237981557846,
-0.00030491783400066197,
-0.0006437045522034168,
0.02691369690001011,
-0.009062771685421467,
0.09504786878824234,
0.06026136502623558,
-0.007550010923296213,
0.013768384233117104,
-0.015581300482153893,
0.061850257217884064,
0.08363184332847595,
-0.021152788773179054,
0.037214245647192,
0.01916860230267048,
0.03996380418539047,
-0.029794994741678238,
0.028463466092944145,
-0.020527735352516174,
0.007247671019285917,
-0.11895982176065445,
0.04884044826030731,
-0.016033900901675224,
0.11007431894540787,
-0.04629937931895256,
-0.007629904896020889,
-0.021541627123951912,
-0.023419607430696487,
-0.021424662321805954,
0.017294691875576973,
-0.1649194359779358,
-0.07919855415821075,
-0.004930516239255667,
0.0756249725818634,
-0.10336755961179733,
-0.01627114601433277,
-0.10474524646997452,
-0.03585783392190933,
0.04439013451337814,
-0.01624193973839283,
-0.05341998115181923,
0.08694881200790405,
-8.915238438449126e-34,
0.006579049862921238,
-0.04384767636656761,
0.05894259735941887,
0.01010510977357626,
-0.029527252539992332,
-0.041665300726890564,
0.04455895349383354,
0.04440043866634369,
-0.096286840736866,
-0.08215203136205673,
-0.015053411945700645,
-0.024097424000501633,
0.059112563729286194,
-0.022245410829782486,
-0.015707731246948242,
0.0631633773446083,
0.04680066928267479,
0.012559348717331886,
0.00870016124099493,
0.07843703031539917,
-0.004349338822066784,
0.02282528579235077,
-0.0433669313788414,
0.030640078708529472,
0.06984125822782516,
0.023700254037976265,
-0.03101443313062191,
0.009833045303821564,
0.04576723277568817,
-0.07614058256149292,
-0.08874921500682831,
-0.0018461780855432153,
-0.11039819568395615,
-0.04085575416684151,
-0.07285290956497192,
0.03098636493086815,
-0.015743214637041092,
-0.0067795454524457455,
0.03517712280154228,
0.13529735803604126,
-0.004077412188053131,
0.03351312875747681,
-0.06627432256937027,
-0.013766919262707233,
-0.007672247476875782,
-0.07300519943237305,
-0.073739193379879,
0.01845688745379448,
0.05345018208026886,
-0.03418630361557007,
0.012722639366984367,
0.009530092589557171,
-0.04114636033773422,
0.02399134263396263,
-0.016997812315821648,
-0.05828774720430374,
0.02358344756066799,
-0.022008858621120453,
0.014952748082578182,
0.008333096280694008,
-0.020200520753860474,
-0.010005613788962364,
-0.02772808074951172,
-0.05400073900818825,
0.03542889654636383,
-0.03706060349941254,
0.031201185658574104,
0.034713540226221085,
-0.05881708115339279,
-0.01667860895395279,
-0.019448963925242424,
-0.04331575334072113,
0.05286606773734093,
-0.04917522519826889,
0.0022494234144687653,
0.024689026176929474,
0.03190436214208603,
-0.047024164348840714,
-0.014921224676072598,
-0.004914197139441967,
0.03343591094017029,
0.06757360696792603,
-0.036702774465084076,
0.04691784828901291,
0.06685439497232437,
0.11670801788568497,
0.004415884148329496,
0.06005194038152695,
0.02610863372683525,
-0.015263632871210575,
-0.016264861449599266,
-0.0018565378850325942,
0.1023591160774231,
0.03388924524188042,
0.0058083562180399895,
-4.582243917639062e-8,
-0.11361035704612732,
0.017632802948355675,
-0.082512766122818,
0.04211641848087311,
0.033410605043172836,
-0.07218960672616959,
-0.059992466121912,
0.10805559158325195,
0.03949422389268875,
0.0019173977198079228,
0.057757485657930374,
-0.03286146745085716,
-0.07498319447040558,
0.03983023762702942,
-0.0008305821684189141,
0.007151317782700062,
0.019199680536985397,
0.037839099764823914,
0.0074575599282979965,
0.023582719266414642,
0.05521884933114052,
0.018704984337091446,
0.08225905895233154,
0.019497103989124298,
0.052336376160383224,
-0.06824801862239838,
-0.03298462554812431,
0.04607551917433739,
0.09493177384138107,
-0.031227707862854004,
-0.0774427205324173,
0.0032905368134379387,
-0.044365882873535156,
-0.03595814108848572,
0.0016502677462995052,
0.08138179033994675,
0.0372513085603714,
-0.00952623225748539,
-0.08391221612691879,
-0.04456840828061104,
0.07644223421812057,
-0.015440973453223705,
-0.038120657205581665,
0.014914111234247684,
-0.016935929656028748,
-0.02083846926689148,
0.049046725034713745,
-0.031084226444363594,
0.04233302175998688,
-0.03631828725337982,
0.02460412308573723,
-0.09910808503627777,
0.04856553301215172,
0.02617109939455986,
0.053709328174591064,
-0.019580408930778503,
-0.009866767562925816,
0.046614184975624084,
0.03253289684653282,
-0.04179023951292038,
0.019055746495723724,
-0.03193060681223869,
-0.051978275179862976,
-0.005845774430781603
] | 0 | 0 | 0 | 0 |
|
Population Transformer: Learning Population-level Representations of Neural Activity | https://openreview.net/forum?id=FVuqJt3c4L | [
"Geeling Chau",
"Christopher Wang",
"Sabera J Talukder",
"Vighnesh Subramaniam",
"Saraswati Soedarmadji",
"Yisong Yue",
"Boris Katz",
"Andrei Barbu"
] | Oral | We present a self-supervised framework that learns population-level codes for arbitrary ensembles of neural recordings at scale. We address key challenges in scaling models with neural time-series data, namely, sparse and variable electrode distribution across subjects and datasets. The Population Transformer (PopT) stacks on top of pretrained temporal embeddings and enhances downstream decoding by enabling learned aggregation of multiple spatially-sparse data channels. The pretrained PopT lowers the amount of data required for downstream decoding experiments, while increasing accuracy, even on held-out subjects and tasks. Compared to end-to-end methods, this approach is computationally lightweight, while achieving similar or better decoding performance. We further show how our framework is generalizable to multiple time-series embeddings and neural data modalities. Beyond decoding, we interpret the pretrained and fine-tuned PopT models to show how they can be used to extract neuroscience insights from large amounts of data. We release our code as well as a pretrained PopT to enable off-the-shelf improvements in multi-channel intracranial data decoding and interpretability. Code is available at https://github.com/czlwang/PopulationTransformer. | representation learning, neuroscience, self supervised learning | Representation learning of neural data | 5,882 | 2406.03044 | [
-0.071141816675663,
-0.1271103024482727,
0.06249323859810829,
0.044814545661211014,
0.027657218277454376,
0.026177681982517242,
-0.036116503179073334,
-0.04613885655999184,
0.024733342230319977,
-0.09072672575712204,
-0.06653272360563278,
-0.10220364481210709,
-0.11196433752775192,
0.07788960635662079,
-0.08217601478099823,
-0.026025306433439255,
0.011961349286139011,
0.021565262228250504,
-0.008007765747606754,
-0.009942368604242802,
0.01284918561577797,
0.0030930896755307913,
0.036657191812992096,
0.09622707962989807,
0.08204317837953568,
0.052633047103881836,
0.0198442991822958,
-0.004105107393115759,
0.020172422751784325,
0.02097254805266857,
0.127811998128891,
-0.00469838734716177,
-0.03771371394395828,
0.03014073520898819,
-0.09479761123657227,
0.057786982506513596,
0.004350508097559214,
0.024697614833712578,
-0.033605851233005524,
0.0012970370007678866,
0.08177627623081207,
0.06261125952005386,
0.01816766895353794,
0.055256254971027374,
0.027777595445513725,
0.010580670088529587,
0.033584680408239365,
-0.14279396831989288,
-0.04518900439143181,
0.0010492618894204497,
0.022281821817159653,
0.05958164110779762,
0.002888342598453164,
0.08201564103364944,
-0.02285783179104328,
0.017329761758446693,
0.007000801153481007,
-0.006546701770275831,
-0.04194889962673187,
0.014589670114219189,
-0.02085127681493759,
-0.013801820576190948,
-0.0009556218283250928,
-0.02082223817706108,
-0.012534263543784618,
0.005393756553530693,
0.032056599855422974,
0.05932716652750969,
-0.0583636499941349,
-0.04395646974444389,
-0.026630032807588577,
0.018289918079972267,
-0.051594581454992294,
0.059014834463596344,
0.06604035198688507,
-0.01339087262749672,
0.028517628088593483,
-0.031760744750499725,
0.11783874779939651,
-0.09797513484954834,
0.005234560463577509,
0.027770740911364555,
-0.002169680316001177,
0.03790168836712837,
0.11310846358537674,
-0.0025623829569667578,
0.018923696130514145,
0.12423811852931976,
-0.042401060461997986,
0.04642007499933243,
-0.04397866129875183,
-0.028240321204066277,
0.06954756379127502,
0.01995101384818554,
-0.025419672951102257,
0.029607484117150307,
-0.050939954817295074,
-0.044188279658555984,
0.05798051133751869,
0.06199388951063156,
-0.05364355444908142,
0.004258033819496632,
0.030005736276507378,
0.041723329573869705,
-0.061284568160772324,
-0.053549181669950485,
0.05648108199238777,
0.07073916494846344,
0.002275643637403846,
0.016170889139175415,
0.06213850528001785,
0.020455313846468925,
0.0024877996183931828,
0.03450464457273483,
0.06928355246782303,
-0.0005949775804765522,
-0.016318250447511673,
0.03698046877980232,
0.07765752077102661,
0.054640378803014755,
-0.025235971435904503,
-0.017811469733715057,
-0.1164715364575386,
0.047954704612493515,
0.04597597196698189,
0.0146781662479043,
-0.056719813495874405,
1.591747553366014e-33,
0.008082865737378597,
0.001846569823101163,
0.04931389167904854,
0.04824402928352356,
-0.049787361174821854,
0.0003861612640321255,
0.010293788276612759,
0.006328158546239138,
0.03197285532951355,
0.08666673302650452,
-0.03175031393766403,
0.06671229749917984,
-0.05275290086865425,
0.08378056436777115,
0.00504198856651783,
0.01058773323893547,
-0.09487906843423843,
0.06308485567569733,
0.022273117676377296,
-0.06961372494697571,
0.052925340831279755,
-0.001229328685440123,
-0.05901828035712242,
-0.030433932319283485,
0.050065163522958755,
-0.040177494287490845,
-0.018093587830662727,
-0.05295863375067711,
-0.040229178965091705,
0.00927888136357069,
-0.15075922012329102,
-0.012552807107567787,
0.017621075734496117,
-0.05659164860844612,
0.04630973935127258,
-0.040361616760492325,
0.08846751600503922,
-0.024832898750901222,
-0.015737829729914665,
0.033236972987651825,
0.0414358489215374,
0.013795213773846626,
0.0069161588326096535,
-0.04925825074315071,
-0.056671325117349625,
-0.07025156170129776,
-0.009760314598679543,
0.03379596769809723,
-0.03185471147298813,
-0.01823338121175766,
-0.011057909578084946,
-0.021017000079154968,
-0.03977428004145622,
-0.03378636762499809,
-0.00767176877707243,
0.03285403177142143,
-0.04654529318213463,
0.023440584540367126,
0.10377903282642365,
0.04162431135773659,
0.021955175325274467,
0.04786040633916855,
0.0062233321368694305,
0.001003499492071569,
0.09328489005565643,
0.00014334487786982208,
0.0015125860227271914,
0.07546575367450714,
0.03696629777550697,
-0.029593026265501976,
-0.053219784051179886,
0.03493567183613777,
-0.009740070439875126,
-0.06252656131982803,
0.04750967398285866,
0.03227448835968971,
0.11562298238277435,
-0.008535735309123993,
-0.04870602488517761,
-0.006516241002827883,
0.02487202361226082,
-0.056155089288949966,
0.0029499810189008713,
-0.00968351773917675,
0.025242704898118973,
0.0411815382540226,
0.09075286239385605,
-0.05428428202867508,
-0.07961170375347137,
-0.004043876193463802,
-0.04012008383870125,
-0.0691647008061409,
0.021526623517274857,
-0.07427456974983215,
-0.08892060071229935,
-1.3118799808504895e-33,
-0.023603057488799095,
0.018024828284978867,
-0.06217697635293007,
-0.03733842074871063,
-0.004847895819693804,
-0.03269719332456589,
0.0010002903873100877,
0.10291077941656113,
0.0041377064771950245,
-0.039289478212594986,
-0.04887347295880318,
-0.08911174535751343,
0.12819191813468933,
-0.0006637443439103663,
0.0320311076939106,
-0.05249590426683426,
-0.10345347225666046,
0.028778430074453354,
0.016956910490989685,
0.026004133746027946,
-0.08006693422794342,
0.09522982686758041,
-0.039324574172496796,
0.021594557911157608,
-0.0715295821428299,
0.04365579038858414,
-0.047294437885284424,
0.06339238584041595,
0.034764282405376434,
-0.0278871338814497,
-0.05594158545136452,
-0.027606628835201263,
0.0021980712190270424,
-0.10340794175863266,
-0.05644720047712326,
0.10077988356351852,
0.05899573490023613,
-0.05645236000418663,
-0.05475889891386032,
-0.0038476847112178802,
0.03801780939102173,
-0.012102345004677773,
-0.037923414260149,
0.05140330642461777,
-0.04145772382616997,
-0.011811682023108006,
-0.08251701295375824,
0.04103498160839081,
-0.014569745399057865,
0.078349269926548,
-0.013590537011623383,
0.06646393984556198,
-0.08783356100320816,
-0.014980820007622242,
-0.033953361213207245,
-0.11119324713945389,
0.043380256742239,
-0.08606791496276855,
0.029554154723882675,
-0.05658433958888054,
-0.05879323557019234,
-0.05449968948960304,
0.0029070309828966856,
-0.08505942672491074,
0.007152148988097906,
-0.04075765982270241,
0.038320574909448624,
-0.03817149996757507,
0.0312657468020916,
0.007956135086715221,
0.029013659805059433,
-0.0427221916615963,
0.042923156172037125,
-0.05527880787849426,
0.03990801051259041,
-0.005673669744282961,
0.005229632835835218,
0.004138929769396782,
-0.0067409323528409,
0.005078529007732868,
0.024070534855127335,
-0.040985263884067535,
-0.01168319582939148,
0.014868560247123241,
0.07061310857534409,
0.06196662411093712,
0.050043538212776184,
0.05992026999592781,
0.022165032103657722,
0.02081114985048771,
-0.022360382601618767,
0.005326838698238134,
-0.021030757576227188,
0.03322092816233635,
-0.014268324710428715,
-4.1347824719650816e-8,
-0.08192423731088638,
0.07254389673471451,
0.033029135316610336,
0.02989456057548523,
-0.05780179426074028,
-0.061609018594026566,
0.04507967084646225,
0.07242882251739502,
-0.0029476142954081297,
-0.01002397108823061,
0.1594436913728714,
-0.04273039475083351,
0.02239159867167473,
-0.0051257433369755745,
0.019900357350707054,
0.0989234521985054,
-0.019802561029791832,
0.024923309683799744,
-0.0157781969755888,
-0.0028599000070244074,
0.0013323421590030193,
0.03417166322469711,
0.06791656464338303,
-0.06831371039152145,
0.06607627123594284,
-0.009802097454667091,
-0.03814017400145531,
0.08761051297187805,
0.021352721378207207,
-0.0498448945581913,
-0.012498256750404835,
0.06923604756593704,
0.026003222912549973,
-0.022019648924469948,
-0.047238241881132126,
0.01300665084272623,
0.032614562660455704,
0.05204370617866516,
-0.055254366248846054,
-0.04688771069049835,
-0.036051277071237564,
0.0013030943227931857,
-0.10020527243614197,
0.054989639669656754,
-0.009433301165699959,
-0.05474548041820526,
0.007712714374065399,
0.000003519005076668691,
0.025851495563983917,
-0.046412836760282516,
-0.04101553559303284,
-0.03200904652476311,
0.0030377754010260105,
0.04832405596971512,
0.05722736939787865,
0.04999540001153946,
-0.06157417967915535,
-0.017521750181913376,
0.0024996108841151,
0.015765318647027016,
-0.04044857248663902,
0.07752450555562973,
-0.11043103784322739,
-0.00866065826267004
] | https://github.com/czlwang/populationtransformer | 4 | 0 | 0 | 0 |
KAN: Kolmogorov–Arnold Networks | https://openreview.net/forum?id=Ozo7qJ5vZi | [
"Ziming Liu",
"Yixuan Wang",
"Sachin Vaidya",
"Fabian Ruehle",
"James Halverson",
"Marin Soljacic",
"Thomas Y. Hou",
"Max Tegmark"
] | Oral | Inspired by the Kolmogorov-Arnold representation theorem, we propose Kolmogorov-Arnold Networks (KANs) as promising alternatives to Multi-Layer Perceptrons (MLPs). While MLPs have fixed activation functions on nodes ("neurons''), KANs have learnable activation functions on edges ("weights''). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. We show that this seemingly simple change makes KANs outperform MLPs in terms of accuracy and interpretability, on small-scale AI + Science tasks. For accuracy, smaller KANs can achieve comparable or better accuracy than larger MLPs in function fitting tasks. Theoretically and empirically, KANs possess faster neural scaling laws than MLPs. For interpretability, KANs can be intuitively visualized and can easily interact with human users. Through two examples in mathematics and physics, KANs are shown to be useful ``collaborators'' helping scientists (re)discover mathematical and physical laws. In summary, KANs are promising alternatives for MLPs. Despite the slow training of KANs, their improved accuracy and interpretability show the potential to improve today's deep learning models which rely heavily on MLPs. More research is necessary to make KANs' training more efficient. | Kolmogorov-Arnold networks, Kolmogorov-Arnold representation theorem, learnable activation functions, interpretability, AI + Science | Inspired by the Kolmogorov-Arnold representation theorem, we propose Kolmogorov-Arnold Networks (KANs) as promising alternatives to Multi-Layer Perceptrons (MLPs). | 5,796 | null | [
-0.08938876539468765,
-0.07097781449556351,
0.027797939255833626,
0.05194469168782234,
-0.030385123565793037,
0.023275433108210564,
0.01809537038207054,
-0.047197505831718445,
0.06521181017160416,
0.03867439553141594,
-0.033130716532468796,
-0.06097961962223053,
-0.007205570116639137,
0.04442787915468216,
-0.06389647722244263,
0.020330652594566345,
-0.030202772468328476,
0.07901312410831451,
-0.14150607585906982,
0.020887380465865135,
-0.045509640127420425,
-0.005474030040204525,
-0.03038083203136921,
-0.003336485708132386,
0.06068361550569534,
-0.01946749910712242,
-0.06804709881544113,
-0.02634415216743946,
-0.019538916647434235,
0.021037405356764793,
0.05679386109113693,
-0.061894986778497696,
0.012676508165895939,
0.017473772168159485,
-0.09821441024541855,
0.07697557657957077,
0.027653224766254425,
-0.008489789441227913,
-0.010101457126438618,
-0.009563645347952843,
0.00023231098020914942,
0.05041852965950966,
-0.0022037094458937645,
0.07049364596605301,
0.09388169646263123,
0.07507689297199249,
0.02818070352077484,
-0.08229683339595795,
0.0013145891716703773,
0.03433417156338692,
-0.07856396585702896,
-0.006806648336350918,
0.044380415230989456,
-0.04366235435009003,
0.014554831199347973,
0.02063600718975067,
0.00988381914794445,
0.0005170516669750214,
0.0056694610975682735,
-0.06069086864590645,
0.01535228081047535,
-0.10160918533802032,
0.013823049142956734,
0.005344159435480833,
-0.038208816200494766,
-0.04267631098628044,
-0.0035289525985717773,
0.007521246559917927,
0.0651049017906189,
0.019870206713676453,
0.059158481657505035,
0.07271377742290497,
-0.04260063171386719,
0.035653263330459595,
0.08367440104484558,
-0.026417326182127,
0.08265812695026398,
0.0017457877984270453,
0.03432728722691536,
-0.04782438650727272,
-0.014268453232944012,
-0.02522091008722782,
-0.04994945228099823,
0.00871996209025383,
0.08549986034631729,
0.06827417016029358,
0.010144563391804695,
0.024627767503261566,
-0.00510428985580802,
0.0391644649207592,
-0.02305031754076481,
0.0005349128041416407,
0.04309173673391342,
-0.11093487590551376,
-0.024521183222532272,
0.01766546443104744,
0.004451143555343151,
-0.07832711189985275,
-0.07213516533374786,
0.13189299404621124,
-0.047229304909706116,
0.03984919935464859,
0.0473456047475338,
-0.02047532983124256,
-0.0027817669324576855,
-0.016125261783599854,
0.010984043590724468,
-0.04030712693929672,
0.026577988639473915,
-0.06232838332653046,
0.050208669155836105,
0.004079931415617466,
-0.005974140949547291,
0.0813397765159607,
0.06947077065706253,
-0.0006801937124691904,
0.03044348768889904,
-0.008735809475183487,
0.07392561435699463,
0.08150480687618256,
-0.07909408211708069,
-0.00721838790923357,
-0.07831285893917084,
0.06462269276380539,
-0.018326906487345695,
0.015038510784506798,
-0.06895341724157333,
3.769724425062672e-33,
-0.03991241380572319,
0.07528635114431381,
0.10729445517063141,
-0.015264599584043026,
0.047556322067976,
-0.0922803059220314,
-0.025132928043603897,
-0.04863934591412544,
0.0026799815241247416,
-0.050193190574645996,
-0.11353465169668198,
0.07732518017292023,
-0.046081673353910446,
0.0663960725069046,
0.047243352979421616,
0.03137759491801262,
-0.02219988778233528,
-0.022178567945957184,
-0.011658201925456524,
-0.08310004323720932,
0.06326550990343094,
-0.12169855833053589,
0.01949234865605831,
-0.06200239807367325,
0.0038242805749177933,
0.024382008239626884,
0.0419672392308712,
-0.03986067324876785,
-0.017874427139759064,
0.011856219731271267,
-0.02812669798731804,
0.07672066241502762,
0.026320284232497215,
0.00978340208530426,
-0.012454668991267681,
-0.029436452314257622,
0.017169591039419174,
-0.06945304572582245,
0.018587883561849594,
0.009886381216347218,
0.004084349609911442,
0.029181187972426414,
0.007005381863564253,
0.007606809958815575,
-0.019426429644227028,
0.0172375850379467,
0.05498991906642914,
0.01077673863619566,
-0.009606956504285336,
-0.02701464295387268,
-0.02162034995853901,
-0.021493157371878624,
0.06058384105563164,
-0.06035012751817703,
0.07426460087299347,
0.042259812355041504,
0.00020610484352800995,
0.033172085881233215,
0.029218800365924835,
0.02646517939865589,
-0.010683922097086906,
0.06624776870012283,
0.05696991831064224,
0.09031617641448975,
-0.02707904390990734,
0.03431253135204315,
-0.06163765862584114,
0.11427586525678635,
0.08453837037086487,
-0.04570231959223747,
-0.07563292235136032,
0.08118720352649689,
0.021082323044538498,
-0.09605813771486282,
0.031179307028651237,
-0.005175933241844177,
-0.0545642264187336,
-0.16829152405261993,
-0.020803960040211678,
0.12416405230760574,
-0.04254265874624252,
-0.04877025634050369,
-0.041188936680555344,
-0.024689681828022003,
0.015552015043795109,
-0.09380656480789185,
0.054144371300935745,
-0.006329583004117012,
-0.014241031371057034,
-0.0299062542617321,
-0.06696280092000961,
-0.05688420310616493,
0.015091695822775364,
-0.05938214436173439,
-0.08177101612091064,
-4.3077448833561715e-33,
-0.129045307636261,
-0.0065299142152071,
-0.056576527655124664,
0.08857037127017975,
-0.012520413845777512,
-0.0281215887516737,
-0.054703209549188614,
-0.04824341833591461,
-0.03664415702223778,
0.01934022456407547,
-0.03507154807448387,
-0.020015211775898933,
-0.020322000607848167,
-0.057980604469776154,
0.08654072135686874,
-0.025635976344347,
-0.0699736699461937,
-0.002009207382798195,
0.01984993740916252,
-0.04040364921092987,
0.019257351756095886,
0.038075514137744904,
-0.08475684374570847,
0.03812991827726364,
0.007371858227998018,
-0.010994741693139076,
-0.015757247805595398,
0.04678483307361603,
-0.016852842643857002,
-0.023168830201029778,
-0.035499658435583115,
-0.08302076160907745,
-0.009138231165707111,
-0.0011786198010668159,
0.024631600826978683,
-0.038443613797426224,
0.03618335351347923,
0.011729842983186245,
-0.003930247388780117,
-0.014207306317985058,
0.02370051108300686,
-0.007345208898186684,
-0.029803959652781487,
-0.033040765672922134,
-0.023603325709700584,
0.004777309484779835,
-0.014938109554350376,
0.020206930115818977,
-0.045628566294908524,
0.03922624886035919,
0.009954491630196571,
0.016417115926742554,
-0.09228545427322388,
-0.006452692206948996,
-0.025251438841223717,
0.00962685327976942,
-0.04254273697733879,
0.10208316892385483,
0.028425943106412888,
0.007947633042931557,
-0.09470340609550476,
-0.09610896557569504,
0.06780856102705002,
0.11178793013095856,
-0.01593087427318096,
-0.037639275193214417,
0.05173950642347336,
0.06145777925848961,
0.002025188645347953,
-0.026537586003541946,
-0.03439895808696747,
-0.05493900924921036,
0.03961064666509628,
0.0036064626183360815,
-0.047885678708553314,
-0.03292389586567879,
-0.025673173367977142,
-0.052188560366630554,
0.0030695756431668997,
-0.027575789019465446,
-0.0003245234547648579,
-0.059574421495199203,
0.05881013721227646,
0.10138437151908875,
0.15443997085094452,
0.09695716947317123,
0.027561957016587257,
0.03694789856672287,
0.050654057413339615,
0.013746354728937149,
0.05140430107712746,
0.08737939596176147,
-0.014125915244221687,
0.04400963708758354,
-0.051929935812950134,
-4.719796464769388e-8,
-0.04086757078766823,
0.022623330354690552,
0.03548114374279976,
-0.013700145296752453,
0.04688580334186554,
0.031482841819524765,
0.010342261753976345,
0.01549492683261633,
-0.09927698224782944,
-0.02771362103521824,
0.09686928987503052,
0.023563558235764503,
-0.04972313717007637,
-0.0551169291138649,
0.04024166613817215,
0.012071889825165272,
0.02990996651351452,
0.012754658237099648,
0.0006697705830447376,
-0.01129993237555027,
-0.08714701980352402,
0.049002811312675476,
-0.036347247660160065,
0.0025635347701609135,
-0.0004796391585841775,
-0.04546799510717392,
-0.08244769275188446,
0.1032966747879982,
0.02284020371735096,
-0.016980262473225594,
-0.0010193842463195324,
-0.04384071007370949,
0.009267253801226616,
0.06540568917989731,
0.01710752584040165,
0.015804000198841095,
-0.018668968230485916,
0.0329609252512455,
-0.043153826147317886,
0.001870587351731956,
0.029004957526922226,
0.06618591398000717,
0.018480239436030388,
0.012159371748566628,
0.020114198327064514,
0.045303456485271454,
0.029125826433300972,
-0.11021192371845245,
0.018023233860731125,
0.07417909801006317,
0.04012487828731537,
0.026266522705554962,
-0.032704517245292664,
0.03378632292151451,
0.04082108289003372,
0.07786864787340164,
-0.06181587651371956,
-0.08954887837171555,
-0.048865947872400284,
0.12495137751102448,
-0.00199644616805017,
0.03226441890001297,
-0.0439206063747406,
0.0018013184890151024
] | 0 | 0 | 0 | 0 |
|
Problem-Parameter-Free Federated Learning | https://openreview.net/forum?id=ZuazHmXTns | [
"Wenjing Yan",
"Kai Zhang",
"Xiaolu Wang",
"Xuanyu Cao"
] | Oral | Federated learning (FL) has garnered significant attention from academia and industry in recent years due to its advantages in data privacy, scalability, and communication efficiency. However, current FL algorithms face a critical limitation: their performance heavily depends on meticulously tuned hyperparameters, particularly the learning rate or stepsize. This manual tuning process is challenging in federated settings due to data heterogeneity and limited accessibility of local datasets. Consequently, the reliance on problem-specific parameters hinders the widespread adoption of FL and potentially compromises its performance in dynamic or diverse environments. To address this issue, we introduce PAdaMFed, a novel algorithm for nonconvex FL that carefully combines adaptive stepsize and momentum techniques. PAdaMFed offers two key advantages: 1) it operates autonomously without relying on problem-specific parameters; and 2) it manages data heterogeneity and partial participation without requiring heterogeneity bounds. Despite these benefits, PAdaMFed provides several strong theoretical guarantees: 1) It achieves state-of-the-art convergence rates with a sample complexity of $\mathcal{O}(\epsilon^{-4})$ and communication complexity of $\mathcal{O}(\epsilon^{-3})$ to obtain an accuracy of $||\nabla f\left(\boldsymbol{\theta}\right)|| \leq \epsilon$, even using constant learning rates; 2) these complexities can be improved to the best-known $\mathcal{O}(\epsilon^{-3})$ for sampling and $\mathcal{O}(\epsilon^{-2})$ for communication when incorporating variance reduction; 3) it exhibits linear speedup with respect to the number of local update steps and participating clients at each global round. These attributes make PAdaMFed highly scalable and adaptable for various real-world FL applications. Extensive empirical evidence on both image classification and sentiment analysis tasks validates the efficacy of our approaches. | Adaptive federated learning, problem-parameter free, arbitrary data heterogeneity, adaptive stepsize | null | 5,729 | null | [
-0.03667658939957619,
-0.0803283229470253,
0.020979508757591248,
0.05160882696509361,
0.03289339318871498,
-0.07648976892232895,
-0.06098907068371773,
-0.0769835039973259,
-0.01946232281625271,
-0.07821519672870636,
-0.002610964933410287,
-0.0012862920993939042,
0.029702799394726753,
-0.05266592279076576,
0.01711399294435978,
0.012652236968278885,
0.06658869981765747,
0.04469696059823036,
-0.04454145208001137,
-0.009341402910649776,
-0.044777125120162964,
-0.0015377586241811514,
0.02404974400997162,
-0.0027212498243898153,
0.05344734340906143,
0.007196848280727863,
0.04133280739188194,
0.0016115546459332108,
0.024008966982364655,
-0.02986723557114601,
0.08543499559164047,
0.019485022872686386,
-0.09039685130119324,
-0.017434677109122276,
0.03161713853478432,
0.04572150111198425,
-0.04331836476922035,
0.06458873301744461,
-0.06824368238449097,
0.051716215908527374,
0.022729258984327316,
-0.01742050051689148,
-0.021200668066740036,
0.009628070518374443,
0.05596788227558136,
0.006896551698446274,
0.030732888728380203,
-0.030386174097657204,
0.020778315141797066,
0.038872674107551575,
-0.08170714229345322,
-0.041698191314935684,
-0.06293927133083344,
0.0861070454120636,
-0.03507808595895767,
0.000539668370038271,
-0.026711998507380486,
0.02353614568710327,
-0.04046054556965828,
-0.014835494570434093,
-0.05159491300582886,
-0.13068787753582,
0.022614026442170143,
-0.004478222690522671,
-0.0860627144575119,
0.02575409784913063,
0.038957349956035614,
0.10243052989244461,
0.022723354399204254,
0.004165699705481529,
-0.045835334807634354,
0.07495810836553574,
-0.02529953420162201,
0.04774865135550499,
0.05328233912587166,
0.05549853667616844,
0.03465982899069786,
0.026550736278295517,
0.06313858181238174,
-0.04968629032373428,
-0.006713028531521559,
0.02079111337661743,
-0.03429052606225014,
-0.04746463522315025,
0.12009408324956894,
-0.086064413189888,
-0.019932327792048454,
0.018433522433042526,
0.0289438646286726,
-0.011065802536904812,
-0.023812664672732353,
0.05893973633646965,
-0.03505216911435127,
-0.001890653627924621,
0.018989531323313713,
-0.016017276793718338,
0.04363815486431122,
-0.06044713035225868,
-0.016090037301182747,
0.12662553787231445,
-0.08700501173734665,
0.10440073907375336,
0.006864471361041069,
0.05155082792043686,
0.009107855148613453,
-0.006406296975910664,
-0.016135992482304573,
0.024989770725369453,
0.058689456433057785,
-0.049975380301475525,
0.030408121645450592,
0.05070662871003151,
0.04565926268696785,
0.002997898729518056,
-0.043220970779657364,
-0.04893648996949196,
-0.030344367027282715,
-0.04987810179591179,
0.0033937906846404076,
0.0493793785572052,
-0.034871265292167664,
-0.08401620388031006,
0.010912146419286728,
0.06047990173101425,
-0.01247694343328476,
-0.06095938757061958,
-0.09623869508504868,
3.901246447214347e-33,
0.024515876546502113,
0.08451247960329056,
0.04105003923177719,
-0.09338659048080444,
-0.004643729422241449,
0.004333081189543009,
0.09101853519678116,
0.023503514006733894,
-0.06962385028600693,
0.05400904640555382,
0.01685074344277382,
0.0027308138087391853,
-0.0018593441927805543,
0.020969929173588753,
0.08570408821105957,
-0.09579402953386307,
-0.019993364810943604,
0.08159992098808289,
0.04466954618692398,
-0.06310120970010757,
0.09206783026456833,
-0.03742007911205292,
-0.03539207577705383,
-0.022768933326005936,
-0.0395328588783741,
-0.02396417036652565,
0.05949648842215538,
0.02720828168094158,
0.024992816150188446,
0.029287399724125862,
-0.020856158807873726,
-0.0428953655064106,
-0.03799407184123993,
0.0706174299120903,
0.08432602137327194,
-0.037018753588199615,
-0.02941623516380787,
-0.08187975734472275,
-0.033876385539770126,
0.005201097112149,
0.002839605091139674,
-0.0453227274119854,
0.1251208782196045,
-0.05257519707083702,
-0.18384577333927155,
-0.07499964535236359,
-0.011353088542819023,
-0.029385438188910484,
0.014237089082598686,
-0.08373178541660309,
0.0020110055338591337,
-0.0794505849480629,
-0.0325457826256752,
-0.037902433425188065,
0.035418085753917694,
-0.01041360106319189,
-0.05469796061515808,
0.031905997544527054,
-0.03282470628619194,
0.06827075034379959,
-0.020918121561408043,
0.007074256893247366,
-0.009401591494679451,
0.06618054956197739,
-0.023326802998781204,
0.055760860443115234,
-0.018659282475709915,
0.020194699987769127,
0.10299133509397507,
-0.08517630398273468,
0.06988642364740372,
0.07028526812791824,
-0.021694365888834,
-0.026662517338991165,
0.12008500844240189,
-0.058804336935281754,
0.10857550799846649,
-0.06551244854927063,
0.01448087114840746,
-0.007191728334873915,
0.00852028839290142,
0.052990321069955826,
-0.009084921330213547,
0.05712892487645149,
-0.03638330101966858,
-0.0719963014125824,
-0.015330626629292965,
0.030299730598926544,
-0.06493905186653137,
-0.005788449198007584,
-0.09548665583133698,
-0.009439181536436081,
-0.03828814998269081,
0.007952631451189518,
-0.018841685727238655,
-2.7695251182937813e-33,
-0.04244111478328705,
-0.016025083139538765,
-0.03921030834317207,
0.036208245903253555,
0.021728627383708954,
0.02082172781229019,
-0.02497965283691883,
-0.011355794034898281,
0.030374040827155113,
-0.08164308965206146,
-0.09176524728536606,
-0.020262470468878746,
0.07078585028648376,
-0.07130864262580872,
0.003365209558978677,
0.06956394761800766,
-0.060190897434949875,
-0.001187608577311039,
-0.00295452494174242,
0.031255125999450684,
-0.04174095019698143,
0.02595568262040615,
-0.050121940672397614,
0.07603570818901062,
0.033049680292606354,
0.02048683539032936,
-0.05044163763523102,
0.11150702834129333,
-0.0021396612282842398,
0.010884898714721203,
-0.013723430223762989,
-0.02816687524318695,
-0.10348404198884964,
-0.08269810676574707,
-0.06967825442552567,
0.04537228122353554,
-0.005695858038961887,
0.044773366302251816,
-0.027261458337306976,
0.08297273516654968,
0.03887677192687988,
-0.035691361874341965,
-0.07050543278455734,
0.013466374017298222,
0.018581904470920563,
0.04372352734208107,
-0.029189227148890495,
-0.006048906594514847,
0.05203870311379433,
0.06437569111585617,
-0.0016036591259762645,
0.06687591969966888,
-0.006743613164871931,
0.019909339025616646,
-0.011567101813852787,
0.02205209992825985,
0.010721514001488686,
0.02053639478981495,
0.04543909430503845,
0.011486985720694065,
0.03643444553017616,
-0.02388552390038967,
0.014282244257628918,
0.056265704333782196,
0.07591863721609116,
0.005277218762785196,
-0.0034338831901550293,
0.03814590722322464,
-0.07900813221931458,
0.07692024856805801,
-0.004317344631999731,
0.03171864151954651,
-0.0327763669192791,
0.005579913500696421,
0.0607757605612278,
0.001816355623304844,
-0.0007584705017507076,
-0.024896588176488876,
-0.06955531984567642,
0.018923291936516762,
0.08281903713941574,
0.038305386900901794,
0.025024056434631348,
0.04915070906281471,
0.06662552058696747,
0.09578616172075272,
0.11386284977197647,
0.06063505634665489,
0.034429676830768585,
-0.04866178706288338,
0.06360076367855072,
0.01379169151186943,
-0.024619925767183304,
0.0005053090862929821,
-0.028268923982977867,
-4.9948514657671694e-8,
-0.049445971846580505,
0.004719110205769539,
-0.01962161250412464,
-0.028462719172239304,
0.06911870837211609,
0.006724510807543993,
-0.01159642357379198,
0.04923703521490097,
-0.03290330618619919,
-0.03742249310016632,
0.08296579867601395,
-0.037808384746313095,
0.002697780029848218,
-0.0590432770550251,
-0.00783719215542078,
0.022698352113366127,
0.007468878757208586,
-0.0648542121052742,
-0.047210510820150375,
0.01705237105488777,
-0.05278433486819267,
0.024178439751267433,
-0.06487882882356644,
-0.007468535099178553,
0.0779145285487175,
-0.08809501677751541,
0.09073606878519058,
0.04635246843099594,
0.1076635867357254,
0.02237936481833458,
-0.06292211264371872,
0.010116997174918652,
0.004228860605508089,
-0.01122561190277338,
-0.07347077876329422,
0.078133724629879,
0.028125502169132233,
0.02127629518508911,
-0.06400922685861588,
-0.05355090647935867,
-0.027564724907279015,
0.08539760112762451,
-0.04762250557541847,
-0.016600241884589195,
0.038544926792383194,
0.0008372865850105882,
-0.02773636020720005,
-0.0829324945807457,
0.030949395149946213,
-0.03154599666595459,
0.06564200669527054,
0.03347311541438103,
-0.008029479533433914,
0.055723097175359726,
0.1266263872385025,
-0.005367421545088291,
-0.03520575538277626,
-0.04111544415354729,
0.07902558892965317,
0.07322695851325989,
0.013820827938616276,
-0.017901450395584106,
-0.0837077870965004,
0.045222584158182144
] | 0 | 0 | 0 | 0 |
|
SymmetricDiffusers: Learning Discrete Diffusion on Finite Symmetric Groups | https://openreview.net/forum?id=EO8xpnW7aX | [
"Yongxing Zhang",
"Donglin Yang",
"Renjie Liao"
] | Oral | The group of permutations $S_n$, also known as the finite symmetric groups, are essential in fields such as combinatorics, physics, and chemistry. However, learning a probability distribution over $S_n$ poses significant challenges due to its intractable size and discrete nature. In this paper, we introduce *SymmetricDiffusers*, a novel discrete diffusion model that simplifies the task of learning a complicated distribution over $S_n$ by decomposing it into learning simpler transitions of the reverse diffusion using deep neural networks. We identify the riffle shuffle as an effective forward transition and provide empirical guidelines for selecting the diffusion length based on the theory of random walks on finite groups. Additionally, we propose a generalized Plackett-Luce (PL) distribution for the reverse transition, which is provably more expressive than the PL distribution. We further introduce a theoretically grounded "denoising schedule" to improve sampling and learning efficiency. Extensive experiments show that our model achieves state-of-the-art or comparable performance on solving tasks including sorting 4-digit MNIST images, jigsaw puzzles, and traveling salesman problems. Our code is released at <https://github.com/DSL-Lab/SymmetricDiffusers>. | Finite Symmetric Groups, Discrete Diffusion, Permutations, Riffle Shuffles, Plackett-Luce Distribution, Sorting, Jigsaw Puzzle | null | 5,686 | 2410.02942 | [
-0.11471308022737503,
-0.13208559155464172,
0.06874005496501923,
-0.012078560888767242,
-0.017545249313116074,
-0.05832900106906891,
0.007233690470457077,
-0.07929114997386932,
0.010843247175216675,
0.0036547966301441193,
0.01969780959188938,
0.035066694021224976,
-0.02802487276494503,
0.06812036037445068,
-0.07919184863567352,
-0.009366718120872974,
0.037036165595054626,
0.0758245587348938,
-0.03490050137042999,
-0.004412616603076458,
-0.021570095792412758,
-0.06991657614707947,
-0.01327279582619667,
-0.04624857008457184,
0.05662554129958153,
0.047064926475286484,
0.009938539005815983,
-0.0731329470872879,
-0.007614632602781057,
-0.024392617866396904,
0.0817425474524498,
0.027232954278588295,
-0.005760917440056801,
-0.022505968809127808,
-0.011523934081196785,
0.022068742662668228,
-0.06057965010404587,
0.03253978118300438,
0.06850586086511612,
0.05930221825838089,
0.02812996320426464,
0.03266351297497749,
0.0049024648033082485,
0.06262822449207306,
0.027779944241046906,
-0.0834023654460907,
0.01837076060473919,
-0.041968073695898056,
0.0014428789727389812,
0.04253455996513367,
0.053457532078027725,
0.048427317291498184,
-0.0738120749592781,
0.08009247481822968,
0.028837580233812332,
-0.08365916460752487,
0.05514196678996086,
-0.02039208449423313,
-0.05653590336441994,
-0.07946594059467316,
0.04223364219069481,
-0.11102427542209625,
-0.014493318274617195,
-0.03909951075911522,
-0.036743875592947006,
-0.012657001614570618,
0.0692158043384552,
0.061554260551929474,
0.043462589383125305,
0.008142678998410702,
0.03508168086409569,
0.0948524996638298,
-0.0979442447423935,
0.11119265109300613,
0.06282985210418701,
-0.025632867589592934,
0.03876515477895737,
-0.025669081136584282,
0.0078037031926214695,
-0.07660651206970215,
-0.04423736035823822,
-0.01811966300010681,
0.08684305101633072,
0.00342169520445168,
0.07386158406734467,
-0.03771628811955452,
0.010630723088979721,
0.07497050613164902,
-0.0287682693451643,
-0.06099659204483032,
-0.05388496816158295,
0.08970137685537338,
-0.019867708906531334,
-0.07128917425870895,
-0.021062493324279785,
0.020754490047693253,
0.025362476706504822,
-0.02139916829764843,
0.04569176584482193,
0.11551002413034439,
-0.0062508233822882175,
0.0024714593309909105,
0.002989297267049551,
-0.0810084193944931,
-0.025893671438097954,
-0.02485731989145279,
-0.005548169370740652,
0.018335901200771332,
-0.03038882464170456,
-0.058907400816679,
-0.014273049309849739,
0.08886434882879257,
0.0184341873973608,
0.08403471112251282,
-0.02272067219018936,
-0.07732182741165161,
-0.023902449756860733,
0.02900514379143715,
0.05373977869749069,
0.0939323753118515,
-0.01514311134815216,
0.07704106718301773,
-0.09037996083498001,
0.03677916154265404,
-0.005208880174905062,
0.05063069611787796,
0.03509572893381119,
1.3563761154489594e-33,
-0.0005150420474819839,
0.01449329499155283,
0.023898951709270477,
-0.017824729904532433,
0.10562334954738617,
-0.024661120027303696,
0.028439968824386597,
0.05432519316673279,
0.02119051292538643,
0.02237757109105587,
-0.033426329493522644,
0.027712585404515266,
-0.028911350294947624,
0.03159681707620621,
-0.026838328689336777,
-0.05399014800786972,
-0.022293856367468834,
-0.017609434202313423,
0.007815350778400898,
-0.14005643129348755,
-0.00362209090963006,
-0.05366906523704529,
-0.03261428326368332,
-0.018686868250370026,
-0.044801656156778336,
0.024639751762151718,
-0.000501162139698863,
-0.02588435262441635,
0.07528424263000488,
0.037006448954343796,
-0.09114448726177216,
0.05964525789022446,
-0.09726448357105255,
-0.013571050949394703,
0.04358305037021637,
0.027340859174728394,
-0.004690423607826233,
-0.02500838041305542,
0.04574628546833992,
-0.05315970629453659,
0.013808970339596272,
0.0014161213766783476,
-0.005896779242902994,
-0.0100272661074996,
-0.05764809250831604,
0.015118907205760479,
0.045784756541252136,
0.015060191042721272,
-0.048866890370845795,
-0.04222797229886055,
-0.004897246602922678,
-0.0878579393029213,
0.01670730486512184,
-0.03449877351522446,
0.016281167045235634,
-0.04883332550525665,
0.08794792741537094,
0.006280283909291029,
0.004192550200968981,
0.07867498695850372,
0.07973459362983704,
0.10456650704145432,
-0.04251101613044739,
0.0889873281121254,
0.047990407794713974,
0.03574541211128235,
-0.03606449440121651,
0.004670570604503155,
0.10431849211454391,
-0.061095066368579865,
0.0329890213906765,
0.055113691836595535,
-0.01747599057853222,
-0.05992262810468674,
0.08586608618497849,
0.003344734199345112,
0.08884134143590927,
-0.024270590394735336,
-0.023644944652915,
-0.05615127831697464,
-0.042706992477178574,
-0.013654850423336029,
-0.05768689885735512,
-0.13166840374469757,
-0.08520647883415222,
-0.008332225494086742,
-0.01757204718887806,
-0.08969968557357788,
-0.02766532264649868,
-0.05849614366889,
-0.03187103196978569,
-0.08084923028945923,
0.01128476858139038,
0.030349666252732277,
0.025342226028442383,
-1.4596108762385536e-33,
-0.07109148800373077,
0.04619075730443001,
0.040352027863264084,
0.03032599948346615,
0.036273542791604996,
0.00035181010025553405,
-0.018406424671411514,
0.06364770233631134,
0.04322545975446701,
-0.02307332307100296,
-0.10910069197416306,
0.024568870663642883,
0.037019696086645126,
-0.021618656814098358,
0.02631915733218193,
0.002021701540797949,
0.06961143016815186,
0.0033519966527819633,
0.00865250825881958,
-0.009881848469376564,
-0.0851423442363739,
0.11645892262458801,
-0.07826586067676544,
0.04390968382358551,
-0.0460454523563385,
0.010937578044831753,
0.021808790042996407,
0.036360934376716614,
0.03670657053589821,
0.03254624083638191,
0.0036406847648322582,
-0.04998166114091873,
-0.049306951463222504,
-0.03175776079297066,
-0.0775923803448677,
0.01140794437378645,
-0.005886751692742109,
0.13489748537540436,
-0.041376836597919464,
0.05509607493877411,
0.03217576444149017,
-0.06858822703361511,
0.00887063704431057,
0.06292155385017395,
0.01760970801115036,
0.003722653491422534,
-0.07277356088161469,
-0.000667547108605504,
-0.03561432659626007,
0.0629284605383873,
0.005119148641824722,
0.058089956641197205,
-0.040749773383140564,
0.0029873340390622616,
-0.0237022265791893,
0.04380662739276886,
-0.07294551283121109,
-0.013856351375579834,
0.07477986067533493,
0.02696928195655346,
-0.0669524222612381,
-0.027291400358080864,
0.03777352720499039,
0.012198183685541153,
-0.10643323510885239,
-0.03323839232325554,
-0.10339255630970001,
-0.006000020075589418,
0.009448439814150333,
0.002533691469579935,
-0.024065736681222916,
-0.04594328626990318,
0.006737590301781893,
-0.06407072395086288,
0.0008840574300847948,
-0.0546049140393734,
0.012859120965003967,
-0.001002926961518824,
-0.00480299536138773,
-0.005561171565204859,
-0.027139538899064064,
-0.029671167954802513,
-0.04086562246084213,
0.016029058024287224,
0.014852523803710938,
-0.02303336374461651,
0.0767415389418602,
0.01521201804280281,
0.07335842400789261,
-0.011334327980875969,
0.034905701875686646,
-0.04141181334853172,
0.07136693596839905,
0.01132750604301691,
0.019614124670624733,
-4.406663123290855e-8,
-0.017471855506300926,
0.06752519309520721,
0.10436119139194489,
0.04046015813946724,
0.125070720911026,
0.03066122904419899,
0.059401415288448334,
0.06773879379034042,
-0.01058536022901535,
0.022664008662104607,
0.11050010472536087,
0.03974658623337746,
0.008739814162254333,
-0.008096079342067242,
-0.03745017200708389,
0.1455117166042328,
0.05596731975674629,
-0.03271862491965294,
0.008191956207156181,
0.03218817338347435,
-0.01805424503982067,
-0.0005725069786421955,
0.054273538291454315,
0.012285725213587284,
-0.0011320766061544418,
-0.02745763026177883,
-0.07318668067455292,
-0.09708834439516068,
0.08001907169818878,
-0.02283436246216297,
-0.007166713941842318,
0.03270474821329117,
0.04515262320637703,
0.002640519291162491,
-0.02680273726582527,
0.09599319845438004,
-0.01586596481502056,
-0.0292733833193779,
0.04926132783293724,
-0.04850093275308609,
-0.00173629866912961,
-0.021615562960505486,
0.02781546302139759,
0.022107701748609543,
-0.07015430927276611,
-0.08500795811414719,
-0.023388752713799477,
-0.02617594599723816,
-0.004925482906401157,
-0.000872197502758354,
0.005953838583081961,
-0.08882825821638107,
-0.008664015680551529,
-0.002172908978536725,
0.05463939532637596,
0.03372333198785782,
-0.010754778981208801,
-0.019748350605368614,
0.02396455593407154,
0.07294929027557373,
-0.0022728152107447386,
0.011122522875666618,
-0.07866699248552322,
0.03855917230248451
] | https://github.com/nickzhang53/symmetricdiffusers | 13 | 0 | 0 | 0 |
Language Representations Can be What Recommenders Need: Findings and Potentials | https://openreview.net/forum?id=eIJfOIMN9z | [
"Leheng Sheng",
"An Zhang",
"Yi Zhang",
"Yuxin Chen",
"Xiang Wang",
"Tat-Seng Chua"
] | Oral | Recent studies empirically indicate that language models (LMs) encode rich world knowledge beyond mere semantics, attracting significant attention across various fields.
However, in the recommendation domain, it remains uncertain whether LMs implicitly encode user preference information. Contrary to prevailing understanding that LMs and traditional recommenders learn two distinct representation spaces due to the huge gap in language and behavior modeling objectives, this work re-examines such understanding and explores extracting a recommendation space directly from the language representation space.
Surprisingly, our findings demonstrate that item representations, when linearly mapped from advanced LM representations, yield superior recommendation performance.
This outcome suggests the possible homomorphism between the advanced language representation space and an effective item representation space for recommendation, implying that collaborative signals may be implicitly encoded within LMs.
Motivated by the finding of homomorphism, we explore the possibility of designing advanced collaborative filtering (CF) models purely based on language representations without ID-based embeddings.
To be specific, we incorporate several crucial components (i.e., a multilayer perceptron (MLP), graph convolution, and contrastive learning (CL) loss function) to build a simple yet effective model, with the language representations of item textual metadata (i.e., title) as the input.
Empirical results show that such a simple model can outperform leading ID-based CF models on multiple datasets, which sheds light on using language representations for better recommendation.
Moreover, we systematically analyze this simple model and find several key features for using advanced language representations:
a good initialization for item representations, superior zero-shot recommendation abilities in new datasets, and being aware of user intention.
Our findings highlight the connection between language modeling and behavior modeling, which can inspire both natural language processing and recommender system communities. | Collaborative filtering, Language-representation-based recommendation, Language models, Language model representations | null | 5,613 | 2407.05441 | [
-0.030027316883206367,
-0.14339227974414825,
0.04748520627617836,
0.06144518405199051,
0.09097427129745483,
0.058217935264110565,
0.03191759064793587,
0.021071886643767357,
0.005595213733613491,
-0.062171123921871185,
-0.037874363362789154,
0.007463300600647926,
0.0650131106376648,
0.10080619156360626,
0.005711488891392946,
0.07778292894363403,
0.09448780119419098,
0.035147931426763535,
-0.08946433663368225,
-0.06848215311765671,
0.04676146060228348,
0.007565180771052837,
-0.007854506373405457,
-0.07144806534051895,
-0.03288838639855385,
0.045689478516578674,
-0.04802205413579941,
-0.024235470220446587,
-0.02697310596704483,
0.03145601600408554,
0.08146655559539795,
0.06602383404970169,
-0.016614574939012527,
-0.006151487585157156,
-0.08934018015861511,
0.05796588584780693,
-0.039248526096343994,
-0.0441238097846508,
-0.03965519741177559,
-0.046899229288101196,
-0.03682257607579231,
0.024616664275527,
0.032740071415901184,
0.09404360502958298,
0.02771252766251564,
-0.024976806715130806,
-0.015712538734078407,
-0.036810025572776794,
-0.059384893625974655,
0.1069013923406601,
-0.056680675595998764,
-0.0035237562842667103,
-0.000985940801911056,
0.02580379880964756,
-0.046941742300987244,
-0.034182559698820114,
-0.08511736243963242,
0.06832564622163773,
-0.013849519193172455,
-0.07444586604833603,
-0.08750982582569122,
-0.09881091117858887,
0.04545753076672554,
-0.006749385502189398,
0.02016039751470089,
-0.06909604370594025,
-0.03841715306043625,
0.010757752694189548,
-0.07299965620040894,
0.07337310165166855,
-0.0004684068262577057,
0.0714205726981163,
-0.004773988854140043,
-0.007282340433448553,
0.08160843700170517,
0.04745713248848915,
0.04424707219004631,
-0.07533546537160873,
-0.006802961230278015,
-0.06145502254366875,
0.013768183067440987,
0.060044921934604645,
0.012771492823958397,
-0.033121053129434586,
0.12735845148563385,
0.005199038423597813,
-0.006532721221446991,
-0.0616905651986599,
-0.022463595494627953,
0.00421905517578125,
-0.017207492142915726,
-0.03035632148385048,
-0.0645400732755661,
-0.0253842044621706,
0.058304429054260254,
0.055196795612573624,
0.02117984928190708,
-0.017431532964110374,
-0.0455034114420414,
0.047658275812864304,
-0.001080091344192624,
0.08656751364469528,
-0.00910324789583683,
-0.05230362340807915,
-0.016135158017277718,
-0.0484548881649971,
-0.00012467842316254973,
0.08059655129909515,
0.010274563916027546,
-0.05751867592334747,
-0.06646468490362167,
0.0995437428355217,
0.014121764339506626,
0.04555331543087959,
0.04458591341972351,
-0.04674265533685684,
0.11178353428840637,
-0.06253284960985184,
0.09197460114955902,
0.05599013715982437,
-0.06783975660800934,
0.05376935377717018,
0.04777483269572258,
-0.028986549004912376,
-0.022456249222159386,
0.048504188656806946,
-0.04179592803120613,
1.423800267360218e-33,
0.016004225239157677,
-0.008052989840507507,
-0.020845551043748856,
0.04572161287069321,
0.01784958876669407,
-0.04424625635147095,
-0.018516767770051956,
0.06098930165171623,
-0.04265932738780975,
-0.03990761935710907,
-0.01177307777106762,
0.09793318808078766,
-0.014482286758720875,
0.12855759263038635,
0.008514164946973324,
0.052509795874357224,
-0.032340340316295624,
0.006368111819028854,
-0.05117454379796982,
-0.11585298180580139,
-0.031084207817912102,
0.04068468511104584,
-0.012424806132912636,
-0.042209744453430176,
-0.028446989133954048,
-0.026045938953757286,
0.03560296818614006,
-0.08194548636674881,
0.02798490971326828,
-0.013237867504358292,
-0.11094779521226883,
0.0040864525362849236,
0.024759728461503983,
-0.01979019120335579,
-0.025936095044016838,
-0.02027275785803795,
-0.044321201741695404,
-0.033928632736206055,
0.052253659814596176,
-0.08891484886407852,
0.03683917969465256,
-0.0011145356111228466,
-0.019539132714271545,
0.005191255826503038,
-0.07804247736930847,
-0.05672014132142067,
-0.018387213349342346,
-0.021720943972468376,
-0.016219893470406532,
-0.06937571614980698,
0.06584897637367249,
-0.05233834683895111,
-0.0034395873080939054,
0.058594461530447006,
0.0177927128970623,
-0.04339294880628586,
-0.02522152103483677,
0.03541279584169388,
0.006169956177473068,
-0.048674389719963074,
-0.007156411185860634,
0.058607254177331924,
0.06484818458557129,
0.023450035601854324,
0.013154077343642712,
0.0012630654964596033,
0.033096831291913986,
0.06681128591299057,
0.05463899299502373,
0.010648226365447044,
-0.018619447946548462,
0.04074082896113396,
0.027354376390576363,
-0.04236253350973129,
0.012615072540938854,
-0.007928675971925259,
-0.001052384846843779,
-0.1434137225151062,
0.032458189874887466,
0.00496722012758255,
0.01896810159087181,
-0.09000056236982346,
0.04568050429224968,
-0.08151290565729141,
-0.09289301931858063,
-0.023865222930908203,
0.04573478922247887,
-0.1282755434513092,
0.053961630910634995,
-0.09568449854850769,
-0.026367075741291046,
-0.013595296069979668,
0.01172690000385046,
-0.01011310052126646,
0.03823867812752724,
-2.3770077180870494e-33,
-0.048900727182626724,
0.002927416702732444,
0.08642260730266571,
0.030018985271453857,
-0.026158131659030914,
-0.08003613352775574,
-0.012266027741134167,
-0.0670088455080986,
0.004725282080471516,
0.004852591548115015,
-0.04652496054768562,
-0.11202646791934967,
0.04431210085749626,
0.019003644585609436,
0.12354909628629684,
-0.0005115363164804876,
-0.023515772074460983,
0.07452886551618576,
0.046074870973825455,
0.0015133186243474483,
0.015127169899642467,
0.08131489902734756,
-0.03876755014061928,
0.05551321431994438,
-0.0386015884578228,
0.05107738450169563,
0.004787190817296505,
-0.00803442019969225,
0.03239148482680321,
-0.04660779610276222,
-0.03549840301275253,
-0.029699355363845825,
0.008339324034750462,
-0.0747823640704155,
-0.045209553092718124,
0.06770016252994537,
-0.021663228049874306,
-0.002964380197227001,
-0.04964382201433182,
0.04371367022395134,
0.05543763190507889,
-0.02218586392700672,
-0.053853586316108704,
0.004550260026007891,
0.002670754212886095,
0.041476771235466,
0.0012418278492987156,
-0.02416480891406536,
-0.009351410903036594,
-0.022329119965434074,
0.021298423409461975,
0.05862424150109291,
-0.005746809765696526,
0.016427531838417053,
-0.07371310144662857,
-0.022743798792362213,
-0.00101497583091259,
-0.04413127526640892,
0.09201277792453766,
0.027066588401794434,
-0.00416464451700449,
-0.04875829815864563,
0.037320997565984726,
-0.09851516038179398,
-0.05615534633398056,
-0.07323849201202393,
0.016550172120332718,
0.0005113657098263502,
-0.023684589192271233,
0.053170159459114075,
-0.009531778283417225,
-0.0006127524538896978,
-0.06522579491138458,
0.02196510136127472,
0.011315787211060524,
-0.021873023360967636,
0.06800908595323563,
-0.033771008253097534,
-0.04321882128715515,
-0.026438681408762932,
0.06877221167087555,
0.008281986229121685,
-0.001453000120818615,
-0.009715608321130276,
0.14788900315761566,
0.0176312904804945,
-0.016405358910560608,
0.048000846058130264,
-0.04291769489645958,
-0.0474364310503006,
0.05216880515217781,
-0.030445508658885956,
-0.04986187815666199,
0.04311253875494003,
-0.03140466287732124,
-4.3269075433727266e-8,
-0.10869420319795609,
0.019150665029883385,
0.011002696119248867,
0.07019995898008347,
0.010444147512316704,
-0.06531413644552231,
-0.034577228128910065,
0.10728706419467926,
-0.002527398755773902,
-0.012885850854218006,
0.06534120440483093,
0.01391760352998972,
-0.05764109641313553,
-0.09329266846179962,
0.02399100922048092,
0.04410518333315849,
0.05706625059247017,
-0.005424422677606344,
0.04679865017533302,
0.03820202872157097,
0.01661457121372223,
0.013244250789284706,
0.010744819417595863,
-0.015498093329370022,
0.007012247573584318,
0.029580477625131607,
0.05114968866109848,
0.08446172624826431,
0.08393663913011551,
-0.00890349317342043,
-0.062409911304712296,
0.02955092489719391,
0.0427204929292202,
-0.047065600752830505,
-0.026045726612210274,
0.0823378935456276,
0.006677123252302408,
-0.08161067217588425,
-0.06898010522127151,
-0.008932282216846943,
0.07668750733137131,
0.10276416689157486,
0.021035324782133102,
0.024137523025274277,
-0.012149704620242119,
0.015059933997690678,
0.012336467392742634,
-0.10341307520866394,
-0.005767821334302425,
0.031164294108748436,
-0.004193610046058893,
-0.05849343538284302,
0.055159565061330795,
0.03736865147948265,
0.07590629905462265,
-0.04594473913311958,
-0.030575130134820938,
0.0024471229407936335,
0.05448450520634651,
0.01969040557742119,
0.02120756171643734,
0.045985348522663116,
-0.027217412367463112,
0.07096939533948898
] | https://github.com/lehengthu/alpharec | 69 | 0 | 0 | 0 |
HiRA: Parameter-Efficient Hadamard High-Rank Adaptation for Large Language Models | https://openreview.net/forum?id=TwJrTz9cRS | [
"Qiushi Huang",
"Tom Ko",
"Zhan Zhuang",
"Lilian Tang",
"Yu Zhang"
] | Oral | We propose Hadamard High-Rank Adaptation (HiRA), a parameter-efficient fine-tuning (PEFT) method that enhances the adaptability of Large Language Models (LLMs). While Low-rank Adaptation (LoRA) is widely used to reduce resource demands, its low-rank updates may limit its expressiveness for new tasks. HiRA addresses this by using a Hadamard product to retain high-rank update parameters, improving the model capacity. Empirically, HiRA outperforms LoRA and its variants on several tasks, with extensive ablation studies validating its effectiveness. Our code is available at https://github.com/hqsiswiliam/hira. | Parametric-efficient fine-tuning, Large Language Model | null | 5,572 | null | [
-0.04082508385181427,
-0.10737151652574539,
0.026272781193256378,
0.04809293523430824,
0.028000859543681145,
0.047559335827827454,
-0.039197083562612534,
-0.040264274924993515,
-0.02319495566189289,
-0.037529051303863525,
-0.08698543906211853,
-0.022605931386351585,
0.06139867380261421,
0.0071058375760912895,
0.020134447142481804,
0.019345108419656754,
0.034691303968429565,
-0.0034358231350779533,
-0.06961531192064285,
-0.07301215827465057,
0.03809625282883644,
0.07514151185750961,
0.04974071681499481,
0.0025933876167982817,
0.08861903101205826,
-0.036447081714868546,
-0.042117055505514145,
0.010114713571965694,
-0.00540292588993907,
-0.09950854629278183,
-0.009481817483901978,
0.04633137583732605,
0.04516742005944252,
0.0016551479930058122,
-0.05305453762412071,
0.06387274712324142,
-0.08300209045410156,
0.019445203244686127,
-0.01331987977027893,
-0.010867740027606487,
0.017632944509387016,
0.04654442146420479,
0.012296652421355247,
0.014406376518309116,
0.07639238983392715,
-0.04697219282388687,
-0.022841308265924454,
-0.06141088157892227,
-0.00957769900560379,
0.0734187662601471,
-0.08078934997320175,
-0.05368325859308243,
-0.03194214403629303,
0.06886051595211029,
-0.030278267338871956,
-0.049508072435855865,
0.04130629077553749,
0.0026964470744132996,
0.03510315343737602,
-0.09030549973249435,
-0.04438696429133415,
-0.03969992697238922,
0.0027980811428278685,
-0.040307268500328064,
0.01817932166159153,
-0.08724064379930496,
0.0381791926920414,
0.0021474368404597044,
-0.044057998806238174,
0.04567845165729523,
-0.0878751128911972,
0.06218045949935913,
-0.005771816708147526,
0.0653209239244461,
0.009814613498747349,
0.014509109780192375,
0.03704602271318436,
0.0025630949530750513,
0.09483213722705841,
0.02772330678999424,
0.020933162420988083,
-0.008232316002249718,
0.03294990584254265,
-0.024922896176576614,
0.12045804411172867,
0.018509315326809883,
0.007353867869824171,
-0.04192070662975311,
-0.03578728064894676,
-0.038046885281801224,
0.047609541565179825,
-0.05301035940647125,
0.10429719090461731,
0.050699908286333084,
-0.07642170041799545,
0.07198873907327652,
0.014323301613330841,
-0.0238714050501585,
-0.014812586829066277,
0.08605669438838959,
0.04467495158314705,
0.013655136339366436,
-0.025958744809031487,
-0.026928935199975967,
-0.04048538953065872,
-0.05199503153562546,
0.0572427362203598,
0.027788285166025162,
-0.00532951857894659,
-0.11037060618400574,
0.014972840435802937,
0.05065255984663963,
0.012471163645386696,
0.033861950039863586,
-0.04845966398715973,
0.006581324152648449,
0.028216272592544556,
-0.09727668762207031,
-0.03620954230427742,
0.004664959851652384,
-0.0073795258067548275,
-0.011071957647800446,
0.07974175363779068,
0.03014986962080002,
0.020411111414432526,
0.048707664012908936,
-0.03349687531590462,
2.6629197192887004e-33,
0.10249537974596024,
0.06583287566900253,
-0.003565335413441062,
0.011794496327638626,
-0.008588256314396858,
-0.033549223095178604,
0.04584983363747597,
-0.00997670367360115,
-0.08011163771152496,
-0.08031637221574783,
-0.05028247833251953,
0.07323265820741653,
-0.03129316866397858,
0.026541218161582947,
0.016132891178131104,
-0.06969715654850006,
0.036701615899801254,
0.08806929737329483,
0.0016493112780153751,
-0.06346473097801208,
0.03214140608906746,
0.0022405863273888826,
0.016656655818223953,
-0.06380220502614975,
0.04109776392579079,
0.07480205595493317,
0.0750068724155426,
-0.05467769876122475,
-0.024722063913941383,
0.020965781062841415,
-0.031955644488334656,
-0.06032931059598923,
0.01992580108344555,
0.016122201457619667,
-0.04760231450200081,
-0.10413172841072083,
-0.07142268866300583,
-0.02497635968029499,
-0.004274321254342794,
-0.06729432940483093,
0.0018593902932479978,
0.043670255690813065,
0.00950311403721571,
-0.038268815726041794,
-0.029354380443692207,
-0.013744673691689968,
-0.02183712087571621,
0.030469536781311035,
-0.060243286192417145,
-0.05105270817875862,
0.01748105324804783,
0.026634570211172104,
0.012766474857926369,
-0.00944725051522255,
-0.02778666839003563,
-0.004949870985001326,
0.055153317749500275,
0.04474080726504326,
0.03981423005461693,
0.03554544225335121,
0.030238892883062363,
0.02326176129281521,
-0.0018296446651220322,
0.07135279476642609,
0.09444093704223633,
-0.019246922805905342,
0.007736257743090391,
-0.007098563946783543,
0.07034385204315186,
0.03793317452073097,
0.02274351380765438,
0.0018394275102764368,
-0.009194741025567055,
0.031050357967615128,
-0.04510516673326492,
-0.05422249063849449,
0.004313548561185598,
-0.038425594568252563,
0.00865281093865633,
-0.012089277617633343,
0.042653340846300125,
0.008257757872343063,
0.028380874544382095,
-0.07113785296678543,
-0.009280731901526451,
-0.07898727804422379,
0.022757435217499733,
-0.08711326122283936,
-0.04281926527619362,
0.018070289865136147,
0.01797119341790676,
0.018699316307902336,
0.012248226441442966,
-0.05577905476093292,
0.06089853122830391,
-7.439987058859905e-34,
0.003195703960955143,
-0.04867800697684288,
-0.03290828317403793,
0.09791377186775208,
-0.01435068529099226,
-0.0614817813038826,
0.01540434267371893,
0.09762343764305115,
-0.0750027596950531,
-0.03649976849555969,
0.020973961800336838,
-0.09085965156555176,
0.02468874305486679,
-0.046035174280405045,
0.07792773097753525,
-0.010346734896302223,
-0.0579870268702507,
0.005392179358750582,
-0.012883239425718784,
0.06658913940191269,
-0.02473258413374424,
0.08438779413700104,
-0.041863713413476944,
0.054911788552999496,
-0.030322642996907234,
0.002875133417546749,
-0.00253102695569396,
0.023009393364191055,
-0.039233606308698654,
-0.06185323745012283,
-0.08330689370632172,
0.010766107589006424,
-0.08921845257282257,
0.04004259034991264,
-0.0652957633137703,
0.029877297580242157,
0.006029762327671051,
-0.02818232774734497,
-0.025201600044965744,
0.13517433404922485,
0.028209427371621132,
0.024948883801698685,
-0.0703832134604454,
-0.0331580750644207,
0.0023874687030911446,
0.051491860300302505,
-0.1487744003534317,
-0.04116265848278999,
0.009600693359971046,
-0.005108469165861607,
0.02870260737836361,
-0.02528122253715992,
0.00813185703009367,
0.008714285679161549,
0.04570398107171059,
-0.053469665348529816,
-0.05924621596932411,
-0.06503748148679733,
-0.04355849698185921,
0.006729750894010067,
-0.03321244940161705,
-0.03684002906084061,
0.06535083055496216,
-0.08832424879074097,
0.050871022045612335,
0.042345721274614334,
-0.03865616023540497,
-0.04477594792842865,
-0.010547108016908169,
0.007811953313648701,
0.07654504477977753,
-0.05532745271921158,
-0.03013385646045208,
0.06168590858578682,
-0.045777030289173126,
-0.013533614575862885,
0.0000467542813566979,
-0.054079752415418625,
0.004012869670987129,
-0.06486400961875916,
-0.016850262880325317,
0.07042897492647171,
0.012376233004033566,
0.05425749719142914,
0.027554431930184364,
0.11029841005802155,
0.10373073816299438,
0.08108597248792648,
0.022786755114793777,
0.005438356194645166,
-0.010328000411391258,
0.02554258145391941,
0.05422605574131012,
0.005366397090256214,
-0.0666201040148735,
-4.123777941344997e-8,
-0.09623857587575912,
0.004481514450162649,
-0.07652568817138672,
0.131749227643013,
-0.007236691191792488,
-0.060542937368154526,
-0.05222313106060028,
0.037793152034282684,
0.032952070236206055,
0.02342366613447666,
0.00006138565368019044,
-0.032100025564432144,
0.03316906467080116,
0.0191566850990057,
0.07945059984922409,
0.06552795320749283,
0.031797412782907486,
0.14319223165512085,
-0.02326769195497036,
-0.04990723729133606,
0.016238536685705185,
0.16023966670036316,
0.033098310232162476,
0.061287686228752136,
0.08849405497312546,
-0.06406951695680618,
-0.020330673083662987,
0.05791444703936577,
0.11195050925016403,
-0.006481640972197056,
0.005585185252130032,
-0.0375358946621418,
-0.03232773020863533,
-0.019500408321619034,
-0.006647062487900257,
0.06750722974538803,
0.0013238341780379415,
-0.031530629843473434,
-0.022385314106941223,
0.042629897594451904,
0.10033947974443436,
0.06432969123125076,
-0.08379470556974411,
0.04892420023679733,
0.017978526651859283,
-0.028714105486869812,
-0.07246286422014236,
-0.12988673150539398,
0.024199528619647026,
-0.05977252125740051,
0.03941763937473297,
-0.05457311496138573,
0.019821736961603165,
-0.02296624705195427,
0.049996063113212585,
0.06641688197851181,
-0.00406979164108634,
-0.0024721482768654823,
0.10116873681545258,
0.040375467389822006,
0.08908534049987793,
-0.022393682971596718,
-0.008144349791109562,
0.03559538349509239
] | 0 | 0 | 0 | 0 |
|
A Theoretically-Principled Sparse, Connected, and Rigid Graph Representation of Molecules | https://openreview.net/forum?id=OIvg3MqWX2 | [
"Shih-Hsin Wang",
"Yuhao Huang",
"Justin M. Baker",
"Yuan-En Sun",
"Qi Tang",
"Bao Wang"
] | Oral | Graph neural networks (GNNs) -- learn graph representations by exploiting graph's sparsity, connectivity, and symmetries -- have become indispensable for learning geometric data like molecules. However, the most used graphs (e.g., radial cutoff graphs) in molecular modeling lack theoretical guarantees for achieving connectivity and sparsity simultaneously, which are essential for the performance and scalability of GNNs. Furthermore, existing widely used graph construction methods for molecules lack rigidity, limiting GNNs' ability to exploit graph nodes' spatial arrangement. In this paper, we introduce a new hyperparameter-free graph construction of molecules and beyond with sparsity, connectivity, and rigidity guarantees. Remarkably, our method consistently generates connected and sparse graphs with the edge-to-node ratio being bounded above by 3. Our graphs' rigidity guarantees that edge distances and dihedral angles are sufficient to uniquely determine the general spatial arrangements of atoms. We substantiate the effectiveness and efficiency of our proposed graphs in various molecular modeling benchmarks. Code is available at \url{https://github.com/Utah-Math-Data-Science/UnitSphere}. | Graph representation, sparsity, connectivity, rigidity, molecules, learning | We introduce a new sparse, connected, and rigid graph representation for molecules. | 5,512 | null | [
-0.031822193413972855,
-0.09568814188241959,
-0.040629059076309204,
-0.014082053676247597,
-0.0019865846261382103,
0.007665673270821571,
-0.05395674332976341,
0.005668822675943375,
0.024866929277777672,
-0.015519760549068451,
0.013155204243957996,
0.0456649586558342,
-0.050229474902153015,
0.0917518213391304,
-0.00535255391150713,
0.02009868435561657,
0.001750839757733047,
0.03864370658993721,
-0.06383456289768219,
0.018224338069558144,
0.010292202234268188,
0.0026639909483492374,
0.04447806999087334,
0.03993501886725426,
0.09929592162370682,
0.03507760912179947,
0.06104609742760658,
0.04749003052711487,
0.026354199275374413,
0.005661849398165941,
0.11799226701259613,
-0.0666753351688385,
-0.07285217195749283,
0.00871499814093113,
0.041466522961854935,
-0.007786561269313097,
-0.05090513452887535,
0.06672481447458267,
0.004362907260656357,
-0.005360043607652187,
0.0030536435078829527,
0.03769140690565109,
-0.002139364369213581,
0.08775133639574051,
0.06398360431194305,
-0.009028390981256962,
-0.07724682241678238,
0.024879146367311478,
0.0710696130990982,
-0.017250465229153633,
-0.051436182111501694,
-0.02517162263393402,
-0.053229428827762604,
-0.007550002541393042,
0.03368169069290161,
-0.0017490037716925144,
-0.07937008887529373,
-0.03284589946269989,
-0.02586919628083706,
-0.021899711340665817,
0.10279826074838638,
-0.11041916906833649,
0.02911999449133873,
-0.011783941648900509,
-0.030615808442234993,
0.03692782297730446,
0.09318315982818604,
0.0627850666642189,
0.025437481701374054,
0.0842769593000412,
-0.015007410198450089,
0.04093655198812485,
-0.07235276699066162,
0.061510633677244186,
0.027961062267422676,
0.014653706923127174,
0.03885625675320625,
0.036220937967300415,
0.022354239597916603,
-0.05123323202133179,
0.03256410360336304,
0.03062707930803299,
0.0355348214507103,
0.056114718317985535,
0.0057479082606732845,
-0.03151456639170647,
0.0189213864505291,
0.040937814861536026,
-0.010496187955141068,
-0.012687419541180134,
-0.009753809310495853,
0.01830485090613365,
-0.018270105123519897,
-0.09716245532035828,
-0.06806794553995132,
0.008123277686536312,
0.08777087926864624,
0.005776762031018734,
-0.029178645461797714,
0.08770616352558136,
0.009867551736533642,
-0.05527212843298912,
0.02547246217727661,
-0.014059905894100666,
-0.007201910484582186,
0.027178406715393066,
-0.02852553129196167,
0.056866519153118134,
0.013043280690908432,
-0.04474712535738945,
0.008387631736695766,
0.10015439987182617,
-0.03359109163284302,
0.04671424254775047,
-0.0288112573325634,
-0.042760927230119705,
0.02325582690536976,
-0.07682940363883972,
0.07625727355480194,
0.11982491612434387,
-0.0966549664735794,
0.009012514725327492,
-0.05133054777979851,
0.015748467296361923,
-0.0035175702068954706,
-0.016738448292016983,
-0.10787658393383026,
1.0508961740679483e-33,
0.09406541287899017,
0.09937378019094467,
0.09612306952476501,
0.0016513594891875982,
0.03533821180462837,
-0.01010562852025032,
-0.0162881501019001,
0.014466841705143452,
-0.03763168305158615,
0.033394183963537216,
-0.13009321689605713,
-0.048955801874399185,
0.042390160262584686,
0.01999022252857685,
-0.01333634927868843,
-0.010838965885341167,
0.03926412761211395,
-0.0069849626161158085,
0.006269301753491163,
-0.08065249770879745,
-0.06042720377445221,
-0.04339878261089325,
0.028867198154330254,
0.01581842266023159,
-0.08011657744646072,
0.03468496352434158,
-0.04107533022761345,
0.0019480942282825708,
-0.019499219954013824,
-0.010900399647653103,
-0.10178656876087189,
0.026732679456472397,
-0.0042692553251981735,
0.07413896918296814,
0.06363022327423096,
-0.003515013959258795,
-0.0509333573281765,
-0.08737645298242569,
0.04919956251978874,
-0.12373051047325134,
0.012263240292668343,
0.039805956184864044,
0.05257092043757439,
-0.019823309034109116,
-0.02267732471227646,
0.05002398416399956,
-0.004387132357805967,
-0.03802623599767685,
-0.0551738515496254,
0.007082559634000063,
-0.0254177525639534,
0.021598801016807556,
0.014757662080228329,
0.036999136209487915,
0.023704249411821365,
-0.02489873208105564,
0.045368898659944534,
0.016231123358011246,
0.03191942721605301,
0.040433451533317566,
-0.0010683906730264425,
0.02061006799340248,
-0.021808680146932602,
-0.012164240702986717,
0.013003753498196602,
0.029026620090007782,
-0.1084449291229248,
0.013155486434698105,
0.09370838105678558,
0.01894773170351982,
-0.0251940730959177,
0.08889993280172348,
-0.02986367978155613,
-0.04018155485391617,
0.004357555415481329,
0.06934249401092529,
0.003407040610909462,
-0.14615149796009064,
-0.02389204315841198,
0.028751470148563385,
-0.02291610650718212,
-0.07020092010498047,
0.01624329201877117,
-0.04584624618291855,
-0.10214773565530777,
-0.10392706096172333,
0.031160783022642136,
-0.058060258626937866,
0.04894532263278961,
-0.08208095282316208,
-0.053094953298568726,
-0.10711102932691574,
0.0019711339846253395,
-0.03556257486343384,
0.010826067067682743,
-1.1925576615149307e-33,
-0.006147488951683044,
0.07392474263906479,
0.05721783638000488,
0.05750896781682968,
0.030964767560362816,
-0.024619359523057938,
-0.07079783082008362,
-0.03137224167585373,
-0.03349914029240608,
-0.011611406691372395,
0.01892329752445221,
-0.029050804674625397,
0.0023685675114393234,
-0.05007849633693695,
0.00904983188956976,
0.07439690828323364,
-0.06173758953809738,
-0.035475876182317734,
0.007822481915354729,
0.01381379272788763,
-0.006694058421999216,
0.02852206863462925,
-0.01780409924685955,
0.04822273179888725,
0.018266772851347923,
0.025794290006160736,
0.027653053402900696,
-0.052341412752866745,
0.022243281826376915,
0.07786652445793152,
-0.006230010185390711,
-0.052126843482255936,
-0.10856305062770844,
-0.04417337849736214,
-0.00834624096751213,
0.05487311631441116,
-0.003959387540817261,
0.026750804856419563,
0.015291305258870125,
-0.048510123044252396,
-0.065854012966156,
-0.02001931518316269,
-0.09889113903045654,
0.07798852026462555,
0.06249498948454857,
0.06430751830339432,
0.022394491359591484,
0.017301443964242935,
-0.06925901025533676,
0.02220364287495613,
-0.050311628729104996,
0.04701622575521469,
0.07431073486804962,
-0.0346415713429451,
-0.013907352462410927,
-0.08025416731834412,
-0.016867518424987793,
0.019949747249484062,
0.07562222331762314,
-0.03178739175200462,
-0.07390879839658737,
-0.12602323293685913,
-0.00928487814962864,
0.024312708526849747,
-0.06829380244016647,
-0.0692921057343483,
0.020247194916009903,
-0.001279659802094102,
0.013453695923089981,
0.05010886862874031,
-0.016262440010905266,
0.005651806481182575,
0.06485913693904877,
0.03429948166012764,
0.058942921459674835,
0.003283303463831544,
0.053971849381923676,
-0.04306687042117119,
-0.040332213044166565,
-0.009363928809762001,
0.026485878974199295,
0.030902421101927757,
-0.05024223029613495,
0.03570348396897316,
0.07341045141220093,
0.06496862322092056,
0.0473492331802845,
0.03878571465611458,
0.0019178506918251514,
0.03315731883049011,
0.017850155010819435,
0.07782816886901855,
-0.0389760360121727,
0.06306542456150055,
0.03827599436044693,
-3.9109629312861216e-8,
-0.017218971624970436,
0.007284916937351227,
0.044353533536195755,
-0.07031402736902237,
0.06073607876896858,
0.049548715353012085,
0.13414980471134186,
0.08421153575181961,
-0.06054330989718437,
0.03391804173588753,
0.08979775756597519,
-0.009983869269490242,
-0.07241032272577286,
-0.015118751674890518,
-0.03041755221784115,
0.06096780300140381,
0.009851057082414627,
-0.011550884693861008,
-0.003317104885354638,
0.08519083261489868,
-0.05875864624977112,
0.011805775575339794,
-0.014074662700295448,
0.025529038161039352,
0.08177762478590012,
-0.10950896888971329,
-0.028676653280854225,
0.03156474232673645,
0.020761489868164062,
-0.001645342563278973,
-0.036810874938964844,
0.007049792446196079,
0.12475840002298355,
0.05579933524131775,
-0.04352475702762604,
0.07045552879571915,
-0.00708931777626276,
0.026322005316615105,
-0.1357697695493698,
-0.05665639042854309,
-0.028032802045345306,
0.07967827469110489,
-0.02062106877565384,
-0.02701801434159279,
0.011933700181543827,
0.060544464737176895,
0.05849625915288925,
0.0647631362080574,
0.006726703606545925,
-0.02388511784374714,
-0.023062003776431084,
-0.03594110906124115,
-0.053716473281383514,
-0.06196055933833122,
-0.054486341774463654,
0.03547096252441406,
-0.08583812415599823,
-0.027651969343423843,
0.05585554614663124,
-0.04476430267095566,
-0.05321529135107994,
-0.02962818555533886,
-0.07964085787534714,
-0.0263175368309021
] | 0 | 0 | 0 | 0 |
|
How much of my dataset did you use? Quantitative Data Usage Inference in Machine Learning | https://openreview.net/forum?id=EUSkm2sVJ6 | [
"Yao Tong",
"Jiayuan Ye",
"Sajjad Zarifzadeh",
"Reza Shokri"
] | Oral | How much of my data was used to train a machine learning model? This is a critical question for data owners assessing the risk of unauthorized usage of their data to train models. However, previous work mistakenly treats this as a binary problem—inferring whether all-or-none or any-or-none of the data was used—which is fragile when faced with real, non-binary data usage risks. To address this, we propose a fine-grained analysis called Dataset Usage Cardinality Inference (DUCI), which estimates the exact proportion of data used. Our algorithm, leveraging debiased membership guesses, matches the performance of the optimal MLE approach (with a maximum error <0.1) but with significantly lower (e.g., $300 \times$ less) computational cost. | Machine Learning, Privacy, Dataset Usage Inference, Dataset Ownership, Membership Inference Attack, Dataset Copyright | The first method to quantitatively and non-binarily answer the question ``How much has a dataset been used in the training of a given model?'' | 5,454 | null | [
0.024088704958558083,
-0.04266609624028206,
-0.0009394037188030779,
0.044779203832149506,
0.07855402678251266,
-0.02839662693440914,
0.05867154896259308,
0.038219764828681946,
0.03143744170665741,
0.010179117321968079,
-0.018586745485663414,
-0.0326526053249836,
0.07901076227426529,
-0.02028135396540165,
-0.054117586463689804,
0.01692843809723854,
0.0642506331205368,
-0.040301524102687836,
-0.06637268513441086,
-0.0511186309158802,
-0.03689180314540863,
-0.012612704187631607,
-0.010258406400680542,
0.07187251001596451,
-0.020455682650208473,
-0.01927589252591133,
0.05500689521431923,
0.0024863816797733307,
-0.040849700570106506,
-0.028689928352832794,
-0.011107238940894604,
0.055834367871284485,
0.03581621125340462,
0.017364785075187683,
0.022311942651867867,
-0.02536487579345703,
-0.010006183758378029,
-0.02822774462401867,
0.03272993490099907,
0.0190407857298851,
-0.002265969756990671,
-0.0014005211414769292,
0.016306260600686073,
0.05965588241815567,
0.04663562774658203,
0.00943036563694477,
0.012187592685222626,
-0.006365554872900248,
0.0006767254672013223,
0.0787261500954628,
0.01245065126568079,
-0.0400908961892128,
0.07459397614002228,
-0.009511422365903854,
-0.08029810339212418,
-0.1251387596130371,
0.026581846177577972,
-0.016681263223290443,
0.0014329085825011134,
0.02817991003394127,
-0.07204444706439972,
-0.04869703948497772,
-0.03940499946475029,
-0.009964882396161556,
-0.07391144335269928,
0.049575429409742355,
-0.08666432648897171,
0.012535267509520054,
0.006485299207270145,
0.04260456934571266,
0.02511250041425228,
0.04496888071298599,
-0.11530552059412003,
0.02832142449915409,
-0.015971163287758827,
0.040739841759204865,
0.08705127984285355,
-0.02526366151869297,
0.11959079653024673,
-0.09495380520820618,
-0.04355330020189285,
-0.04732217639684677,
0.010715642012655735,
0.06245590001344681,
0.06840498745441437,
-0.018343985080718994,
0.006419727113097906,
-0.011577890254557133,
0.046973712742328644,
0.005941981449723244,
-0.066941998898983,
0.02653457410633564,
0.06902866810560226,
0.02024572715163231,
0.0775090903043747,
0.07006937265396118,
-0.05348128825426102,
-0.0019964089151471853,
-0.03475043550133705,
0.089292511343956,
-0.12406708300113678,
0.10360603779554367,
0.01049015112221241,
-0.04869477450847626,
0.0294355358928442,
-0.009260354563593864,
0.0635150596499443,
-0.04644237831234932,
0.03941874951124191,
-0.04279332235455513,
0.04433584213256836,
0.016628842800855637,
-0.09104268997907639,
-0.01300483476370573,
0.09596386551856995,
-0.03614755719900131,
-0.06310521811246872,
0.033576980233192444,
-0.11660192161798477,
0.12041068077087402,
-0.02657221257686615,
0.06511750817298889,
0.05262073501944542,
-0.002813829807564616,
-0.07003486156463623,
-0.05418657138943672,
-0.12209942936897278,
1.9319075168321595e-33,
-0.004474703688174486,
-0.01320852804929018,
0.04643759876489639,
-0.022725405171513557,
0.024459293112158775,
-0.08021677285432816,
-0.02033325284719467,
0.06526435911655426,
-0.010838232934474945,
-0.00788951013237238,
-0.0202170517295599,
0.08533523976802826,
-0.04210390895605087,
-0.006119545083492994,
0.12757930159568787,
0.10948527604341507,
0.010318273678421974,
0.03251179680228233,
-0.06577359139919281,
-0.031412556767463684,
0.034537047147750854,
-0.09075760096311569,
-0.025658192113041878,
-0.0008137907134369016,
0.00501917814835906,
-0.06854542344808578,
-0.017142686992883682,
-0.010675910860300064,
0.010387244634330273,
0.029987284913659096,
0.00767137948423624,
0.03014633245766163,
0.034794874489307404,
0.033236924558877945,
0.03898137807846069,
0.020495854318141937,
0.010007348842918873,
0.051356129348278046,
0.005334365181624889,
0.07756698131561279,
-0.013736804015934467,
-0.04112239181995392,
0.10675008594989777,
-0.032612819224596024,
-0.09377319365739822,
-0.04861277714371681,
-0.012174505740404129,
-0.0896196961402893,
-0.05547378957271576,
0.03289046511054039,
0.04310925677418709,
-0.0015165178338065743,
-0.05080709606409073,
0.016632717102766037,
-0.11436177790164948,
-0.00005093361687613651,
-0.03186063468456268,
0.030201848596334457,
0.1092023104429245,
0.04452750086784363,
-0.03574378788471222,
0.010018701665103436,
0.10312578827142715,
0.0441504642367363,
-0.04910646751523018,
0.009955157525837421,
-0.03636102378368378,
-0.050059910863637924,
0.0037506252992898226,
0.02083112671971321,
-0.025037236511707306,
-0.028855744749307632,
0.012592003680765629,
-0.05382249131798744,
0.03917638957500458,
-0.016955764964222908,
0.054152194410562515,
0.018412116914987564,
-0.004924665670841932,
0.04487380012869835,
0.06906171888113022,
-0.06374973058700562,
0.08699951320886612,
-0.031060880050063133,
-0.02291945181787014,
0.005673599895089865,
0.047915391623973846,
-0.018800867721438408,
-0.07407937943935394,
0.06043088063597679,
-0.06076207756996155,
0.004601921886205673,
-0.062107741832733154,
-0.05287793651223183,
-0.019789788872003555,
-1.7536316762444667e-33,
-0.07443083822727203,
0.05094388499855995,
0.06904006004333496,
0.02890937589108944,
-0.02934315800666809,
-0.012205616571009159,
-0.024818580597639084,
-0.015564635396003723,
0.07178518921136856,
-0.036328427493572235,
-0.02521783672273159,
0.010837579146027565,
0.042360637336969376,
0.01848308928310871,
-0.08105625957250595,
0.037923458963632584,
-0.07070975005626678,
-0.022192588075995445,
-0.008891270495951176,
-0.0065993801690638065,
-0.03751774877309799,
0.07674326002597809,
-0.055969368666410446,
-0.027873089537024498,
-0.001444202964194119,
0.021305356174707413,
-0.12253747880458832,
0.02889351360499859,
0.05919213965535164,
-0.008068729192018509,
-0.07784456759691238,
0.014091399498283863,
0.00986381247639656,
-0.0693453922867775,
-0.055314552038908005,
-0.013860918581485748,
0.07115469872951508,
0.08616780489683151,
0.022822322323918343,
0.005095425993204117,
-0.036383792757987976,
0.04192386940121651,
-0.12253676354885101,
0.00991648156195879,
-0.003305060788989067,
-0.05092766880989075,
0.00026110123144462705,
-0.06660635024309158,
0.0974467322230339,
-0.005729556083679199,
0.024778379127383232,
0.023064585402607918,
0.031063416972756386,
0.03569633141160011,
-0.0035056958440691233,
-0.012537145987153053,
0.03491257131099701,
-0.005990295205265284,
0.061450716108083725,
0.05068870261311531,
-0.03409259766340256,
0.04790760204195976,
-0.01808038353919983,
-0.005968581885099411,
-0.011082408018410206,
-0.059431884437799454,
-0.02838188409805298,
0.003860849654302001,
-0.025205984711647034,
0.018887469545006752,
-0.0067994012497365475,
-0.0035522254183888435,
0.051009420305490494,
-0.012248475104570389,
-0.05470592901110649,
0.018507884815335274,
0.0025316374376416206,
-0.010320709086954594,
0.008462825790047646,
0.05238921567797661,
0.06736235320568085,
-0.07624366134405136,
0.010050838813185692,
-0.04438117891550064,
0.0987875908613205,
-0.0008761297794990242,
0.0719284936785698,
-0.043184977024793625,
-0.04323726147413254,
0.06366769224405289,
-0.11109650135040283,
-0.02560269832611084,
-0.07455381751060486,
0.04340701922774315,
-0.008165514096617699,
-5.212178066926754e-8,
0.04766996204853058,
0.04851890727877617,
0.11384937912225723,
0.004296695347875357,
0.09380962699651718,
0.024617623537778854,
-0.036894142627716064,
0.15099318325519562,
-0.08139799535274506,
-0.056582748889923096,
0.06523360311985016,
0.010804502293467522,
-0.06667011976242065,
-0.03778279572725296,
-0.022101886570453644,
0.05963848531246185,
0.03380417823791504,
-0.05338573828339577,
-0.018511159345507622,
0.04502635449171066,
0.004986331798136234,
0.07950466871261597,
0.007018532603979111,
-0.08184513449668884,
0.02943059802055359,
-0.08566742390394211,
0.04792917147278786,
0.07474028319120407,
-0.030158594250679016,
0.006977335084229708,
-0.07175412774085999,
-0.015999499708414078,
0.03783198446035385,
-0.03383413702249527,
-0.027039235457777977,
0.058475691825151443,
-0.0721559002995491,
-0.03131746128201485,
-0.015511736273765564,
-0.010921635664999485,
0.005189664661884308,
0.04273687303066254,
0.03215622156858444,
0.012605725787580013,
0.07232532650232315,
0.04396453872323036,
-0.0013675071531906724,
0.04123380780220032,
0.08757377415895462,
-0.022345583885908127,
0.02133658155798912,
0.0015681436052545905,
0.03929275646805763,
0.11640037596225739,
0.03825080767273903,
0.06826553493738174,
0.027868686243891716,
-0.023778855800628662,
-0.04322146996855736,
-0.012970120646059513,
0.0031673891935497522,
-0.1059931218624115,
-0.06889016926288605,
-0.0370272658765316
] | 0 | 0 | 0 | 0 |
|
LARP: Tokenizing Videos with a Learned Autoregressive Generative Prior | https://openreview.net/forum?id=Wr3UuEx72f | [
"Hanyu Wang",
"Saksham Suri",
"Yixuan Ren",
"Hao Chen",
"Abhinav Shrivastava"
] | Oral | We present LARP, a novel video tokenizer designed to overcome limitations in current video tokenization methods for autoregressive (AR) generative models. Unlike traditional patchwise tokenizers that directly encode local visual patches into discrete tokens, LARP introduces a holistic tokenization scheme that gathers information from the visual content using a set of learned holistic queries. This design allows LARP to capture more global and semantic representations, rather than being limited to local patch-level information. Furthermore, it offers flexibility by supporting an arbitrary number of discrete tokens, enabling adaptive and efficient tokenization based on the specific requirements of the task. To align the discrete token space with downstream AR generation tasks, LARP integrates a lightweight AR transformer as a training-time prior model that predicts the next token on its discrete latent space. By incorporating the prior model during training, LARP learns a latent space that is not only optimized for video reconstruction but is also structured in a way that is more conducive to autoregressive generation. Moreover, this process defines a sequential order for the discrete tokens, progressively pushing them toward an optimal configuration during training, ensuring smoother and more accurate AR generation at inference time. Comprehensive experiments demonstrate LARPs strong performance, achieving state-of-the-art FVD on the UCF101 class-conditional video generation benchmark. LARP enhances the compatibility of AR models with videos and opens up the potential to build unified high-fidelity multimodal large language models (MLLMs). Project page: https://hywang66.github.io/larp/ | Video Generation, Visual Tokenization | A holistic video tokenizer with a learned autoregressive generative prior. | 5,428 | 2410.21264 | [
-0.036000806838274,
-0.13468103110790253,
0.02736418880522251,
-0.007050794083625078,
0.08875705301761627,
0.04876575618982315,
-0.02409866824746132,
-0.04030432179570198,
0.08303754031658173,
-0.03463870659470558,
-0.03482953831553459,
-0.03753947094082832,
-0.06243666261434555,
0.05313771218061447,
-0.015312212519347668,
0.012712516821920872,
0.06033520773053169,
0.11360649019479752,
-0.07277927547693253,
-0.06633201241493225,
0.042129069566726685,
-0.036000240594148636,
0.015718627721071243,
-0.0006361678242683411,
0.06703642010688782,
0.07127706706523895,
0.029466187581419945,
-0.03763549402356148,
0.05132057145237923,
-0.0381961315870285,
-0.025615179911255836,
-0.010941185988485813,
-0.06589492410421371,
0.04256291314959526,
-0.03271293640136719,
0.04809228703379631,
-0.06779956817626953,
0.03671768307685852,
0.013034386560320854,
-0.05175061151385307,
0.05410800874233246,
-0.01860607974231243,
0.04658357799053192,
-0.05364294350147247,
0.027413494884967804,
0.026113687083125114,
0.03271600976586342,
-0.058826152235269547,
0.004181630443781614,
0.035303112119436264,
-0.06725005805492401,
0.012698780745267868,
-0.05220605432987213,
0.021030103787779808,
-0.07971951365470886,
0.01248928252607584,
0.04400903731584549,
-0.0517878420650959,
0.05585423484444618,
-0.017070787027478218,
-0.03558864817023277,
-0.095999576151371,
0.0372326523065567,
-0.01299383956938982,
-0.030341092497110367,
-0.10447254031896591,
0.08441994339227676,
0.011502663604915142,
0.042033739387989044,
0.02668413333594799,
-0.051738083362579346,
0.07734724134206772,
-0.028242290019989014,
0.029246343299746513,
0.01844736747443676,
0.04258183762431145,
0.03221100568771362,
-0.07859522104263306,
0.01741056703031063,
-0.09727402031421661,
0.13175204396247864,
0.06600591540336609,
0.0722280889749527,
0.06744832545518875,
0.10711860656738281,
-0.01413000002503395,
-0.004596243612468243,
0.06596909463405609,
-0.012439917773008347,
0.0165402851998806,
-0.11918225884437561,
-0.02806166000664234,
0.029164636507630348,
-0.03187321871519089,
0.016728907823562622,
-0.025100227445364,
0.025605734437704086,
-0.0695892795920372,
-0.024333877488970757,
0.07674649357795715,
0.017831582576036453,
0.03968018293380737,
0.04393276199698448,
-0.006819412112236023,
0.0602719746530056,
-0.04282904416322708,
0.011952055618166924,
0.07524802535772324,
0.06399176269769669,
-0.047133732587099075,
0.0004859804466832429,
0.05640275403857231,
0.010315698571503162,
-0.03554719313979149,
-0.00013945475802756846,
0.021302517503499985,
-0.03967602550983429,
-0.05672242492437363,
-0.05004790425300598,
0.038296107202768326,
0.048187531530857086,
0.004629457835108042,
-0.06053470820188522,
0.048115622252225876,
-0.015775030478835106,
-0.08074028789997101,
0.01217396929860115,
2.2661282950907756e-33,
0.03092280775308609,
0.029191281646490097,
0.05024297535419464,
0.01604158617556095,
0.03666842728853226,
0.01836172305047512,
0.019561544060707092,
0.038633693009614944,
-0.006859686691313982,
-0.05338217318058014,
0.006912136450409889,
-0.0688936859369278,
-0.016460880637168884,
0.11785802990198135,
-0.04689537733793259,
-0.02207237295806408,
-0.11586742103099823,
0.00967975053936243,
0.01587533950805664,
0.018899383023381233,
-0.03031441755592823,
-0.01358396839350462,
-0.04596341773867607,
0.03450101241469383,
-0.05771953612565994,
-0.013341759331524372,
0.09865610301494598,
-0.04881899431347847,
-0.0212380550801754,
-0.018755324184894562,
-0.11635162681341171,
-0.018332982435822487,
0.056356530636548996,
0.01303064450621605,
0.08423227071762085,
-0.02714208886027336,
-0.0006504248594865203,
-0.0809423178434372,
0.00018296836060471833,
-0.06200437620282173,
-0.013670644722878933,
0.07849148660898209,
-0.04060123860836029,
-0.06836465001106262,
-0.08848804980516434,
-0.07547128945589066,
0.02562454715371132,
0.046618957072496414,
-0.08320493996143341,
0.019222909584641457,
0.11257001012563705,
0.013460556976497173,
-0.00787358358502388,
-0.07676895707845688,
-0.005443535279482603,
-0.05253616347908974,
0.02977820858359337,
0.009462387301027775,
0.01410567294806242,
-0.009533043950796127,
0.0041705225594341755,
0.015309182927012444,
0.055526670068502426,
0.06514837592840195,
-0.0369693823158741,
-0.016772674396634102,
0.10998968780040741,
0.03574373573064804,
0.08525921404361725,
0.02299325168132782,
-0.040248431265354156,
0.02838941477239132,
-0.02695910632610321,
-0.08122248947620392,
0.06388186663389206,
0.022381532937288284,
0.042408619076013565,
-0.009360107593238354,
-0.036896828562021255,
0.07633170485496521,
-0.06879916042089462,
-0.051577452570199966,
0.03114248998463154,
-0.04257562756538391,
-0.0021454067900776863,
-0.04142371937632561,
0.055446721613407135,
-0.06872234493494034,
0.0025277603417634964,
-0.08956746011972427,
-0.05478362366557121,
-0.03550449013710022,
-0.03416861593723297,
0.028585489839315414,
0.052879542112350464,
-1.2859880647346978e-33,
0.07364606857299805,
0.07194072008132935,
-0.02128366008400917,
0.00986831821501255,
-0.03558997064828873,
-0.09083446115255356,
0.0032413448207080364,
0.02046923339366913,
-0.012071112170815468,
-0.06807174533605576,
-0.054028574377298355,
-0.039740048348903656,
-0.01477016881108284,
-0.04002033546566963,
0.009813192300498486,
0.009325215592980385,
-0.026008324697613716,
-0.014392203651368618,
-0.03765583783388138,
0.034130532294511795,
-0.0001446967653464526,
0.03370070084929466,
-0.056028708815574646,
0.03481730818748474,
-0.016849374398589134,
0.07630421966314316,
-0.01154773123562336,
0.07450924813747406,
0.03221038356423378,
-0.025016162544488907,
0.056018829345703125,
-0.05687786266207695,
0.009524269960820675,
-0.03484869375824928,
-0.10957068204879761,
0.03337734565138817,
0.08119475096464157,
0.032887838780879974,
-0.06874136626720428,
0.06986537575721741,
0.07236549258232117,
-0.061339352279901505,
-0.024882495403289795,
0.03933693468570709,
-0.012639353051781654,
-0.03976551070809364,
0.02742105722427368,
-0.001277805888094008,
-0.02965846285223961,
0.06300230324268341,
-0.015138841234147549,
0.0471552312374115,
-0.10001940280199051,
-0.002305726520717144,
-0.0311481561511755,
-0.09917618334293365,
-0.03360782191157341,
0.009014018811285496,
0.04085063189268112,
-0.031340256333351135,
0.013408619910478592,
-0.010905563831329346,
0.03912225738167763,
-0.09792382270097733,
0.011283056810498238,
-0.020878348499536514,
-0.012371229939162731,
-0.011684914119541645,
-0.05794527381658554,
0.011470319703221321,
0.0655943900346756,
0.004882724490016699,
-0.0037035569548606873,
0.06418586522340775,
0.07180514931678772,
-0.03798691928386688,
0.047910891473293304,
0.0015534640988335013,
0.040182363241910934,
-0.06707282364368439,
-0.04814017564058304,
-0.04768913984298706,
0.04025503620505333,
0.09096898138523102,
0.04234880954027176,
0.09303879737854004,
0.06481954455375671,
0.06408119946718216,
0.04944484308362007,
0.04242970421910286,
-0.054293014109134674,
0.06385678052902222,
-0.004860812798142433,
0.011854812502861023,
-0.023638132959604263,
-3.617186195015165e-8,
-0.08551260083913803,
0.05788811668753624,
-0.05230516567826271,
0.012844632379710674,
-0.033417657017707825,
-0.026253050193190575,
0.006504861172288656,
0.0666983425617218,
0.009952353313565254,
-0.046494003385305405,
0.06881020963191986,
-0.0360388346016407,
-0.008183399215340614,
-0.0699659138917923,
-0.03326541185379028,
-0.016848603263497353,
0.04545833170413971,
-0.0007036795723251998,
-0.026682965457439423,
0.0422222763299942,
-0.057122595608234406,
0.05869588255882263,
-0.00833718292415142,
-0.08345290273427963,
0.01434267871081829,
-0.01809168979525566,
0.061424076557159424,
0.043432701379060745,
0.07841548323631287,
-0.033210329711437225,
-0.009128437377512455,
0.05632251873612404,
0.047176800668239594,
-0.11691009253263474,
-0.06965598464012146,
0.055663298815488815,
0.08797573298215866,
-0.04674078896641731,
0.015399690717458725,
-0.07371199876070023,
-0.019041189923882484,
0.04823439195752144,
-0.030659491196274757,
-0.0061094556003808975,
-0.05922422558069229,
0.026453426107764244,
0.023420017212629318,
-0.09796998649835587,
0.023916294798254967,
-0.011384022422134876,
-0.03605048730969429,
-0.05402617156505585,
0.0032649512868374586,
0.10331472009420395,
0.09039392322301865,
-0.0047451164573431015,
0.08638928085565567,
-0.038237251341342926,
0.09534166008234024,
0.05859832093119621,
-0.014679260551929474,
0.03757606819272041,
-0.02213152125477791,
0.0743047371506691
] | 0 | 0 | 0 | 0 |
|
MOS: Model Synergy for Test-Time Adaptation on LiDAR-Based 3D Object Detection | https://openreview.net/forum?id=Y6aHdDNQYD | [
"Zhuoxiao Chen",
"Junjie Meng",
"Mahsa Baktashmotlagh",
"Yonggang Zhang",
"Zi Huang",
"Yadan Luo"
] | Oral | LiDAR-based 3D object detection is crucial for various applications but often experiences performance degradation in real-world deployments due to domain shifts. While most studies focus on cross-dataset shifts, such as changes in environments and object geometries, practical corruptions from sensor variations and weather conditions remain underexplored. In this work, we propose a novel online test-time adaptation framework for 3D detectors that effectively tackles these shifts, including a challenging $\textit{cross-corruption}$ scenario where cross-dataset shifts and corruptions co-occur. By leveraging long-term knowledge from previous test batches, our approach mitigates catastrophic forgetting and adapts effectively to diverse shifts. Specifically, we propose a Model Synergy (MOS) strategy that dynamically selects historical checkpoints with diverse knowledge and assembles them to best accommodate the current test batch. This assembly is directed by our proposed Synergy Weights (SW), which perform a weighted averaging of the selected checkpoints, minimizing redundancy in the composite model. The SWs are computed by evaluating the similarity of predicted bounding boxes on the test data and the independence of features between checkpoint pairs in the model bank. To maintain an efficient and informative model bank, we discard checkpoints with the lowest average SW scores, replacing them with newly updated models. Our method was rigorously tested against existing test-time adaptation strategies across three datasets and eight types of corruptions, demonstrating superior adaptability to dynamic scenes and conditions. Notably, it achieved a 67.3% improvement in a challenging cross-corruption scenario, offering a more comprehensive benchmark for adaptation. Source code: https://github.com/zhuoxiao-chen/MOS. | Test-Time Adaptation, 3D Object Detection | null | 5,340 | 2406.14878 | [
-0.04502519220113754,
-0.09255241602659225,
0.01534615270793438,
0.1065048947930336,
0.11156056076288223,
-0.07894112914800644,
-0.018358763307332993,
-0.12228378653526306,
-0.013636521995067596,
-0.04295108839869499,
-0.05431516468524933,
-0.0020863248500972986,
0.0293237566947937,
0.04055711627006531,
-0.027025477960705757,
-0.005541578400880098,
0.048127297312021255,
0.05719095841050148,
-0.04566929489374161,
0.06068319082260132,
-0.05156894028186798,
0.057862747460603714,
-0.028136445209383965,
0.09562632441520691,
-0.03487531840801239,
0.003543395781889558,
0.01956840790808201,
-0.08293558657169342,
-0.05048278346657753,
-0.029826119542121887,
0.004219204653054476,
-0.0072173369117081165,
-0.1877191960811615,
0.02457791194319725,
0.03283815458416939,
-0.06884066760540009,
-0.05003417655825615,
-0.003982914611697197,
0.0059746005572378635,
-0.0381258949637413,
0.06665582209825516,
-0.015532034449279308,
0.0160723514854908,
0.037570770829916,
-0.00904815923422575,
0.049268633127212524,
0.03276890888810158,
-0.02367263473570347,
0.05108991265296936,
-0.0300564207136631,
0.005158149171620607,
-0.014132777228951454,
-0.06638006865978241,
0.01809808798134327,
0.00527433305978775,
0.03562421351671219,
0.07835511118173599,
0.007634034380316734,
-0.008952449075877666,
0.009043028578162193,
0.005296488292515278,
-0.08272118866443634,
0.027445675805211067,
-0.017970947548747063,
0.024590039625763893,
-0.0355866514146328,
0.03689683973789215,
0.00895177386701107,
0.027530129998922348,
0.06221534684300423,
-0.039933253079652786,
0.1362835317850113,
0.018688470125198364,
-0.03622019290924072,
0.02646304853260517,
0.08285076171159744,
0.07938666641712189,
-0.013878371566534042,
0.010374506935477257,
-0.07655200362205505,
-0.002197988098487258,
0.030561815947294235,
0.006418028846383095,
0.028787728399038315,
0.08026479184627533,
-0.025342650711536407,
0.021690212190151215,
0.07148043811321259,
0.06256752461194992,
-0.043518852442502975,
-0.017987679690122604,
-0.027806103229522705,
0.004784060176461935,
-0.019918207079172134,
0.023971060290932655,
-0.03879889100790024,
0.09861966222524643,
-0.03447052463889122,
0.0555499903857708,
0.06635641306638718,
-0.049460720270872116,
0.01202419027686119,
0.03443169221282005,
0.016015011817216873,
0.022465651854872704,
0.008858529850840569,
0.04389191418886185,
-0.029886795207858086,
-0.002187360543757677,
-0.01762656681239605,
0.06885045766830444,
0.06004342436790466,
0.06437820941209793,
-0.004249753896147013,
0.007410753984004259,
-0.02974926307797432,
-0.051846422255039215,
-0.01257818378508091,
-0.07491715252399445,
0.05189495161175728,
-0.02421531081199646,
-0.02588699199259281,
0.09314573556184769,
0.017461152747273445,
0.01501214038580656,
0.053347572684288025,
0.002326172310858965,
-2.24268365719435e-35,
0.023689668625593185,
-0.00221668416634202,
0.021384382620453835,
0.020313922315835953,
0.1013728529214859,
-0.032795462757349014,
0.04341727867722511,
0.0346989706158638,
-0.0831194818019867,
0.00799549650400877,
-0.08205211162567139,
-0.03160906210541725,
-0.04381372034549713,
0.034622058272361755,
0.04684149846434593,
0.03196427598595619,
-0.05846285820007324,
0.004334930330514908,
-0.060898710042238235,
0.04006829485297203,
-0.021282758563756943,
-0.10392554104328156,
-0.025306036695837975,
-0.016472257673740387,
-0.00335518061183393,
0.012602303177118301,
0.02814868465065956,
0.04437832161784172,
-0.0008922931156121194,
0.0024008089676499367,
-0.020467005670070648,
0.08013249188661575,
-0.03415612503886223,
0.10214626789093018,
0.06395828723907471,
-0.03841099515557289,
-0.04431499168276787,
-0.044281307607889175,
-0.04032038897275925,
-0.026853814721107483,
0.02157377265393734,
-0.01510169729590416,
-0.05289126932621002,
-0.0897471085190773,
-0.026057105511426926,
-0.11150462180376053,
0.03441682830452919,
0.031867410987615585,
-0.034931231290102005,
-0.023344503715634346,
0.039108023047447205,
0.01005768682807684,
-0.09051492810249329,
-0.042725782841444016,
-0.11276373267173767,
-0.04205702245235443,
0.021391283720731735,
-0.023028137162327766,
0.11860411614179611,
0.033105432987213135,
-0.0014428810682147741,
-0.019309306517243385,
-0.027155736461281776,
0.08393590152263641,
0.0596790686249733,
0.05979381874203682,
-0.0052384366281330585,
0.05450977385044098,
0.018019000068306923,
0.05162277817726135,
0.02034660242497921,
-0.031815532594919205,
-0.0639050304889679,
0.057800401002168655,
0.054434698075056076,
-0.03304394334554672,
0.08331151306629181,
0.047729697078466415,
0.022378072142601013,
0.026340140029788017,
0.06689310818910599,
-0.039210833609104156,
-0.005817957688122988,
-0.09463102370500565,
-0.09231724590063095,
-0.07259584963321686,
-0.0009106779471039772,
-0.07005523890256882,
-0.0784900039434433,
0.006297624204307795,
-0.04622168093919754,
-0.05595967918634415,
0.00794786587357521,
-0.06984743475914001,
-0.004762200638651848,
-1.0381877944386132e-33,
0.032372698187828064,
-0.047214534133672714,
0.02513963356614113,
0.057379912585020065,
-0.0010901341447606683,
-0.07162486016750336,
-0.08435714244842529,
-0.004417491611093283,
-0.04558515548706055,
-0.06988918036222458,
-0.08068166673183441,
-0.038776595145463943,
0.009150306694209576,
0.007323715835809708,
-0.05082349479198456,
0.01914285495877266,
-0.07002056390047073,
-0.07663974165916443,
-0.06871195137500763,
-0.029724448919296265,
0.09686695784330368,
0.022514542564749718,
-0.06984122842550278,
-0.020585941150784492,
-0.1336379498243332,
0.01757770963013172,
-0.007076497655361891,
0.09084758907556534,
-0.009562975727021694,
-0.029920591041445732,
-0.031524721533060074,
0.026562249287962914,
0.009146580472588539,
0.002612223383039236,
0.006313995458185673,
0.02777058444917202,
0.09715783596038818,
-0.05904914066195488,
-0.02928735874593258,
0.10658852756023407,
0.051669664680957794,
0.006140148267149925,
-0.03382111340761185,
0.023479292169213295,
0.03635711595416069,
0.03267804533243179,
-0.06429008394479752,
0.02901666797697544,
-0.00678950035944581,
0.01660684123635292,
-0.045308392494916916,
-0.010948315262794495,
-0.06660338491201401,
0.043702948838472366,
0.007536476477980614,
-0.013614323921501637,
-0.024028949439525604,
-0.011158326640725136,
0.03742287680506706,
0.07476182281970978,
-0.015305424109101295,
-0.02200378105044365,
0.008741342462599277,
0.009599435143172741,
0.007381923962384462,
-0.02250857464969158,
0.030716439709067345,
-0.02372651733458042,
-0.012836460024118423,
0.09827765822410583,
0.03366570547223091,
-0.02520882338285446,
-0.05073283985257149,
-0.06993130594491959,
-0.0023554957006126642,
-0.0840715765953064,
-0.005041357129812241,
-0.01417818944901228,
0.03009507805109024,
-0.017132487148046494,
-0.029100926592946053,
-0.08480247110128403,
0.05129347741603851,
0.14094983041286469,
0.040640462189912796,
0.11322872340679169,
0.0622049905359745,
0.045465923845767975,
-0.006601122673600912,
-0.033059436827898026,
-0.09655309468507767,
-0.024607844650745392,
-0.03300454840064049,
0.0271467138081789,
0.014240569435060024,
-3.957267935561504e-8,
-0.06885452568531036,
0.10453154146671295,
-0.036699291318655014,
0.0378849022090435,
0.00647733872756362,
-0.015910036861896515,
0.016332168132066727,
0.09535529464483261,
-0.022613627836108208,
-0.021180596202611923,
0.04247194156050682,
-0.08368505537509918,
0.021252496168017387,
-0.012085037305951118,
-0.014524172060191631,
0.05403931066393852,
-0.033848561346530914,
-0.012822101823985577,
-0.009719965048134327,
0.04469166696071625,
0.01014456246048212,
0.04534286633133888,
0.10249391943216324,
0.05569695308804512,
-0.007170028053224087,
0.010883904993534088,
0.03819960355758667,
0.1320532262325287,
0.0707428827881813,
-0.05087767913937569,
-0.023394936695694923,
-0.01954164169728756,
0.06672884523868561,
0.02124190516769886,
-0.04462345689535141,
0.05096413940191269,
0.043474894016981125,
0.008360526524484158,
0.050121501088142395,
0.019623851403594017,
-0.026637528091669083,
0.049167729914188385,
-0.01574752852320671,
-0.02424655295908451,
-0.03726355731487274,
0.03941164165735245,
-0.052382055670022964,
-0.02131226472556591,
-0.05315622314810753,
-0.062828429043293,
0.0371905192732811,
0.0027886831667274237,
-0.04363388568162918,
0.14894577860832214,
0.04511158540844917,
0.027283165603876114,
0.027714252471923828,
-0.0571650005877018,
0.052092522382736206,
0.0597674623131752,
0.03373049199581146,
-0.05712383612990379,
-0.10507605224847794,
-0.025018133223056793
] | 0 | 0 | 0 | 0 |
|
Synthetic continued pretraining | https://openreview.net/forum?id=07yvxWDSla | [
"Zitong Yang",
"Neil Band",
"Shuangping Li",
"Emmanuel Candes",
"Tatsunori Hashimoto"
] | Oral | Pretraining on large-scale, unstructured internet text enables language models to acquire a significant amount of world knowledge.
However, this knowledge acquisition is data-inefficient---to learn a fact, models must be trained on hundreds to thousands of diverse representations of it.
This poses a challenge when adapting a pretrained model to a small corpus of domain-specific documents, where each fact may appear rarely or only once.
We propose to bridge this gap with synthetic continued pretraining: using the small domain-specific corpus to synthesize a large corpus more amenable to learning, and then performing continued pretraining on the synthesized corpus.
We instantiate this proposal with EntiGraph, a synthetic data augmentation algorithm that extracts salient entities from the source corpus and then generates diverse text by drawing connections between those entities.
Synthetic continued pretraining with EntiGraph enables a language model to answer questions and follow generic instructions related to the source documents without access to them.
If the source documents are instead available at inference time, we show that the knowledge acquired through our approach compounds with retrieval-augmented generation.
To better understand these results, we build a simple mathematical model of EntiGraph, and show how synthetic data augmentation can "rearrange" knowledge to enable more data-efficient learning. | large language model, synthetic data, continued pretraining | null | 5,336 | 2409.07431 | [
-0.04962855577468872,
-0.09045513719320297,
0.025939837098121643,
0.06309178471565247,
0.11166145652532578,
0.0313040092587471,
-0.029517650604248047,
-0.0008493047207593918,
0.003626378485932946,
0.003884422592818737,
0.0027464129962027073,
-0.0120999775826931,
0.04254244267940521,
0.004927680362015963,
0.0037027723155915737,
0.06246446445584297,
0.09728428721427917,
0.06624767184257507,
-0.07685855031013489,
-0.03335268795490265,
0.03932721167802811,
0.02365105412900448,
0.09608390927314758,
0.02610543556511402,
0.04364114999771118,
0.015552708879113197,
0.002092577749863267,
-0.04042590782046318,
0.056744031608104706,
-0.04964541643857956,
0.07010087370872498,
0.07242053002119064,
-0.05179845169186592,
0.025050686672329903,
-0.003555477363988757,
0.07585277408361435,
-0.08215887099504471,
0.09015714377164841,
0.009798336774110794,
-0.00792913231998682,
-0.00008757990872254595,
-0.041857339441776276,
-0.029713544994592667,
0.05568132922053337,
0.0809512510895729,
0.011513176374137402,
-0.06025295332074165,
0.011843808926641941,
0.0756830945611,
0.061104774475097656,
-0.08734074234962463,
-0.022702721878886223,
-0.03285571187734604,
0.06153937056660652,
0.04182352498173714,
0.005715646781027317,
-0.02463444508612156,
-0.015543046407401562,
-0.021790040656924248,
-0.031793296337127686,
-0.012578906491398811,
-0.11126994341611862,
-0.005173546262085438,
-0.05512682721018791,
-0.07960487902164459,
-0.04544811323285103,
0.03604540973901749,
0.10492376983165741,
0.010238023474812508,
0.06821785122156143,
-0.006579516921192408,
0.13677345216274261,
-0.03278301656246185,
0.08473366498947144,
0.056047920137643814,
0.06193762645125389,
0.018651332706212997,
-0.02126680128276348,
0.03450508043169975,
-0.055042561143636703,
0.02601131796836853,
0.07552482932806015,
0.08777032047510147,
-0.02326664887368679,
0.09438696503639221,
0.048045482486486435,
0.06580798327922821,
0.023301998153328896,
-0.004295823164284229,
0.029666993767023087,
-0.07033682614564896,
-0.06205906346440315,
0.07998985797166824,
-0.008118737488985062,
0.03796372562646866,
0.055772144347429276,
0.018346130847930908,
-0.03703582286834717,
0.025844722986221313,
0.07689150422811508,
0.04725927859544754,
0.09774842858314514,
-0.005252355244010687,
-0.04878878593444824,
-0.04779436066746712,
-0.02674376592040062,
0.029758520424365997,
0.06141677498817444,
-0.022563043981790543,
-0.05837741866707802,
0.046184077858924866,
0.032886702567338943,
-0.00033655916922725737,
-0.06433241069316864,
0.02695821039378643,
0.0062601747922599316,
0.04653715342283249,
-0.0030499270651489496,
-0.02215654030442238,
0.03880946338176727,
-0.03187759220600128,
0.0676843523979187,
-0.014386391267180443,
0.010106857866048813,
-0.02492653764784336,
-0.04981958121061325,
-0.012042292393743992,
5.403672225623206e-34,
0.11765522509813309,
0.04534463584423065,
0.0352163128554821,
0.10227614641189575,
-0.021629905328154564,
-0.004312887787818909,
-0.01025413628667593,
0.09219114482402802,
-0.09818360954523087,
-0.038316741585731506,
0.00014327105600386858,
0.04297606647014618,
-0.04543665796518326,
0.08267074823379517,
0.001008249120786786,
-0.04031352698802948,
-0.053857848048210144,
0.022367876023054123,
0.03650669753551483,
-0.01832471415400505,
-0.04552469775080681,
-0.0228889100253582,
0.011625035665929317,
0.0013998717768117785,
0.0002958237600978464,
-0.005692117847502232,
0.04920186847448349,
-0.04719267785549164,
-0.05164550989866257,
-0.012160436250269413,
-0.08491897583007812,
-0.055277738720178604,
0.01067223958671093,
0.015333441086113453,
0.006028377916663885,
-0.07616330683231354,
0.0309490654617548,
-0.07356797158718109,
-0.014974890276789665,
-0.023832611739635468,
-0.0165141262114048,
0.09718330949544907,
0.058782558888196945,
-0.0025730766355991364,
-0.0837399959564209,
-0.0655936598777771,
-0.004592567682266235,
0.025744693353772163,
0.008426916785538197,
-0.03336525708436966,
0.07894816249608994,
0.022480621933937073,
-0.032332029193639755,
-0.07353980094194412,
-0.011873362585902214,
0.001684460905380547,
0.024138761684298515,
0.021066252142190933,
0.05679772049188614,
0.03155841305851936,
0.017217472195625305,
0.036095842719078064,
0.008566407486796379,
0.1387515664100647,
0.0010075437603518367,
0.02427719160914421,
-0.027541736140847206,
0.03668670728802681,
0.043892860412597656,
-0.02035386487841606,
-0.07993631809949875,
0.023988261818885803,
-0.04642695188522339,
-0.05634652078151703,
0.002893385011702776,
0.03607483208179474,
-0.053951699286699295,
-0.11607509851455688,
0.010161344893276691,
0.027101777493953705,
-0.0001555853959871456,
-0.03281635046005249,
-0.03414201736450195,
-0.06361741572618484,
-0.011264897882938385,
-0.04292161390185356,
0.060514990240335464,
-0.06572632491588593,
0.03452359512448311,
-0.012436583638191223,
-0.051714517176151276,
-0.059165239334106445,
-0.078914575278759,
-0.04737010970711708,
0.06493263691663742,
5.7726745960257074e-34,
0.015132094733417034,
-0.02069925144314766,
-0.042040832340717316,
0.029530655592679977,
-0.04092530533671379,
0.016750184819102287,
0.007956361398100853,
0.06141353026032448,
-0.0788339376449585,
-0.04748431220650673,
-0.027626173570752144,
-0.07710907608270645,
0.054273542016744614,
-0.03515326976776123,
0.0005408602301031351,
-0.03487924858927727,
0.005680731497704983,
-0.0020647647324949503,
-0.04596365988254547,
0.09453661739826202,
-0.017617622390389442,
0.03056074120104313,
-0.07695095986127853,
0.009463793598115444,
0.0134450513869524,
0.07269769161939621,
-0.04346298426389694,
0.08453891426324844,
-0.06011616438627243,
-0.00103668961673975,
-0.03063483163714409,
-0.0556180439889431,
-0.052105169743299484,
-0.04983123019337654,
-0.115085169672966,
0.06350086629390717,
0.05365428701043129,
-0.062231823801994324,
-0.05360579118132591,
0.02559613063931465,
0.08582346141338348,
-0.009895478375256062,
-0.06293061375617981,
-0.026793450117111206,
-0.07568491250276566,
-0.04974283650517464,
-0.1310538351535797,
0.04253538325428963,
0.022005470469594002,
0.017801938578486443,
0.008033493533730507,
-0.00001963332761079073,
-0.03130221739411354,
-0.036490581929683685,
-0.0040982561185956,
-0.15228216350078583,
-0.04280850291252136,
-0.037234824150800705,
0.015763912349939346,
-0.07130865007638931,
-0.022639760747551918,
-0.04123159497976303,
0.02243560180068016,
-0.050546132028102875,
0.03377039358019829,
-0.10074076056480408,
-0.03860338404774666,
0.01409664936363697,
0.0008716468000784516,
-0.03232358396053314,
0.038504842668771744,
-0.03542284667491913,
-0.004045384470373392,
0.004738952498883009,
0.043607328087091446,
-0.007837427780032158,
0.04362911731004715,
-0.005455957725644112,
-0.03010641410946846,
-0.03124813176691532,
-0.011046594008803368,
-0.0036068533081561327,
0.03684457018971443,
0.061671994626522064,
0.10857628285884857,
0.056265875697135925,
0.026828764006495476,
0.07635807245969772,
0.061189204454422,
0.07396180182695389,
-0.014097924344241619,
0.039343103766441345,
-0.05134841799736023,
0.1053958460688591,
-0.058012302964925766,
-3.689815741836355e-8,
-0.13019254803657532,
0.01850919984281063,
-0.028504548594355583,
0.04993613436818123,
0.021871497854590416,
-0.05435114726424217,
-0.018888453021645546,
0.05391734838485718,
-0.07813312858343124,
-0.03934803605079651,
0.07972697168588638,
0.028592843562364578,
0.03555732220411301,
-0.05323564633727074,
-0.0004633456701412797,
0.08399001508951187,
0.06124401465058327,
-0.0356777049601078,
0.00668276147916913,
-0.03237059339880943,
0.03913566842675209,
0.0389266237616539,
0.01895579695701599,
0.01783841848373413,
0.07869543880224228,
-0.07594677060842514,
0.00521385483443737,
0.03264036402106285,
0.042341139167547226,
-0.04745560139417648,
-0.0432683564722538,
-0.07025686651468277,
-0.051310740411281586,
-0.047555167227983475,
0.031867869198322296,
0.09581860899925232,
0.0655352994799614,
-0.010102065280079842,
-0.02331145852804184,
-0.08850980550050735,
0.026322897523641586,
0.05908333137631416,
-0.06714393198490143,
0.030138909816741943,
0.01804751716554165,
-0.060150932520627975,
-0.03264980390667915,
-0.029874341562390327,
0.014002151787281036,
-0.059278517961502075,
-0.05788494274020195,
-0.0809439942240715,
0.06877802312374115,
0.0076435948722064495,
0.01866013929247856,
0.048720378428697586,
-0.019336199387907982,
0.050850313156843185,
0.034407347440719604,
0.0046024322509765625,
-0.03253030776977539,
-0.00612748833373189,
0.021174373105168343,
0.06029067561030388
] | https://github.com/zitongyang/synthetic_continued_pretraining | 106 | 0 | 0 | 0 |
EmbodiedSAM: Online Segment Any 3D Thing in Real Time | https://openreview.net/forum?id=XFYUwIyTxQ | [
"Xiuwei Xu",
"Huangxing Chen",
"Linqing Zhao",
"Ziwei Wang",
"Jie Zhou",
"Jiwen Lu"
] | Oral | Embodied tasks require the agent to fully understand 3D scenes simultaneously with its exploration, so an online, real-time, fine-grained and highly-generalized 3D perception model is desperately needed. Since high-quality 3D data is limited, directly training such a model in 3D is infeasible. Meanwhile, vision foundation models (VFM) has revolutionized the field of 2D computer vision with superior performance, which makes the use of VFM to assist embodied 3D perception a promising direction. However, most existing VFM-assisted 3D perception methods are either offline or too slow that cannot be applied in practical embodied tasks. In this paper, we aim to leverage Segment Anything Model (SAM) for real-time 3D instance segmentation in an online setting. This is a challenging problem since future frames are not available in the input streaming RGB-D video, and an instance may be observed in several frames so efficient object matching between frames is required. To address these challenges, we first propose a geometric-aware query lifting module to represent the 2D masks generated by SAM by 3D-aware queries, which is then iteratively refined by a dual-level query decoder. In this way, the 2D masks are transferred to fine-grained shapes on 3D point clouds. Benefit from the query representation for 3D masks, we can compute the similarity matrix between the 3D masks from different views by efficient matrix operation, which enables real-time inference. Experiments on ScanNet, ScanNet200, SceneNN and 3RScan show our method achieves state-of-the-art performance among online 3D perception models, even outperforming offline VFM-assisted 3D instance segmentation methods by a large margin. Our method also demonstrates great generalization ability in several zero-shot dataset transferring experiments and show great potential in data-efficient setting. | 3d instance segmentation; online 3d scene segmentation | We presented EmbodiedSAM, an efficient framework that leverages vision foundation models for online, real-time, fine-grained and generalized 3D instance segmentation. | 5,293 | 2408.11811 | [
0.009136312641203403,
-0.12716980278491974,
0.005725135561078787,
0.0017497761873528361,
0.05772509425878525,
-0.0021341617684811354,
-0.014628823846578598,
-0.0507042296230793,
0.006391294300556183,
-0.05180056020617485,
-0.05188903957605362,
-0.06698121875524521,
-0.03506418690085411,
0.08256242424249649,
0.07786809653043747,
-0.01954394020140171,
0.17161068320274353,
0.09229924529790878,
-0.07325486838817596,
0.07313057035207748,
0.08467376232147217,
-0.024161791428923607,
0.02044910192489624,
-0.010037203319370747,
0.05683378875255585,
-0.019904758781194687,
0.02113182283937931,
-0.11711879074573517,
0.03151698410511017,
-0.00003252663373132236,
-0.02978912554681301,
-0.027143746614456177,
-0.026855017989873886,
-0.0036225004587322474,
-0.013629142194986343,
-0.01787322573363781,
0.015526656061410904,
0.007973093539476395,
-0.1376628428697586,
-0.09935122728347778,
-0.05681968480348587,
0.03017229214310646,
-0.013194963335990906,
-0.0012137586018070579,
0.015034887939691544,
0.0650038793683052,
0.024035528302192688,
-0.03443710505962372,
0.06874958425760269,
-0.0354551300406456,
-0.06812106817960739,
-0.10911714285612106,
-0.03367181122303009,
0.01589970476925373,
-0.040882740169763565,
0.07051657885313034,
0.042002156376838684,
-0.04046789929270744,
0.0511312298476696,
0.037026260048151016,
0.019531728699803352,
-0.04266133904457092,
0.09810236096382141,
0.020642653107643127,
-0.027081191539764404,
-0.07358864694833755,
0.014672833494842052,
-0.07359252870082855,
0.017194774001836777,
-0.01948218047618866,
-0.08970098942518234,
0.001698159845545888,
-0.01643153466284275,
-0.048664432018995285,
0.01489644218236208,
-0.002892692107707262,
0.04790684953331947,
-0.09729039669036865,
0.041167546063661575,
-0.013221029192209244,
0.013143991120159626,
0.012745082378387451,
-0.00723013561218977,
0.06383979320526123,
0.03492094203829765,
0.05897127836942673,
0.03860434517264366,
0.030770577490329742,
-0.015618656761944294,
0.029880329966545105,
-0.036182790994644165,
-0.025330647826194763,
0.009958318434655666,
-0.0015853530494496226,
0.0362691730260849,
-0.12081605941057205,
0.009361100383102894,
-0.028070615604519844,
0.05165548250079155,
0.05142393708229065,
-0.006045668385922909,
0.011369121260941029,
0.023094050586223602,
-0.005181238986551762,
0.027843322604894638,
0.049847789108753204,
0.03808861970901489,
0.03698533773422241,
0.05244624242186546,
-0.01469876803457737,
0.05097929388284683,
-0.0067283096723258495,
0.03041522018611431,
0.015521569177508354,
-0.040107209235429764,
0.026234708726406097,
0.015910964459180832,
-0.046335238963365555,
0.00929135363548994,
0.022207610309123993,
0.026478370651602745,
0.0325571671128273,
0.0006200321367941797,
-0.022138044238090515,
0.039793021976947784,
-0.03833574056625366,
-0.02000412344932556,
-4.526872940849155e-35,
0.048622164875268936,
0.01867588609457016,
0.0706024020910263,
-0.06451964378356934,
0.01096807699650526,
0.020491616800427437,
0.04002935439348221,
0.10364823043346405,
-0.013979420065879822,
0.037984855473041534,
-0.007347204256802797,
-0.056367114186286926,
-0.005181466694921255,
0.10184546560049057,
-0.01861725188791752,
-0.0660829246044159,
-0.04802878201007843,
0.002103676088154316,
-0.01733540929853916,
0.06849859654903412,
0.006508548744022846,
-0.03215853497385979,
-0.05844276025891304,
0.02250867336988449,
0.01787957176566124,
-0.006749722175300121,
0.019691193476319313,
-0.03021109476685524,
0.033045291900634766,
0.0021498019341379404,
-0.035791825503110886,
-0.0354047492146492,
-0.07537256926298141,
-0.023856325075030327,
0.030495576560497284,
-0.05672203376889229,
0.0066870711743831635,
-0.022143622860312462,
-0.004970394540578127,
-0.04552735760807991,
0.018015556037425995,
0.036820102483034134,
0.0027429135516285896,
-0.09261579066514969,
-0.13896527886390686,
-0.060802243649959564,
0.050388600677251816,
0.12279633432626724,
-0.15345340967178345,
0.009036797098815441,
0.08332133293151855,
0.05567219480872154,
-0.03787689283490181,
-0.07156631350517273,
-0.0791984423995018,
-0.048159997910261154,
0.0629744753241539,
0.09674584865570068,
0.09066793322563171,
0.014872550964355469,
-0.06487702578306198,
-0.0027173750568181276,
-0.005190434865653515,
0.06936489045619965,
-0.04127496853470802,
-0.006101764272898436,
0.030845729634165764,
0.03217034786939621,
0.020476533100008965,
0.017013592645525932,
-0.05261624604463577,
-0.02413029968738556,
0.04625539109110832,
-0.06103299558162689,
0.028158964589238167,
-0.05108308047056198,
0.0007891333079896867,
-0.032504260540008545,
-0.021652601659297943,
0.06017802655696869,
-0.04142586886882782,
0.008734848350286484,
0.00562283443287015,
-0.0286426842212677,
-0.015587244182825089,
-0.043871089816093445,
0.06198618561029434,
-0.08967024832963943,
0.0027668289840221405,
-0.03280050680041313,
-0.07479868829250336,
-0.10461562126874924,
-0.054688360542058945,
-0.044609539210796356,
0.02399931102991104,
6.54378832908307e-34,
0.11498676240444183,
-0.01282719150185585,
-0.021795451641082764,
0.01839609444141388,
0.028625061735510826,
-0.10690007358789444,
0.018752099946141243,
0.009098738431930542,
-0.037740256637334824,
-0.08564247190952301,
-0.020773259922862053,
0.03151840344071388,
-0.021921001374721527,
0.049614518880844116,
0.06592331826686859,
0.001329426420852542,
-0.038719698786735535,
-0.056895919144153595,
-0.05288209021091461,
0.05612728372216225,
-0.009431261569261551,
0.008339449763298035,
-0.07184285670518875,
0.010066612623631954,
-0.08669094741344452,
0.10376447439193726,
-0.018376123160123825,
0.07051559537649155,
-0.035132262855768204,
0.04546558856964111,
0.009978572838008404,
-0.05549917370080948,
0.01405212190002203,
-0.027505481615662575,
-0.05488190799951553,
0.036070626229047775,
0.034785229712724686,
0.015388192608952522,
-0.02715316414833069,
-0.017084762454032898,
0.006666902452707291,
-0.05045754089951515,
-0.021669358015060425,
0.08934630453586578,
-0.021297113969922066,
0.026345329359173775,
-0.03666142374277115,
0.02947288565337658,
-0.0799289345741272,
0.06817829608917236,
-0.09391860663890839,
0.12663938105106354,
-0.06522808969020844,
-0.014177534729242325,
0.008583354763686657,
-0.044027041643857956,
-0.09297408163547516,
-0.009557131677865982,
0.06536025553941727,
-0.01639353297650814,
0.023415582254529,
0.024033958092331886,
-0.04815637320280075,
0.026627182960510254,
0.04272972792387009,
0.002320043044164777,
-0.05353786423802376,
-0.04115338250994682,
-0.04566435143351555,
0.0015337482327595353,
-0.020752159878611565,
0.0007133072940632701,
0.04866394028067589,
0.019315462559461594,
-0.03243981674313545,
0.019662581384181976,
0.009584628976881504,
0.0007192158955149353,
0.08433045446872711,
0.029299994930624962,
-0.032619841396808624,
-0.0716337040066719,
0.08427123725414276,
0.11438342928886414,
0.08121256530284882,
0.0883120447397232,
0.005748126655817032,
0.06374318152666092,
-0.0046785869635641575,
-0.005535750184208155,
-0.052154842764139175,
0.015553225763142109,
0.004860213957726955,
0.06937836110591888,
0.020963724702596664,
-4.21674180017817e-8,
-0.07222216576337814,
0.049195028841495514,
0.09992004185914993,
-0.030332662165164948,
-0.01471642218530178,
0.036387573927640915,
0.004458313807845116,
0.09899015724658966,
-0.03745070472359657,
-0.02240574173629284,
0.05614916980266571,
-0.08378435671329498,
-0.001021013013087213,
-0.01636761799454689,
0.06981808692216873,
0.0818224623799324,
-0.004597317427396774,
-0.005747364833950996,
-0.020822182297706604,
-0.001486258814111352,
-0.03127383068203926,
0.023061728104948997,
0.015452356077730656,
0.0074084000661969185,
0.03018878400325775,
0.04808146134018898,
-0.043275415897369385,
0.10559120029211044,
0.11973536014556885,
-0.005704981274902821,
0.0030348629225045443,
0.03385036811232567,
0.03764449059963226,
-0.001431474694982171,
0.01709107495844364,
-0.008645450696349144,
-0.018028106540441513,
0.03791337087750435,
-0.018642010167241096,
-0.06360278278589249,
0.06896864622831345,
0.04792476445436478,
0.04299144446849823,
0.007601369172334671,
0.07060258090496063,
0.07645625621080399,
0.023889295756816864,
-0.07416845858097076,
0.0394759401679039,
0.034185100346803665,
-0.043532125651836395,
-0.03249024227261543,
-0.0472838319838047,
0.13976560533046722,
0.03495018556714058,
0.02251611463725567,
0.018109500408172607,
-0.04552532359957695,
0.04819135740399361,
0.08361698687076569,
-0.03638974204659462,
-0.05984340235590935,
-0.06129324436187744,
0.01709512248635292
] | 0 | 0 | 0 | 0 |
|
Tractable Multi-Agent Reinforcement Learning through Behavioral Economics | https://openreview.net/forum?id=stUKwWBuBm | [
"Eric Mazumdar",
"Kishan Panaganti",
"Laixi Shi"
] | Oral | A significant roadblock to the development of principled multi-agent reinforcement learning (MARL) algorithms is the fact that desired solution concepts like Nash equilibria may be intractable to compute. We show how one can overcome this obstacle by introducing concepts from behavioral economics into MARL. To do so, we imbue agents with two key features of human decision-making: risk aversion and bounded rationality. We show that introducing these two properties into games gives rise to a class of equilibria---risk-averse quantal response equilibria (RQE)---which are tractable to compute in \emph{all} $n$-player matrix and finite-horizon Markov games. In particular, we show that they emerge as the endpoint of no-regret learning in suitably adjusted versions of the games. Crucially, the class of computationally tractable RQE is independent of the underlying game structure and only depends on agents' degrees of risk-aversion and bounded rationality. To validate the expressivity of this class of solution concepts we show that it captures peoples' patterns of play in a number of 2-player matrix games previously studied in experimental economics. Furthermore, we give a first analysis of the sample complexity of computing these equilibria in finite-horizon Markov games when one has access to a generative model. We validate our findings on a simple multi-agent reinforcement learning benchmark. Our results open the doors for to the principled development of new decentralized multi-agent reinforcement learning algorithms. | behavioral economics, risk-aversion, multi-agent reinforcement learning, quantal response, bounded rationality | By incorporating risk aversion and bounded rationality into agents' decision-making processes, we introduced a computationally tractable equilibria class for matrix and Markov games which aligns with observed human behaviors. | 5,242 | null | [
-0.008942156098783016,
-0.014574315398931503,
-0.08982554078102112,
-0.0636376291513443,
-0.0036769944708794355,
0.08326777070760727,
0.027801642194390297,
-0.030872385948896408,
0.07399290055036545,
0.08776859194040298,
-0.042277004569768906,
-0.045746345072984695,
0.053125280886888504,
0.06541213393211365,
-0.00634244829416275,
-0.0349428728222847,
0.01447634119540453,
-0.01313482504338026,
-0.010898361913859844,
0.0037059152964502573,
0.01658724993467331,
0.014029351063072681,
0.02072243206202984,
-0.05904616042971611,
-0.04635002464056015,
-0.016665134578943253,
-0.020278166979551315,
0.05854185298085213,
-0.024952128529548645,
0.012273218482732773,
0.07547581940889359,
0.019374700263142586,
0.045805785804986954,
-0.0756782740354538,
-0.036615438759326935,
0.0712646022439003,
-0.0868239626288414,
0.00006036388367647305,
-0.06444662064313889,
0.021472901105880737,
0.005967872217297554,
0.0037024863995611668,
-0.039899300783872604,
0.027485119178891182,
0.04077853634953499,
-0.049335796386003494,
0.006777611095458269,
-0.08223370462656021,
0.03292149677872658,
-0.12256794422864914,
-0.026334194466471672,
0.02072267234325409,
-0.02456211857497692,
0.0640384629368782,
0.03299040347337723,
0.0010178386000916362,
-0.02475862205028534,
0.014286208897829056,
-0.039280444383621216,
-0.11702677607536316,
-0.02989882044494152,
-0.0911998301744461,
0.03249439597129822,
-0.03431209176778793,
-0.028619268909096718,
-0.01377642061561346,
-0.03962143138051033,
0.041756145656108856,
-0.03959955647587776,
0.035803649574518204,
-0.044134750962257385,
-0.07438183575868607,
-0.08962484449148178,
0.005854527931660414,
0.028667718172073364,
0.023013565689325333,
-0.03176109492778778,
-0.03079429641366005,
-0.0016064513474702835,
0.005330528132617474,
-0.0592692531645298,
-0.05205092951655388,
-0.010803132317960262,
0.007932441309094429,
0.06011999025940895,
-0.023215418681502342,
-0.023146094754338264,
0.037395283579826355,
0.18411490321159363,
-0.04141603037714958,
-0.05233624950051308,
0.041395317763090134,
0.0352344736456871,
-0.006162125617265701,
0.02212279476225376,
-0.011970290914177895,
-0.03650035336613655,
-0.058484192937612534,
-0.046407997608184814,
0.08886967599391937,
-0.015684159472584724,
0.002359910635277629,
-0.04910821467638016,
0.05473463609814644,
0.03441307693719864,
0.009661514312028885,
-0.02496226690709591,
0.006059112958610058,
0.06399324536323547,
-0.05981897935271263,
-0.020602786913514137,
0.0287270899862051,
0.024019205942749977,
0.06735870987176895,
-0.0034562929067760706,
0.05070079490542412,
-0.014674502424895763,
0.023307550698518753,
-0.0020185511093586683,
0.0680745393037796,
0.01584683172404766,
-0.0231681689620018,
0.07787679135799408,
0.11230078339576721,
0.03832446411252022,
-0.03099152259528637,
-0.0400034599006176,
7.134356690936187e-34,
0.042065128684043884,
-0.05222440883517265,
0.06276567280292511,
-0.02035520039498806,
0.05242272838950157,
0.010776380077004433,
0.07797221094369888,
-0.03830496966838837,
0.01257288083434105,
0.003529045730829239,
0.0664968490600586,
0.07989559322595596,
-0.0402180440723896,
0.03477705270051956,
0.09015525132417679,
0.021758053451776505,
-0.15418066084384918,
-0.016783827915787697,
0.01193629764020443,
-0.08697443455457687,
0.0795443132519722,
-0.022572876885533333,
0.03913946822285652,
-0.015709396451711655,
-0.010338686406612396,
-0.014510747976601124,
0.008893433026969433,
0.003338921582326293,
0.09637397527694702,
-0.003444383153691888,
0.00421183742582798,
0.013385527767241001,
-0.08323407173156738,
0.051031533628702164,
-0.024285612627863884,
0.030826348811388016,
-0.1397581398487091,
0.044910285621881485,
-0.024336058646440506,
-0.007796767167747021,
-0.02614463120698929,
-0.023379258811473846,
0.10116618126630783,
-0.010126547887921333,
-0.062243688851594925,
-0.06265540421009064,
-0.0012168013490736485,
-0.05456763505935669,
-0.10185910761356354,
-0.07899429649114609,
0.04496298357844353,
-0.029514841735363007,
0.03724433481693268,
-0.04555431008338928,
-0.06287974119186401,
0.007879006676375866,
0.007846372202038765,
0.06266515702009201,
-0.12874263525009155,
-0.007924053817987442,
-0.0028971582651138306,
-0.060336098074913025,
0.03805328160524368,
-0.019454045221209526,
-0.064475879073143,
0.12654027342796326,
-0.0552472323179245,
0.021870292723178864,
0.06194451078772545,
0.015096506103873253,
0.0269085094332695,
0.05874260142445564,
0.09834857285022736,
-0.06148657202720642,
0.04624025896191597,
-0.041313767433166504,
0.059790585190057755,
-0.02715456858277321,
0.04540036991238594,
0.01750199869275093,
-0.026008019223809242,
0.00017712698900140822,
-0.06959007680416107,
0.00309367710724473,
-0.01602485403418541,
-0.07759367674589157,
0.0033442876301705837,
-0.08215262740850449,
-0.01916680671274662,
0.06387046724557877,
-0.08092115819454193,
-0.00965873897075653,
-0.014903686009347439,
0.1008668839931488,
-0.003533931216225028,
-3.6202841955846194e-33,
-0.04969605803489685,
-0.09568005055189133,
-0.031153228133916855,
0.017686905339360237,
-0.000440794974565506,
-0.004487781785428524,
-0.0008898467058315873,
-0.02963520959019661,
0.05660261958837509,
-0.09225771576166153,
-0.08942526578903198,
0.08419393748044968,
0.12112962454557419,
0.04490574821829796,
-0.036486800760030746,
-0.05653686448931694,
0.0520632229745388,
0.012061547487974167,
0.017851654440164566,
0.0020986252930015326,
-0.015104000456631184,
0.045326948165893555,
-0.05004297196865082,
0.026285206899046898,
0.06578653305768967,
0.029787292703986168,
0.03758993372321129,
0.003073047613725066,
-0.018004508689045906,
0.07052620500326157,
-0.060794398188591,
-0.0445728674530983,
-0.036562830209732056,
-0.01341215055435896,
-0.025058496743440628,
0.06556650996208191,
0.038159213960170746,
0.02067563310265541,
-0.11556164920330048,
0.10126498341560364,
-0.004264242481440306,
-0.10039257258176804,
-0.04805782809853554,
0.04394729807972908,
0.014619836583733559,
-0.017238818109035492,
-0.0006869249627925456,
0.0031957365572452545,
0.0034685509745031595,
0.03699246048927307,
0.04453501105308533,
0.07234861701726913,
-0.046117812395095825,
0.06742686778306961,
-0.050455622375011444,
-0.016673512756824493,
0.010796374641358852,
0.008852191269397736,
0.09134440869092941,
0.001432235585525632,
-0.026274345815181732,
0.06534533947706223,
-0.07335535436868668,
0.05707873776555061,
0.01442298386245966,
0.04265088215470314,
-0.10162518173456192,
-0.03677191957831383,
0.04486165568232536,
-0.0518069863319397,
-0.01509436871856451,
-0.020295925438404083,
0.015526932664215565,
-0.054496146738529205,
0.11199359595775604,
0.03820164501667023,
-0.06750882416963577,
0.05820199474692345,
-0.0026302202604711056,
-0.041226454079151154,
-0.046168431639671326,
0.033632710576057434,
0.05855410918593407,
0.042806532233953476,
-0.01409298088401556,
-0.03872841224074364,
0.07509057968854904,
0.028090864419937134,
-0.00931395310908556,
-0.06179678812623024,
0.07852942496538162,
0.057386986911296844,
0.011058286763727665,
0.014016068540513515,
-0.11124707013368607,
-4.6589825331011525e-8,
0.0051097264513373375,
-0.051438283175230026,
0.05773574858903885,
-0.03941734507679939,
0.08039511740207672,
-0.013821288011968136,
-0.01052931696176529,
-0.0002890362811740488,
-0.011364572681486607,
0.05118763446807861,
0.04647297039628029,
0.10497409105300903,
0.05304303020238876,
-0.008780228905379772,
-0.07366348803043365,
0.06503153592348099,
-0.05937248095870018,
-0.027729932218790054,
-0.010712279006838799,
0.026774441823363304,
0.03526510298252106,
-0.0034620855003595352,
-0.045894306153059006,
-0.09029701352119446,
-0.050450459122657776,
-0.03554106503725052,
-0.05894772335886955,
0.0030328556895256042,
0.03477229177951813,
0.07848770171403885,
0.03792969509959221,
-0.007861962541937828,
0.05378195643424988,
0.04374400153756142,
0.025444025173783302,
0.021727925166487694,
0.010091410018503666,
-0.0240597203373909,
-0.004749078303575516,
-0.02798551507294178,
-0.016515983268618584,
0.053890522569417953,
-0.02070060558617115,
-0.0002281997149111703,
0.10873749852180481,
0.019287580624222755,
-0.10746436566114426,
-0.018245717510581017,
0.05180550739169121,
-0.06443236768245697,
0.03244329243898392,
0.019369300454854965,
0.012517794035375118,
0.0023249369114637375,
0.025383347645401955,
-0.04992913082242012,
-0.002122338395565748,
-0.08276401460170746,
0.07719472795724869,
0.009599987417459488,
0.009214324876666069,
-0.02231040596961975,
-0.024396969005465508,
-0.00972480233758688
] | 0 | 0 | 0 | 0 |
|
Improved Finite-Particle Convergence Rates for Stein Variational Gradient Descent | https://openreview.net/forum?id=sbG8qhMjkZ | [
"Sayan Banerjee",
"Krishna Balasubramanian",
"PROMIT GHOSAL"
] | Oral | We provide finite-particle convergence rates for the Stein Variational Gradient Descent (SVGD) algorithm in the Kernelized Stein Discrepancy ($\KSD$) and Wasserstein-2 metrics. Our key insight is that the time derivative of the relative entropy between the joint density of $N$ particle locations and the $N$-fold product target measure, starting from a regular initial distribution, splits into a dominant 'negative part' proportional to $N$ times the expected $\KSD^2$ and a smaller 'positive part'. This observation leads to $\KSD$ rates of order $1/\sqrt{N}$, in both continuous and discrete time, providing a near optimal (in the sense of matching the corresponding i.i.d. rates) double exponential improvement over the recent result by~\cite{shi2024finite}. Under mild assumptions on the kernel and potential, these bounds also grow polynomially in the dimension $d$. By adding a bilinear component to the kernel, the above approach is used to further obtain Wasserstein-2 convergence in continuous time. For the case of `bilinear + Mat\'ern' kernels, we derive Wasserstein-2 rates that exhibit a curse-of-dimensionality similar to the i.i.d. setting. We also obtain marginal convergence and long-time propagation of chaos results for the time-averaged particle laws. | Stein Variational Gradient Descent, Non-asymptotic Rates, Variational Inference | Near-optimal finite-particle, discrete-time rates for SVGD | 5,180 | 2409.08469 | [
-0.03790741041302681,
-0.060016315430402756,
0.04915113374590874,
0.009482678957283497,
0.06820549815893173,
0.0022559422068297863,
-0.01457817479968071,
0.045147597789764404,
0.11478909105062485,
-0.003484283806756139,
-0.03530089184641838,
0.11092063784599304,
-0.04745398461818695,
-0.028658119961619377,
-0.027415649965405464,
-0.02671801671385765,
0.0710320919752121,
0.05275927111506462,
-0.03900982812047005,
0.0629938542842865,
-0.004081603605300188,
0.06689789146184921,
0.04384934902191162,
-0.04968061298131943,
0.08916198462247849,
-0.11939781904220581,
0.04011069983243942,
0.021122315898537636,
0.019842052832245827,
-0.0023448930587619543,
0.0635092556476593,
-0.004194293636828661,
-0.03519652783870697,
0.02732206881046295,
-0.06374747306108475,
0.05678745359182358,
0.04130447655916214,
0.03311551734805107,
-0.037327274680137634,
0.05231088399887085,
-0.01345397811383009,
0.1044163629412651,
-0.03561476245522499,
-0.0316152460873127,
0.004851370584219694,
0.01283713337033987,
-0.022962575778365135,
-0.06580423563718796,
-0.00073202297789976,
-0.018778858706355095,
0.006657063029706478,
0.06116427481174469,
0.027840854600071907,
-0.030337020754814148,
-0.009614459238946438,
-0.04176829382777214,
0.014493541792035103,
0.03277624398469925,
0.041973862797021866,
0.01884344406425953,
0.05764785036444664,
-0.12846653163433075,
-0.015231812372803688,
-0.08010000735521317,
-0.03095163032412529,
-0.11818324029445648,
0.09375038743019104,
-0.06179856136441231,
-0.04926811531186104,
0.06368790566921234,
-0.024382032454013824,
0.09114450961351395,
-0.11366595327854156,
0.007557479664683342,
0.03196370229125023,
-0.0047651310451328754,
-0.005442274734377861,
-0.012909610755741596,
0.02809078060090542,
-0.11868615448474884,
0.028254644945263863,
0.022867612540721893,
0.018080057576298714,
-0.06749273836612701,
0.008847684599459171,
0.010915141552686691,
0.011007791385054588,
0.07563038170337677,
0.07322491705417633,
-0.0228588730096817,
-0.006589529570192099,
-0.008739356882870197,
-0.05332097411155701,
0.0036156626883894205,
-0.008148673921823502,
-0.02913287654519081,
0.007294377777725458,
0.07318070530891418,
0.07020004093647003,
0.050700843334198,
-0.01728510856628418,
-0.016984570771455765,
0.022266795858740807,
-0.045197922736406326,
0.03455526381731033,
-0.1041460931301117,
0.10289130359888077,
0.02146127074956894,
0.047535914927721024,
-0.016635840758681297,
0.08762013167142868,
-0.04279499128460884,
0.0007831047405488789,
0.09433396905660629,
-0.016617055982351303,
-0.027698690071702003,
0.039017677307128906,
0.0019273185171186924,
0.02848103642463684,
0.07055111229419708,
-0.015457114204764366,
0.05252359062433243,
-0.015692846849560738,
0.05587131902575493,
-0.047360505908727646,
0.1407535970211029,
0.0031678942032158375,
9.61130558157862e-34,
0.03676643595099449,
0.050851017236709595,
-0.027110060676932335,
-0.060574062168598175,
0.009575306437909603,
0.03962314501404762,
-0.031178612262010574,
-0.03698936477303505,
-0.009371209889650345,
-0.00770482886582613,
-0.031204136088490486,
-0.018933068960905075,
-0.06144285574555397,
0.0017812673468142748,
0.01555647049099207,
-0.06259671598672867,
-0.019268136471509933,
0.02809857577085495,
0.05757272243499756,
-0.016482315957546234,
0.09606993943452835,
-0.06691868603229523,
-0.024914389476180077,
0.06697800010442734,
0.021839264780282974,
-0.09212995320558548,
0.028816409409046173,
-0.04039716348052025,
-0.014322533272206783,
0.005470736883580685,
-0.07372815907001495,
0.04711364582180977,
-0.028622189536690712,
-0.03967985510826111,
0.07232926785945892,
-0.028623295947909355,
0.022750312462449074,
0.06578602641820908,
-0.051640499383211136,
-0.0864100530743599,
-0.020408662036061287,
0.03902510926127434,
-0.0061616962775588036,
-0.053821299225091934,
-0.09599777311086655,
-0.09929229319095612,
0.05498165637254715,
0.03648919612169266,
0.06336188316345215,
-0.09130992740392685,
-0.025044476613402367,
0.01879807934165001,
0.028726449236273766,
-0.051550425589084625,
0.08242427557706833,
0.0040277442894876,
0.018426083028316498,
-0.0020851807203143835,
0.054854005575180054,
0.08397610485553741,
-0.02464897558093071,
0.08562161028385162,
0.040740396827459335,
0.0014079113025218248,
0.011209411546587944,
0.015462635084986687,
0.07178240269422531,
-0.03691493719816208,
0.0016459167236462235,
-0.01105506531894207,
0.02976667322218418,
0.038145773112773895,
-0.018937960267066956,
-0.15312696993350983,
0.03890123963356018,
-0.05638614296913147,
0.0383000373840332,
-0.009986127726733685,
-0.006150119937956333,
-0.018559696152806282,
-0.036923330277204514,
-0.039395734667778015,
-0.04280237480998039,
-0.07564162462949753,
-0.043050654232501984,
-0.09521381556987762,
-0.023878851905465126,
-0.07648006081581116,
-0.1525917500257492,
-0.026339171454310417,
-0.011581544764339924,
-0.007065099198371172,
0.04149568825960159,
0.01768646016716957,
-0.042808905243873596,
-2.1505474457041895e-33,
-0.017507679760456085,
-0.025602860376238823,
0.03877618908882141,
0.07297192513942719,
-0.0016483075451105833,
0.06458171457052231,
-0.04572858661413193,
0.03881421685218811,
-0.04895506799221039,
-0.13795676827430725,
-0.007511206436902285,
-0.011548326350748539,
0.04013233631849289,
0.07174050807952881,
0.012941241264343262,
0.007506989408284426,
0.0022513261064887047,
0.07398199290037155,
0.03897852078080177,
-0.03793038800358772,
0.0037320428527891636,
0.05198471620678902,
0.040671203285455704,
0.03962766006588936,
-0.01608937606215477,
0.008254972286522388,
0.04048563539981842,
0.12528173625469208,
-0.051405832171440125,
-0.01854277029633522,
-0.0070495898835361,
-0.020724844187498093,
-0.04837482422590256,
0.01828443445265293,
-0.06931028515100479,
0.06396497786045074,
0.018602553755044937,
0.01675875298678875,
-0.006062971893697977,
0.050494346767663956,
0.02271031029522419,
0.046567369252443314,
-0.02431204356253147,
-0.0221287589520216,
0.004073353484272957,
-0.07702361047267914,
0.03114551678299904,
0.0124061806127429,
0.0373777337372303,
0.011527961120009422,
-0.016884297132492065,
0.01605614647269249,
-0.07823513448238373,
0.1524713784456253,
-0.07366053014993668,
0.023090146481990814,
-0.004452606663107872,
-0.0029451085720211267,
-0.016607198864221573,
-0.0329299122095108,
-0.11537625640630722,
0.018497411161661148,
0.022038230672478676,
0.0630379244685173,
-0.005641341675072908,
-0.03989371284842491,
-0.017891481518745422,
-0.05495036020874977,
0.054264482110738754,
0.018251335248351097,
-0.028379961848258972,
-0.05832832679152489,
0.03722451627254486,
0.054396118968725204,
-0.050803862512111664,
-0.06638047099113464,
0.1253776103258133,
0.006602017674595118,
-0.04022941365838051,
-0.05794932320713997,
-0.03354474902153015,
0.100645512342453,
-0.02116650901734829,
0.012088827788829803,
0.09238119423389435,
-0.054016247391700745,
0.019938647747039795,
-0.0394979752600193,
0.01580384559929371,
-0.022613413631916046,
0.019483773037791252,
-0.04075208678841591,
0.03467964753508568,
0.020073216408491135,
-0.010480131022632122,
-4.7975579064996055e-8,
0.07473086565732956,
-0.013144701719284058,
-0.005838396027684212,
0.00117674027569592,
0.1304568201303482,
-0.0736902728676796,
-0.012472058646380901,
0.04400693625211716,
-0.0870150476694107,
0.02103361487388611,
0.03980233147740364,
0.07394477725028992,
-0.04905526712536812,
-0.04258732497692108,
0.03819945827126503,
-0.0025496541056782007,
-0.05000700801610947,
-0.018408823758363724,
-0.009386472404003143,
-0.027685057371854782,
0.005792242009192705,
0.009724424220621586,
0.013689310289919376,
-0.03150448575615883,
-0.05696997418999672,
-0.03860964998602867,
-0.023739134892821312,
-0.08698855340480804,
0.03727933391928673,
-0.039088550955057144,
-0.07620298117399216,
0.07168889790773392,
-0.028405049815773964,
-0.007643986959010363,
-0.03646058216691017,
-0.048585664480924606,
-0.02741621434688568,
0.04622728377580643,
-0.015881942585110664,
0.04623465985059738,
0.026657138019800186,
0.013380494900047779,
0.002242741407826543,
-0.037012435495853424,
-0.023245224729180336,
-0.025107119232416153,
-0.06081521883606911,
0.0060555459931492805,
0.02272380329668522,
0.06097806245088577,
-0.012320498935878277,
0.059989701956510544,
0.0605076439678669,
0.06145628169178963,
0.06831485033035278,
-0.018128059804439545,
-0.027650706470012665,
-0.10381359606981277,
0.022504234686493874,
0.06425170600414276,
0.03103923797607422,
-0.03343488648533821,
-0.09818917512893677,
0.008239375427365303
] | 0 | 0 | 0 | 0 |
|
Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning | https://openreview.net/forum?id=gc8QAQfXv6 | [
"Gangwei Jiang",
"Caigao JIANG",
"Zhaoyi Li",
"Siqiao Xue",
"JUN ZHOU",
"Linqi Song",
"Defu Lian",
"Ying Wei"
] | Oral | Catastrophic forgetting (CF) poses a significant challenge in machine learning, where a model forgets previously learned information upon learning new tasks.
Despite the advanced capabilities of Large Language Models (LLMs), they continue to face challenges with CF during continual learning. The majority of existing research focuses on analyzing forgetting patterns through a singular training sequence, thereby overlooking the intricate effects that diverse tasks have on model behavior.
Our study explores CF across various settings, discovering that model forgetting is influenced by both the specific training tasks and the models themselves. To this end, we interpret forgetting by examining the function vector (FV), a compact representation of functions in LLMs, offering a model-dependent indicator for the occurrence of CF. Through theoretical and empirical analyses, we demonstrated that CF in LLMs primarily stems from biases in function activation rather than the overwriting of task processing functions.
Leveraging these insights, we propose a novel function vector guided training methodology, incorporating a regularization technique to stabilize the FV and mitigate forgetting. Empirical tests on four benchmarks confirm the effectiveness of our proposed training method, substantiating our theoretical framework concerning CF and model function dynamics. | Catastrophic forgetting; Large language model; Instruction tuning | null | 5,157 | 2502.11019 | [
0.025781307369470596,
-0.06675877422094345,
0.007749659474939108,
0.07144119590520859,
0.03885781392455101,
0.07205383479595184,
-0.020270181819796562,
-0.004217925947159529,
0.0407327301800251,
-0.05051298066973686,
-0.05050823837518692,
0.07197913527488708,
0.0020726420916616917,
0.0018462856533005834,
-0.04805370792746544,
-0.02497025579214096,
-0.07314948737621307,
0.05070353299379349,
-0.0786627009510994,
-0.04678862541913986,
0.024624861776828766,
0.08467501401901245,
0.0500832200050354,
0.021532779559493065,
-0.03986048325896263,
-0.012697955593466759,
-0.045724064111709595,
-0.04752407968044281,
0.011123978532850742,
-0.03386693075299263,
0.03389360010623932,
0.046620287001132965,
-0.05013053119182587,
0.0433671809732914,
-0.10917001217603683,
0.048749156296253204,
-0.0952828899025917,
0.03869616240262985,
-0.0003789085312746465,
-0.03941504284739494,
-0.05491696298122406,
0.001841220771893859,
0.0056452578864991665,
0.042075347155332565,
0.06667762994766235,
-0.01985977590084076,
0.030910233035683632,
-0.08289458602666855,
0.0017718685558065772,
0.04684273898601532,
-0.02254905365407467,
0.0012384895235300064,
-0.04378082603216171,
0.043828900903463364,
0.0629894882440567,
0.05631357058882713,
0.04812312498688698,
0.038083966821432114,
-0.06321348249912262,
0.005534552037715912,
-0.09628602862358093,
-0.13291451334953308,
-0.02008250541985035,
-0.022404218092560768,
-0.05339799076318741,
0.01219338271766901,
0.05055127292871475,
0.0366678424179554,
-0.04210037365555763,
0.0764756128191948,
-0.0657026544213295,
0.06562772393226624,
-0.040093064308166504,
0.03589978069067001,
0.0513896569609642,
0.08841720223426819,
-0.003076825523748994,
-0.04655318707227707,
0.049209799617528915,
0.01763172261416912,
0.003920070361346006,
0.025284167379140854,
0.014361472800374031,
-0.04189123585820198,
0.13209420442581177,
0.06369398534297943,
0.066013403236866,
-0.058775968849658966,
0.08912976831197739,
0.009801181964576244,
-0.017840461805462837,
-0.06804187595844269,
0.007909455336630344,
-0.02805294655263424,
-0.02015094831585884,
0.0005651218816637993,
0.0643232986330986,
-0.054308418184518814,
-0.030503585934638977,
0.021176910027861595,
0.015507585369050503,
0.037703048437833786,
-0.01035426277667284,
-0.018982475623488426,
-0.02828354388475418,
-0.00829765759408474,
-0.015268298797309399,
0.09155695140361786,
0.03743722662329674,
-0.07477373629808426,
-0.021543897688388824,
0.03373287618160248,
0.01998612843453884,
0.05821765586733818,
0.0015352205373346806,
-0.05574903264641762,
0.03677281364798546,
-0.08011074364185333,
-0.006741726770997047,
0.11292202770709991,
0.0038587902672588825,
0.046429071575403214,
0.03520963713526726,
0.06321128457784653,
-0.07220281660556793,
-0.008671346120536327,
-0.05051461234688759,
2.3088393313149264e-33,
-0.03867911174893379,
0.03620287775993347,
-0.00114184629637748,
-0.0030979884322732687,
0.0581367164850235,
-0.0461445115506649,
0.02697562240064144,
0.02376500703394413,
-0.052666377276182175,
0.002035219222307205,
-0.022382695227861404,
0.002856493927538395,
-0.04046538472175598,
0.06930826604366302,
0.028258154168725014,
0.008184045553207397,
0.014934559352695942,
0.03986210748553276,
-0.05060989037156105,
-0.04275038465857506,
0.07773111760616302,
-0.02875322848558426,
0.04735032096505165,
-0.09793447703123093,
0.07660233974456787,
0.006262184586375952,
0.06955187767744064,
0.030760115012526512,
-0.05464457720518112,
-0.0012264439137652516,
-0.11165463924407959,
0.014003868214786053,
-0.010578141547739506,
0.051436129957437515,
-0.009944263845682144,
-0.051058582961559296,
0.04126521944999695,
-0.05490408092737198,
-0.004824346397072077,
-0.05681255832314491,
0.00013638862583320588,
0.048854369670152664,
0.006087732035666704,
-0.013098337687551975,
-0.033616889268159866,
-0.12877905368804932,
0.1025436595082283,
-0.031034765765070915,
-0.013042495585978031,
-0.08287239819765091,
-0.013133295811712742,
0.020484596490859985,
-0.020533697679638863,
-0.02743658423423767,
-0.02627568319439888,
0.023275993764400482,
-0.016226232051849365,
-0.03120291978120804,
-0.07768663018941879,
0.0022468683309853077,
0.08163116872310638,
0.038569603115320206,
-0.023135656490921974,
0.0678868219256401,
0.026887282729148865,
-0.011333873495459557,
0.01033344678580761,
0.07577130943536758,
0.011171451769769192,
-0.015155509114265442,
0.04906436800956726,
-0.039825208485126495,
-0.00033148025977425277,
-0.04795171320438385,
0.05982638895511627,
-0.001091902842745185,
0.03285907208919525,
-0.20073042809963226,
0.01428360864520073,
0.00008400758815696463,
0.10557284951210022,
-0.06298008561134338,
-0.02175145596265793,
-0.043315786868333817,
-0.05451377108693123,
-0.105010487139225,
0.088272325694561,
-0.10434459894895554,
-0.04316937178373337,
0.020573316141963005,
-0.05328550562262535,
-0.024190932512283325,
0.02432500571012497,
-0.03553073853254318,
-0.022230975329875946,
-3.978292020387086e-33,
-0.07683715969324112,
0.0031892277766019106,
-0.03550717607140541,
0.036654483526945114,
-0.02079113945364952,
-0.04034411162137985,
-0.040033865720033646,
-0.006401637569069862,
-0.047967541962862015,
-0.16191211342811584,
0.0009365953737869859,
-0.06768186390399933,
-0.014621354639530182,
0.05223826318979263,
-0.022650327533483505,
0.014916225336492062,
-0.0547451488673687,
0.00848236121237278,
-0.03931192308664322,
0.02616444230079651,
0.002134336158633232,
0.052066586911678314,
0.001462670974433422,
0.028773389756679535,
0.008947172202169895,
-0.006682618521153927,
0.023734543472528458,
0.08192168176174164,
0.031544819474220276,
-0.02557568997144699,
-0.04788335785269737,
0.04647665098309517,
0.027772290632128716,
-0.03973821550607681,
0.031809426844120026,
0.009818975813686848,
0.056052565574645996,
-0.0023006396368145943,
-0.006507523823529482,
0.06420893967151642,
0.1178072988986969,
0.02754034847021103,
-0.03437909111380577,
0.012499739415943623,
0.0464954674243927,
-0.03357752785086632,
-0.011571589857339859,
0.01255950890481472,
0.05877980217337608,
0.010285930708050728,
-0.03038359433412552,
-0.0006710452726110816,
-0.12122856825590134,
0.00662721274420619,
-0.09118534624576569,
-0.062665194272995,
0.01194847747683525,
-0.08626484125852585,
0.014548766426742077,
-0.01850127801299095,
-0.0430111438035965,
-0.006927988026291132,
0.04434711113572121,
-0.07722306251525879,
0.04469475522637367,
-0.013016365468502045,
-0.020372841507196426,
0.0706414133310318,
0.021049147471785545,
0.03221576660871506,
0.04171132296323776,
0.03174116834998131,
-0.02363949455320835,
-0.01216065976768732,
-0.04284410923719406,
0.012659437023103237,
-0.055842939764261246,
-0.09268494695425034,
-0.050829458981752396,
0.02295362763106823,
0.03661965951323509,
0.02092774771153927,
-0.009550240822136402,
0.08339965343475342,
0.01435252744704485,
0.09419327974319458,
0.02652762085199356,
0.11098217964172363,
-0.0015280278166756034,
-0.03913603723049164,
-0.020380165427923203,
0.01121152751147747,
-0.02102806232869625,
0.005105039104819298,
-0.09237021952867508,
-4.294279776217991e-8,
-0.07759557664394379,
0.023492947220802307,
-0.0036882557906210423,
0.011082764714956284,
0.08685384690761566,
-0.07527130097150803,
-0.013246707618236542,
0.05260948836803436,
-0.07002513855695724,
-0.03579472377896309,
-0.0007470046984963119,
0.00918507855385542,
-0.013816341757774353,
-0.07131657749414444,
0.039191171526908875,
0.07184822112321854,
0.06259331852197647,
0.08514593541622162,
-0.016161523759365082,
-0.024222299456596375,
0.012912707403302193,
0.023054130375385284,
-0.008591534569859505,
-0.02151402272284031,
-0.01969214156270027,
-0.062053341418504715,
0.022183669731020927,
0.13250571489334106,
0.1251203715801239,
-0.008914906531572342,
0.019696848466992378,
0.014689533971250057,
0.005282021127641201,
-0.023014118894934654,
-0.030329130589962006,
0.08272256702184677,
0.1034773588180542,
-0.07216808199882507,
-0.0055472939275205135,
0.000017170850696857087,
0.004609707277268171,
0.0351821705698967,
0.05306631699204445,
0.05128471925854683,
-0.013198264874517918,
0.02197025902569294,
-0.05188348889350891,
-0.052558474242687225,
0.019843297079205513,
-0.056329626590013504,
0.013266376219689846,
0.09018177539110184,
-0.05174374580383301,
0.12745998799800873,
0.10732030123472214,
0.03126419708132744,
-0.011425269767642021,
-0.0146727804094553,
-0.06812667101621628,
0.06443073600530624,
0.019220037385821342,
0.0471174456179142,
-0.028569960966706276,
-0.006355922203511
] | 0 | 0 | 0 | 0 |
|
One Step Diffusion via Shortcut Models | https://openreview.net/forum?id=OlzB6LnXcS | [
"Kevin Frans",
"Danijar Hafner",
"Sergey Levine",
"Pieter Abbeel"
] | Oral | Diffusion models and flow matching models have enabled generating diverse and realistic images by learning to transfer noise to data. However, sampling from these models involves iterative denoising over many neural network passes, making generation slow and expensive. Previous approaches for speeding up sampling require complex training regimes, such as multiple training phases, multiple networks, or fragile scheduling. We introduce Shortcut Models, a family of generative models that use a single network and training phase to produce high-quality samples in a single or multiple sampling steps. Shortcut models condition the network not only on the current noise level but also on the desired step size, allowing the model to skip ahead in the generation process. Across a wide range of sampling step budgets, shortcut models consistently produce higher quality samples than previous approaches, such as consistency models and reflow. Compared to distillation, shortcut models reduce complexity to a single network and training phase and additionally allow varying step budgets at inference time. | diffusion, flow-matching, fast inference, distillation | null | 5,115 | 2410.12557 | [
-0.04025096818804741,
-0.11531088501214981,
0.12205927073955536,
0.037564005702733994,
0.041082389652729034,
-0.060439594089984894,
-0.09416024386882782,
-0.07750524580478668,
0.018887698650360107,
-0.0648561492562294,
-0.07657978683710098,
-0.04603038355708122,
0.005941299255937338,
-0.016402466222643852,
-0.026596296578645706,
-0.023357464000582695,
-0.009654732421040535,
0.07327235490083694,
-0.0644545778632164,
-0.06181631609797478,
-0.025539588183164597,
-0.0017913696356117725,
-0.029943833127617836,
-0.022140977904200554,
0.04465319216251373,
-0.006509318482130766,
0.03714396432042122,
-0.046074215322732925,
0.07676329463720322,
-0.054592955857515335,
0.07794380187988281,
0.00021191313862800598,
-0.05109735578298569,
0.002485991222783923,
-0.0037392782978713512,
0.03226195648312569,
-0.02120554819703102,
0.035388194024562836,
0.08334043622016907,
0.029347391799092293,
0.06686827540397644,
0.05062451213598251,
0.01731116883456707,
-0.006561600137501955,
0.0755961611866951,
-0.035689592361450195,
0.003900873940438032,
-0.040514297783374786,
0.028965305536985397,
0.051251932978630066,
-0.041221246123313904,
0.003943517338484526,
-0.0509677454829216,
0.0593314990401268,
0.017291130498051643,
0.007730059325695038,
0.054383620619773865,
0.05462529882788658,
-0.019321298226714134,
-0.008050799369812012,
0.002273841528221965,
-0.1205819770693779,
-0.07458114624023438,
-0.030523665249347687,
-0.02184743992984295,
-0.012532798573374748,
0.006306818686425686,
0.04491716995835304,
0.07200119644403458,
0.03426302596926689,
0.006962135899811983,
0.05193428322672844,
-0.08221707493066788,
0.037608761340379715,
-0.05293932184576988,
0.03256414458155632,
0.0721917524933815,
-0.010350065305829048,
0.01693110726773739,
-0.09174765646457672,
0.011652244254946709,
-0.009610123001039028,
0.07150377333164215,
-0.03294798359274864,
0.07573942840099335,
0.004022915381938219,
-0.010862123221158981,
0.054788921028375626,
0.0029066770803183317,
-0.04820500686764717,
-0.08564994484186172,
0.06946761906147003,
-0.04099311679601669,
-0.02087067998945713,
0.045161549001932144,
0.06089978665113449,
0.03282729908823967,
-0.038342494517564774,
-0.03544032573699951,
0.08743178844451904,
-0.008657807484269142,
0.06111807003617287,
0.05180469900369644,
-0.0367085337638855,
0.043622005730867386,
-0.07381495088338852,
0.11204207688570023,
0.04110710695385933,
-0.01934441365301609,
0.010005605407059193,
-0.01352072786539793,
0.04967713728547096,
0.04393116012215614,
0.0005735089653171599,
0.04716196656227112,
-0.05378974601626396,
-0.04622450843453407,
-0.07581917196512222,
0.009000652469694614,
0.05389327555894852,
-0.055247120559215546,
0.016803843900561333,
-0.10430051386356354,
-0.002079575788229704,
-0.008869959972798824,
-0.04853683337569237,
-0.01675248332321644,
4.615493345163277e-33,
-0.01877181977033615,
0.00836378987878561,
0.03212568908929825,
0.0074577415362000465,
0.08356107026338577,
0.04241611063480377,
-0.00019881762273143977,
-0.002884818473830819,
0.0074410103261470795,
-0.02236061729490757,
-0.030766697600483894,
-0.08552160859107971,
-0.12674151360988617,
0.12147609144449234,
-0.031668826937675476,
-0.058338943868875504,
-0.004352645482867956,
0.04527651518583298,
0.023402445018291473,
0.04751499742269516,
0.003032580018043518,
-0.08669471740722656,
-0.0641063004732132,
0.036207690834999084,
-0.012044855393469334,
0.040584951639175415,
0.07513664662837982,
0.05144105851650238,
0.009652486070990562,
0.00021396452211774886,
-0.09393980354070663,
0.0006598051986657083,
-0.010008517652750015,
0.024132966995239258,
0.06579221785068512,
-0.014790755696594715,
0.09059073030948639,
0.016044380143284798,
-0.006943315267562866,
-0.0036522317677736282,
-0.026215095072984695,
0.0474051795899868,
0.007448130287230015,
-0.02243877574801445,
-0.10963781177997589,
-0.05575723946094513,
-0.03140692040324211,
-0.02932087332010269,
-0.023965192958712578,
0.013983780518174171,
0.033233270049095154,
-0.007360060699284077,
-0.007178148720413446,
-0.06585406512022018,
-0.06681064516305923,
-0.00627683661878109,
0.061820369213819504,
-0.04720400646328926,
0.05453215166926384,
0.05007452890276909,
-0.01656172424554825,
0.02853880263864994,
-0.07992418855428696,
0.12439566850662231,
0.09649201482534409,
0.05824778974056244,
-0.02234957367181778,
0.060109447687864304,
0.05397333949804306,
-0.0420769602060318,
-0.017653021961450577,
-0.03585443273186684,
-0.11248350888490677,
-0.08232797682285309,
0.07728506624698639,
0.012314954772591591,
0.009899396449327469,
-0.040216293185949326,
0.01495011243969202,
0.03867698088288307,
-0.10772259533405304,
0.002583997091278434,
-0.056305043399333954,
-0.07128480076789856,
0.03992493450641632,
0.04252785071730614,
-0.023732801899313927,
-0.07123670727014542,
-0.029630078002810478,
-0.06112918630242348,
-0.08522532135248184,
0.016517596319317818,
0.003507314482703805,
0.042241547256708145,
0.03380754962563515,
-3.7976313956351625e-33,
0.004856538958847523,
0.0871051773428917,
0.02450234815478325,
0.11587199568748474,
0.0006969033274799585,
0.014901557005941868,
-0.011075966991484165,
-0.005608349107205868,
0.1261865794658661,
-0.012339268811047077,
-0.023342447355389595,
-0.0072988662868738174,
0.0024158537853509188,
-0.09920778125524521,
0.008486250415444374,
-0.04444398730993271,
0.03298649564385414,
-0.04618797451257706,
0.010724042542278767,
-0.023218313232064247,
-0.004422815982252359,
0.007901886478066444,
-0.05797567591071129,
0.010022320784628391,
-0.11994348466396332,
0.08310548216104507,
-0.04650113731622696,
0.08376670628786087,
0.03674076125025749,
-0.01750284805893898,
-0.04060163348913193,
-0.008640359155833721,
0.05748090147972107,
-0.07292599231004715,
-0.052788980305194855,
0.12846986949443817,
0.030265217646956444,
0.07366747409105301,
-0.04821917414665222,
0.07937424629926682,
0.049172889441251755,
-0.06585215032100677,
-0.050892431288957596,
0.07209503650665283,
-0.012346995994448662,
-0.00899029802531004,
-0.04003513976931572,
-0.005186930764466524,
-0.059128452092409134,
0.06163781136274338,
0.046355146914720535,
0.06396627426147461,
-0.07434555143117905,
0.06143752858042717,
-0.012611313723027706,
-0.07011286169290543,
-0.06604354083538055,
-0.0066399360075592995,
0.0510391965508461,
-0.0034668827429413795,
-0.04606406018137932,
-0.058235861361026764,
-0.042833976447582245,
-0.05736752972006798,
-0.0054231127724051476,
-0.05644937977194786,
-0.023644985631108284,
0.03314339742064476,
0.008781016804277897,
0.06362546980381012,
-0.008603361435234547,
-0.003731889883056283,
0.038336072117090225,
0.012064292095601559,
-0.019066281616687775,
-0.12887568771839142,
-0.0031837441492825747,
-0.028033332899212837,
0.057960424572229385,
-0.03478051722049713,
0.05353931337594986,
-0.018636291846632957,
-0.03717537224292755,
-0.014378254301846027,
0.10929567366838455,
0.012203366495668888,
-0.01557881385087967,
-0.025219453498721123,
0.010557818226516247,
0.05031981319189072,
-0.00656380457803607,
-0.009215754456818104,
0.057479389011859894,
0.007609170861542225,
-0.12420401722192764,
-4.4085865624765574e-8,
-0.02773841843008995,
0.10377378016710281,
0.06335882097482681,
0.03345571085810661,
0.048627015203237534,
-0.0626976415514946,
0.033720314502716064,
0.11350920796394348,
0.055330295115709305,
-0.03417513892054558,
0.1113182082772255,
0.009899883531033993,
-0.017694754526019096,
-0.00932983122766018,
0.004312239587306976,
0.0093730129301548,
0.025775527581572533,
0.0392998531460762,
0.012647490948438644,
0.0040297918021678925,
-0.030330104753375053,
-0.009392397478222847,
-0.01923443004488945,
-0.0402955561876297,
0.1032099798321724,
-0.04692849889397621,
0.029599210247397423,
0.03535831347107887,
0.09414789080619812,
-0.022937403991818428,
-0.012482655234634876,
0.044291410595178604,
0.04041818529367447,
-0.0033855601213872433,
-0.025151483714580536,
0.02179836481809616,
0.04414604976773262,
0.003726460039615631,
-0.024312861263751984,
-0.10136918723583221,
-0.034220196306705475,
0.021914727985858917,
-0.037367381155490875,
-0.03516067564487457,
-0.08588199317455292,
-0.0003835690440610051,
0.03636891767382622,
-0.058163903653621674,
0.00479830102995038,
0.057392366230487823,
0.00430503161624074,
-0.028297586366534233,
0.06461097300052643,
-0.008748425170779228,
-0.006428608670830727,
-0.03269944712519646,
0.03160998970270157,
-0.0212921854108572,
0.06552442163228989,
0.06492504477500916,
0.04440397024154663,
0.013108870945870876,
-0.01021107193082571,
0.01517170574516058
] | https://github.com/kvfrans/shortcut-models | 472 | 0 | 0 | 0 |
Judge Decoding: Faster Speculative Sampling Requires Going Beyond Model Alignment | https://openreview.net/forum?id=mtSSFiqW6y | [
"Gregor Bachmann",
"Sotiris Anagnostidis",
"Albert Pumarola",
"Markos Georgopoulos",
"Artsiom Sanakoyeu",
"Yuming Du",
"Edgar Schönfeld",
"Ali Thabet",
"Jonas K Kohler"
] | Oral | The performance of large language models (LLMs) is closely linked to their underlying size, leading to ever-growing networks and hence slower inference. Speculative decoding has been proposed as a technique to accelerate autoregressive generation, leveraging a fast draft model to propose candidate tokens, which are then verified in parallel based on their likelihood under the target model. While this approach guarantees to reproduce the target output, it incurs a substantial penalty: many high-quality draft tokens are rejected, even when they represent objectively valid continuations. Indeed, we show that even powerful draft models such as GPT-4o, as well as human text cannot achieve high acceptance rates under the standard verification scheme. This severely limits the speedup potential of current speculative decoding methods, as an early rejection becomes overwhelmingly likely when solely relying on alignment of draft and target.
We thus ask the following question: Can we adapt verification to recognize correct, but non-aligned replies? To this end, we draw inspiration from the LLM-as-a-judge framework, which demonstrated that LLMs are able to rate answers in a versatile way. We carefully design a dataset coined TokenCourt to elicit the same capability in the target model by training a compact module on top of the embeddings to produce ``judgements" of the current continuation. We showcase our strategy on the Llama-3.1 family, where our 8B/405B-Judge achieves a speedup of $9\times$ over Llama-405B, while maintaining its quality on a large range of benchmarks. These benefits remain present even in optimized inference frameworks, where our method reaches up to $141$ tokens/s for 8B/70B-Judge and $129$ tokens/s for 8B/405B on $2$ and $8$ H100s respectively. | LLM inference, speculative decoding | null | 5,114 | 2501.19309 | [
-0.10726683586835861,
-0.07039820402860641,
0.010234213434159756,
0.029484333470463753,
0.08962767571210861,
-0.0031325826421380043,
-0.009805622510612011,
-0.006968858186155558,
0.114081472158432,
-0.008963501080870628,
-0.06018055975437164,
-0.03172854706645012,
0.03400784358382225,
-0.0156539436429739,
-0.02846274897456169,
-0.0057881781831383705,
0.10649793595075607,
0.06942597776651382,
-0.05786888673901558,
-0.04205610603094101,
-0.02825985476374626,
0.014125286601483822,
0.053510647267103195,
0.018033603206276894,
0.034522660076618195,
-0.06917519867420197,
-0.03642638027667999,
-0.010403691790997982,
-0.006206800229847431,
0.016104048117995262,
0.047915976494550705,
0.09194806963205338,
-0.058720577508211136,
0.021402252838015556,
-0.030903976410627365,
0.0647830069065094,
-0.02647414430975914,
-0.010746818035840988,
0.06789412349462509,
-0.06525566428899765,
-0.044440917670726776,
-0.0051650675013661385,
0.011740491725504398,
0.03861086070537567,
0.09223680198192596,
-0.045089107006788254,
-0.02815130539238453,
0.01054796390235424,
-0.05133695900440216,
0.016391580924391747,
-0.07968456298112869,
-0.006122450344264507,
0.0049844710156321526,
0.004316533450037241,
-0.07849276065826416,
-0.0348697267472744,
-0.0017377479234710336,
0.03957725688815117,
0.02052852138876915,
-0.0599399209022522,
-0.08335641771554947,
-0.1279822587966919,
-0.078447625041008,
-0.023951219394803047,
-0.0035274357069283724,
-0.04671880602836609,
0.024622134864330292,
0.06473132222890854,
-0.039244651794433594,
0.08172793686389923,
-0.01946237124502659,
0.05340178683400154,
-0.0033428126480430365,
0.06704138964414597,
0.0011015785858035088,
0.09026957303285599,
0.05678583309054375,
-0.03782149404287338,
0.07946845144033432,
-0.07736305147409439,
-0.006188629195094109,
-0.010948221199214458,
0.07482238113880157,
-0.03880012780427933,
0.03298638015985489,
-0.017672337591648102,
0.058082327246665955,
0.07528121769428253,
0.0314861498773098,
0.0251917764544487,
-0.08289755880832672,
-0.11166220158338547,
0.06916555762290955,
0.014615542255342007,
0.06495087593793869,
0.06560954451560974,
0.002919895574450493,
-0.06189846247434616,
-0.0022572402376681566,
0.09802141785621643,
0.00984773226082325,
0.1445252150297165,
0.0005338785122148693,
-0.09963420778512955,
-0.011585268191993237,
-0.07927358895540237,
-0.017980553209781647,
0.05903380364179611,
0.04123232141137123,
-0.07989373803138733,
-0.0031830393709242344,
0.03701189532876015,
0.05608184635639191,
0.02850678190588951,
0.03259444609284401,
-0.05954256281256676,
0.011134127154946327,
-0.03448184207081795,
-0.00879593100398779,
0.03957975655794144,
-0.03308532014489174,
0.05579760670661926,
-0.02196769416332245,
0.009660976007580757,
0.026673397049307823,
-0.00028186978306621313,
0.007991029880940914,
4.668614026712433e-34,
0.05206387862563133,
0.06402316689491272,
0.05560553818941116,
0.07158712297677994,
-0.052785832434892654,
-0.01566741243004799,
0.0019549261778593063,
0.007256573531776667,
-0.022531405091285706,
-0.04730508103966713,
-0.020446650683879852,
-0.008155607618391514,
-0.010614041239023209,
0.049765389412641525,
-0.008350546471774578,
0.0036807896103709936,
-0.06559138745069504,
0.035031527280807495,
-0.06359101831912994,
-0.001555230119265616,
-0.00015108802472241223,
-0.013346023857593536,
-0.008078238926827908,
-0.08352075517177582,
0.0007787770009599626,
-0.014594090171158314,
0.03290655463933945,
-0.04625421762466431,
-0.06568453460931778,
0.04556556046009064,
-0.14054842293262482,
-0.05533510819077492,
0.00722520612180233,
0.05970252305269241,
0.047895338386297226,
-0.05784458667039871,
0.013170111924409866,
-0.07808370143175125,
0.004362908191978931,
-0.04284624010324478,
0.004660590086132288,
0.06414543837308884,
0.02144472487270832,
0.0729871466755867,
-0.0767771303653717,
-0.06717314571142197,
-0.029711086302995682,
-0.02947271429002285,
-0.026484673842787743,
-0.042538780719041824,
0.09450074285268784,
0.06344783306121826,
-0.10264994949102402,
-0.010936789214611053,
-0.030222611501812935,
-0.048713795840740204,
-0.032584115862846375,
-0.0018442420987412333,
0.025237008929252625,
0.042540937662124634,
0.03641163930296898,
0.06694979965686798,
-0.014198179356753826,
0.04155505448579788,
-0.0016596399946138263,
0.037133462727069855,
-0.08679240942001343,
0.06613234430551529,
0.022201454266905785,
0.029119154438376427,
-0.00025526894023641944,
-0.01848210208117962,
-0.02853972092270851,
-0.013962563127279282,
0.03364034742116928,
-0.001137309125624597,
0.0929967388510704,
-0.001302964286878705,
0.014189657755196095,
-0.01270452793687582,
0.026931678876280785,
-0.017456060275435448,
-0.02792561799287796,
-0.020201902836561203,
-0.035525254905223846,
-0.11202403157949448,
0.060251276940107346,
-0.05321146547794342,
-0.022562244907021523,
-0.061731040477752686,
-0.09058532118797302,
-0.03904132917523384,
-0.03070175088942051,
-0.0165533609688282,
0.018851444125175476,
-9.36655571912292e-34,
-0.03797515481710434,
-0.03167170286178589,
-0.04490727186203003,
0.0856509804725647,
-0.0912395790219307,
-0.09591399878263474,
-0.04892367497086525,
-0.026749027892947197,
-0.020920036360621452,
-0.07768144458532333,
-0.07733796536922455,
-0.059617556631565094,
0.0733676478266716,
-0.013615588657557964,
0.03770381584763527,
-0.059279754757881165,
0.020564353093504906,
0.008665450848639011,
0.025145871564745903,
0.10889608412981033,
0.029517490416765213,
0.0359930582344532,
-0.1393929123878479,
0.07154097408056259,
0.045985642820596695,
0.06435465812683105,
0.0007302752346731722,
0.03769630193710327,
0.014643575996160507,
-0.02300477772951126,
-0.04871751368045807,
0.027125360444188118,
-0.023571481928229332,
-0.022487351670861244,
-0.04783277586102486,
0.05141586437821388,
0.08582758903503418,
0.017379816621541977,
0.0018591436091810465,
0.0844612792134285,
0.05221904069185257,
-0.009073554538190365,
-0.047422144562006,
0.019576529040932655,
-0.042664509266614914,
-0.0008803137461654842,
-0.0479598306119442,
0.0077051338739693165,
0.020407311618328094,
0.032392535358667374,
0.020934078842401505,
0.054685767740011215,
-0.04392170533537865,
0.0749485045671463,
-0.015833478420972824,
-0.13527826964855194,
-0.07678704708814621,
-0.04583953320980072,
0.02530692331492901,
0.037645235657691956,
-0.043305523693561554,
0.01918794773519039,
0.08288874477148056,
-0.07716687768697739,
0.06302548199892044,
-0.05475284904241562,
-0.012410413473844528,
-0.009584472514688969,
0.02906622365117073,
-0.005821313709020615,
0.10713976621627808,
-0.0856580063700676,
-0.025553591549396515,
0.08930835872888565,
0.0638829916715622,
-0.0072148204781115055,
0.06960349529981613,
-0.024390844628214836,
-0.0358453206717968,
-0.05154450610280037,
0.016565199941396713,
0.04925601929426193,
0.02235962077975273,
0.059777237474918365,
0.08091975748538971,
-0.019683636724948883,
0.023163553327322006,
0.0485408641397953,
-0.03171154484152794,
0.04178082197904587,
0.006027453113347292,
-0.010501896031200886,
0.047401148825883865,
0.06076379492878914,
-0.07708002626895905,
-4.530330599550325e-8,
-0.030864505097270012,
0.04388130456209183,
0.0008133179508149624,
0.11431895941495895,
0.006265053991228342,
-0.005336908623576164,
-0.06467033177614212,
-0.016406454145908356,
0.014969076961278915,
-0.07077810168266296,
0.102973572909832,
-0.007980825379490852,
-0.03863287717103958,
-0.06583888083696365,
-0.008550727739930153,
0.08904930204153061,
-0.007248636335134506,
-0.033626738935709,
-0.057986319065093994,
0.013593239709734917,
0.002144705271348357,
0.08916414529085159,
-0.035842739045619965,
-0.04086991026997566,
0.05734546482563019,
0.011631375178694725,
-0.01179988868534565,
0.06675644218921661,
0.017808591946959496,
0.0033581901807338,
-0.01927824132144451,
0.0025635119527578354,
-0.0527741014957428,
-0.05245685577392578,
0.007871055975556374,
0.08006023615598679,
0.055118490010499954,
0.020290834829211235,
0.034042902290821075,
-0.01927546039223671,
0.05356089770793915,
0.03420432657003403,
-0.12548936903476715,
0.012008361518383026,
0.019235458225011826,
-0.043431684374809265,
0.004787227604538202,
-0.0713033452630043,
-0.015278898179531097,
-0.031077580526471138,
0.043029773980379105,
-0.09577327221632004,
0.02468791790306568,
0.05151505768299103,
0.0027056264225393534,
0.0674794390797615,
-0.025751320645213127,
-0.010340986773371696,
0.02769094705581665,
0.04302698001265526,
0.0687786340713501,
0.02476280741393566,
-0.01811433769762516,
-0.00411644484847784
] | 0 | 0 | 0 | 0 |
|
Robustness Inspired Graph Backdoor Defense | https://openreview.net/forum?id=trKNi4IUiP | [
"Zhiwei Zhang",
"Minhua Lin",
"Junjie Xu",
"Zongyu Wu",
"Enyan Dai",
"Suhang Wang"
] | Oral | Graph Neural Networks (GNNs) have achieved promising results in tasks such as node classification and graph classification. However, recent studies reveal that GNNs are vulnerable to backdoor attacks, posing a significant threat to their real-world adoption. Despite initial efforts to defend against specific graph backdoor attacks, there is no work on defending against various types of backdoor attacks where generated triggers have different properties. Hence, we first empirically verify that prediction variance under edge dropping is a crucial indicator for identifying poisoned nodes. With this observation, we propose using random edge dropping to detect backdoors and theoretically show that it can efficiently distinguish poisoned nodes from clean ones. Furthermore, we introduce a novel robust training strategy to efficiently counteract the impact of the triggers. Extensive experiments on real-world datasets show that our framework can effectively identify poisoned nodes, significantly degrade the attack success rate, and maintain clean accuracy when defending against various types of graph backdoor attacks with different properties. Our code is available at: https://github.com/zzwjames/RIGBD. | Backdoor Defense, Graph Neural Network | null | 5,103 | 2406.09836 | [
-0.07160617411136627,
-0.011845278553664684,
0.019286075606942177,
0.0739273875951767,
0.11751489341259003,
-0.011031421832740307,
0.012472892180085182,
-0.07063207030296326,
-0.046529557555913925,
-0.015270081348717213,
-0.03216235712170601,
-0.03230525180697441,
0.09569871425628662,
-0.029069410637021065,
-0.08883583545684814,
0.001526898704469204,
0.045249246060848236,
0.05172762647271156,
0.008758150041103363,
-0.004272118676453829,
-0.07823031395673752,
0.02950521931052208,
-0.0004848660319112241,
-0.0026295615825802088,
0.043298665434122086,
-0.03767311945557594,
-0.026040008291602135,
-0.0217556394636631,
-0.01695059798657894,
-0.018615586683154106,
0.06630486994981766,
-0.08535457402467728,
-0.05841211602091789,
0.003625488141551614,
-0.06318569928407669,
0.001121977111324668,
-0.002736501395702362,
0.004248656332492828,
0.04557998105883598,
-0.010850556194782257,
0.00975063443183899,
0.04485726356506348,
0.02251146361231804,
0.052672091871500015,
-0.01502501405775547,
0.006340507417917252,
-0.011199076659977436,
-0.0013243088033050299,
-0.023617999628186226,
-0.05680638924241066,
-0.04424407705664635,
-0.003697185078635812,
-0.049807365983724594,
0.022900179028511047,
-0.01577356457710266,
-0.0948149636387825,
-0.006724134087562561,
0.030065497383475304,
-0.01576641947031021,
0.07849754393100739,
0.040442533791065216,
-0.06192722171545029,
-0.002070412505418062,
-0.04098755866289139,
-0.013513561338186264,
0.04928265139460564,
0.014051039703190327,
0.030204782262444496,
0.03632402792572975,
0.0385778471827507,
0.1094168946146965,
0.09120678901672363,
-0.08788846433162689,
0.015127879567444324,
0.019096367061138153,
0.14732690155506134,
-0.038403403013944626,
0.028464045375585556,
-0.04384665563702583,
-0.024186275899410248,
0.011289711110293865,
-0.02518957294523716,
0.027116550132632256,
0.027620330452919006,
0.05146418511867523,
0.058583371341228485,
-0.04460363835096359,
-0.009370631538331509,
0.01965109445154667,
0.06881783157587051,
-0.012958379462361336,
0.03782310336828232,
0.0832955613732338,
0.040868379175662994,
0.05206378176808357,
0.04699588194489479,
0.010511619038879871,
-0.08803535252809525,
-0.033419616520404816,
0.0956309586763382,
-0.09033698588609695,
-0.07279527187347412,
-0.006942916661500931,
-0.07042162120342255,
0.09863980859518051,
0.0761423408985138,
0.04193338006734848,
-0.04942461475729942,
0.025008317083120346,
-0.06211293861269951,
-0.007464470807462931,
0.09114006906747818,
-0.030274881049990654,
-0.03069409728050232,
0.042291220277547836,
-0.04065973684191704,
-0.01728011481463909,
-0.0385277159512043,
0.00009665665129432455,
0.0623759962618351,
-0.0012079410953447223,
0.04646807536482811,
-0.0028069510590285063,
-0.045312780886888504,
0.004563610069453716,
-0.017876625061035156,
-0.08298135548830032,
2.3362268803211472e-33,
0.040478792041540146,
0.07113880664110184,
-0.011910906061530113,
-0.06060738116502762,
0.06763119995594025,
-0.020348651334643364,
-0.060445889830589294,
0.03856617957353592,
0.03850160166621208,
0.05229436978697777,
-0.09501556307077408,
-0.059497032314538956,
-0.01123408880084753,
0.0411711111664772,
0.04300127923488617,
0.03767487406730652,
0.06081800535321236,
-0.004536524880677462,
-0.011915243230760098,
-0.11275610327720642,
-0.0024942567106336355,
-0.12416655570268631,
0.013077503070235252,
0.028397230431437492,
0.03651038557291031,
0.0537506528198719,
-0.02043454349040985,
0.07790704816579819,
0.038621388375759125,
0.010095864534378052,
-0.05074501410126686,
0.01646513305604458,
0.0626276507973671,
0.06211404129862785,
-0.022387538105249405,
-0.03174746409058571,
-0.08651279658079147,
-0.006006979383528233,
-0.015007883310317993,
-0.03643583506345749,
-0.05007006973028183,
0.035172607749700546,
-0.03763444349169731,
-0.0211240965873003,
0.06349778175354004,
-0.06464382261037827,
-0.1129760816693306,
-0.054204195737838745,
-0.04928863048553467,
-0.09883800148963928,
0.011096836067736149,
-0.01590018719434738,
0.032474592328071594,
0.055827829986810684,
-0.09113164991140366,
0.030763952061533928,
0.0617210678756237,
0.08601322025060654,
-0.03094959259033203,
0.10067382454872131,
0.04383331537246704,
0.011156890541315079,
-0.0493328832089901,
-0.019462190568447113,
-0.022746829316020012,
0.006440507713705301,
-0.017432846128940582,
0.025564847514033318,
-0.01730974018573761,
-0.018644258379936218,
-0.015972601249814034,
0.06850815564393997,
-0.035443373024463654,
-0.03610369935631752,
-0.08756915479898453,
0.008051054552197456,
0.045649755746126175,
-0.03804421052336693,
0.002142587210983038,
0.014275985769927502,
0.002665273379534483,
-0.0697908028960228,
0.057216864079236984,
0.03484136983752251,
-0.07335997372865677,
0.010169542394578457,
0.0025330495554953814,
-0.08909472078084946,
-0.013150418177247047,
0.0956922248005867,
-0.03538692742586136,
-0.079513318836689,
0.02049495279788971,
-0.013226817362010479,
-0.05366625636816025,
-3.133428598278599e-33,
-0.09910791367292404,
0.10922425240278244,
-0.012238667346537113,
0.03025948628783226,
-0.04024406149983406,
-0.025981470942497253,
-0.041025541722774506,
0.043846242129802704,
0.00024600367760285735,
0.048768188804388046,
0.011576350778341293,
-0.0036424819845706224,
0.060792580246925354,
-0.010812596417963505,
0.006166267674416304,
-0.02530963532626629,
0.0011106550227850676,
0.03248150274157524,
-0.0553770512342453,
-0.01396536361426115,
0.014737920835614204,
0.07079850882291794,
-0.0878731831908226,
-0.010946279391646385,
0.05058262497186661,
-0.034663401544094086,
-0.0007333427201956511,
0.0688246339559555,
0.018361639231443405,
0.02947673387825489,
-0.07293052226305008,
0.02019667439162731,
0.02305475063621998,
-0.013762734830379486,
0.031179629266262054,
0.02182294800877571,
0.053239502012729645,
-0.03923049941658974,
-0.002489972161129117,
0.020926583558321,
0.0389116108417511,
0.039198096841573715,
-0.09415736049413681,
0.035559602081775665,
-0.02001841366291046,
0.0988868921995163,
-0.03228829801082611,
0.07992994785308838,
-0.01660194993019104,
-0.03274552896618843,
-0.007335389964282513,
0.017969314008951187,
0.004204653203487396,
0.028424471616744995,
-0.06388864666223526,
-0.08765599876642227,
0.07353570312261581,
0.01297339890152216,
0.06055421382188797,
0.04177558794617653,
-0.06692787259817123,
-0.06272367388010025,
-0.07229320704936981,
0.042569443583488464,
-0.024862511083483696,
-0.05508210510015488,
-0.023795802146196365,
0.02040114812552929,
0.026415565982460976,
-0.027692291885614395,
0.10944779962301254,
0.037223950028419495,
-0.05108901485800743,
0.010178507305681705,
-0.0637732520699501,
-0.019199619069695473,
-0.033127088099718094,
-0.08475451916456223,
-0.01529879029840231,
0.05764073505997658,
0.05435506999492645,
-0.040163297206163406,
0.016211265698075294,
-0.007306304294615984,
0.09862134605646133,
0.039021413773298264,
0.004201721865683794,
0.02983560413122177,
0.060557037591934204,
0.06128377467393875,
-0.08623679727315903,
0.01070121768862009,
-0.06366033107042313,
0.03232676163315773,
-0.01825130358338356,
-3.8090913534460924e-8,
-0.03924749791622162,
0.0616995133459568,
0.002022114582359791,
0.03082626312971115,
0.07794457674026489,
0.07755199819803238,
-0.0004424944054335356,
0.08552245795726776,
-0.009204001165926456,
-0.034708473831415176,
0.08762984722852707,
-0.028118781745433807,
-0.04456844553351402,
-0.08541936427354813,
-0.034423261880874634,
-0.0031197883654385805,
-0.0775320827960968,
0.04309777915477753,
0.003909197170287371,
0.0016419299645349383,
-0.022987941280007362,
-0.008400509133934975,
0.022563965991139412,
0.03810817375779152,
0.06440999358892441,
-0.1435200273990631,
-0.05815209448337555,
0.04676809161901474,
0.00497896783053875,
0.09147458523511887,
-0.06469867378473282,
-0.05784369632601738,
0.07947415858507156,
0.01163374725729227,
-0.032725751399993896,
0.19595152139663696,
-0.015214103274047375,
-0.03358560428023338,
0.025155896320939064,
0.026072418317198753,
-0.01939498446881771,
0.0517364926636219,
-0.030754446983337402,
-0.05426594987511635,
-0.11739146709442139,
0.0012904936447739601,
0.07956255972385406,
-0.01387897226959467,
0.08500177413225174,
-0.04802032932639122,
0.028180072084069252,
-0.06154698505997658,
-0.008070852607488632,
0.00288861314766109,
-0.0065470049157738686,
0.040227800607681274,
-0.05303903669118881,
-0.07236433029174805,
0.0007304183673113585,
0.055954985320568085,
-0.001317086978815496,
0.0006074173143133521,
-0.010522277094423771,
-0.049489766359329224
] | 0 | 0 | 0 | 0 |
|
Proxy Denoising for Source-Free Domain Adaptation | https://openreview.net/forum?id=FIj9IEPCKr | [
"Song Tang",
"Wenxin Su",
"Yan Gan",
"Mao Ye",
"Jianwei Dr. Zhang",
"Xiatian Zhu"
] | Oral | Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained source model to an unlabeled target domain with no access to the source data. Inspired by the success of large Vision-Language (ViL) models in many applications, the latest research has validated ViL's benefit for SFDA by using their predictions as pseudo supervision. However, we observe that ViL's supervision could be noisy and inaccurate at an unknown rate, potentially introducing additional negative effects during adaption. To address this thus-far ignored challenge, we introduce a novel Proxy Denoising (__ProDe__) approach. The key idea is to leverage the ViL model as a proxy to facilitate the adaptation process towards the latent domain-invariant space. Concretely, we design a proxy denoising mechanism to correct ViL's predictions. This is grounded on a proxy confidence theory that models the dynamic effect of proxy's divergence against the domain-invariant space during adaptation. To capitalize the corrected proxy, we further derive a mutual knowledge distilling regularization. Extensive experiments show that ProDe significantly outperforms the current state-of-the-art alternatives under both conventional closed-set setting and the more challenging open-set, partial-set, generalized SFDA, multi-target, multi-source, and test-time settings. Our code and data are available at https://github.com/tntek/source-free-domain-adaptation. | Domain adaptation, source-free, multimodal proxy space, proxy confidence theory | null | 5,075 | 2406.01658 | [
-0.0712432786822319,
-0.1238275021314621,
0.05066178739070892,
0.05095476657152176,
0.1268313229084015,
-0.02273336984217167,
0.016385573893785477,
-0.07354508340358734,
0.03074382245540619,
-0.04068597033619881,
-0.04073334485292435,
-0.07182077318429947,
0.08686225861310959,
-0.014932730235159397,
0.03416290879249573,
0.004056009463965893,
0.1005006954073906,
0.08480750024318695,
-0.12038785964250565,
-0.08948677778244019,
0.04281114041805267,
-0.0012769998284056783,
-0.009100213646888733,
0.0016748231137171388,
0.042637601494789124,
-0.007110128179192543,
0.023905787616968155,
0.02143084444105625,
0.035606749355793,
-0.021140048280358315,
0.002065272070467472,
0.05336635559797287,
-0.1017056554555893,
-0.038921669125556946,
-0.014339565299451351,
0.01986389234662056,
-0.05724339187145233,
0.05759118124842644,
0.007863271981477737,
-0.02215459570288658,
0.026509422808885574,
0.05991044268012047,
0.01546219177544117,
0.04440968856215477,
-0.0015882840380072594,
-0.009876195341348648,
0.03530506417155266,
-0.045978665351867676,
-0.030118189752101898,
-0.008977183140814304,
-0.06333387643098831,
-0.03834092989563942,
-0.03172408789396286,
0.06280253827571869,
-0.03443164378404617,
0.013423525728285313,
0.032061539590358734,
0.03379205986857414,
-0.000436449859989807,
0.0234803669154644,
-0.037543270736932755,
-0.10225933790206909,
-0.033852674067020416,
0.021569563075900078,
-0.060431938618421555,
-0.006398087367415428,
-0.031919121742248535,
0.10636036098003387,
-0.028339048847556114,
0.006778736598789692,
-0.09686166793107986,
0.09158425778150558,
-0.020381487905979156,
0.031538963317871094,
0.05043259635567665,
0.06881097704172134,
0.017305703833699226,
-0.024431314319372177,
0.05297256261110306,
-0.13872838020324707,
0.09691988676786423,
0.05506138131022453,
0.04242982715368271,
0.04096087068319321,
0.14753177762031555,
0.02562115341424942,
0.03762024641036987,
-0.030011635273694992,
0.03383111581206322,
0.01528574526309967,
-0.08751316368579865,
-0.11906453222036362,
0.046435337513685226,
0.06556934118270874,
-0.010015839710831642,
0.037956636399030685,
0.008771602064371109,
-0.08896433562040329,
-0.014145145192742348,
0.10841072350740433,
-0.012822928838431835,
0.059246908873319626,
0.022310776636004448,
-0.03742652013897896,
-0.022985758259892464,
-0.039449118077754974,
0.07343214750289917,
0.10088064521551132,
0.022091515362262726,
-0.04123934730887413,
0.01219177432358265,
0.08846944570541382,
0.021175559610128403,
-0.02121473290026188,
0.004209398292005062,
-0.02174798958003521,
-0.01863890141248703,
-0.06999503076076508,
-0.02793804183602333,
-0.038230184465646744,
-0.018757255747914314,
0.0073114922270178795,
0.05710853636264801,
-0.018848825246095657,
0.049123313277959824,
0.025604864582419395,
-0.034841038286685944,
-5.886937697877922e-35,
0.048771608620882034,
0.048846159130334854,
0.04480656236410141,
-0.017180468887090683,
0.02568349428474903,
-0.0024065719917416573,
-0.0027795631904155016,
0.017551321536302567,
-0.03979792073369026,
-0.036334749311208725,
0.027442211285233498,
-0.05600552633404732,
-0.07155454903841019,
0.04958450049161911,
0.008069201372563839,
0.03665176033973694,
-0.02437138371169567,
0.04346180334687233,
-0.017695920541882515,
0.03808613121509552,
0.018847700208425522,
0.018634112551808357,
-0.07816310226917267,
-0.048354748636484146,
0.00798734836280346,
-0.0019020832842215896,
-0.012680375017225742,
-0.03068562224507332,
-0.00838657096028328,
0.02169160172343254,
-0.11447303742170334,
0.04322414472699165,
0.05171222984790802,
0.051399115473032,
0.0630408227443695,
-0.03924338519573212,
0.07166080176830292,
-0.06657621264457703,
-0.03257236257195473,
-0.009083542041480541,
0.03852951526641846,
0.01504332572221756,
0.021531876176595688,
-0.043932873755693436,
-0.018863024190068245,
-0.08637450635433197,
-0.004939087200909853,
-0.005271255504339933,
-0.08499281108379364,
-0.026422662660479546,
0.09485261887311935,
0.04847254976630211,
-0.058400411158800125,
-0.07887540012598038,
-0.07052095234394073,
-0.02463189885020256,
0.007698692847043276,
-0.008406554348766804,
0.04367610439658165,
0.027962887659668922,
-0.0026008549612015486,
0.02459166757762432,
-0.03218580409884453,
0.09701579809188843,
0.0780135989189148,
0.06834474205970764,
-0.000763110991101712,
-0.022765841335058212,
-0.011846961453557014,
-0.008525591343641281,
-0.03499874845147133,
0.04751014709472656,
-0.09513138234615326,
-0.024019217118620872,
0.03494720160961151,
-0.05245184153318405,
-0.026945769786834717,
-0.07593532651662827,
0.054282888770103455,
0.061467647552490234,
-0.015183824114501476,
-0.030497975647449493,
-0.08505338430404663,
-0.04110344499349594,
-0.07594826072454453,
-0.029476957395672798,
0.03172988072037697,
-0.09401827305555344,
-0.027122322469949722,
-0.11090409755706787,
0.03511863574385643,
-0.04042178764939308,
-0.005295888055115938,
0.010402713902294636,
0.02211715653538704,
-1.3351396161399047e-33,
-0.047408901154994965,
0.02536216750741005,
0.048269931226968765,
0.02989891916513443,
-0.04869937524199486,
-0.03483939543366432,
0.053911253809928894,
0.06370844691991806,
0.058262333273887634,
-0.07212024927139282,
-0.04388568922877312,
-0.08722605556249619,
-0.027122633531689644,
-0.014734048396348953,
-0.023541880771517754,
0.004945277702063322,
-0.014168789610266685,
-0.027875635772943497,
-0.0567251592874527,
0.04699437692761421,
0.06882742792367935,
0.060289982706308365,
-0.04822168126702309,
0.11299064010381699,
0.01010521873831749,
0.06038054823875427,
0.025733809918165207,
0.10939712822437286,
-0.04344791918992996,
0.003364281728863716,
-0.013135231100022793,
0.00034419193980284035,
-0.06557051092386246,
-0.06002392992377281,
-0.07676727324724197,
0.064956434071064,
-0.013653039000928402,
0.0168315377086401,
-0.06563512235879898,
0.02689642645418644,
0.007167475763708353,
-0.02529703639447689,
-0.07554435729980469,
-0.03353823348879814,
-0.02644417993724346,
-0.04115689918398857,
-0.08685876429080963,
-0.0005279487231746316,
0.08691848069429398,
0.011309926398098469,
-0.0061029656790196896,
0.04963283985853195,
-0.028975659981369972,
0.007836296223104,
-0.05116711184382439,
-0.08112513273954391,
-0.006469608284533024,
0.02307673543691635,
0.08864413946866989,
0.00011759359767893329,
-0.05930504575371742,
0.002635217970237136,
-0.00914760958403349,
-0.01461938675493002,
0.03127489611506462,
-0.002657526871189475,
-0.05643767490983009,
0.07730354368686676,
0.09716890007257462,
-0.03797812759876251,
0.059618085622787476,
0.007442358881235123,
-0.09355761855840683,
-0.028475625440478325,
-0.0026854132302105427,
-0.07131560891866684,
0.08432692289352417,
-0.05839547887444496,
-0.024095037952065468,
0.03709113597869873,
0.059371475130319595,
-0.03811706230044365,
0.09310418367385864,
0.024736808612942696,
0.019116878509521484,
0.11402574926614761,
0.020004432648420334,
0.02597268670797348,
0.02109687216579914,
-0.030775588005781174,
-0.07835740596055984,
-0.035066455602645874,
-0.04651637747883797,
0.014883490279316902,
-0.06461983919143677,
-4.160648003903589e-8,
-0.0843808576464653,
0.05934285372495651,
0.01758628338575363,
0.031087927520275116,
-0.0032394856680184603,
-0.09122440963983536,
-0.06477343291044235,
0.022519052028656006,
-0.015141613781452179,
-0.007063183933496475,
0.026910779997706413,
0.011891857720911503,
-0.03869997337460518,
-0.008738943375647068,
-0.00822138600051403,
0.04951990023255348,
0.07840251177549362,
0.1072327047586441,
-0.03948519378900528,
-0.0264226496219635,
-0.01777239516377449,
0.03777731955051422,
-0.018086377531290054,
-0.03200910985469818,
0.07226357609033585,
-0.09047190099954605,
0.010690638795495033,
0.05502188205718994,
0.11081260442733765,
0.04695883020758629,
-0.004526795819401741,
0.011472426354885101,
0.006041310727596283,
-0.01793571561574936,
-0.05004829168319702,
0.059948887676000595,
0.07239840924739838,
-0.046448905020952225,
-0.00258735497482121,
-0.036652788519859314,
-0.013704943470656872,
0.059086933732032776,
-0.042855724692344666,
-0.015187527984380722,
-0.040391381829977036,
0.00852582510560751,
0.03997769579291344,
-0.10750109702348709,
0.06493927538394928,
-0.00670792767778039,
0.03701893612742424,
0.01541993860155344,
0.01042756624519825,
0.07985341548919678,
0.04335522651672363,
0.02537583000957966,
0.011339779943227768,
-0.020732393488287926,
-0.04460234194993973,
0.03374800458550453,
0.01328317541629076,
0.01571512781083584,
0.036609143018722534,
0.04750969260931015
] | https://github.com/tntek/source-free-domain-adaptation | 175 | 0 | 0 | 0 |
Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risks of Language Models | https://openreview.net/forum?id=tc90LV0yRL | [
"Andy K Zhang",
"Neil Perry",
"Riya Dulepet",
"Joey Ji",
"Celeste Menders",
"Justin W Lin",
"Eliot Jones",
"Gashon Hussein",
"Samantha Liu",
"Donovan Julian Jasper",
"Pura Peetathawatchai",
"Ari Glenn",
"Vikram Sivashankar",
"Daniel Zamoshchin",
"Leo Glikbarg",
"Derek Askaryar",
"Haoxiang Yang",
"Aolin Zhang",
"Rishi Alluri",
"Nathan Tran",
"et al. (5 additional authors not shown)"
] | Oral | Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause real-world impact. Policymakers, model providers, and researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. We include 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structured bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. Anonymized code and data are available at https://drive.google.com/file/d/1kp3H0pw1WMAH-Qyyn9WA0ZKmEa7Cr4D4 and https://drive.google.com/file/d/1BcTQ02BBR0m5LYTiK-tQmIK17_TxijIy. | Language Model Agents, Benchmark, Cybersecurity, Risk | Cybench is a cybersecurity agent benchmark with 40 professional-level Capture the Flag tasks that are recent, meaningful, and difficult with subtasks. | 5,074 | 2408.08926 | [
-0.04413054138422012,
-0.03163900226354599,
-0.08739134669303894,
-0.04505855590105057,
0.06375610828399658,
0.03621784970164299,
0.05270831659436226,
0.0162751916795969,
0.005595236085355282,
-0.04454416036605835,
-0.04960346966981888,
-0.12552785873413086,
0.10187991708517075,
0.07798464596271515,
-0.006575257517397404,
0.007771761622279882,
0.07930555939674377,
0.018011461943387985,
-0.06443000584840775,
-0.049728117883205414,
0.00007315954280784354,
0.005625873804092407,
0.047079119831323624,
-0.052983563393354416,
-0.1076892837882042,
-0.027733515948057175,
-0.04932699725031853,
-0.09655540436506271,
-0.08180630952119827,
0.01989622414112091,
0.06440985202789307,
0.015762409195303917,
0.007401505019515753,
0.08434483408927917,
0.07007323205471039,
-0.0014380436623468995,
-0.013849725015461445,
-0.03408527374267578,
0.005167631898075342,
-0.054611992090940475,
-0.05004960298538208,
-0.061376817524433136,
-0.01727624051272869,
0.0076753245666623116,
-0.003622545627877116,
-0.05827989801764488,
-0.009431685321033001,
-0.04293233901262283,
0.0007563659455627203,
-0.0703112781047821,
-0.058524250984191895,
-0.07927138358354568,
0.04307781904935837,
0.039221473038196564,
0.0028683505952358246,
-0.01315699890255928,
-0.017687562853097916,
-0.0759158805012703,
0.027918348088860512,
-0.03217658773064613,
-0.03313600644469261,
0.008401689119637012,
-0.025644680485129356,
0.051327306777238846,
-0.0053912620060145855,
-0.003300119424238801,
-0.031706370413303375,
0.037648219615221024,
0.0027898859698325396,
-0.002622133819386363,
-0.10747699439525604,
0.036391809582710266,
-0.03493262454867363,
0.042497023940086365,
0.009917013347148895,
0.039537351578474045,
0.004985449369996786,
0.012932674959301949,
0.03499596565961838,
-0.1399945169687271,
0.01675715111196041,
-0.0033349439036101103,
0.030506232753396034,
0.0949924886226654,
0.021681156009435654,
0.08647511154413223,
0.04717240110039711,
0.042789027094841,
0.12182603031396866,
0.06904420256614685,
0.0290076807141304,
0.02431165985763073,
0.07748138904571533,
0.05727323144674301,
0.05502663180232048,
-0.0012102172477170825,
0.03568980097770691,
-0.08444654941558838,
-0.05408986285328865,
0.09718699753284454,
-0.059819284826517105,
-0.026091983541846275,
0.03867669031023979,
-0.10431011766195297,
-0.030706003308296204,
0.009279218502342701,
0.043707314878702164,
-0.013989310711622238,
0.012283575721085072,
-0.019206475466489792,
-0.013573612086474895,
0.008962045423686504,
0.03137541934847832,
-0.06666010618209839,
0.08775343000888824,
0.04560070112347603,
-0.02413717657327652,
0.04179216921329498,
0.0018125476781278849,
0.02464168146252632,
0.08222655206918716,
0.022291872650384903,
0.025014175102114677,
0.02121615782380104,
0.07637125998735428,
0.07302086800336838,
-0.05597963184118271,
1.9316506070316575e-34,
0.04740918427705765,
0.08047641068696976,
-0.00616415124386549,
0.03999929130077362,
0.05404023081064224,
-0.05416138097643852,
0.07046646624803543,
0.08100249618291855,
-0.0737185999751091,
-0.011121165938675404,
-0.07029560208320618,
0.0453820638358593,
-0.01620643585920334,
0.00798262283205986,
0.05573585629463196,
0.054873932152986526,
-0.039843276143074036,
0.04508666321635246,
-0.05157084763050079,
-0.008876164443790913,
0.044648345559835434,
-0.022643378004431725,
0.018201127648353577,
-0.0028514459263533354,
0.07592464983463287,
-0.01698640175163746,
-0.0024918760173022747,
-0.07803968340158463,
0.03275711089372635,
0.02808009274303913,
-0.12031571567058563,
0.012733140960335732,
-0.012034351006150246,
0.07074015587568283,
0.016105759888887405,
0.013546799309551716,
-0.06808765977621078,
-0.07528580725193024,
0.024369636550545692,
0.014690166339278221,
-0.05953500047326088,
-0.0036122037563472986,
0.01734156534075737,
-0.04820529371500015,
0.029657889157533646,
-0.11943288147449493,
-0.060238976031541824,
-0.00656184833496809,
0.02159876748919487,
-0.05065368860960007,
0.04014291986823082,
0.046505533158779144,
0.023145152255892754,
0.00560385687276721,
-0.05784233286976814,
0.0006979881436564028,
0.017464596778154373,
-0.003082564566284418,
0.03141235560178757,
0.06359798461198807,
-0.0073677534237504005,
0.05773606151342392,
-0.008452055044472218,
0.024599753320217133,
0.025381669402122498,
-0.05652719736099243,
-0.02957797609269619,
0.08934779465198517,
0.02980339713394642,
0.05012689530849457,
-0.10272404551506042,
-0.006627046503126621,
0.0582057386636734,
-0.016489127650856972,
0.006194572430104017,
-0.051330871880054474,
0.032848622649908066,
-0.11051047593355179,
0.028511231765151024,
0.028039291501045227,
-0.014760389924049377,
-0.047650255262851715,
0.044667985290288925,
0.007187318056821823,
-0.02335226722061634,
-0.07053391635417938,
-0.028525102883577347,
-0.04027573764324188,
-0.027655161917209625,
-0.010423568077385426,
-0.05391378700733185,
-0.08450743556022644,
-0.04520320147275925,
0.01563214883208275,
0.008624757640063763,
-2.0890938867401465e-33,
-0.04919541999697685,
-0.060415539890527725,
-0.12573844194412231,
0.048114605247974396,
-0.012806720100343227,
-0.07378736138343811,
-0.03150450065732002,
-0.03741210326552391,
-0.0340404286980629,
0.011674408800899982,
-0.06480693072080612,
-0.01631820946931839,
0.06667530536651611,
-0.05611461400985718,
0.0302564837038517,
-0.0664055272936821,
-0.036443598568439484,
-0.03704272210597992,
-0.01914949156343937,
0.05271840840578079,
0.007716415915638208,
0.10872337967157364,
-0.035762351006269455,
-0.012048468925058842,
0.006853524129837751,
0.038256555795669556,
0.05336455628275871,
0.04632256552577019,
0.06612719595432281,
0.047192174941301346,
0.01432881224900484,
0.03181075304746628,
-0.021013636142015457,
0.025013795122504234,
-0.0033422759734094143,
0.05679028108716011,
0.09281650185585022,
-0.05672673508524895,
0.030290715396404266,
0.08521483838558197,
0.07604115456342697,
-0.03588538244366646,
-0.04882274940609932,
-0.015609904192388058,
-0.026339340955018997,
0.08180755376815796,
-0.05919911339879036,
0.0336160734295845,
-0.003949947655200958,
-0.07757537811994553,
-0.008475125767290592,
0.03164806589484215,
-0.049225904047489166,
-0.017247065901756287,
-0.0652727484703064,
-0.01016093511134386,
-0.010488098487257957,
-0.0678144097328186,
-0.013635161332786083,
0.029000625014305115,
-0.025522226467728615,
0.015123559162020683,
-0.002397227333858609,
0.04543793201446533,
0.02341213822364807,
-0.03828050568699837,
-0.04548059403896332,
0.08528345823287964,
-0.047650307416915894,
-0.039304446429014206,
0.031557388603687286,
-0.03916337713599205,
-0.05435369908809662,
-0.02155131846666336,
0.022942351177334785,
-0.04167182371020317,
-0.03117084503173828,
-0.03616395592689514,
-0.034886281937360764,
0.04626225307583809,
0.01970464549958706,
-0.000033782875107135624,
0.007786372676491737,
-0.0017127766041085124,
0.05205487459897995,
0.05575554072856903,
-0.007983420975506306,
0.09260647743940353,
0.007658977061510086,
0.028220752254128456,
-0.047628916800022125,
0.007370638661086559,
-0.022827519103884697,
0.09037462621927261,
-0.0356791615486145,
-4.235700146182353e-8,
0.03064640238881111,
0.020405368879437447,
0.02854638360440731,
0.008977867662906647,
0.011525415815412998,
0.061145320534706116,
-0.0972381979227066,
-0.047679051756858826,
-0.004484268836677074,
-0.02854062244296074,
0.09108034521341324,
-0.040458500385284424,
-0.024133991450071335,
-0.04585139453411102,
-0.017690302804112434,
0.07160305976867676,
-0.07788167893886566,
-0.002794425468891859,
-0.03456159308552742,
-0.021134866401553154,
0.09776217490434647,
0.03657291829586029,
-0.02722099982202053,
-0.005853106267750263,
-0.022312678396701813,
-0.026143118739128113,
-0.0445680171251297,
0.0551237054169178,
-0.01424566749483347,
0.060794226825237274,
0.021994179114699364,
-0.03791842609643936,
0.0010465881787240505,
0.036214035004377365,
0.021238146349787712,
0.13649675250053406,
-0.032502204179763794,
-0.09771853685379028,
0.06757338345050812,
0.0004180855758022517,
-0.012157664634287357,
0.057753320783376694,
-0.03438446670770645,
-0.005979191977530718,
-0.01765546202659607,
-0.05169498175382614,
-0.10836362838745117,
-0.14672128856182098,
0.09838318079710007,
-0.09048958122730255,
-0.025434093549847603,
-0.00008179090946214274,
-0.03453048691153526,
0.16654366254806519,
0.08229836076498032,
0.08305098116397858,
-0.00686208950355649,
-0.07534635812044144,
0.04519013315439224,
0.07390962541103363,
0.004152639303356409,
-0.02029980719089508,
0.023947475478053093,
-0.004716264549642801
] | https://github.com/andyzorigin/cybench | 97 | 0 | 0 | 0 |
Open-YOLO 3D: Towards Fast and Accurate Open-Vocabulary 3D Instance Segmentation | https://openreview.net/forum?id=CRmiX0v16e | [
"Mohamed El Amine Boudjoghra",
"Angela Dai",
"Jean Lahoud",
"Hisham Cholakkal",
"Rao Muhammad Anwer",
"Salman Khan",
"Fahad Shahbaz Khan"
] | Oral | Recent works on open-vocabulary 3D instance segmentation show strong promise but at the cost of slow inference speed and high computation requirements. This high computation cost is typically due to their heavy reliance on aggregated clip features from multi-view, which require computationally expensive 2D foundation models like Segment Anything (SAM) and CLIP. Consequently, this hampers their applicability in many real-world applications that require both fast and accurate predictions. To this end, we propose a novel open-vocabulary 3D instance segmentation approach, named Open-YOLO 3D, that efficiently leverages only 2D object detection from multi-view RGB images for open-vocabulary 3D instance segmentation.
We demonstrate that our proposed Multi-View Prompt Distribution (MVPDist) method makes use of multi-view information to account for misclassification from the object detector to predict a reliable label for 3D instance masks. Furthermore, since projections of 3D object instances are already contained within the 2D bounding boxes, we show that our proposed low granularity label maps, which require only a 2D object detector to construct, are sufficient and very fast to predict prompt IDs for 3D instance masks when used with our proposed MVPDist.
We validate our Open-YOLO 3D on two benchmarks, ScanNet200 and Replica,
under two scenarios: (i) with ground truth masks, where labels are required for given object proposals, and (ii) with class-agnostic 3D proposals generated from a 3D proposal network.
Our Open-YOLO 3D achieves state-of-the-art performance on both datasets while obtaining up to $\sim$16$\times$ speedup compared to the best existing method in literature. On ScanNet200 val. set, our Open-YOLO 3D achieves mean average precision (mAP) of 24.7% while operating at 22 seconds per scene. github.com/aminebdj/OpenYOLO3D | Open Vocabulary, 3D point cloud instance segmentation | null | 4,987 | 2406.02548 | [
0.031549468636512756,
-0.07444337010383606,
-0.022558610886335373,
-0.02271619811654091,
0.12661948800086975,
-0.08064988255500793,
-0.01751069724559784,
-0.0500720739364624,
0.054017093032598495,
-0.06769195944070816,
-0.02859467640519142,
-0.10089678317308426,
-0.01825263537466526,
0.09886303544044495,
0.08682062476873398,
-0.01153609436005354,
0.08931618183851242,
0.07316475361585617,
-0.09628914296627045,
0.016292033717036247,
0.05180579423904419,
-0.011807250790297985,
0.03800265118479729,
0.011487992480397224,
0.008159436285495758,
-0.0099767055362463,
0.013218501582741737,
-0.08650362491607666,
0.006569093093276024,
-0.02113650180399418,
0.02228138968348503,
0.000015945321138133295,
0.054899848997592926,
0.08742145448923111,
0.016419991850852966,
-0.04756675288081169,
0.01622963324189186,
0.023365773260593414,
-0.05826781317591667,
-0.013856333680450916,
0.0030322163365781307,
0.02630462683737278,
-0.004995326977223158,
0.0010034922743216157,
0.021593663841485977,
0.05731992423534393,
0.012816655449569225,
-0.07386457920074463,
0.05771390348672867,
-0.017763644456863403,
-0.02083463780581951,
-0.0920993983745575,
-0.058637116104364395,
0.040022045373916626,
-0.033331796526908875,
0.016903825104236603,
0.0611814521253109,
-0.061164114624261856,
0.052001964300870895,
0.058310963213443756,
-0.0027253914158791304,
-0.07977177202701569,
0.0471467450261116,
0.0255696102976799,
-0.018564771860837936,
-0.045332811772823334,
0.001914069289341569,
-0.03561494126915932,
-0.019242949783802032,
-0.01914631389081478,
-0.007455226965248585,
0.06280532479286194,
0.09431116282939911,
-0.04758600518107414,
-0.024553822353482246,
0.01877642050385475,
-0.008280993439257145,
-0.06325362622737885,
-0.00519421836361289,
-0.09872294962406158,
0.02920779213309288,
0.009746728464961052,
0.01780043914914131,
0.021655721589922905,
0.05832471325993538,
0.02532780170440674,
-0.0025820655282586813,
0.02791152149438858,
-0.07516681402921677,
0.0386127233505249,
-0.042732883244752884,
-0.049858324229717255,
-0.029258793219923973,
0.018732864409685135,
-0.024100178852677345,
-0.09726123511791229,
-0.01412808708846569,
-0.11192940920591354,
0.025937529280781746,
0.060475461184978485,
-0.02924799732863903,
-0.013647903688251972,
0.09056974202394485,
-0.03396457061171532,
0.016986964270472527,
0.02548578940331936,
0.03695176914334297,
-0.0299405325204134,
0.06597328186035156,
0.055938590317964554,
0.07696623355150223,
0.03172938525676727,
0.05091387778520584,
-0.006534466985613108,
-0.06171755492687225,
0.028402404859662056,
-0.016196416690945625,
-0.014384562149643898,
0.023233555257320404,
0.022747205570340157,
-0.010439791716635227,
0.0027336273342370987,
0.03528840094804764,
-0.021150777116417885,
-0.0036536448169499636,
-0.009540134109556675,
-0.024736275896430016,
2.1144957686484392e-33,
0.04316534101963043,
0.008046110160648823,
-0.0213288813829422,
0.024055784568190575,
0.04174663871526718,
0.0071119326166808605,
0.006918600760400295,
0.07117650657892227,
-0.009550130926072598,
0.007696288637816906,
0.020131085067987442,
-0.027302969247102737,
-0.01902233436703682,
0.07672032713890076,
0.006369682494550943,
-0.003777983132749796,
-0.004409120883792639,
0.02182440273463726,
-0.09733372181653976,
0.08435618877410889,
0.0011775150196626782,
-0.049934688955545425,
-0.051786284893751144,
0.03193337470293045,
0.022839266806840897,
-0.007101790048182011,
0.05311249941587448,
-0.030342169106006622,
0.061030805110931396,
0.03796873986721039,
0.011396437883377075,
0.009654383175075054,
0.0519234836101532,
0.06322290748357773,
-0.03343263268470764,
-0.03033006191253662,
-0.05515811964869499,
-0.017550606280565262,
0.00700615206733346,
-0.06249865144491196,
-0.0036781278904527426,
0.023713594302535057,
0.03580276295542717,
-0.0571306049823761,
-0.07187236100435257,
-0.04801877588033676,
0.07966091483831406,
0.08377201110124588,
-0.10616738349199295,
0.04869337007403374,
0.1088140457868576,
-0.06126733496785164,
-0.07265328615903854,
-0.07205985486507416,
-0.09344077110290527,
-0.07747442275285721,
0.0000498613590025343,
0.010956857353448868,
0.07523583620786667,
0.058839987963438034,
-0.1056567057967186,
0.07369357347488403,
-0.06448692083358765,
0.06294804811477661,
-0.09001828730106354,
0.0062328833155334,
0.010569008067250252,
0.03968886658549309,
-0.019986383616924286,
0.10461729019880295,
-0.032535724341869354,
0.0008920286782085896,
0.0083119822666049,
-0.047210149466991425,
0.06618236750364304,
-0.003531387075781822,
0.01974976249039173,
-0.030801940709352493,
0.05095845088362694,
0.02906487137079239,
-0.033661190420389175,
-0.022545529529452324,
0.053701844066381454,
-0.0033419891260564327,
-0.0675233006477356,
-0.01947793923318386,
0.07918258011341095,
-0.04521668329834938,
0.0032151041086763144,
0.057483889162540436,
-0.06897849589586258,
-0.04988100752234459,
-0.056673843413591385,
0.010861664079129696,
-0.00050206552259624,
-1.1284631891878724e-33,
0.09479045867919922,
0.019848639145493507,
-0.028845621272921562,
0.010653925128281116,
0.003019917057827115,
-0.07228830456733704,
-0.03804821893572807,
0.05553146079182625,
-0.025695502758026123,
-0.12387973070144653,
-0.007800113409757614,
0.050259631127119064,
0.03256567195057869,
0.07604700326919556,
0.04595950245857239,
-0.07322170585393906,
-0.041696369647979736,
-0.025273896753787994,
-0.06418969482183456,
0.0669742301106453,
0.02370183728635311,
0.010641275905072689,
-0.03405966982245445,
-0.058422986418008804,
-0.07048958539962769,
0.07101569324731827,
-0.040974151343107224,
0.01568981446325779,
0.009567844681441784,
0.0060513936914503574,
-0.05986568704247475,
0.015045695006847382,
-0.005445535760372877,
-0.04902462661266327,
-0.07338888198137283,
0.024431342259049416,
0.06378322094678879,
-0.025294674560427666,
-0.03006604313850403,
-0.03386140614748001,
-0.02434518374502659,
-0.0252192672342062,
-0.05072817578911781,
0.034467361867427826,
-0.09085938334465027,
0.02313331700861454,
0.0413123294711113,
0.07548324763774872,
0.02698066085577011,
0.0869155153632164,
-0.1768796443939209,
0.07252373546361923,
-0.07335832715034485,
0.0410723052918911,
-0.0038319025188684464,
-0.04321281611919403,
-0.16480670869350433,
-0.07467273622751236,
0.029379550367593765,
-0.022880330681800842,
-0.023810258135199547,
-0.02512422949075699,
-0.05551864951848984,
0.04785338416695595,
0.040319014340639114,
-0.0353328213095665,
-0.035296231508255005,
0.02093544229865074,
0.012493318878114223,
0.054741162806749344,
0.014574993401765823,
0.02965386211872101,
0.032850462943315506,
-0.04546439275145531,
-0.06547682732343674,
0.021065302193164825,
-0.011636735871434212,
-0.08974600583314896,
0.08980346471071243,
0.009633262641727924,
0.07155927270650864,
-0.03599812462925911,
0.04724525660276413,
0.08598305284976959,
0.0896231085062027,
0.04413008317351341,
0.048048291355371475,
0.050523269921541214,
0.01612919010221958,
-0.00044755052658729255,
-0.04556112363934517,
0.006534799933433533,
0.042336877435445786,
0.08861411362886429,
0.035385116934776306,
-3.575993900994945e-8,
-0.03907464072108269,
0.02131766267120838,
0.06320938467979431,
-0.05016223341226578,
-0.007962247356772423,
0.004348595626652241,
-0.03518961742520332,
0.15564925968647003,
0.035762470215559006,
-0.005036670248955488,
0.030145954340696335,
-0.02061288431286812,
-0.0643015131354332,
-0.057361263781785965,
-0.0020191301591694355,
0.10608986020088196,
-0.012398574501276016,
0.009622644633054733,
0.024014336988329887,
0.0037193240132182837,
-0.02855018526315689,
-0.011042336001992226,
0.10521695762872696,
0.017928527668118477,
-0.01940799504518509,
-0.0210102591663599,
-0.013321811333298683,
0.06001489236950874,
0.07759547233581543,
-0.09842421114444733,
-0.023076944053173065,
0.015926413238048553,
0.010166804306209087,
0.02532254345715046,
0.024901384487748146,
0.06299375742673874,
-0.06720618903636932,
0.08180974423885345,
0.06917744129896164,
-0.04003596678376198,
0.04706780984997749,
-0.014687704853713512,
0.004690439905971289,
0.003228525398299098,
-0.003625189885497093,
0.039322689175605774,
0.0524640791118145,
-0.05203359201550484,
-0.002339318161830306,
-0.02935202047228813,
-0.049949631094932556,
-0.036371760070323944,
-0.017320042476058006,
0.11021842807531357,
-0.015504295006394386,
0.035933587700128555,
-0.006153583526611328,
-0.06050975248217583,
0.03570470213890076,
0.12248670309782028,
-0.03105834312736988,
-0.035090431571006775,
-0.05304332822561264,
-0.012948846444487572
] | https://github.com/aminebdj/openyolo3d | 138 | 0 | 0 | 0 |
Safety Alignment Should be Made More Than Just a Few Tokens Deep | https://openreview.net/forum?id=6Mxhg9PtDE | [
"Xiangyu Qi",
"Ashwinee Panda",
"Kaifeng Lyu",
"Xiao Ma",
"Subhrajit Roy",
"Ahmad Beirami",
"Prateek Mittal",
"Peter Henderson"
] | Oral | The safety alignment of current Large Language Models (LLMs) is vulnerable. Simple attacks, or even benign fine-tuning, can jailbreak aligned models. We note that many of these vulnerabilities are related to a shared underlying issue: safety alignment can take shortcuts, wherein the alignment adapts a model's generative distribution primarily over only its very first few output tokens. We unifiedly refer to this issue as shallow safety alignment. In this paper, we present case studies to explain why shallow safety alignment can exist and show how this issue universally contributes to multiple recently discovered vulnerabilities in LLMs, including the susceptibility to adversarial suffix attacks, prefilling attacks, decoding parameter attacks, and fine-tuning attacks. The key contribution of this work is that we demonstrate how this consolidated notion of shallow safety alignment sheds light on promising research directions for mitigating these vulnerabilities. We show that deepening the safety alignment beyond the first few tokens can meaningfully improve robustness against some common exploits. We also design a regularized fine-tuning objective that makes the safety alignment more persistent against fine-tuning attacks by constraining updates on initial tokens. Overall, we advocate that future safety alignment should be made more than just a few tokens deep. | Safety Alignment, AI Safety, LLM | We identify an underlying problem (shallow safety alignment) tha makes current safety alignment vulnerable, and we also propose approaches for mitigations. | 4,914 | 2406.05946 | [
-0.09337154030799866,
-0.047051046043634415,
-0.011600438505411148,
-0.022927146404981613,
0.0395076610147953,
0.021026073023676872,
-0.01913151517510414,
-0.02525174990296364,
0.013989116996526718,
-0.025969449430704117,
-0.006882615853101015,
-0.0705384835600853,
0.06775060296058655,
0.004948597401380539,
-0.011114608496427536,
0.0030868956819176674,
0.011335054412484169,
0.1376550793647766,
-0.11596468836069107,
0.01642138883471489,
-0.027177896350622177,
0.07244449853897095,
-0.028847534209489822,
0.006083983927965164,
-0.05837003514170647,
-0.008684211410582066,
-0.04434431716799736,
-0.031243085861206055,
-0.010485283099114895,
-0.04007747024297714,
0.014544041827321053,
0.02413710206747055,
-0.03501642495393753,
0.0472390316426754,
-0.002375571057200432,
0.021776879206299782,
-0.058262307196855545,
-0.03300691395998001,
0.048714037984609604,
-0.09876181930303574,
-0.057835038751363754,
0.009578765369951725,
-0.014722614549100399,
0.041266318410634995,
-0.05311926454305649,
-0.061883345246315,
0.012689641676843166,
-0.019050197675824165,
-0.05818473547697067,
-0.04940200597047806,
-0.05956362187862396,
0.006841854192316532,
0.010389504954218864,
0.04758885130286217,
0.011265379376709461,
-0.05000416561961174,
0.02188923954963684,
0.04337555542588234,
0.016385572031140327,
0.04990001767873764,
-0.0014246056089177728,
-0.03384542837738991,
-0.040549468249082565,
0.027528569102287292,
-0.007473039906471968,
-0.0492209792137146,
0.015726380050182343,
0.05997044965624809,
-0.01478434819728136,
0.10653234273195267,
-0.02338555082678795,
0.05303221568465233,
-0.028871243819594383,
0.023078592494130135,
0.017417777329683304,
0.13507835566997528,
-0.020600341260433197,
-0.029371675103902817,
0.02027365379035473,
-0.07307976484298706,
0.06051420420408249,
0.0006754535133950412,
0.08776047080755234,
0.07105153799057007,
0.11599966883659363,
0.05115094408392906,
-0.03821345791220665,
-0.040460005402565,
0.12571372091770172,
0.04296617954969406,
-0.053199417889118195,
-0.05501529946923256,
0.12405523657798767,
0.07699798047542572,
0.09486792236566544,
-0.0018057895358651876,
0.018944481387734413,
-0.057412516325712204,
-0.08512043207883835,
0.06863030046224594,
-0.008091944269835949,
0.06323885172605515,
0.02208947204053402,
-0.043041493743658066,
0.09067201614379883,
-0.018054429441690445,
0.00659393984824419,
-0.030065452679991722,
0.0018887522164732218,
-0.0378810279071331,
0.03992900624871254,
0.06441432237625122,
0.04616781324148178,
-0.019933346658945084,
-0.023110169917345047,
-0.048184849321842194,
-0.06453916430473328,
-0.028236806392669678,
-0.01926128752529621,
0.030634760856628418,
-0.003907341975718737,
0.05155640468001366,
0.018928516656160355,
0.012712934985756874,
0.004617797676473856,
-0.010313736274838448,
-0.03651105612516403,
3.190319953464505e-33,
0.002833198755979538,
0.07807448506355286,
-0.08187836408615112,
0.030356459319591522,
0.02626298926770687,
-0.030426418408751488,
-0.04019774869084358,
0.05008893832564354,
-0.03640680015087128,
0.015834227204322815,
-0.021030861884355545,
-0.00384080340154469,
-0.005928275641053915,
0.038123901933431625,
0.04221484065055847,
0.04978509992361069,
-0.017139634117484093,
0.0643828809261322,
-0.08825219422578812,
-0.026428116485476494,
0.03268321603536606,
-0.04011524096131325,
-0.000788208853919059,
-0.06586981564760208,
0.010535080917179585,
0.02901531383395195,
0.011751759797334671,
-0.02836001105606556,
-0.026669811457395554,
0.04926792159676552,
-0.13975933194160461,
-0.038809213787317276,
0.04794345796108246,
0.007718430832028389,
0.06894876807928085,
-0.01779232919216156,
-0.039987169206142426,
-0.07210318744182587,
-0.040657587349414825,
0.0016594236949458718,
-0.036643847823143005,
0.057327646762132645,
-0.007046937942504883,
0.020617414265871048,
0.05611278861761093,
-0.013569408096373081,
-0.09251336753368378,
0.00916958786547184,
-0.0037459551822394133,
-0.11730938404798508,
-0.0441456064581871,
0.038405947387218475,
0.026348169893026352,
0.0005544614396058023,
-0.11049987375736237,
-0.01925559900701046,
0.004268968477845192,
0.002917134203016758,
-0.05348419025540352,
0.11447104066610336,
-0.07663960009813309,
0.03841264545917511,
-0.05535894259810448,
-0.014028169214725494,
0.014784613624215126,
-0.06986396014690399,
0.009762047789990902,
0.02102702669799328,
0.027586670592427254,
0.022816119715571404,
-0.05573219060897827,
-0.008440900593996048,
-0.04350006952881813,
0.04583871364593506,
-0.10842521488666534,
0.00803857296705246,
0.10895717889070511,
-0.012591241858899593,
0.02415323443710804,
-0.04713933914899826,
0.02690809592604637,
-0.006187135819345713,
0.06118156760931015,
0.02470923215150833,
-0.08305495977401733,
-0.08284977078437805,
-0.012314966879785061,
-0.03955753892660141,
-0.05382873862981796,
0.07371693849563599,
0.009687000885605812,
-0.05666257441043854,
0.04597675800323486,
0.015268075279891491,
-0.03732030466198921,
-4.024749026499521e-33,
-0.08048027753829956,
-0.06183674931526184,
-0.0808350145816803,
0.06041068583726883,
-0.08337253332138062,
-0.07066585123538971,
-0.015775859355926514,
0.04138615354895592,
-0.0370679534971714,
-0.011184207163751125,
-0.09732547402381897,
-0.015617519617080688,
0.036687787622213364,
-0.04590819031000137,
0.06033570319414139,
-0.0482344813644886,
-0.020985061302781105,
0.05693701654672623,
0.0038951796013861895,
0.05322890728712082,
0.11312396824359894,
0.0894743949174881,
-0.14596150815486908,
0.07879267632961273,
0.018059194087982178,
0.034826479852199554,
-0.03502529114484787,
0.02352502942085266,
0.02270307019352913,
0.09587879478931427,
-0.020926743745803833,
0.06150149926543236,
-0.028789497911930084,
-0.0035189453046768904,
-0.039703987538814545,
0.02132384292781353,
0.06110134348273277,
-0.019797127693891525,
0.07503360509872437,
0.044163770973682404,
0.07234397530555725,
0.030467741191387177,
-0.07222356647253036,
-0.01351816114038229,
0.01635671593248844,
-0.017152240499854088,
-0.010656503960490227,
-0.02433941885828972,
0.029032953083515167,
-0.07644619047641754,
0.004955577198415995,
-0.01726539619266987,
0.008084624074399471,
0.040521152317523956,
-0.014283597469329834,
-0.07172733545303345,
-0.0027392967604100704,
-0.04714030399918556,
0.006172184366732836,
0.031250450760126114,
-0.023384468629956245,
-0.018303753808140755,
0.017680389806628227,
-0.041991982609033585,
0.012838870286941528,
-0.0037049599923193455,
-0.037353143095970154,
-0.004692999646067619,
0.012686791829764843,
-0.0885254517197609,
0.07442855834960938,
-0.04113774374127388,
-0.06239365413784981,
0.03803803771734238,
-0.05498139560222626,
-0.06505542248487473,
-0.0066996789537370205,
-0.10392459481954575,
0.003916279412806034,
0.03264104947447777,
0.07203742116689682,
0.025202402845025063,
0.006043179426342249,
0.07232311367988586,
0.03003835305571556,
0.007737413980066776,
0.08507280796766281,
0.06688004732131958,
0.0027676608879119158,
0.01761266589164734,
-0.021952668204903603,
0.017324751242995262,
-0.04162786900997162,
0.04812390357255936,
-0.04776392877101898,
-4.191429070488084e-8,
-0.032486483454704285,
-0.0032969205640256405,
-0.035126641392707825,
0.07896419614553452,
0.022277699783444405,
0.02497679367661476,
-0.10760711878538132,
-0.004379331134259701,
0.05080326646566391,
-0.035330742597579956,
0.041092172265052795,
0.023715833202004433,
-0.0658070519566536,
-0.05537888780236244,
-0.04712070897221565,
0.08236392587423325,
-0.10517638921737671,
0.07239928841590881,
-0.08422672748565674,
0.01818474382162094,
-0.023577963933348656,
0.10590573400259018,
-0.03253554925322533,
-0.038027651607990265,
0.03444793075323105,
-0.05953773111104965,
-0.014197033829987049,
0.0703725591301918,
0.06156505271792412,
0.07082384824752808,
-0.02950231544673443,
-0.014165187254548073,
0.0005343999364413321,
0.0028667135629802942,
-0.07258883863687515,
0.17156101763248444,
-0.030990974977612495,
-0.031482093036174774,
0.038285840302705765,
0.0007257119286805391,
0.027318725362420082,
0.036708686500787735,
-0.0013723981101065874,
0.0378444604575634,
-0.04623373597860336,
-0.035621680319309235,
0.003807761939242482,
-0.044587068259716034,
0.07362864166498184,
-0.04259578883647919,
0.03867058455944061,
-0.040877051651477814,
0.04963590204715729,
0.08441223204135895,
-0.023861821740865707,
0.029245683923363686,
-0.04590646177530289,
-0.039019521325826645,
0.06523855775594711,
0.06968259811401367,
0.04415260627865791,
-0.0018110348610207438,
0.0683620423078537,
-0.025992851704359055
] | https://github.com/unispac/shallow-vs-deep-alignment | 102 | 0 | 0 | 0 |
On the Identification of Temporal Causal Representation with Instantaneous Dependence | https://openreview.net/forum?id=2efNHgYRvM | [
"Zijian Li",
"Yifan Shen",
"Kaitao Zheng",
"Ruichu Cai",
"Xiangchen Song",
"Mingming Gong",
"Guangyi Chen",
"Kun Zhang"
] | Oral | Temporally causal representation learning aims to identify the latent causal process from time series observations, but most methods require the assumption that the latent causal processes do not have instantaneous relations. Although some recent methods achieve identifiability in the instantaneous causality case, they require either interventions on the latent variables or grouping of the observations, which are in general difficult to obtain in real-world scenarios. To fill this gap, we propose an \textbf{ID}entification framework for instantane\textbf{O}us \textbf{L}atent dynamics (\textbf{IDOL}) by imposing a sparse influence constraint that the latent causal processes have sparse time-delayed and instantaneous relations. Specifically, we establish identifiability results of the latent causal process based on sufficient variability and the sparse influence constraint by employing contextual information of time series data. Based on these theories, we incorporate a temporally variational inference architecture to estimate the latent variables and a gradient-based sparsity regularization to identify the latent causal process. Experimental results on simulation datasets illustrate that our method can identify the latent causal process. Furthermore, evaluations on multiple human motion forecasting benchmarks with instantaneous dependencies indicate the effectiveness of our method in real-world settings. | Causal Representation Learning, Instantaneous Dependency, Identification | null | 4,912 | 2405.15325 | [
-0.058109309524297714,
-0.07283610105514526,
0.021104011684656143,
0.07691814750432968,
0.09068708121776581,
0.07189693301916122,
0.003409913508221507,
-0.07509136945009232,
0.038307253271341324,
-0.03819858655333519,
-0.010609940625727177,
-0.02278096042573452,
0.013897943310439587,
0.048425257205963135,
-0.06721443682909012,
-0.046232111752033234,
0.06803350895643234,
-0.0016604914562776685,
-0.07830016314983368,
-0.04999038204550743,
0.03202119842171669,
0.02488817274570465,
-0.08432740718126297,
0.013080436736345291,
0.044174421578645706,
-0.02750040404498577,
0.037860870361328125,
0.037505872547626495,
0.05264032632112503,
0.025258883833885193,
0.016437795013189316,
0.02783537283539772,
-0.0856427252292633,
-0.13268277049064636,
0.000046191420551622286,
0.058987196534872055,
0.016361404210329056,
0.05488540977239609,
-0.06749292463064194,
0.00003093259147135541,
0.09492399543523788,
0.00306706759147346,
-0.05869509279727936,
0.05916810408234596,
-0.00162351387552917,
-0.03509623929858208,
0.031075673177838326,
-0.00965619832277298,
-0.06830872595310211,
0.004820053931325674,
-0.04231543093919754,
0.04562210291624069,
0.02321183681488037,
-0.059715040028095245,
0.045806366950273514,
-0.034239135682582855,
0.00883474387228489,
0.023953543975949287,
0.0051084598526358604,
0.05718202143907547,
-0.0452192947268486,
-0.07609490305185318,
0.0027033276855945587,
-0.04951939731836319,
-0.006900155451148748,
0.016953373327851295,
0.0050183371640741825,
0.059227354824543,
-0.02135554514825344,
0.03006894513964653,
-0.00724453292787075,
0.05256599932909012,
-0.10763494670391083,
0.0007592604379169643,
-0.01183922030031681,
0.0061505199410021305,
0.04433450102806091,
-0.02469950169324875,
0.06488051265478134,
-0.0584857240319252,
0.0058869849890470505,
0.07461541891098022,
-0.027550524100661278,
-0.024140754714608192,
0.052098654210567474,
-0.013514181599020958,
0.044192783534526825,
0.07727037370204926,
0.02075672708451748,
0.08816104382276535,
0.053847357630729675,
-0.002498817862942815,
0.0014775180025026202,
-0.011511806398630142,
0.08432308584451675,
-0.036452554166316986,
0.08519703894853592,
-0.00040525264921598136,
0.10198070853948593,
0.12986423075199127,
-0.03673020377755165,
0.024157816544175148,
-0.05706970766186714,
0.10577309876680374,
0.024395842105150223,
-0.0360601469874382,
0.00099466135725379,
0.05090772360563278,
-0.016879217699170113,
-0.030470484867691994,
-0.034969769418239594,
-0.025598658248782158,
0.029163585975766182,
0.024101650342345238,
-0.010102924890816212,
-0.12397371977567673,
0.02741522714495659,
-0.0541970320045948,
0.029493598267436028,
-0.025348396971821785,
0.018165133893489838,
-0.07634303718805313,
0.037331875413656235,
0.04431471601128578,
0.035735227167606354,
-0.028102576732635498,
-0.062039319425821304,
1.8009095080215622e-33,
0.055383723229169846,
0.002325392095372081,
0.05926389619708061,
-0.039347246289253235,
-0.037708789110183716,
-0.025878604501485825,
-0.07925578206777573,
-0.043879471719264984,
0.032741885632276535,
-0.0025684875436127186,
0.02472444623708725,
-0.003304483136162162,
-0.11622646450996399,
-0.0003963060153182596,
-0.023370670154690742,
-0.012120883911848068,
-0.022094523534178734,
0.0780002698302269,
0.01835455372929573,
-0.06775659322738647,
-0.0256575308740139,
-0.07324012368917465,
-0.05412526801228523,
-0.014856403693556786,
0.018679765984416008,
-0.05033770576119423,
0.07897303253412247,
-0.031827718019485474,
0.005189463961869478,
-0.0062857260927557945,
-0.04463253542780876,
-0.03669361397624016,
0.041112758219242096,
-0.010295593179762363,
-0.017834655940532684,
-0.005673923995345831,
0.000049159298214362934,
-0.04347086325287819,
-0.015340276062488556,
-0.05417436733841896,
0.023555120453238487,
-0.022600525990128517,
0.03621736913919449,
-0.039761193096637726,
-0.13629716634750366,
-0.031037267297506332,
0.05787227302789688,
0.04537396505475044,
0.03253864496946335,
-0.036037247627973557,
-0.00006910630327183753,
0.01825190708041191,
0.08759834617376328,
-0.16370034217834473,
-0.03466536104679108,
0.018648944795131683,
-0.014962374232709408,
-0.012997797690331936,
0.004175771027803421,
-0.0033011315390467644,
-0.018036790192127228,
0.019946303218603134,
0.04420981556177139,
0.02436121553182602,
0.051341552287340164,
0.06333540380001068,
-0.010354289785027504,
0.02213877998292446,
0.044361770153045654,
-0.021721189841628075,
0.008691243827342987,
0.03925182670354843,
-0.009808381088078022,
-0.09215424209833145,
0.058529071509838104,
-0.0062933252193033695,
0.006334413308650255,
-0.033794041723012924,
-0.02274228259921074,
0.06867130845785141,
-0.09241126477718353,
-0.091529980301857,
0.028332199901342392,
-0.007519443053752184,
-0.0685771107673645,
0.014246468432247639,
0.044871725142002106,
-0.006482316181063652,
-0.01485576294362545,
-0.09856101870536804,
-0.033203888684511185,
-0.00025889225071296096,
-0.009243512526154518,
0.04317713528871536,
-0.01182570867240429,
-2.4549170959366505e-33,
-0.008135712705552578,
-0.01920236274600029,
0.09867901355028152,
-0.06067721173167229,
0.11673729121685028,
-0.03865106403827667,
-0.0658089816570282,
0.021640386432409286,
0.06711453199386597,
-0.036383192986249924,
-0.010049469769001007,
-0.10764991492033005,
0.0017533397767692804,
-0.027445664629340172,
0.021646415814757347,
0.06260880827903748,
0.06589395552873611,
-0.000028638034564210102,
0.038393810391426086,
0.04477136954665184,
0.03149699419736862,
-0.027555575594305992,
-0.00035667396150529385,
-0.012911101803183556,
0.019335418939590454,
-0.03458580747246742,
0.0265450831502676,
0.0912010669708252,
-0.08554882556200027,
0.0181899331510067,
-0.019588416442275047,
0.009767424315214157,
-0.06193609535694122,
-0.014566443860530853,
-0.0831974595785141,
0.10781640559434891,
0.07813679426908493,
0.007700187619775534,
-0.029982928186655045,
0.015455164946615696,
0.05816977471113205,
0.006091869901865721,
-0.037724368274211884,
0.020068855956196785,
-0.021655719727277756,
-0.011575663462281227,
-0.013834057375788689,
0.06622644513845444,
0.04259476810693741,
0.03466923162341118,
0.008454944007098675,
0.07911739498376846,
0.02048986777663231,
0.05797399580478668,
-0.08357227593660355,
-0.05662141740322113,
0.07644997537136078,
-0.08944036066532135,
0.06911956518888474,
0.004449722822755575,
-0.01719076931476593,
-0.034896597266197205,
-0.05051066353917122,
-0.024439383298158646,
0.010828506201505661,
-0.04555586352944374,
0.03671212121844292,
-0.08578948676586151,
0.014385645277798176,
-0.0070253475569188595,
0.06943011283874512,
0.056083984673023224,
-0.06270905584096909,
-0.0055878981947898865,
0.03991309925913811,
0.003673022612929344,
0.01361605804413557,
-0.017346970736980438,
-0.017813486978411674,
-0.09518466889858246,
0.057431019842624664,
0.0344151072204113,
0.0389493890106678,
-0.044305723160505295,
-0.09642450511455536,
0.11300539970397949,
0.08004473894834518,
0.05922780558466911,
0.06844574958086014,
0.06448733806610107,
-0.06997378170490265,
-0.03477371484041214,
-0.007592332549393177,
0.05693325400352478,
-0.10224980115890503,
-4.295739230997242e-8,
-0.013371489010751247,
0.027712514623999596,
0.04756895825266838,
-0.029631560668349266,
0.04363318160176277,
-0.05319543555378914,
0.019987434148788452,
-0.01850075088441372,
0.039820339530706406,
-0.06650286167860031,
0.0860547125339508,
-0.0370706282556057,
0.021447712555527687,
-0.09845370799303055,
-0.012636315077543259,
0.016824547201395035,
0.009874122217297554,
-0.06188765913248062,
-0.022952577099204063,
0.04962718486785889,
0.042623311281204224,
0.06089982017874718,
-0.030290907248854637,
-0.03840065002441406,
0.09688571840524673,
-0.08861853182315826,
-0.03233526647090912,
0.08470182120800018,
0.07516945898532867,
-0.002156567992642522,
-0.051173023879528046,
0.0699596181511879,
-0.021451564505696297,
-0.07098463922739029,
-0.01510072685778141,
0.07567852735519409,
0.028042178601026535,
-0.06842318177223206,
-0.11714372038841248,
-0.04490143060684204,
-0.03431059792637825,
0.04510454088449478,
-0.021733781322836876,
-0.008866085670888424,
0.05352797359228134,
0.004740629345178604,
-0.08156505227088928,
-0.10141781717538834,
0.05329998955130577,
0.0011558635160326958,
0.027274644002318382,
-0.04763193055987358,
0.029945502057671547,
0.03509688004851341,
0.07655558735132217,
0.04377681389451027,
0.08310500532388687,
0.020231032744050026,
0.0622127465903759,
-0.021430538967251778,
0.06547968089580536,
0.07155897468328476,
0.011096361093223095,
0.0382121279835701
] | 0 | 0 | 0 | 0 |
|
WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct | https://openreview.net/forum?id=mMPMHWOdOy | [
"Haipeng Luo",
"Qingfeng Sun",
"Can Xu",
"Pu Zhao",
"Jian-Guang Lou",
"Chongyang Tao",
"Xiubo Geng",
"Qingwei Lin",
"Shifeng Chen",
"Yansong Tang",
"Dongmei Zhang"
] | Oral | Large language models (LLMs), such as GPT-4, have shown remarkable performance in natural language processing (NLP) tasks, including challenging mathematical reasoning. However, most existing open-source models are only pre-trained on large-scale internet data and without math-related optimization. In this paper, we present WizardMath, which enhances the mathematical reasoning abilities of LLMs, by applying our proposed Reinforcement Learning from Evol-Instruct Feedback (RLEIF) method to the domain of math. Through extensive experiments on two mathematical reasoning benchmarks, namely GSM8k and MATH, we reveal the extraordinary capabilities of our model. Remarkably, WizardMath-Mistral 7B surpasses all other open-source LLMs by a substantial margin. Furthermore, WizardMath 70B even outperforms ChatGPT-3.5, Claude Instant, Gemini Pro and Mistral Medium. Additionally, our preliminary exploration highlights the pivotal role of instruction evolution and process supervision in achieving exceptional math performance. | Mathematical Reasoning, Evol-Instruct, Reinforcement Learning | null | 4,894 | 2308.09583 | [
-0.04734252765774727,
-0.05010021850466728,
0.04070835933089256,
0.012156574986875057,
0.008873308077454567,
-0.04711185768246651,
-0.02018440142273903,
0.03202417865395546,
-0.02586517296731472,
-0.01257726177573204,
-0.07273939996957779,
-0.03543056920170784,
0.02739098109304905,
0.057508669793605804,
0.05059579759836197,
0.02872774377465248,
0.0395999550819397,
-0.0013890782138332725,
-0.09843970835208893,
-0.138326957821846,
-0.023912206292152405,
0.0001970674638869241,
0.07438024878501892,
-0.02489256113767624,
0.06191357970237732,
0.00006104179919930175,
0.0351378358900547,
-0.04242207854986191,
0.07352939248085022,
-0.03478814661502838,
-0.0027160642202943563,
-0.012934375554323196,
-0.04347313195466995,
-0.006629527546465397,
-0.09981057047843933,
0.10917098075151443,
-0.09235591441392899,
0.03486494719982147,
0.016586873680353165,
-0.03520970046520233,
-0.07874570786952972,
-0.03505798056721687,
0.0458885058760643,
-0.006263540126383305,
0.06570117175579071,
-0.018264349550008774,
-0.013541635125875473,
-0.001039686263538897,
-0.01246813964098692,
0.02214132808148861,
-0.11660701781511307,
-0.015658292919397354,
-0.04280734807252884,
0.01881204918026924,
-0.006193482782691717,
0.05816785618662834,
0.010430701076984406,
-0.007549210451543331,
0.008987185545265675,
-0.08411797136068344,
-0.09452180564403534,
-0.06798644363880157,
-0.050433944910764694,
-0.008636401034891605,
-0.03847864642739296,
-0.03597111627459526,
0.03721323981881142,
0.08892273902893066,
-0.010729871690273285,
0.021618859842419624,
-0.008647221140563488,
0.05558770149946213,
-0.011048496700823307,
0.08284550160169601,
0.0201378483325243,
0.10271470248699188,
0.0154131343588233,
-0.042271099984645844,
0.07673291116952896,
0.0055309804156422615,
0.014403009787201881,
-0.007317912299185991,
0.012411298230290413,
0.026483116671442986,
0.07719535380601883,
-0.041290391236543655,
0.015255067497491837,
0.07432599365711212,
0.10952790081501007,
0.03700179234147072,
0.09827477484941483,
-0.018364442512392998,
0.02426961436867714,
0.051895011216402054,
0.013846829533576965,
0.08177303522825241,
-0.03697296231985092,
-0.09091180562973022,
-0.07929778099060059,
0.14765100181102753,
0.03362734243273735,
0.10726480185985565,
0.040562231093645096,
-0.03840332478284836,
-0.04701610282063484,
-0.038482170552015305,
0.0407615527510643,
0.013276265934109688,
0.04335059970617294,
-0.12701815366744995,
0.09475989639759064,
0.0013917165342718363,
0.037849556654691696,
0.03959308937191963,
-0.025178156793117523,
0.010999641381204128,
-0.013362035155296326,
-0.029484614729881287,
0.013919258490204811,
0.011565068736672401,
0.017936870455741882,
0.04365415871143341,
0.006495046429336071,
0.016983143985271454,
0.06294777244329453,
0.006372765172272921,
-0.06499136984348297,
7.504662666206651e-34,
0.08391355723142624,
0.0468309111893177,
0.04595649242401123,
0.02991638332605362,
0.06359118968248367,
0.005142369773238897,
0.023096827790141106,
0.052587203681468964,
-0.04549095034599304,
0.028779063373804092,
-0.012696019373834133,
0.005336062051355839,
-0.07812557369470596,
0.03107776679098606,
0.033802330493927,
-0.017481258139014244,
-0.08169984817504883,
0.02530547045171261,
-0.09470774978399277,
-0.041842058300971985,
0.019009051844477654,
0.005448357667773962,
-0.0028430481906980276,
-0.0082799531519413,
0.007869944907724857,
0.12236043810844421,
0.08532264828681946,
-0.047327011823654175,
-0.018994655460119247,
0.010225009173154831,
-0.037081871181726456,
-0.017740579321980476,
-0.09511862695217133,
0.08524658530950546,
0.0351937972009182,
-0.0444176122546196,
-0.04728489741683006,
-0.1297968477010727,
0.10249434411525726,
-0.02322690561413765,
-0.011153584346175194,
-0.0030345062259584665,
0.048542361706495285,
0.03273739665746689,
-0.08003687113523483,
-0.05338814854621887,
0.006678328383713961,
0.01200193539261818,
0.03341834992170334,
-0.00011497525701997802,
0.031935229897499084,
0.00907898135483265,
-0.030085530132055283,
-0.016702162101864815,
-0.007881090976297855,
0.006260274443775415,
-0.03530653193593025,
0.10927875339984894,
0.03882836550474167,
0.06789689511060715,
-0.006943264044821262,
0.1043383926153183,
0.014146123081445694,
0.04447239264845848,
-0.03902539610862732,
0.024222826585173607,
-0.13212141394615173,
0.05267655476927757,
0.08889854699373245,
0.04246583580970764,
-0.011493642814457417,
-0.01709425449371338,
0.04994523152709007,
-0.02569418027997017,
0.10984721779823303,
-0.05850204825401306,
0.041112206876277924,
-0.0691111609339714,
0.011498321779072285,
-0.01450327504426241,
-0.01629055291414261,
0.05652965232729912,
-0.05120650306344032,
-0.02193637192249298,
-0.014569461345672607,
-0.10989570617675781,
0.02273683063685894,
-0.028813006356358528,
0.0003394476661924273,
-0.02802225574851036,
-0.04409557953476906,
-0.0497191920876503,
-0.007282651495188475,
0.020191067829728127,
-0.03287927061319351,
-1.0595892302987676e-33,
-0.029653625562787056,
-0.053618498146533966,
-0.04797566682100296,
0.06685449928045273,
-0.043282996863126755,
-0.08435426652431488,
-0.027216175571084023,
0.0044103460386395454,
0.008303100243210793,
-0.07499505579471588,
0.004332155454903841,
0.07206670194864273,
0.027754899114370346,
0.001989874755963683,
0.022741852328181267,
-0.09288839995861053,
-0.03775420039892197,
-0.022164886817336082,
-0.004060132894665003,
0.05345964804291725,
-0.020036906003952026,
0.09991850703954697,
-0.07504039257764816,
0.047801483422517776,
0.003360812785103917,
0.010443413630127907,
0.0439445935189724,
0.04314671456813812,
-0.006730152294039726,
0.0999121367931366,
-0.012956888414919376,
-0.030114775523543358,
-0.038244232535362244,
-0.03990113362669945,
-0.06613913178443909,
0.020924082025885582,
0.13909104466438293,
0.04072752222418785,
-0.08541195094585419,
0.01809462159872055,
0.0674942210316658,
-0.08328507095575333,
-0.0036591568496078253,
-0.004506614524871111,
-0.003753817407414317,
0.08553359657526016,
-0.06048613786697388,
-0.01896629109978676,
-0.015370046719908714,
0.04736695811152458,
0.021297376602888107,
-0.0202636681497097,
-0.00882404949516058,
-0.018178103491663933,
-0.024694574996829033,
-0.07811278104782104,
-0.049544624984264374,
-0.07243514806032181,
-0.02283693104982376,
0.016250543296337128,
-0.07849187403917313,
0.022128574550151825,
0.07497087121009827,
-0.041295453906059265,
-0.0034933960996568203,
-0.014021639712154865,
-0.07320298254489899,
0.046221278607845306,
0.007427142933011055,
-0.022453386336565018,
0.009735440835356712,
0.019354654476046562,
0.014976921491324902,
0.04805251583456993,
0.05853639915585518,
0.052719034254550934,
-0.007493261713534594,
-0.011962194927036762,
-0.014132780954241753,
-0.06918448954820633,
0.09331360459327698,
0.0012469299836084247,
0.04921584203839302,
0.02000642940402031,
0.02959316410124302,
-0.020365890115499496,
0.07104907184839249,
0.10175306349992752,
-0.008757230825722218,
0.04825875535607338,
-0.05692864954471588,
0.056421585381031036,
0.025897083804011345,
0.0749218687415123,
-0.06948501616716385,
-4.2559364032968006e-8,
-0.04563932865858078,
0.05096897482872009,
0.029838357120752335,
0.01483799796551466,
0.06340879946947098,
-0.051182981580495834,
-0.045134659856557846,
0.020430009812116623,
-0.03946737200021744,
-0.040032122284173965,
0.1004226878285408,
0.01830192841589451,
-0.02071358822286129,
-0.020565614104270935,
0.011886347085237503,
0.07128863781690598,
0.0008209893712773919,
-0.04622127115726471,
-0.0628189817070961,
-0.03419908881187439,
0.06992053985595703,
0.013165296986699104,
0.02114187739789486,
0.00012101049651391804,
0.03408446162939072,
-0.08763193339109421,
-0.015160320326685905,
0.12174548208713531,
0.023252464830875397,
-0.018097560852766037,
-0.014815892092883587,
-0.03607624024152756,
0.03447309881448746,
0.013249287381768227,
0.08733857423067093,
0.02520599588751793,
-0.04040566086769104,
-0.025644389912486076,
0.04757906123995781,
-0.07060892879962921,
-0.024952108040452003,
0.010059118270874023,
-0.02110624685883522,
-0.03178199753165245,
0.028419051319360733,
-0.017014289274811745,
-0.1063128337264061,
-0.12197842448949814,
0.004318433813750744,
-0.020504778251051903,
-0.02694850228726864,
-0.004181040450930595,
-0.05638597160577774,
0.01096164621412754,
0.07366614788770676,
-0.03674175217747688,
-0.04249602556228638,
-0.022300632670521736,
-0.021960925310850143,
-0.0007978950161486864,
0.0014148185728117824,
0.026140153408050537,
-0.004455041140317917,
0.014164929278194904
] | 0 | 0 | 0 | 0 |
|
Faster Cascades via Speculative Decoding | https://openreview.net/forum?id=vo9t20wsmd | [
"Harikrishna Narasimhan",
"Wittawat Jitkrittum",
"Ankit Singh Rawat",
"Seungyeon Kim",
"Neha Gupta",
"Aditya Krishna Menon",
"Sanjiv Kumar"
] | Oral | Cascades and speculative decoding are two common approaches to improving language models' inference efficiency. Both approaches interleave two models, but via fundamentally distinct mechanisms: deferral rule that invokes the larger model only for “hard” inputs, while speculative decoding uses speculative execution to primarily invoke the larger model in parallel scoring mode. These mechanisms offer different benefits: empirically, cascades offer compelling cost-quality trade-offs, often even outperforming the large model; speculative cascades offer impressive speed-ups, while guaranteeing quality-neutrality. In this paper, we leverage the best of both these approaches by designing new speculative cascading techniques that implement their deferral rule through speculative execution. We characterize the optimal deferral rule for our speculative cascades, and employ a plug-in approximation to the optimal rule. Experiments with Gemma and T5 models on a range of language benchmarks show that our approach yields better cost quality trade-offs than cascading and speculative decoding baselines. | Cascades, Speculative Decoding, Speculative execution, LLM, Inference, Adaptive Inference | Faster language model cascades through the use of speculative execution | 4,871 | 2405.19261 | [
-0.06018581613898277,
-0.03956982120871544,
0.03946373239159584,
0.016702737659215927,
0.07173783332109451,
-0.06039239838719368,
-0.04839146509766579,
-0.033993709832429886,
0.03003801591694355,
-0.01930706575512886,
-0.05743800848722458,
-0.043224625289440155,
0.03643105924129486,
0.022123385220766068,
0.032704152166843414,
-0.021513918414711952,
0.04189032316207886,
0.06680496037006378,
-0.06705199182033539,
-0.0958428755402565,
-0.006590501870959997,
-0.034945499151945114,
-0.02552873082458973,
0.008852646686136723,
0.07431657612323761,
-0.03986026346683502,
-0.04505586251616478,
0.038073889911174774,
-0.02418486215174198,
0.0473538376390934,
0.033951789140701294,
0.02815242111682892,
-0.09569155424833298,
-0.04697898030281067,
-0.06622503697872162,
0.03067549131810665,
0.0053346301428973675,
-0.021899374201893806,
0.06479867547750473,
-0.013944478705525398,
-0.020260479301214218,
0.004510099999606609,
0.038100581616163254,
0.034084074199199677,
0.08487390726804733,
-0.010110897943377495,
0.009960655122995377,
-0.021712439134716988,
-0.05053512752056122,
-0.02238222397863865,
-0.1356690376996994,
-0.08972997963428497,
-0.019727107137441635,
0.02544173039495945,
-0.04604969173669815,
0.022920256480574608,
0.019488487392663956,
0.008655802346765995,
-0.01504831574857235,
-0.004717816598713398,
-0.08743420243263245,
-0.12455026805400848,
-0.04048756882548332,
0.008428624831140041,
-0.014226263388991356,
0.018454071134328842,
0.028101924806833267,
0.044168468564748764,
0.01777484267950058,
0.071538046002388,
-0.018477456644177437,
0.039631880819797516,
0.0158747099339962,
0.06603245437145233,
-0.001373379840515554,
0.057096414268016815,
0.04901941120624542,
-0.0540037602186203,
0.06037420779466629,
-0.0819745883345604,
-0.000023220793082145974,
-0.014175716787576675,
-0.003598801326006651,
-0.1032654270529747,
0.09603363275527954,
-0.028815221041440964,
0.04220101609826088,
-0.02381293475627899,
0.03324487432837486,
0.044293127954006195,
-0.03036309778690338,
-0.1276642084121704,
0.077793650329113,
0.08294279873371124,
-0.03690630570054054,
0.08357167989015579,
-0.00891081802546978,
0.0030449668411165476,
-0.004871773067861795,
0.09023917466402054,
0.002644831780344248,
0.021042723208665848,
-0.019354714080691338,
-0.0731201246380806,
0.0005647344514727592,
-0.06115158274769783,
0.022282423451542854,
0.061415329575538635,
-0.019944297149777412,
-0.03951264172792435,
-0.06581941246986389,
0.09218937158584595,
0.05725396052002907,
-0.04264166206121445,
-0.03941059485077858,
0.06040501222014427,
0.04722217842936516,
-0.012090551666915417,
0.057503409683704376,
0.07288473844528198,
0.022445347160100937,
0.042283497750759125,
0.0021902259904891253,
0.012642506510019302,
0.007493182551115751,
0.004905758425593376,
-0.02370644360780716,
3.0558081348999106e-33,
0.024648519232869148,
0.060244668275117874,
0.02878710627555847,
0.048277921974658966,
-0.01581292785704136,
-0.015018061734735966,
-0.006080832332372665,
-0.03228475898504257,
-0.09893830865621567,
-0.044850483536720276,
0.031589481979608536,
-0.009679136797785759,
-0.0531911626458168,
-0.024591008201241493,
-0.022667743265628815,
-0.01857512630522251,
-0.027320705354213715,
0.04870336875319481,
-0.07687205821275711,
-0.062303923070430756,
0.012745763175189495,
-0.013010953553020954,
-0.043480824679136276,
-0.04255835711956024,
0.025792311877012253,
0.06102510169148445,
0.05404926463961601,
-0.04041421785950661,
-0.04138907790184021,
0.05141216516494751,
-0.08704625815153122,
-0.03835076466202736,
0.01781492307782173,
0.06392655521631241,
0.032576341181993484,
-0.025671793147921562,
-0.0025986318942159414,
-0.08053193241357803,
0.04121095687150955,
-0.0037822381127625704,
-0.019910825416445732,
0.021191243082284927,
-0.025573939085006714,
0.006033235229551792,
-0.044904861599206924,
-0.06354109197854996,
-0.07334944605827332,
-0.029745982959866524,
-0.050119951367378235,
-0.04291623830795288,
0.12772612273693085,
0.04545756056904793,
-0.02118009142577648,
0.08183999359607697,
-0.06817138940095901,
-0.011728393845260143,
-0.012843813747167587,
-0.01923365145921707,
0.03377106785774231,
0.12298174947500229,
0.006635874509811401,
0.04521147534251213,
-0.0005586620536632836,
0.03366576507687569,
-0.02638702280819416,
0.0573437474668026,
-0.04750622808933258,
0.060198087245225906,
0.02796739526093006,
0.01580202206969261,
-0.039245814085006714,
0.013137880712747574,
-0.05127657204866409,
-0.029499303549528122,
0.04466372728347778,
0.001724790083244443,
0.07234244048595428,
-0.06033625081181526,
0.01821136474609375,
-0.025877689942717552,
-0.017010020092129707,
-0.016684746369719505,
-0.002735659945756197,
0.06922997534275055,
-0.04541691765189171,
-0.028621364384889603,
-0.011861455626785755,
-0.009080804884433746,
0.011738412082195282,
-0.032762374728918076,
-0.07946578413248062,
-0.0317782498896122,
-0.048265598714351654,
-0.03092334046959877,
0.07550066709518433,
-3.710297675498836e-33,
-0.026860442012548447,
-0.05398758128285408,
-0.09878460317850113,
0.04314204677939415,
-0.05680404603481293,
-0.13161760568618774,
-0.05031174421310425,
-0.05136624351143837,
0.004716914147138596,
-0.05895019695162773,
-0.07484753429889679,
-0.05492013320326805,
0.016040915623307228,
0.012786855921149254,
0.06327684968709946,
-0.0624108612537384,
0.060078926384449005,
-0.0008972491486929357,
0.03381812945008278,
0.0827050507068634,
0.07496557384729385,
0.0486283116042614,
-0.1441190093755722,
0.05957810580730438,
0.03184253349900246,
0.0512927807867527,
-0.087430439889431,
0.09182388335466385,
0.08665236830711365,
-0.04773696884512901,
-0.07751121371984482,
-0.00488405954092741,
0.01445731334388256,
-0.07028660178184509,
-0.027566727250814438,
0.0601113885641098,
-0.0053574806079268456,
-0.016058817505836487,
0.027410728856921196,
0.05518417805433273,
0.03847930207848549,
0.019702419638633728,
-0.04585230350494385,
0.0024804428685456514,
0.0011874126503244042,
0.0689074769616127,
-0.03760244697332382,
-0.02561664581298828,
-0.002044870052486658,
0.04367195814847946,
0.025087615475058556,
0.045979227870702744,
-0.014074430800974369,
0.06564582139253616,
-0.047500692307949066,
-0.12937524914741516,
-0.08164890855550766,
-0.046275585889816284,
0.029401075094938278,
0.023678507655858994,
-0.009005299769341946,
0.07329177856445312,
0.003810812719166279,
-0.05687336251139641,
0.05188495293259621,
-0.07421107590198517,
-0.012344432063400745,
-0.08063655346632004,
0.04397887736558914,
-0.0317043699324131,
0.0820135846734047,
-0.01737249083817005,
-0.1379871815443039,
0.050447773188352585,
0.03750986605882645,
0.05783836171030998,
0.014497854746878147,
0.021691737696528435,
0.004192361142486334,
0.05496399849653244,
0.0049387505277991295,
0.07604246586561203,
0.09585628658533096,
0.008758476004004478,
0.0720217376947403,
0.07010643184185028,
0.01601792499423027,
0.07122234255075455,
0.022461391985416412,
-0.014792247675359249,
0.013038566336035728,
-0.011959057301282883,
0.04019961133599281,
0.02533210627734661,
-0.06821317970752716,
-4.238916773147139e-8,
-0.00788886845111847,
0.010626548901200294,
-0.03000146523118019,
0.08769521117210388,
-0.007967782206833363,
-0.018413705751299858,
-0.042302027344703674,
0.05944329500198364,
0.04784893989562988,
-0.034313928335905075,
0.11318308860063553,
-0.009392229840159416,
-0.03812037780880928,
-0.020018361508846283,
-0.0359315350651741,
0.0772501602768898,
-0.01949165016412735,
0.016094792634248734,
-0.05343703180551529,
0.025590406730771065,
-0.021190064027905464,
0.0713043063879013,
-0.06904124468564987,
-0.005458584055304527,
0.0468708835542202,
-0.01722247712314129,
0.06087959185242653,
0.07794645428657532,
0.08911407738924026,
-0.008097824640572071,
-0.09739968180656433,
0.0006580949993804097,
-0.04404053837060928,
-0.00777689041569829,
0.0011201133020222187,
0.12637588381767273,
0.023157142102718353,
0.03606336563825607,
-0.02060513384640217,
0.009776203893125057,
0.09940037131309509,
-0.04314497113227844,
-0.08796969056129456,
0.045454300940036774,
0.0550018846988678,
-0.03816740959882736,
-0.03999438136816025,
-0.07893010973930359,
0.05713026598095894,
-0.04743363335728645,
0.038360774517059326,
-0.08895557373762131,
-0.023327507078647614,
0.03917156159877777,
-0.05291382223367691,
0.07064815610647202,
-0.030609117820858955,
-0.00849554967135191,
-0.03969212993979454,
0.041575729846954346,
0.05493779852986336,
-0.010851647704839706,
-0.054691433906555176,
-0.012119300663471222
] | 0 | 0 | 0 | 0 |
|
The Hidden Cost of Waiting for Accurate Predictions | https://openreview.net/forum?id=A3YUPeJTNR | [
"Ali Shirali",
"Ariel D. Procaccia",
"Rediet Abebe"
] | Oral | Algorithmic predictions are increasingly informing societal resource allocations by identifying individuals for targeting. Policymakers often build these systems with the assumption that by gathering more observations on individuals, they can improve predictive accuracy and, consequently, allocation efficiency. An overlooked yet consequential aspect of prediction-driven allocations is that of timing. The planner has to trade off relying on earlier and potentially noisier predictions to intervene before individuals experience undesirable outcomes, or they may wait to gather more observations to make more precise allocations. We examine this tension using a simple mathematical model, where the planner collects observations on individuals to improve predictions over time. We analyze both the ranking induced by these predictions and optimal resource allocation. We show that though individual prediction accuracy improves over time, counter-intuitively, the average ranking loss can worsen. As a result, the planner's ability to improve social welfare can decline. We identify inequality as a driving factor behind this phenomenon. Our findings provide a nuanced perspective and challenge the conventional wisdom that it is preferable to wait for more accurate predictions to ensure the most efficient allocations. | Algorithmic Decision Making, Prediction, Resource Allocation, Social Welfare, Limits of Prediction | null | 4,828 | 2503.00650 | [
-0.037730004638433456,
-0.03433731570839882,
0.008846732787787914,
0.11696669459342957,
0.10970215499401093,
0.02026151679456234,
-0.02841525711119175,
-0.0070663932710886,
0.00177295645698905,
0.039557136595249176,
-0.07260257005691528,
0.008628378622233868,
0.03237007185816765,
0.04678122326731682,
-0.034719955176115036,
-0.04797274246811867,
0.05463843047618866,
0.023416860029101372,
0.01138562336564064,
-0.05289636179804802,
-0.01982739008963108,
-0.03282661363482475,
0.02155398204922676,
-0.004898924380540848,
0.08262655138969421,
-0.07938416302204132,
-0.023000327870249748,
-0.02037191018462181,
-0.007524437271058559,
0.04901434853672981,
0.04459470137953758,
-0.0579984113574028,
0.05390092357993126,
-0.0505501851439476,
-0.07161185145378113,
0.0015727895079180598,
-0.05979277566075325,
-0.03646744787693024,
0.005887466482818127,
-0.0005987502518109977,
0.025181863456964493,
-0.009864368475973606,
-0.0375342071056366,
0.029153162613511086,
-0.056988369673490524,
0.07268144190311432,
0.058669984340667725,
-0.0003947287332266569,
-0.08492307364940643,
-0.07574840635061264,
-0.05274955555796623,
-0.00047919980715960264,
0.02935098297894001,
-0.022917121648788452,
-0.05580659210681915,
0.04693352058529854,
0.046205710619688034,
0.03944393992424011,
-0.019069340080022812,
0.012542327865958214,
-0.07084903866052628,
-0.13881345093250275,
-0.02579403668642044,
-0.05560027435421944,
0.01981368288397789,
-0.011012188158929348,
0.01333710178732872,
0.0723423883318901,
-0.0786665603518486,
0.0709373727440834,
0.05496535450220108,
0.0559162013232708,
-0.05916953086853027,
0.029116248711943626,
0.11073795706033707,
0.046755388379096985,
0.08260820060968399,
0.08578890562057495,
0.032965037971735,
-0.03160145506262779,
-0.015626216307282448,
0.004704172257333994,
-0.038988735526800156,
0.0431024543941021,
0.04293369874358177,
0.0025288620963692665,
0.03864004462957382,
0.0056785824708640575,
0.10268596559762955,
-0.018996713683009148,
0.004257341846823692,
-0.010566175915300846,
-0.009255296550691128,
0.0025340698193758726,
-0.007578957360237837,
0.08891033381223679,
-0.011678420007228851,
-0.04607896879315376,
-0.076360322535038,
0.054780278354883194,
-0.07582346349954605,
0.012641719542443752,
0.0038220302667468786,
-0.025707893073558807,
0.0420810841023922,
-0.08071085065603256,
-0.033239953219890594,
0.013143706135451794,
-0.02534881792962551,
-0.01753014512360096,
0.08384625613689423,
0.014767398126423359,
0.03772995248436928,
0.005749562755227089,
-0.030162321403622627,
0.027242889627814293,
-0.038803260773420334,
0.033296942710876465,
-0.01847909949719906,
0.07660985738039017,
0.020075155422091484,
0.0033405227586627007,
0.06333661824464798,
-0.035869013518095016,
0.01928451843559742,
0.030194133520126343,
-0.10573755204677582,
3.45013193931607e-34,
0.0108278077095747,
-0.03907771781086922,
0.08788254112005234,
0.015143567696213722,
-0.05109655484557152,
-0.05587529391050339,
-0.05907829850912094,
-0.025043297559022903,
0.011244062334299088,
-0.007844124920666218,
0.000838799518533051,
0.011647501029074192,
0.01865636557340622,
0.10271541774272919,
0.07466647028923035,
0.0016435525612905622,
-0.003287702100351453,
0.13020235300064087,
-0.06295547634363174,
-0.012682939879596233,
0.08751954883337021,
-0.024725763127207756,
-0.038090962916612625,
-0.11653819680213928,
0.02236737310886383,
-0.018188275396823883,
-0.07212749868631363,
-0.022167213261127472,
0.04380537196993828,
-0.022368887439370155,
0.006470244377851486,
0.04095589369535446,
0.059014592319726944,
-0.03959139063954353,
0.029791925102472305,
-0.035773199051618576,
-0.022770840674638748,
0.010300032794475555,
0.015379922464489937,
-0.03139514476060867,
-0.05863293260335922,
-0.009476828388869762,
0.07721782475709915,
0.018697811290621758,
0.008621553890407085,
-0.08825474232435226,
-0.00013757777924183756,
-0.0005011854227632284,
-0.07923970371484756,
0.05488167330622673,
0.046156879514455795,
0.07154664397239685,
-0.03492246940732002,
-0.011271107010543346,
-0.08856123685836792,
-0.05479910969734192,
0.0305793359875679,
-0.03766131028532982,
-0.07001055032014847,
-0.08882211893796921,
0.02952640876173973,
-0.06484591215848923,
0.02848110906779766,
-0.015379129908978939,
-0.040561068803071976,
0.02623896673321724,
-0.00045317516196519136,
0.035061199218034744,
0.07580074667930603,
0.014952697791159153,
0.029625065624713898,
0.013047701679170132,
0.010252355597913265,
-0.0035139143001288176,
0.004098871257156134,
0.08902789652347565,
0.04475812241435051,
0.005046087317168713,
0.007535355631262064,
-0.0364692248404026,
0.07110876590013504,
-0.04300906881690025,
0.02555721439421177,
-0.04144664108753204,
0.017857009544968605,
-0.03922594338655472,
0.04277163743972778,
0.03586491197347641,
0.011172608472406864,
0.0049317749217152596,
-0.09040802717208862,
0.02356342412531376,
-0.054562799632549286,
0.07506482303142548,
-0.06599494069814682,
-2.8292356395407456e-33,
-0.04838642477989197,
-0.03937077522277832,
0.07254692912101746,
-0.058586254715919495,
0.04750906303524971,
-0.04945491626858711,
-0.04720083251595497,
-0.11581187695264816,
-0.022779889404773712,
-0.05356073006987572,
-0.10809732228517532,
-0.009537882171571255,
0.10263385623693466,
0.06116807833313942,
-0.013873239979147911,
-0.03350706025958061,
0.027304930612444878,
-0.04642133042216301,
0.005387844983488321,
-0.0656978115439415,
0.04753762483596802,
0.03840329125523567,
-0.04851150140166283,
0.05811840668320656,
-0.01643458567559719,
-0.007109784986823797,
0.0331885926425457,
0.05072477087378502,
-0.052886538207530975,
-0.015633022412657738,
-0.11341256648302078,
-0.11288799345493317,
-0.02886859141290188,
-0.07989390194416046,
0.012924146838486195,
0.08024275302886963,
0.03636464849114418,
-0.015428503975272179,
-0.05359620973467827,
0.09497944265604019,
0.041823238134384155,
-0.04155072197318077,
-0.04750387370586395,
-0.029622353613376617,
-0.04019849747419357,
0.03705102950334549,
-0.014813471585512161,
-0.00991121120750904,
0.039216529577970505,
0.010366206988692284,
0.013361988589167595,
0.09681505709886551,
-0.024327870458364487,
0.025642458349466324,
-0.018510622903704643,
-0.06334085762500763,
-0.07500164210796356,
-0.03640328720211983,
0.12148212641477585,
0.1057879850268364,
-0.08943945169448853,
0.07662910223007202,
-0.006762253120541573,
0.034354936331510544,
-0.060942865908145905,
-0.022560587152838707,
-0.05277711898088455,
-0.1094047799706459,
0.07612209767103195,
-0.06451798975467682,
0.010990077629685402,
-0.009986831806600094,
-0.04253128543496132,
0.004277013707906008,
-0.05152313783764839,
0.053561218082904816,
0.02753681316971779,
0.06165248528122902,
-0.01200124528259039,
0.02560948021709919,
-0.007563486695289612,
0.030253272503614426,
0.059159524738788605,
-0.0966532900929451,
-0.0156612079590559,
0.02121582441031933,
0.09788006544113159,
0.03520090878009796,
-0.008923530578613281,
0.023316195234656334,
-0.07098589837551117,
-0.03929750248789787,
-0.01496459823101759,
0.014163578860461712,
-0.07095621526241302,
-4.747995419052131e-8,
0.05442453548312187,
0.06274022161960602,
0.05586590990424156,
0.06288185715675354,
0.026393495500087738,
-0.036192670464515686,
-0.015244204550981522,
0.03997155651450157,
0.00878637470304966,
0.02339925989508629,
0.08921980112791061,
0.03554460406303406,
0.015074389055371284,
0.005769359413534403,
0.044777415692806244,
0.04376589134335518,
-0.06113605573773384,
-0.05736880376935005,
-0.05151446908712387,
-0.03316820040345192,
0.06552954763174057,
0.012712299823760986,
0.0009459819411858916,
-0.06995491683483124,
0.08227409422397614,
-0.00547818886116147,
-0.027511287480592728,
0.08073783665895462,
0.019287098199129105,
0.09663637727499008,
-0.010134708136320114,
-0.015660248696804047,
-0.01722472347319126,
-0.020052731037139893,
0.022741902619600296,
0.027738304808735847,
-0.09219232946634293,
-0.06620746105909348,
0.051069509238004684,
-0.08204296231269836,
0.05101177468895912,
0.061834223568439484,
-0.029306521639227867,
0.04487529769539833,
0.059015464037656784,
-0.035731274634599686,
-0.033176541328430176,
0.002572671975940466,
0.036935798823833466,
-0.1041366457939148,
0.009911670349538326,
-0.05767986923456192,
0.015645047649741173,
0.01666194386780262,
0.13641348481178284,
-0.058347832411527634,
0.011935343965888023,
-0.03748156130313873,
0.007378388196229935,
0.03293880447745323,
-0.001778424484655261,
-0.04767997935414314,
-0.1040310263633728,
0.024989554658532143
] | https://github.com/alishiraliGit/hidden-costs-of-waiting-for-accurate-predictions | 1 | 0 | 0 | 0 |
Learning Dynamics of LLM Finetuning | https://openreview.net/forum?id=tPNHOoZFl9 | [
"Yi Ren",
"Danica J. Sutherland"
] | Oral | Learning dynamics, which describes how the learning of specific training examples influences the model's predictions on other examples,
gives us a powerful tool for understanding the behavior of deep learning systems. We study the learning dynamics of large language models during different types of finetuning, by analyzing the step-wise decomposition of how influence accumulates among different potential responses. Our framework allows a uniform interpretation of many interesting observations about the training of popular algorithms for both instruction tuning and preference tuning. In particular, we propose a hypothetical explanation of why specific types of hallucination are strengthened after finetuning, e.g., the model might use phrases or facts in the response for question B to answer question A, or the model might keep repeating similar simple phrases when generating responses. We also extend our framework and highlight a unique ``squeezing effect'' to explain a previously observed phenomenon in off-policy direct preference optimization (DPO), where running DPO for too long makes even the desired outputs less likely. This framework also provides insights into where the benefits of on-policy DPO and other variants come from. The analysis not only provides a novel perspective of understanding LLM's finetuning but also inspires a simple, effective method to improve alignment performance. | Learning dynamics, LLM, finetuning, DPO | The paper propose a novel learning dynamics framework to understand LLM's behavior during finetuning (e.g., SFT, DPO, and other variants). Some counter-intuitive behavior can be well explained by the proposed framework. | 4,818 | 2407.10490 | [
-0.042106255888938904,
-0.07338003814220428,
0.07968934625387192,
0.02995339408516884,
0.05061090737581253,
0.03212369978427887,
0.0036155020352452993,
0.007583836559206247,
0.0352928526699543,
-0.027520328760147095,
-0.07440584897994995,
0.020612504333257675,
-0.030317477881908417,
-0.02611747942864895,
0.010545333847403526,
0.020177273079752922,
0.0858929306268692,
0.04129956290125847,
-0.16373211145401,
-0.07547377794981003,
0.07588157802820206,
0.007819569669663906,
0.024633437395095825,
-0.032540783286094666,
-0.003490676637738943,
0.006025823764503002,
-0.029490584507584572,
0.01365711074322462,
0.07912237197160721,
-0.040678516030311584,
0.01422051526606083,
0.053640954196453094,
-0.027797846123576164,
-0.019254224374890327,
-0.07282581180334091,
0.06658399850130081,
-0.1497751623392105,
0.01681804656982422,
-0.017081886529922485,
-0.017932072281837463,
-0.041016798466444016,
0.02726743184030056,
0.03448755294084549,
0.027806099504232407,
0.06960321217775345,
0.0011098120594397187,
-0.007299219258129597,
0.009361272677779198,
-0.08274070173501968,
0.002092666458338499,
-0.06424979865550995,
-0.012936240993440151,
-0.0036803751718252897,
0.018718263134360313,
0.011778958141803741,
0.03606307506561279,
0.034642454236745834,
0.08194591850042343,
-0.01200800109654665,
-0.055994175374507904,
-0.12847092747688293,
-0.1285497099161148,
-0.07333046942949295,
-0.03753116726875305,
-0.04429124295711517,
-0.04732471704483032,
-0.0025041489861905575,
0.08964642882347107,
-0.04863547161221504,
0.06873106956481934,
0.03361380100250244,
0.08402230590581894,
0.00643916754052043,
0.02960960753262043,
0.06720812618732452,
0.005808798596262932,
0.03290671855211258,
-0.01177313644438982,
0.037711381912231445,
-0.08375052362680435,
0.004799887537956238,
-0.009632299654185772,
0.006411367561668158,
0.014945215545594692,
0.0881214290857315,
-0.051597047597169876,
-0.000529918703250587,
-0.028559153899550438,
0.08915285021066666,
0.033500734716653824,
-0.058659788221120834,
-0.07779958099126816,
0.0025978891644626856,
-0.04242659732699394,
0.046771373599767685,
0.10347385704517365,
-0.03223438188433647,
-0.08639770746231079,
-0.06892679631710052,
0.0743815153837204,
0.02404637075960636,
0.08148585259914398,
-0.026668788865208626,
-0.038406506180763245,
0.0027338466607034206,
-0.06518232822418213,
-0.001306345802731812,
0.07921725511550903,
0.06335325539112091,
-0.078293077647686,
0.03153010457754135,
0.07863527536392212,
0.031756844371557236,
0.039538078010082245,
-0.024155249819159508,
-0.05786711350083351,
0.003888744628056884,
-0.08906279504299164,
0.013496180064976215,
0.04818283021450043,
-0.044817619025707245,
-0.04956326261162758,
-0.005007714498788118,
0.014593376778066158,
-0.032726068049669266,
-0.015487122349441051,
-0.07274499535560608,
3.987793320819592e-33,
0.03610733523964882,
-0.02691464312374592,
-0.04549843817949295,
0.006638087332248688,
0.0507861003279686,
-0.0413348488509655,
0.004650854505598545,
0.00684090331196785,
-0.025626441463828087,
0.019954975694417953,
-0.028239618986845016,
-0.015866722911596298,
-0.03657595440745354,
0.10695349425077438,
0.017449110746383667,
-0.020076947286725044,
-0.01865239255130291,
0.04927442967891693,
-0.057020459324121475,
-0.03588022664189339,
0.07949508726596832,
0.07427443563938141,
-0.009309808723628521,
-0.04213957488536835,
-0.007520186714828014,
-0.022622639313340187,
0.03167225793004036,
-0.019546320661902428,
-0.04320381581783295,
0.020149827003479004,
-0.044004350900650024,
-0.019007451832294464,
-0.054559189826250076,
0.02481743134558201,
0.045205820351839066,
-0.03253679350018501,
-0.022179117426276207,
0.006384741049259901,
0.02245626226067543,
-0.020582690834999084,
0.02033216319978237,
0.030916238203644753,
0.043256036937236786,
0.0029964835848659277,
-0.06057451292872429,
-0.03587972745299339,
0.03709835931658745,
0.01016390509903431,
-0.07060367614030838,
-0.0577397421002388,
0.0394342802464962,
0.027853908017277718,
0.044401269406080246,
0.018998682498931885,
-0.0003597439208533615,
-0.024089984595775604,
-0.03098786622285843,
0.032271359115839005,
0.008622555062174797,
0.02317020669579506,
0.036066051572561264,
0.061073191463947296,
-0.002078095218166709,
0.07094301283359528,
0.02678568847477436,
0.03080661967396736,
-0.04288419708609581,
0.03632698953151703,
0.04659067839384079,
-0.06247169151902199,
-0.016550039872527122,
-0.0012663612142205238,
-0.061360303312540054,
-0.08186442404985428,
0.05382770672440529,
0.01252409815788269,
0.038933802396059036,
-0.04904364049434662,
0.054402340203523636,
-0.05147097259759903,
0.07617492973804474,
0.010935249738395214,
0.02604045160114765,
-0.042331844568252563,
-0.03677072748541832,
-0.06576477736234665,
-0.007127550430595875,
-0.1258329153060913,
0.006638938561081886,
-0.03446565195918083,
-0.03232588991522789,
-0.03366456925868988,
-0.015752037987113,
0.029804298654198647,
0.019655603915452957,
-4.571017414373962e-33,
-0.07700268179178238,
-0.024913225322961807,
0.03666505217552185,
0.07921013236045837,
-0.028962159529328346,
0.007709215860813856,
-0.04213246330618858,
-0.021881330758333206,
0.011106790043413639,
-0.11351117491722107,
-0.04662363976240158,
-0.0055839125998318195,
0.0336487703025341,
0.0016967632109299302,
0.03424374386668205,
-0.04517897963523865,
-0.0003950276877731085,
0.0018025959143415093,
0.033756546676158905,
0.06932811439037323,
0.00535444775596261,
0.030393386259675026,
-0.10737289488315582,
0.02457091212272644,
-0.06363086402416229,
0.04007919505238533,
-0.06378393620252609,
0.10587393492460251,
0.012043504975736141,
0.050434328615665436,
-0.051431912928819656,
0.01329316571354866,
-0.07847873866558075,
0.004608732648193836,
-0.020089609548449516,
0.0499078705906868,
-0.006471538450568914,
-0.003144875168800354,
-0.0569603368639946,
0.12105251103639603,
0.060163211077451706,
-0.05575423315167427,
-0.01775417849421501,
0.01649473048746586,
-0.0019437685841694474,
-0.025166140869259834,
-0.05460767075419426,
-0.012858475558459759,
0.03068723902106285,
-0.013165169395506382,
0.02688182145357132,
0.028506141155958176,
-0.006838216912001371,
0.05284716561436653,
-0.08409245312213898,
-0.06279190629720688,
0.03981262817978859,
-0.003003145335242152,
0.09901978820562363,
0.05065157264471054,
-0.01960121840238571,
0.014889353886246681,
0.010768075473606586,
-0.08677385747432709,
-0.015326590277254581,
-0.008519585244357586,
-0.0484011173248291,
-0.024230916053056717,
0.06953223049640656,
-0.02209583856165409,
0.03135692700743675,
-0.04818092659115791,
-0.022226903587579727,
-0.027062879875302315,
-0.02878381684422493,
0.007039625197649002,
0.03826994076371193,
-0.10749085992574692,
-0.060811664909124374,
-0.0031636757776141167,
-0.012076035141944885,
0.06831449270248413,
-0.03307654708623886,
0.03693593293428421,
0.01919265277683735,
0.046675216406583786,
0.0022871517576277256,
0.10201524198055267,
-0.0004556668864097446,
0.006993961054831743,
0.02900303341448307,
-0.011175017803907394,
0.012105514295399189,
0.03716009482741356,
-0.09777729958295822,
-5.4539263771857804e-8,
-0.060532961040735245,
0.01249033771455288,
0.04881950840353966,
0.1328006088733673,
0.10289319604635239,
-0.06679277122020721,
0.02368372306227684,
0.10767970979213715,
-0.05301983654499054,
-0.026322945952415466,
0.10233552753925323,
0.10384518653154373,
-0.010918325744569302,
-0.09750374406576157,
0.02748904749751091,
0.13867871463298798,
0.02215135656297207,
-0.01667775958776474,
-0.037201374769210815,
-0.017892520874738693,
0.06998945027589798,
-0.02808827906847,
0.04688186198472977,
0.005566920153796673,
0.0008541041752323508,
-0.03441515937447548,
-0.012213924899697304,
0.09377247840166092,
0.07387570291757584,
0.05038983002305031,
0.053758617490530014,
0.014094271697103977,
-0.05218655616044998,
-0.04073905944824219,
0.002296998631209135,
0.14546003937721252,
-0.010230081155896187,
-0.09028598666191101,
0.08386171609163284,
-0.04551070183515549,
0.04059555009007454,
0.0484878271818161,
0.001384672592394054,
0.06906253099441528,
-0.014197813346982002,
-0.029961973428726196,
-0.04329625144600868,
-0.10637956112623215,
0.012336049228906631,
-0.038064826279878616,
0.0023131519556045532,
0.01594005897641182,
-0.012176372110843658,
0.0677843987941742,
0.0601346455514431,
0.012901484034955502,
-0.016468025743961334,
-0.06803005188703537,
-0.021541284397244453,
0.11115384101867676,
0.05112934485077858,
0.04471916705369949,
-0.09229213744401932,
0.0022317247930914164
] | https://github.com/joshua-ren/learning_dynamics_llm | 61 | 0 | 0 | 0 |
Root Cause Analysis of Anomalies in Multivariate Time Series through Granger Causal Discovery | https://openreview.net/forum?id=k38Th3x4d9 | [
"Xiao Han",
"Saima Absar",
"Lu Zhang",
"Shuhan Yuan"
] | Oral | Identifying the root causes of anomalies in multivariate time series is challenging due to the complex dependencies among the series. In this paper, we propose a comprehensive approach called AERCA that inherently integrates Granger causal discovery with root cause analysis. By defining anomalies as interventions on the exogenous variables of time series, AERCA not only learns the Granger causality among time series but also explicitly models the distributions of exogenous variables under normal conditions. AERCA then identifies the root causes of anomalies by highlighting exogenous variables that significantly deviate from their normal states. Experiments on multiple synthetic and real-world datasets demonstrate that AERCA can accurately capture the causal relationships among time series and effectively identify the root causes of anomalies. | root cause analysis, Granger causality, multivariate time series | null | 4,815 | null | [
-0.12040255963802338,
-0.09741491079330444,
0.08374623954296112,
0.09345606714487076,
0.09084536880254745,
-0.05997480824589729,
-0.07197344303131104,
-0.07422394305467606,
0.001384426956064999,
-0.013722382485866547,
0.022856533527374268,
0.002560473745688796,
-0.015426934696733952,
0.013313183560967445,
-0.09896362572908401,
-0.07013371586799622,
-0.05813833326101303,
-0.02659086138010025,
0.02943546138703823,
-0.041876234114170074,
0.002592539181932807,
-0.009589002467691898,
-0.07093074172735214,
0.03298071026802063,
0.017577802762389183,
0.001008981023915112,
0.08012605458498001,
0.01372147724032402,
-0.04386380687355995,
-0.03543952852487564,
0.041558638215065,
0.07156989723443985,
-0.055324409157037735,
-0.08076180517673492,
-0.034806717187166214,
0.048106495290994644,
0.09323132783174515,
0.09718678891658783,
0.010949911549687386,
-0.06771405041217804,
0.0764167308807373,
-0.01574113219976425,
0.04356995224952698,
0.05349833890795708,
-0.00000929195994103793,
-0.024688493460416794,
0.04939313977956772,
-0.04054228216409683,
-0.013291764073073864,
-0.00921807810664177,
-0.018187925219535828,
-0.018092842772603035,
-0.0008103835280053318,
0.016718223690986633,
0.0390065461397171,
-0.027184344828128815,
-0.008574118837714195,
-0.0019362203311175108,
-0.007939893752336502,
0.04911007732152939,
-0.004581979475915432,
0.03284893184900284,
-0.06508032977581024,
-0.007639086805284023,
0.011608325876295567,
0.0682007297873497,
0.014391707256436348,
0.15850883722305298,
-0.07997284084558487,
0.02496030181646347,
-0.02637123130261898,
0.03518059477210045,
-0.050276704132556915,
-0.019654177129268646,
0.036882609128952026,
0.03865534067153931,
0.009755655191838741,
-0.04211389273405075,
0.03948897868394852,
-0.03663422912359238,
-0.005570843815803528,
0.06644821166992188,
0.013168497011065483,
-0.07342128455638885,
0.1252439320087433,
0.009656635113060474,
0.08808480203151703,
-0.015816742554306984,
0.02979949116706848,
0.10015296190977097,
0.10181573033332825,
0.012717260047793388,
0.10510185360908508,
0.014878750778734684,
0.028050096705555916,
0.07703271508216858,
-0.004406000021845102,
0.010117944329977036,
0.055738091468811035,
0.03263106942176819,
0.018471281975507736,
0.020754018798470497,
-0.005553742870688438,
0.07140413671731949,
0.005602697841823101,
-0.04239943251013756,
0.005347734782844782,
-0.06513220816850662,
0.02940860018134117,
-0.034264516085386276,
-0.049224723130464554,
0.016879163682460785,
-0.010658805258572102,
-0.04850984737277031,
-0.02246365137398243,
-0.02079816535115242,
-0.032632410526275635,
-0.03834065422415733,
-0.03311994671821594,
0.018041299656033516,
0.002232061931863427,
0.07480449229478836,
-0.005606471560895443,
0.062202151864767075,
0.0500132255256176,
-0.03822975605726242,
-0.1100042536854744,
4.0652246030761944e-33,
0.005103932693600655,
-0.006616870407015085,
-0.022084452211856842,
0.02663169614970684,
-0.054812949150800705,
0.05960458889603615,
-0.08947622776031494,
-0.025657210499048233,
0.02284213714301586,
0.07564930617809296,
0.010720242746174335,
0.03741106018424034,
-0.06549793481826782,
-0.05355394259095192,
0.0313832089304924,
0.019046254456043243,
-0.04812731221318245,
0.029679208993911743,
0.026127949357032776,
-0.027462365105748177,
0.008550360798835754,
-0.0837893858551979,
-0.03587329760193825,
-0.06281401216983795,
-0.007300031371414661,
0.011751960963010788,
0.025603674352169037,
-0.010167880915105343,
0.015373185276985168,
0.02280254289507866,
0.04705314710736275,
-0.017838452011346817,
0.027530331164598465,
0.059680867940187454,
-0.01564490981400013,
-0.04723425954580307,
-0.056463733315467834,
-0.06052803620696068,
-0.010726472362875938,
-0.003354550339281559,
-0.04189887270331383,
-0.009276606142520905,
-0.0005898373783566058,
0.019981607794761658,
-0.0633959248661995,
-0.04005289077758789,
0.06316322833299637,
-0.037341173738241196,
0.02724510058760643,
-0.0032690109219402075,
-0.02650315687060356,
-0.014246797189116478,
0.04018007591366768,
-0.10981250554323196,
-0.02800404466688633,
0.09524302184581757,
-0.0762411430478096,
-0.04338553547859192,
-0.04610423743724823,
-0.007939386181533337,
-0.032092489302158356,
-0.005730736069381237,
-0.0021800489630550146,
-0.013664993457496166,
-0.032663628458976746,
0.06512629985809326,
0.022606516256928444,
0.0020551132038235664,
-0.04304688423871994,
0.12661269307136536,
-0.027307402342557907,
-0.026625225320458412,
-0.015237481333315372,
-0.012005931697785854,
0.022038545459508896,
-0.04344681277871132,
-0.0037982126232236624,
0.015200109221041203,
-0.0303556676954031,
0.019886838272213936,
-0.06712277233600616,
-0.11884326487779617,
0.0823885127902031,
0.02889789640903473,
-0.022162148728966713,
-0.004797023255378008,
-0.005796945188194513,
0.02946608141064644,
-0.03992117568850517,
-0.02730731852352619,
0.036346811801195145,
0.051903337240219116,
-0.014908215031027794,
0.01099711749702692,
0.018718060106039047,
-5.973352420204196e-33,
-0.003566392930224538,
-0.008946303278207779,
0.04423554241657257,
-0.10624498128890991,
-0.005627789534628391,
0.004611694719642401,
-0.03333748131990433,
-0.01538771204650402,
0.0004638662503566593,
-0.03338468447327614,
-0.006895849946886301,
-0.03285779058933258,
0.010274172760546207,
-0.06234518066048622,
0.010470021516084671,
0.043546006083488464,
0.05949981510639191,
-0.02974720485508442,
0.021711230278015137,
-0.03131214156746864,
-0.033120620995759964,
0.0003476083802524954,
-0.015811942517757416,
-0.05886809155344963,
-0.009994669817388058,
-0.03639652952551842,
0.003793998621404171,
0.10821343213319778,
-0.07911087572574615,
0.06187252327799797,
-0.08430849760770798,
0.028007272630929947,
-0.022013317793607712,
0.026685675606131554,
-0.08795640617609024,
0.11910907924175262,
0.034091874957084656,
-0.06887494772672653,
-0.04900360107421875,
0.012592465616762638,
0.07366558909416199,
0.09763515740633011,
-0.015335851348936558,
-0.013288166373968124,
-0.01266053318977356,
0.05594140291213989,
-0.05186488851904869,
0.1648324877023697,
0.050749171525239944,
-0.018162976950407028,
-0.006645713932812214,
0.04429641738533974,
0.01395469531416893,
-0.012619279325008392,
-0.05785041302442551,
0.010609697550535202,
0.06963951140642166,
-0.05745420977473259,
0.0017568327020853758,
-0.019145403057336807,
-0.04852971062064171,
-0.015071159228682518,
0.004967721179127693,
-0.046095192432403564,
-0.019731245934963226,
-0.06341002136468887,
0.009928080253303051,
-0.1266799122095108,
0.08308972418308258,
-0.04442157968878746,
0.11798489838838577,
0.049750685691833496,
-0.12719513475894928,
-0.08391044288873672,
-0.022553781047463417,
0.01673823595046997,
-0.009784040041267872,
-0.11704596877098083,
-0.08880088478326797,
-0.06650928407907486,
0.07792205363512039,
0.03734096512198448,
0.13218195736408234,
0.006202442571520805,
-0.08792418986558914,
0.06536483019590378,
0.05254955217242241,
0.017520491033792496,
0.021696854382753372,
-0.023741841316223145,
-0.03908417373895645,
0.020344989374279976,
-0.04253631830215454,
0.06832238286733627,
0.0005459637613967061,
-3.6124141900018e-8,
-0.04883372411131859,
0.008929158560931683,
-0.009568572044372559,
-0.037653472274541855,
0.05240456387400627,
-0.02242436073720455,
-0.035017527639865875,
0.059054963290691376,
0.014091861434280872,
0.000200351991225034,
0.057225313037633896,
-0.024086518213152885,
-0.0030037083197385073,
-0.015931183472275734,
0.02566482499241829,
-0.04781036823987961,
-0.003960806876420975,
0.08529119193553925,
-0.07212284952402115,
-0.040188033133745193,
0.06721584498882294,
0.0359528623521328,
-0.03125558793544769,
0.02542744018137455,
0.09594707190990448,
-0.05599181354045868,
-0.014406496658921242,
0.03958606719970703,
0.06285031884908676,
-0.00004687027103500441,
0.005663653369992971,
-0.01706950180232525,
0.05916964262723923,
-0.0761629119515419,
-0.042450204491615295,
0.113430455327034,
0.015234769321978092,
-0.046617936342954636,
-0.08084941655397415,
-0.014256210066378117,
-0.01956336945295334,
0.06928438693284988,
-0.050825782120227814,
0.03811560943722725,
-0.04264746978878975,
-0.03175704926252365,
-0.002331013325601816,
-0.0710282251238823,
0.11107594519853592,
-0.0348779559135437,
0.0827927216887474,
-0.028995530679821968,
0.052498359233140945,
0.05291978269815445,
0.11789249628782272,
-0.005786588881164789,
-0.017290247604250908,
0.005786888301372528,
0.011513035744428635,
-0.031341880559921265,
0.08246500045061111,
0.005565576255321503,
-0.010884897783398628,
-0.0029573177453130484
] | 0 | 0 | 0 | 0 |
|
ProtComposer: Compositional Protein Structure Generation with 3D Ellipsoids | https://openreview.net/forum?id=0ctvBgKFgc | [
"Hannes Stark",
"Bowen Jing",
"Tomas Geffner",
"Jason Yim",
"Tommi Jaakkola",
"Arash Vahdat",
"Karsten Kreis"
] | Oral | We develop ProtComposer to generate protein structures conditioned on spatial protein layouts that are specified via a set of 3D ellipsoids capturing substructure shapes and semantics. At inference time, we condition on ellipsoids that are hand-constructed, extracted from existing proteins, or from a statistical model, with each option unlocking new capabilities. Hand-specifying ellipsoids enables users to control the location, size, orientation, secondary structure, and approximate shape of protein substructures. Conditioning on ellipsoids of existing proteins enables redesigning their substructure's connectivity or editing substructure properties. By conditioning on novel and diverse ellipsoid layouts from a simple statistical model, we improve protein generation with expanded Pareto frontiers between designability, novelty, and diversity. Further, this enables sampling designable proteins with a helix-fraction that matches PDB proteins, unlike existing generative models that commonly oversample conceptually simple helix bundles. Code is available at https://github.com/NVlabs/protcomposer. | protein design, diffusion model, controllable generation, drug discovery, proteins, biology | We develop a framework to generate protein structures conditioned on spatial protein layouts that are specified via a set of 3D ellipsoids. | 4,802 | 2503.05025 | [
-0.06704650074243546,
-0.13042928278446198,
0.02809673361480236,
-0.045788563787937164,
0.056138474494218826,
-0.016790878027677536,
-0.06842411309480667,
0.007588489446789026,
0.10412406176328659,
-0.03918766230344772,
-0.00466907350346446,
-0.046465497463941574,
-0.0024958776775747538,
0.049401480704545975,
0.0625884011387825,
0.03733953833580017,
0.018196599557995796,
0.09811225533485413,
-0.058968186378479004,
-0.013272867538034916,
0.07262421399354935,
0.04425791651010513,
0.03807410970330238,
0.030314665287733078,
-0.01038864254951477,
-0.02510724402964115,
0.06836298853158951,
0.033817436546087265,
0.03811588138341904,
-0.06100501865148544,
0.07965259999036789,
-0.01710713654756546,
-0.020512254908680916,
-0.05110650509595871,
0.07984823733568192,
0.013808324001729488,
-0.05835613235831261,
-0.010858267545700073,
0.00033218966564163566,
-0.054456211626529694,
0.05498964339494705,
-0.008895743638277054,
0.041231486946344376,
0.01846921443939209,
0.014559240080416203,
0.034996066242456436,
-0.0408192053437233,
0.024651847779750824,
0.01016108226031065,
0.04487880691885948,
-0.020647183060646057,
-0.08689097315073013,
-0.07652966678142548,
-0.00787799246609211,
-0.007877349853515625,
0.03605734556913376,
-0.02554800920188427,
-0.03061479702591896,
0.029740827158093452,
-0.038438744843006134,
0.04317934811115265,
-0.06024112179875374,
0.046103835105895996,
-0.04352661222219467,
0.03419298678636551,
-0.04106362909078598,
0.014734884724020958,
0.06969708204269409,
0.01964140683412552,
0.04651758074760437,
-0.00194924499373883,
0.07333434373140335,
-0.059962257742881775,
0.10651122033596039,
0.04198228567838669,
0.11218127608299255,
-0.03447350859642029,
0.009945427998900414,
0.0057522268034517765,
-0.0950087234377861,
-0.010708353482186794,
0.08534500747919083,
0.07232426851987839,
0.0746772289276123,
0.06252636760473251,
0.01644205115735531,
-0.008180108852684498,
0.07768987864255905,
-0.0025220373645424843,
-0.0009646836551837623,
-0.012061350978910923,
-0.009376361034810543,
0.0032105848658829927,
-0.06744355708360672,
-0.07444629073143005,
0.03847561031579971,
0.06386560201644897,
-0.08263328671455383,
0.020840052515268326,
0.051149725914001465,
-0.0012041645823046565,
0.01844804547727108,
0.15818876028060913,
-0.060995880514383316,
0.06217770278453827,
-0.026034899055957794,
-0.01688707061111927,
0.06804589927196503,
0.007645012810826302,
-0.0035770703107118607,
0.0452602356672287,
0.014804013073444366,
0.00981622189283371,
-0.010111101903021336,
0.019663043320178986,
0.018577760085463524,
-0.06750447303056717,
-0.043782152235507965,
0.055595461279153824,
0.033855121582746506,
-0.004223220515996218,
0.0235904548317194,
-0.0514400489628315,
0.007997868582606316,
-0.05483616515994072,
-0.0010022983187809587,
-0.11840110272169113,
2.857669194848206e-33,
0.10440754145383835,
0.03831157088279724,
0.027249349281191826,
0.11426017433404922,
0.004124511033296585,
-0.03548112139105797,
-0.008803891949355602,
0.006560125853866339,
-0.0315609909594059,
0.023972589522600174,
-0.05213233828544617,
-0.03691348060965538,
0.02789195440709591,
0.08916853368282318,
0.10036475956439972,
-0.01688770204782486,
-0.022840401157736778,
0.07996398955583572,
-0.0663367286324501,
0.02181403338909149,
-0.05078895017504692,
-0.005651131272315979,
-0.030654193833470345,
-0.034287165850400925,
-0.02361016348004341,
0.0150837991386652,
-0.06146590784192085,
-0.05710946023464203,
-0.09317746758460999,
0.0040373243391513824,
-0.10263024270534515,
0.002430570777505636,
-0.03943426534533501,
0.05657336860895157,
0.06413529813289642,
-0.05640653520822525,
0.04322606697678566,
-0.1281866878271103,
-0.04451008886098862,
-0.05185709521174431,
0.0449700653553009,
0.026929667219519615,
0.007364279590547085,
0.0001749279472278431,
-0.0440853014588356,
-0.03068920597434044,
0.07628852128982544,
0.03156554698944092,
-0.026776650920510292,
-0.071479432284832,
0.12720990180969238,
0.05205988138914108,
0.09968560189008713,
-0.028824929147958755,
-0.049693599343299866,
-0.0075865089893341064,
-0.06053002178668976,
-0.014252553693950176,
0.10488715767860413,
0.03757942095398903,
-0.006981304381042719,
0.0820620208978653,
-0.002673674840480089,
-0.018388735130429268,
0.044188350439071655,
0.06319133937358856,
-0.030168090015649796,
-0.0323413722217083,
0.060097042471170425,
0.05703800916671753,
-0.015196377411484718,
0.012698742561042309,
-0.038682010024785995,
-0.014758190140128136,
-0.011859053745865822,
-0.025364084169268608,
-0.007921639829874039,
-0.022373586893081665,
0.015549649484455585,
0.03923080861568451,
-0.05318961665034294,
0.027813062071800232,
-0.05319790542125702,
-0.03996076062321663,
-0.05017312243580818,
-0.0737384483218193,
0.1404060274362564,
-0.04511325806379318,
-0.0635012686252594,
-0.049707841128110886,
-0.0030987937934696674,
-0.10264105349779129,
0.021327557042241096,
0.01856672205030918,
-0.009362511336803436,
-2.5403531273803683e-33,
-0.03186431899666786,
-0.037066757678985596,
0.027518045157194138,
-0.028045298531651497,
-0.07653503865003586,
-0.0844501480460167,
-0.013990560546517372,
-0.030285455286502838,
-0.014596306718885899,
-0.06948932260274887,
-0.003037065966054797,
-0.018830806016921997,
0.061027709394693375,
-0.08851860463619232,
0.015928048640489578,
0.06674817949533463,
-0.14748884737491608,
-0.05053352564573288,
0.0066433376632630825,
0.01708056405186653,
-0.02339955046772957,
0.0415753498673439,
-0.007460025604814291,
0.009821506217122078,
-0.02594630792737007,
0.09464752674102783,
-0.04054401442408562,
-0.03220100328326225,
0.014204559847712517,
0.029518108814954758,
-0.12477318197488785,
0.0015632639406248927,
-0.050843745470047,
-0.00509704789146781,
-0.08656716346740723,
0.020436054095625877,
-0.04197084158658981,
0.011633027344942093,
0.09700954705476761,
-0.06495480984449387,
-0.009685608558356762,
0.017285244539380074,
-0.056353919208049774,
0.10689570009708405,
0.056707900017499924,
-0.009085224941372871,
0.019010256975889206,
0.022634347900748253,
-0.01929311640560627,
-0.010441338643431664,
-0.047662194818258286,
0.04897717386484146,
-0.007484592031687498,
0.0033314996398985386,
-0.031020576134324074,
-0.05779862031340599,
-0.07797130197286606,
-0.05896058306097984,
0.045006196945905685,
-0.0455092154443264,
-0.10447660088539124,
-0.07367949932813644,
0.008430804125964642,
0.04420412331819534,
0.07751590013504028,
-0.008130502887070179,
-0.00861021876335144,
-0.04850861057639122,
-0.031555283814668655,
0.0050606392323970795,
0.04236217215657234,
-0.005191270262002945,
0.044324249029159546,
0.04342234134674072,
0.11418271064758301,
-0.07545439153909683,
-0.014905823394656181,
-0.012655377388000488,
-0.029251445084810257,
0.002518081571906805,
-0.0394337922334671,
0.004245923366397619,
0.0068827057257294655,
0.058316830545663834,
0.12965036928653717,
0.0004405697691254318,
-0.011963116936385632,
0.04009394720196724,
0.030084650963544846,
0.031669069081544876,
-0.006398085970431566,
0.024516087025403976,
0.009027387946844101,
0.15319271385669708,
0.06799614429473877,
-4.083000249011093e-8,
0.01976209320127964,
0.08177167922258377,
0.032970625907182693,
-0.015779582783579826,
0.007828340865671635,
0.040384940803050995,
-0.07155773043632507,
0.04349677637219429,
-0.0031345661263912916,
-0.0004972335300408304,
0.032426103949546814,
-0.012376523576676846,
-0.07676370441913605,
-0.01872589997947216,
0.013577516190707684,
0.05824074521660805,
0.04101857915520668,
0.05382012575864792,
-0.04665615037083626,
0.006545602809637785,
-0.020570766180753708,
0.08671595901250839,
-0.016295181587338448,
-0.022671081125736237,
0.050040945410728455,
-0.04315722733736038,
-0.025969941169023514,
-0.04288599267601967,
0.02211281843483448,
0.016934236511588097,
0.014892125502228737,
0.0018419072730466723,
0.019743412733078003,
0.027323873713612556,
-0.01393344160169363,
0.04212559014558792,
-0.04404501989483833,
-0.02603461593389511,
-0.07087568193674088,
-0.07168333232402802,
0.022512421011924744,
0.01102630514651537,
-0.09515125304460526,
-0.024891706183552742,
0.04117485508322716,
-0.029692640528082848,
0.04481177031993866,
-0.0021059520076960325,
-0.01985812932252884,
0.036031387746334076,
-0.06708449125289917,
-0.05963455140590668,
0.017087332904338837,
-0.0344269834458828,
-0.048888739198446274,
0.03497293218970299,
-0.02885306626558304,
0.0021753248292952776,
0.06330358237028122,
-0.03617289289832115,
-0.03131626546382904,
0.038269512355327606,
0.030699709430336952,
-0.07728132605552673
] | https://github.com/nvlabs/protcomposer | 36 | 0 | 0 | 0 |
More RLHF, More Trust? On The Impact of Preference Alignment On Trustworthiness | https://openreview.net/forum?id=FpiCLJrSW8 | [
"Aaron Jiaxun Li",
"Satyapriya Krishna",
"Himabindu Lakkaraju"
] | Oral | The trustworthiness of Large Language Models (LLMs) refers to the extent to which their outputs are reliable, safe, and ethically aligned, and it has become a crucial consideration alongside their cognitive performance. In practice, Reinforcement Learning From Human Feedback (RLHF) has been widely used to align LLMs with labeled human preferences, but its assumed effect on model trustworthiness hasn't been rigorously evaluated. To bridge this knowledge gap, this study investigates how models aligned with general-purpose preference data perform across five trustworthiness verticals: toxicity, stereotypical bias, machine ethics, truthfulness, and privacy. Our results demonstrate that RLHF on human preferences doesn't automatically guarantee trustworthiness, and reverse effects are often observed. Furthermore, we propose to adapt efficient influence function based data attribution methods to the RLHF setting to better understand the influence of fine-tuning data on individual trustworthiness benchmarks, and show its feasibility by providing our estimated attribution scores. Together, our results underscore the need for more nuanced approaches for model alignment from both the data and framework perspectives, and we hope this research will guide the community towards developing language models that are increasingly capable without sacrificing trustworthiness. | Large Language Model, Trustworthy ML, Data Attribution | Evaluating the Impact of RLHF on Trustworthiness Aspects | 4,767 | 2404.18870 | [
-0.04316132888197899,
-0.1029890924692154,
-0.0456826314330101,
0.012686679139733315,
0.023175986483693123,
0.036136507987976074,
0.0037600737996399403,
0.012651091441512108,
0.05636511370539665,
0.0193984005600214,
0.019530726596713066,
-0.033802956342697144,
0.07725534588098526,
0.016163967549800873,
-0.004141364246606827,
0.004621449392288923,
0.11072051525115967,
0.030285922810435295,
-0.16881154477596283,
-0.0473473034799099,
-0.011806238442659378,
0.03148878738284111,
0.0756756141781807,
0.011702385731041431,
-0.021574217826128006,
-0.0636657103896141,
-0.03166162967681885,
-0.024669626727700233,
-0.03871627897024155,
-0.02211766317486763,
0.011912667192518711,
0.009135989472270012,
0.03335876017808914,
-0.07375475019216537,
-0.0629735067486763,
0.01526659820228815,
-0.0732678472995758,
-0.0531386062502861,
-0.004368187859654427,
-0.044637516140937805,
-0.08861611783504486,
-0.03886529803276062,
0.07301893085241318,
-0.017059560865163803,
0.003754311939701438,
-0.017337990924715996,
0.0667584091424942,
-0.02739567495882511,
-0.05616180971264839,
-0.002690868219360709,
-0.06546587496995926,
-0.030852464959025383,
0.005050431005656719,
0.03901160508394241,
-0.025727946311235428,
-0.020810937508940697,
0.019509900361299515,
0.055471401661634445,
-0.044563665986061096,
-0.018224790692329407,
-0.06815605610609055,
-0.11802452802658081,
-0.08421075344085693,
0.047416891902685165,
-0.07509823143482208,
0.018085408955812454,
-0.04324721917510033,
0.014248775318264961,
-0.11464281380176544,
0.018765291199088097,
0.0036991313099861145,
0.029086105525493622,
0.042813029140233994,
-0.028280064463615417,
-0.0178005900233984,
0.02277511917054653,
0.06240134686231613,
0.020156649872660637,
0.04661582037806511,
-0.050748471170663834,
-0.019320713356137276,
0.0669284537434578,
0.07256921380758286,
0.036794744431972504,
0.02847650833427906,
-0.025249935686588287,
0.05182969570159912,
0.0022850383538752794,
0.07768123596906662,
0.060329314321279526,
0.01972264051437378,
0.03577176108956337,
0.043058864772319794,
0.027269655838608742,
-0.014976625330746174,
0.045537110418081284,
-0.09228691458702087,
-0.08285830914974213,
-0.06001228094100952,
0.03663833439350128,
-0.01647663302719593,
0.10261708498001099,
-0.06284065544605255,
-0.08309265971183777,
0.061822738498449326,
0.01722170040011406,
-0.0668172836303711,
0.04793163761496544,
0.02017737179994583,
-0.03698982298374176,
-0.061332739889621735,
0.07514842599630356,
0.0026290719397366047,
0.04238031432032585,
0.07296136766672134,
-0.03393051400780678,
-0.01179198082536459,
-0.005192573647946119,
0.043036166578531265,
-0.02256241999566555,
0.011329005472362041,
-0.04428340494632721,
0.12675410509109497,
0.013274140655994415,
-0.01851199008524418,
-0.03283621743321419,
-0.02684675343334675,
2.3837434843882534e-33,
-0.0016666892915964127,
0.05398891493678093,
0.04536448419094086,
0.010734331794083118,
0.06442579627037048,
-0.02907075360417366,
-0.015285463072359562,
0.058443259447813034,
-0.0776500478386879,
0.027354717254638672,
0.020149478688836098,
0.06706106662750244,
-0.015186098404228687,
0.037884555757045746,
-0.028548061847686768,
0.05700339749455452,
-0.04448313266038895,
0.08039581775665283,
-0.08384779840707779,
-0.023572415113449097,
0.15181532502174377,
-0.021091856062412262,
0.056179679930210114,
-0.04139500856399536,
0.02573966793715954,
-0.05452670529484749,
0.006682874634861946,
0.01383890863507986,
-0.02293667383491993,
0.017127536237239838,
-0.04851740598678589,
-0.04872346296906471,
0.06317679584026337,
-0.0759975016117096,
-0.005691636353731155,
0.008752838708460331,
-0.07184892147779465,
-0.030136700719594955,
0.04449610039591789,
-0.02299356274306774,
0.032207559794187546,
0.012443198822438717,
0.04268649220466614,
0.08094954490661621,
-0.03217051550745964,
-0.0160489771515131,
0.007045801263302565,
-0.047558803111314774,
-0.008844527415931225,
-0.013796106912195683,
0.02430783025920391,
0.035637032240629196,
0.06540359556674957,
0.006434697192162275,
-0.09981787204742432,
0.025271983817219734,
-0.007863874547183514,
0.05232314020395279,
-0.0725615993142128,
-0.010021021589636803,
0.002077129203826189,
0.03473901376128197,
-0.017209960147738457,
-0.05878923088312149,
0.047582950443029404,
0.028188003227114677,
-0.04885613173246384,
-0.057339880615472794,
0.01653730310499668,
-0.010570118203759193,
0.033744920045137405,
0.06261695176362991,
-0.06074688583612442,
-0.013349190354347229,
-0.012996511533856392,
-0.010374649427831173,
-0.027325091883540154,
0.01324665267020464,
0.13652512431144714,
-0.04334496334195137,
0.010644493624567986,
-0.012314440682530403,
-0.006557420361787081,
-0.02505885250866413,
-0.057696133852005005,
-0.10126993805170059,
0.026769187301397324,
-0.016335414722561836,
0.03483700379729271,
0.04622508957982063,
0.04119778424501419,
-0.03642306476831436,
0.03434223681688309,
0.0326041541993618,
-0.04314214736223221,
-2.8642558203036494e-33,
-0.07908497005701065,
-0.04976702481508255,
0.017644641920924187,
0.10402695089578629,
0.004650390241295099,
-0.04366940259933472,
-0.04642743244767189,
0.07446452230215073,
0.04234936833381653,
-0.00011708303645718843,
-0.016741013154387474,
-0.024105532094836235,
0.06091334670782089,
-0.030333712697029114,
0.08253658562898636,
-0.060311757028102875,
0.05854135751724243,
0.00838639959692955,
-0.012909048236906528,
-0.018929576501250267,
0.03848085179924965,
0.05442584306001663,
-0.12815335392951965,
0.10062825679779053,
-0.011966649442911148,
-0.0056215920485556126,
-0.04547147452831268,
-0.04885503277182579,
0.042049869894981384,
0.03678310662508011,
-0.0541839674115181,
0.055512554943561554,
-0.0878249928355217,
-0.01615409553050995,
-0.04598173126578331,
0.03153838962316513,
0.040875911712646484,
-0.021237550303339958,
-0.029332472011446953,
0.0663614347577095,
0.0037862311583012342,
0.020050952211022377,
-0.11722680181264877,
-0.0755220502614975,
-0.013553286902606487,
-0.03713934123516083,
-0.08354974538087845,
-0.03895752504467964,
0.03341880068182945,
-0.05886048078536987,
0.007696604821830988,
0.01798734813928604,
0.013354307971894741,
0.1115180104970932,
-0.040319304913282394,
-0.037485089153051376,
0.004136348608881235,
-0.03191641718149185,
0.1168651431798935,
-0.0006576480227522552,
-0.07238658517599106,
0.07443731278181076,
-0.024399951100349426,
-0.010936690494418144,
-0.03622624650597572,
0.0019409321248531342,
-0.02356705069541931,
0.013764441013336182,
0.028075536713004112,
-0.007367155514657497,
0.0012771354522556067,
-0.07189667224884033,
0.0058384304866194725,
-0.03508772328495979,
-0.01336260512471199,
-0.030578771606087685,
0.015382064506411552,
-0.05786162242293358,
-0.034610554575920105,
-0.00027656034217216074,
0.13439996540546417,
0.027381746098399162,
0.07383109629154205,
0.011318329721689224,
0.02800334431231022,
-0.007891017012298107,
0.1359681338071823,
0.06438232958316803,
-0.03229691833257675,
0.014192835427820683,
0.008345535956323147,
-0.037118297070264816,
-0.057117000222206116,
0.06643577665090561,
-0.037784576416015625,
-5.038510408894581e-8,
0.01095168013125658,
-0.03018205799162388,
0.028433749452233315,
0.10480795055627823,
0.0031406241469085217,
0.005758910905569792,
0.012694078497588634,
0.014380375854671001,
-0.09534382820129395,
-0.013573127798736095,
0.012882081791758537,
0.0908297449350357,
-0.015385995618999004,
-0.0567251592874527,
-0.023209933191537857,
0.09751683473587036,
-0.022048428654670715,
-0.011616735719144344,
0.03414691239595413,
0.09316606819629669,
0.05320249870419502,
0.010913802310824394,
-0.01450550276786089,
-0.08533891290426254,
0.10695830732584,
-0.033722881227731705,
-0.04874498397111893,
0.07178272306919098,
0.05484261363744736,
0.033918797969818115,
0.03226650133728981,
-0.06595348566770554,
-0.014664862304925919,
0.01596824638545513,
0.030322803184390068,
0.06872361898422241,
-0.020987756550312042,
-0.07526005804538727,
0.0537477545440197,
0.04274873808026314,
0.05282818526029587,
0.14563173055648804,
0.03434817120432854,
0.05701812356710434,
0.020060179755091667,
-0.010559864342212677,
-0.06471607089042664,
-0.08752287924289703,
0.025217898190021515,
-0.05245885252952576,
0.0030480490531772375,
-0.04707875847816467,
-0.055832136422395706,
0.03863310068845749,
-0.01539178378880024,
-0.038565848022699356,
-0.023059029132127762,
-0.02993721514940262,
0.04199867695569992,
0.04611552879214287,
0.12398246675729752,
0.009693902917206287,
0.01532706618309021,
-0.009203960187733173
] | https://github.com/ai4life-group/rlhf_trust | 2 | 0 | 0 | 0 |
Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects | https://openreview.net/forum?id=7BLXhmWvwF | [
"Tai Hoang",
"Huy Le",
"Philipp Becker",
"Vien Anh Ngo",
"Gerhard Neumann"
] | Oral | Manipulating objects with varying geometries and deformable objects is a major challenge in robotics. Tasks such as insertion with different objects or cloth hanging require precise control and effective modelling of complex dynamics. In this work, we frame this problem through the lens of a heterogeneous graph that comprises smaller sub-graphs, such as actuators and objects, accompanied by different edge types describing their interactions. This graph representation serves as a unified structure for both rigid and deformable objects tasks, and can be extended further to tasks comprising multiple actuators. To evaluate this setup, we present a novel and challenging reinforcement learning benchmark, including rigid insertion of diverse objects, as well as rope and cloth manipulation with multiple end-effectors. These tasks present a large search space, as both the initial and target configurations are uniformly sampled in 3D space. To address this issue, we propose a novel graph-based policy model, dubbed Heterogeneous Equivariant Policy (HEPi), utilizing $SE(3)$ equivariant message passing networks as the main backbone to exploit the geometric symmetry. In addition, by modeling explicit heterogeneity, HEPi can outperform Transformer-based and non-heterogeneous equivariant policies in terms of average returns, sample efficiency, and generalization to unseen objects. Our project page is available at https://thobotics.github.io/hepi. | Robotic Manipulation, Equivariance, Graph Neural Networks, Reinforcement Learning, Deformable Objects | Geometry-aware RL with heterogeneous SE(3) equivariant back-bone policy for robotic manipulation | 4,674 | 2502.07005 | [
-0.019357137382030487,
-0.08409924805164337,
-0.007204339373856783,
0.04619244113564491,
0.01913372054696083,
0.0017912450712174177,
0.012615453451871872,
-0.04731065034866333,
-0.05170563980937004,
0.08738025277853012,
0.031080763787031174,
-0.048381268978118896,
-0.0318710058927536,
0.1370375007390976,
0.04716974496841431,
0.06591644138097763,
0.04727383330464363,
0.03809953108429909,
-0.07807440310716629,
0.05245271325111389,
0.09002792090177536,
0.0216983612626791,
0.12007994204759598,
0.02064281888306141,
0.017637280747294426,
0.02142263948917389,
0.012032601051032543,
-0.0751330778002739,
-0.027457235381007195,
-0.051602091640233994,
0.04717206209897995,
-0.09939204156398773,
-0.03432070091366768,
-0.07405365258455276,
-0.07210686802864075,
0.060078784823417664,
-0.038505200296640396,
-0.029320374131202698,
0.00702156126499176,
0.00230620545335114,
-0.021902458742260933,
-0.0037767686881124973,
0.03774745389819145,
0.04458038881421089,
0.09901043027639389,
0.09833480417728424,
-0.03199157863855362,
0.00004932669617119245,
0.0644569918513298,
-0.01352598611265421,
0.02187354862689972,
-0.01904960535466671,
0.0367075577378273,
0.006537780165672302,
0.03627486899495125,
-0.016958672553300858,
0.026234542950987816,
-0.025025561451911926,
0.07577227056026459,
-0.06460370868444443,
0.07934899628162384,
-0.0748077929019928,
0.018163228407502174,
-0.002356606535613537,
-0.08945266157388687,
-0.03433599695563316,
0.044597793370485306,
0.024163924157619476,
-0.0279727503657341,
0.05794085934758186,
-0.03611646592617035,
-0.007391675841063261,
-0.04205363988876343,
0.07527163624763489,
0.06045948341488838,
0.04973767697811127,
0.005405402276664972,
0.024751778692007065,
-0.02210448496043682,
-0.0057006534188985825,
0.005922685842961073,
-0.04519009590148926,
-0.000991090782918036,
0.018225867301225662,
0.014497794210910797,
-0.0250397901982069,
-0.04789559170603752,
0.014949052594602108,
0.06398985534906387,
-0.011847672052681446,
0.06976345181465149,
0.06644811481237411,
0.008625706657767296,
-0.0828961506485939,
0.01558059360831976,
-0.027183663100004196,
-0.03299679979681969,
-0.053637389093637466,
-0.04244523495435715,
0.06829662621021271,
0.024930668994784355,
-0.014488563872873783,
0.013042024336755276,
0.09808126091957092,
0.02620954066514969,
-0.024682240560650826,
-0.07497765123844147,
-0.02734929695725441,
-0.0015407606260851026,
-0.020664285868406296,
-0.010599875822663307,
-0.02315627411007881,
0.034227605909109116,
0.056439612060785294,
-0.035082049667835236,
0.02505965530872345,
-0.02411627396941185,
-0.02542257308959961,
0.05763162672519684,
0.08861130475997925,
-0.04857087880373001,
-0.02777693048119545,
0.005815050099045038,
0.01719236932694912,
0.03936110809445381,
-0.0036616127472370863,
-0.06150152161717415,
-6.411356951848439e-34,
0.08433233946561813,
0.04289553686976433,
0.0291307270526886,
0.04301156848669052,
0.016157882288098335,
0.02633395604789257,
-0.031707558780908585,
0.008385864086449146,
0.020202504470944405,
-0.007727750111371279,
-0.11218766868114471,
-0.04145531356334686,
-0.03118058294057846,
0.07858947664499283,
0.08458865433931351,
0.030788416042923927,
-0.043391186743974686,
0.0008042264962568879,
-0.04031379148364067,
0.05005544796586037,
0.018661092966794968,
-0.04681801795959473,
-0.03069784864783287,
-0.029856057837605476,
0.024246536195278168,
0.05798766016960144,
0.019396737217903137,
0.02284967340528965,
-0.000060219739680178463,
-0.01684456132352352,
-0.011208678595721722,
-0.0017363050719723105,
0.00017691217362880707,
0.12845228612422943,
-0.03001910261809826,
0.014087812043726444,
-0.03410958871245384,
-0.06451043486595154,
0.000617029785644263,
-0.060960687696933746,
0.01293485052883625,
-0.01751120574772358,
0.014437051489949226,
0.03401849791407585,
-0.052196770906448364,
-0.0491887629032135,
0.07754237204790115,
0.030845314264297485,
-0.07941038906574249,
-0.000023689030058449134,
-0.0011658662697300315,
0.10203329473733902,
0.04197661206126213,
-0.07214020192623138,
-0.0008241569739766419,
-0.04805796220898628,
0.01746951788663864,
0.047633834183216095,
0.03527713567018509,
0.022129060700535774,
0.017053350806236267,
0.014730322174727917,
-0.032501257956027985,
0.05699167400598526,
0.02228590101003647,
0.07552408427000046,
-0.006438754498958588,
0.10923285037279129,
0.1055987998843193,
-0.05669926851987839,
-0.05598626285791397,
0.036366693675518036,
-0.0236672330647707,
-0.04476723074913025,
0.037531476467847824,
0.003814027179032564,
-0.03702542930841446,
-0.13016656041145325,
0.018499882891774178,
0.0458579882979393,
-0.0769207775592804,
-0.02223820798099041,
-0.05222799628973007,
0.0053640007972717285,
-0.022254427894949913,
-0.092256098985672,
0.016953524202108383,
-0.07334227114915848,
0.04449409246444702,
-0.036555565893650055,
-0.049516428261995316,
-0.10987622290849686,
-0.03686714917421341,
0.018085990101099014,
0.1104736328125,
-2.3998803126779068e-34,
0.07380761206150055,
-0.025441326200962067,
0.057585541158914566,
0.08486151695251465,
0.026083024218678474,
-0.05130436271429062,
-0.07419848442077637,
-0.09952227771282196,
-0.004534150008112192,
-0.024320052936673164,
-0.031266842037439346,
-0.016188479959964752,
0.01663600094616413,
-0.01099421177059412,
0.03400406613945961,
0.05417336896061897,
-0.08422611653804779,
-0.02779022976756096,
-0.03946760669350624,
-0.012267610989511013,
0.048978887498378754,
0.0006764983991160989,
-0.006123320199549198,
-0.011771218851208687,
-0.0049935937859117985,
0.03764544427394867,
-0.027922336012125015,
-0.03479363024234772,
0.033248912543058395,
0.07735100388526917,
-0.07460138201713562,
-0.08576933294534683,
0.008314825594425201,
0.019896581768989563,
0.03826126456260681,
0.03768276795744896,
0.05821003019809723,
0.08341534435749054,
0.027474476024508476,
-0.010442732833325863,
0.012722005136311054,
-0.06501850485801697,
-0.061825841665267944,
0.08971618115901947,
0.022581536322832108,
-0.0015997200971469283,
-0.10881580412387848,
0.015321739949285984,
-0.11096037179231644,
0.006368121597915888,
-0.12365603446960449,
0.08930148929357529,
-0.048379406332969666,
-0.08281633257865906,
-0.011794312857091427,
-0.06705938279628754,
-0.05707539990544319,
-0.0097427349537611,
0.021937677636742592,
-0.022558795288205147,
-0.09618690609931946,
0.008607464842498302,
-0.04320194572210312,
0.08483066409826279,
-0.022523894906044006,
0.043203260749578476,
0.01255036611109972,
-0.08530806005001068,
0.027377262711524963,
0.051792923361063004,
0.02134358510375023,
-0.07019767165184021,
0.026267901062965393,
-0.054882314056158066,
0.09794220328330994,
-0.026068953797221184,
0.019697027280926704,
-0.006215120665729046,
-0.00544428126886487,
0.009828788228332996,
0.04744892194867134,
-0.02870204485952854,
0.06726765632629395,
0.08888603746891022,
-0.019526217132806778,
0.03188198804855347,
-0.054778099060058594,
0.07087554782629013,
-0.04892576485872269,
0.07476142048835754,
0.000510442245285958,
0.031135739758610725,
0.020948827266693115,
0.08400091528892517,
0.006774225737899542,
-4.5498303791191574e-8,
-0.08741648495197296,
0.03981281444430351,
0.0685601532459259,
-0.03376499190926552,
0.07281745225191116,
0.0866594910621643,
-0.04701948165893555,
-0.005321834236383438,
-0.08354490995407104,
0.048851463943719864,
0.08800134807825089,
0.02166622504591942,
0.0543731264770031,
-0.06155257299542427,
-0.07004572451114655,
0.005087744910269976,
-0.02470960095524788,
-0.023512626066803932,
0.01681577041745186,
0.032929036766290665,
-0.008120933547616005,
-0.03856100142002106,
0.055735498666763306,
-0.038989100605249405,
0.006523448508232832,
-0.08663078397512436,
-0.07839714735746384,
0.05707397684454918,
0.07549039274454117,
0.021938133984804153,
0.016566738486289978,
-0.035821009427309036,
0.04032711684703827,
0.02734634280204773,
-0.0467195101082325,
-0.014046215452253819,
0.037343334406614304,
-0.06939093768596649,
0.012833594344556332,
-0.00042373628821223974,
0.05149274319410324,
0.07199644297361374,
-0.014962618239223957,
-0.013082711026072502,
0.04812365025281906,
0.06876730918884277,
-0.0395110622048378,
-0.09005314856767654,
0.05560826510190964,
-0.05926358699798584,
-0.02606557495892048,
-0.06532438844442368,
-0.06864777952432632,
0.08116748183965683,
0.041836172342300415,
0.04810427501797676,
0.012999783270061016,
-0.043463658541440964,
0.06369143724441528,
0.07199574261903763,
-0.015601816587150097,
-0.06431923061609268,
-0.06643185764551163,
-0.060138460248708725
] | https://github.com/thobotics/geometry_rl | 9 | 0 | 0 | 0 |
Topological Blindspots: Understanding and Extending Topological Deep Learning Through the Lens of Expressivity | https://openreview.net/forum?id=EzjsoomYEb | [
"Yam Eitan",
"Yoav Gelberg",
"Guy Bar-Shalom",
"Fabrizio Frasca",
"Michael M. Bronstein",
"Haggai Maron"
] | Oral | Topological deep learning (TDL) is a rapidly growing field that seeks to leverage topological structure in data and facilitate learning from data supported on topological objects, ranging from molecules to 3D shapes. Most TDL architectures can be unified under the framework of higher-order message-passing (HOMP), which generalizes graph message-passing to higher-order domains. In the first part of the paper, we explore HOMP's expressive power from a topological perspective, demonstrating the framework's inability to capture fundamental topological and metric invariants such as diameter, orientability, planarity, and homology. In addition, we demonstrate HOMP's limitations in fully leveraging lifting and pooling methods on graphs. To the best of our knowledge, this is the first work to study the expressivity of TDL from a topological perspective. In the second part of the paper, we develop two new classes of architectures -- multi-cellular networks (MCN) and scalable MCN (SMCN) -- which draw inspiration from expressive GNNs. MCN can reach full expressivity, but scaling it to large data objects can be computationally expansive. Designed as a more scalable alternative, SMCN still mitigates many of HOMP's expressivity limitations. Finally, we design new benchmarks for evaluating models based on their ability to learn topological properties of complexes. We then evaluate SMCN on these benchmarks as well as on real-world graph datasets, demonstrating improvements over both HOMP baselines and expressive graph methods, highlighting the value of expressively leveraging topological information. | Topological Deep Learning, Message Passing, Higher Order Message Passing, Expressivity, Graph Neural Networks, GNNs, Topology, Homology, Symmetry | null | 4,548 | 2408.05486 | [
-0.054338108748197556,
-0.0626794770359993,
0.04891284927725792,
-0.00712617440149188,
-0.02544691227376461,
-0.02829967439174652,
-0.025471124798059464,
-0.07851418107748032,
0.060184307396411896,
-0.045917168259620667,
-0.015969935804605484,
0.007045205682516098,
-0.010552511550486088,
0.07834295928478241,
0.021239863708615303,
0.029742082580924034,
0.009038828313350677,
0.055710092186927795,
-0.06285068392753601,
-0.026497308164834976,
0.07700187712907791,
0.04438629373908043,
-0.021706504747271538,
0.021163886412978172,
0.04209116846323013,
0.0355687141418457,
-0.013685185462236404,
-0.0072886161506175995,
0.03136439248919487,
-0.012764464132487774,
0.08564620465040207,
0.02392096444964409,
-0.09994301199913025,
-0.027560224756598473,
0.027958711609244347,
0.06584528088569641,
-0.06454716622829437,
0.046695783734321594,
0.012344248592853546,
-0.03738153725862503,
0.06120792031288147,
0.0025770626962184906,
0.0564778670668602,
0.1031261458992958,
0.00622530048713088,
0.049125753343105316,
-0.058516208082437515,
-0.04016760364174843,
0.022163309156894684,
-0.05094083026051521,
-0.02393895573914051,
-0.04375318065285683,
-0.06793980300426483,
0.0546419695019722,
0.04567604139447212,
0.09627454727888107,
-0.029340077191591263,
0.0200321264564991,
0.015200180001556873,
-0.04685094952583313,
0.0621621198952198,
-0.068089060485363,
-0.0037311960477381945,
-0.02724309265613556,
-0.01891002431511879,
-0.027633465826511383,
-0.021907789632678032,
0.08953040093183517,
0.03699683025479317,
0.04524301365017891,
0.0829407125711441,
0.06715518981218338,
-0.056414492428302765,
-0.004613579250872135,
0.08133764564990997,
0.058571942150592804,
-0.012919986620545387,
-0.023678990080952644,
-0.009415497072041035,
-0.0486142635345459,
0.04763771966099739,
0.00039475696394219995,
0.01894650049507618,
0.08176777511835098,
0.045575451105833054,
-0.03490966185927391,
-0.09377360343933105,
0.020900031551718712,
-0.08512744307518005,
0.013029035180807114,
-0.000510231067892164,
0.020254263654351234,
-0.06485720723867416,
-0.07611639052629471,
-0.032782215625047684,
-0.053769223392009735,
-0.028554067015647888,
-0.0899963304400444,
-0.03197892755270004,
0.1285211443901062,
-0.02856951579451561,
-0.0055938102304935455,
0.07067210227251053,
-0.04446906968951225,
0.06505263596773148,
-0.0058876872062683105,
0.05293333902955055,
0.050575390458106995,
0.0004681832797359675,
-0.06491048634052277,
-0.029129397124052048,
0.03270993381738663,
0.05383722484111786,
0.030747387558221817,
-0.02918870747089386,
0.004259789362549782,
-0.026069600135087967,
-0.0748962014913559,
0.10705673694610596,
0.06755556911230087,
-0.0907491147518158,
0.000543507281690836,
-0.037708256393671036,
0.0268419049680233,
0.02284490317106247,
0.03614441305398941,
-0.05101421847939491,
1.1112918799517474e-33,
0.05881339684128761,
0.06057186424732208,
0.07056910544633865,
0.03153139352798462,
0.10139542818069458,
-0.006737487856298685,
-0.028062185272574425,
-0.06869093328714371,
0.008236764930188656,
0.09074603766202927,
-0.07322734594345093,
0.05874066427350044,
-0.017121493816375732,
0.09901615977287292,
0.026624131947755814,
-0.03263384476304054,
0.008736209012567997,
-0.018800966441631317,
-0.005104367155581713,
-0.07736746966838837,
-0.051299870014190674,
-0.05117632821202278,
-0.013989091850817204,
0.026534873992204666,
-0.031323324888944626,
0.012840026058256626,
-0.03600026294589043,
-0.00830104574561119,
0.004385380074381828,
0.010294620878994465,
-0.075741246342659,
0.03609756752848625,
-0.010607401840388775,
0.07806247472763062,
0.015417823567986488,
0.022200265899300575,
-0.007566282991319895,
-0.07190921902656555,
0.015773359686136246,
-0.056608349084854126,
-0.033908259123563766,
0.02890169806778431,
-0.01782548986375332,
0.01939285546541214,
-0.010990376584231853,
0.015601279214024544,
0.03866017237305641,
-0.025762174278497696,
-0.06201568990945816,
-0.07536771148443222,
0.0830690786242485,
-0.04956622049212456,
-0.06445860117673874,
-0.009145508520305157,
0.03991200774908066,
-0.07406605035066605,
0.055359262973070145,
0.06930337101221085,
0.10072676837444305,
0.11159024387598038,
0.048274654895067215,
0.013387993909418583,
-0.03362329676747322,
0.039719901978969574,
-0.017542993649840355,
0.011452286504209042,
-0.027836378663778305,
-0.022965244948863983,
0.1044921725988388,
-0.027221689000725746,
-0.005317357834428549,
0.06038081645965576,
-0.03157287836074829,
-0.06345859169960022,
0.02193506434559822,
0.06012599170207977,
-0.046220146119594574,
-0.10366123169660568,
-0.022788869217038155,
0.07363591343164444,
-0.025385212153196335,
-0.07261289656162262,
-0.047593843191862106,
-0.0013974644243717194,
-0.02991856075823307,
-0.008442222140729427,
0.04695059731602669,
-0.06097633019089699,
-0.02631857804954052,
0.03432760760188103,
-0.10249540954828262,
-0.05458104982972145,
0.05999036133289337,
-0.019310807809233665,
-0.05475287884473801,
-6.0525065363687225e-34,
-0.0330156534910202,
0.02995404042303562,
-0.009026514366269112,
0.007835925556719303,
0.027767108753323555,
-0.04974701628088951,
-0.0589180625975132,
-0.037498101592063904,
-0.0432131253182888,
-0.006921089719980955,
0.0026580800767987967,
0.013375022448599339,
0.052577629685401917,
-0.03210259601473808,
0.04134969040751457,
0.017133383080363274,
-0.022040219977498055,
-0.09073706716299057,
-0.018673913553357124,
0.00254017929546535,
0.015180647373199463,
0.03255705162882805,
-0.08790421485900879,
0.07311467826366425,
-0.02333851531147957,
0.07379695773124695,
-0.04462680593132973,
-0.014469839632511139,
0.08961419761180878,
0.10895377397537231,
-0.0772743672132492,
-0.08338429033756256,
-0.05492839589715004,
0.022500239312648773,
0.049534108489751816,
0.03247305378317833,
0.054924458265304565,
0.04386543110013008,
-0.043086592108011246,
-0.055674292147159576,
0.007425541989505291,
-0.07914582639932632,
-0.10918635129928589,
0.04579446464776993,
0.05287608876824379,
0.03964561969041824,
-0.023894766345620155,
-0.015607809647917747,
-0.08325698226690292,
0.05624007806181908,
-0.04137586057186127,
0.02628827840089798,
-0.027457760646939278,
-0.05038151517510414,
-0.0219613928347826,
0.0419992059469223,
0.008419777266681194,
0.011882742866873741,
0.019230224192142487,
-0.00123429368250072,
0.015193839557468891,
-0.14802667498588562,
-0.03613463044166565,
0.02727000042796135,
-0.07288145273923874,
-0.00800502859055996,
-0.0006923924665898085,
-0.023953553289175034,
-0.09023589640855789,
0.061461787670850754,
-0.03695762902498245,
-0.02977125532925129,
-0.03718968480825424,
-0.035102758556604385,
0.022976255044341087,
-0.019051944836974144,
0.06510284543037415,
0.03187005594372749,
0.005976187065243721,
-0.027154503390192986,
0.012916132807731628,
-0.0070022218860685825,
-0.023936361074447632,
0.09584119915962219,
0.14040543138980865,
0.02152753435075283,
0.031139709055423737,
0.07908500730991364,
-0.016805250197649002,
0.018279381096363068,
0.00863391999155283,
-0.00005624823461403139,
-0.08359479159116745,
0.057680439203977585,
-0.0004548310535028577,
-4.24564348122658e-8,
-0.10494040697813034,
-0.0080510089173913,
0.047066155821084976,
-0.09217116981744766,
0.007835948839783669,
-0.0044613308273255825,
0.10684522241353989,
0.08020848035812378,
-0.002323331544175744,
0.08855371922254562,
0.060587264597415924,
0.01762026734650135,
-0.06633222103118896,
-0.0013254203367978334,
0.004729896783828735,
0.06741634011268616,
0.027557184919714928,
-0.06182088330388069,
0.010888613760471344,
-0.020462244749069214,
-0.056645557284355164,
0.06543175131082535,
-0.0026424629613757133,
0.00601276196539402,
0.047629665583372116,
-0.09174339473247528,
-0.05606972053647041,
0.03964180499315262,
0.005913905333727598,
-0.029341813176870346,
-0.0052218325436115265,
0.028315529227256775,
0.0822698250412941,
0.04406779259443283,
-0.0087948152795434,
0.09902025759220123,
0.02148975431919098,
0.015622126869857311,
-0.03545496240258217,
-0.042845986783504486,
-0.02866237238049507,
0.07895875722169876,
-0.010230046696960926,
-0.024857768788933754,
0.00011591580550884828,
0.031857311725616455,
0.18138223886489868,
0.023930508643388748,
0.010141056030988693,
0.03946608677506447,
-0.06840985268354416,
0.045341409742832184,
-0.02229483798146248,
0.038775499910116196,
0.0038767284713685513,
0.0015646001556888223,
-0.05605446919798851,
-0.08577534556388855,
0.006987936794757843,
0.020256072282791138,
-0.032371945679187775,
-0.01776866987347603,
-0.03239584341645241,
-0.05652299523353577
] | https://github.com/yoavgelberg/smcn | 1 | 0 | 0 | 0 |
Loopy: Taming Audio-Driven Portrait Avatar with Long-Term Motion Dependency | https://openreview.net/forum?id=weM4YBicIP | [
"Jianwen Jiang",
"Chao Liang",
"Jiaqi Yang",
"Gaojie Lin",
"Tianyun Zhong",
"Yanbo Zheng"
] | Oral | With the introduction of video diffusion model, audio-conditioned human video generation has recently achieved significant breakthroughs in both the naturalness of motion and the synthesis of portrait details. Due to the limited control of audio signals in driving human motion, existing methods often add auxiliary spatial signals such as movement regions to stabilize movements, which compromise the naturalness and freedom of motion. To address this issue, we propose an end-to-end audio-only conditioned video diffusion model named Loopy. Specifically, we designed two key modules: an inter- and intra-clip temporal module and an audio-to-latents module. These enable the model to better utilize long-term motion dependencies and establish a stronger audio-portrait movement correlation. Consequently, the model can generate more natural and stable portrait videos with subtle facial expressions, without the need for manually setting movement constraints. Extensive experiments show that Loopy outperforms recent audio-driven portrait diffusion models, delivering more lifelike and high-quality results across various scenarios. Video samples are available at https://loopyavataranony.github.io/ | Diffusion Model, Avatar, Portrait Animation, Audio-Condition Video Generation | We propose Loopy, an end-to-end audio-conditioned video diffusion model that uses long-term motion information to learn natural motions and improve audio-portrait correlation, eliminating motion constraints and delivering high-quality results. | 4,292 | 2409.02634 | [
-0.0068893395364284515,
-0.14306244254112244,
0.009534402750432491,
-0.03648928552865982,
0.0722050666809082,
0.01062743179500103,
-0.06696227192878723,
-0.12546218931674957,
0.028882447630167007,
-0.11003170162439346,
-0.002679259516298771,
-0.049129657447338104,
-0.06645908206701279,
0.034497834742069244,
0.025341708213090897,
-0.032454922795295715,
0.03609028458595276,
0.118142269551754,
-0.012562571093440056,
0.005623222328722477,
0.019314700737595558,
-0.03961861506104469,
0.005868860520422459,
-0.00006765213765902445,
0.027611784636974335,
0.045576609671115875,
0.048629458993673325,
-0.04480434209108353,
0.07909748703241348,
-0.03644264116883278,
0.06674489378929138,
0.02498088777065277,
0.048958003520965576,
-0.029119396582245827,
-0.12690182030200958,
-0.01697501353919506,
-0.03849180415272713,
0.013978991657495499,
-0.06024472042918205,
-0.03896302357316017,
-0.01126453559845686,
0.07712092250585556,
0.0302275363355875,
-0.023075833916664124,
-0.0619766041636467,
-0.03504772484302521,
-0.05442911759018898,
-0.033264245837926865,
-0.03541886806488037,
0.05587068572640419,
-0.026118336245417595,
-0.07241122424602509,
0.022624405100941658,
-0.0056971656158566475,
-0.04529643431305885,
0.017524320632219315,
0.06537578999996185,
0.005276501644402742,
0.022150209173560143,
-0.00956896971911192,
0.031684067100286484,
-0.061070095747709274,
-0.0011820385698229074,
-0.04228205978870392,
-0.007944075390696526,
-0.068818598985672,
0.018398016691207886,
-0.02563924714922905,
0.014536586590111256,
0.06314972043037415,
-0.11483679711818695,
-0.04109816625714302,
-0.05401911586523056,
0.017444033175706863,
-0.011667649261653423,
0.021560436114668846,
0.09554871916770935,
-0.05451576039195061,
0.026624413207173347,
-0.010555673390626907,
0.07009334862232208,
-0.03611500933766365,
0.07456152141094208,
-0.006920379586517811,
0.00409073056653142,
-0.001118244370445609,
0.03672802448272705,
0.032495807856321335,
-0.0592113733291626,
0.027203455567359924,
-0.11790557205677032,
0.02823498658835888,
0.0563388466835022,
-0.0764349177479744,
0.0706401988863945,
0.02945028245449066,
-0.02634495124220848,
-0.06337966024875641,
0.04658679664134979,
0.051771316677331924,
0.02526763640344143,
-0.05648506432771683,
-0.03884897008538246,
0.0644797682762146,
0.07834236323833466,
-0.06828631460666656,
0.05378944054245949,
0.021720634773373604,
-0.02796633541584015,
-0.0009849313646554947,
0.05348296836018562,
-0.0426994264125824,
0.0300062857568264,
-0.027685614302754402,
0.09026890993118286,
0.003133171470835805,
-0.00881260447204113,
0.016319135203957558,
0.002201351337134838,
0.009951387532055378,
0.056419067084789276,
0.007644244469702244,
-0.07099924981594086,
-0.02039289101958275,
0.03348948061466217,
-0.025070665404200554,
-0.00022414824343286455,
-7.970466078044406e-34,
0.04122244194149971,
0.022564180195331573,
0.06840267032384872,
0.1202513724565506,
0.07454937696456909,
-0.014409278519451618,
-0.024493208155035973,
-0.034598201513290405,
-0.02228056639432907,
-0.012976464815437794,
0.051296476274728775,
-0.029210148379206657,
-0.0696806013584137,
0.09608327597379684,
-0.00553798908367753,
-0.024719232693314552,
-0.0190991573035717,
0.0369856096804142,
0.06251935660839081,
0.0886869952082634,
0.024097850546240807,
-0.005056674126535654,
-0.014864697121083736,
0.018036721274256706,
0.007144012488424778,
0.02970716916024685,
0.07240621000528336,
-0.046089161187410355,
0.043396592140197754,
0.00441468320786953,
-0.13276945054531097,
-0.02940772846341133,
-0.019870148971676826,
-0.06114048883318901,
0.06390413641929626,
0.0110814543440938,
-0.010965952649712563,
-0.047748416662216187,
0.027521301060914993,
0.025580890476703644,
-0.0021165774669498205,
0.05164402723312378,
-0.049515753984451294,
-0.02284657582640648,
-0.11985787749290466,
-0.026522282510995865,
0.07315271347761154,
0.15512476861476898,
-0.06546112895011902,
0.0678284615278244,
-0.004075588192790747,
0.040591415017843246,
-0.02517550438642502,
-0.07382617890834808,
-0.09905620664358139,
-0.020392533391714096,
0.11449817568063736,
0.011834358796477318,
-0.013370842672884464,
0.014680927619338036,
0.038905054330825806,
0.04410845413804054,
-0.00619507348164916,
-0.05666652321815491,
0.03536194562911987,
-0.014056899584829807,
-0.03571956977248192,
0.07588481158018112,
-0.0031360455323010683,
0.013461570255458355,
-0.02399665117263794,
0.10073070228099823,
-0.04442950338125229,
-0.06375318765640259,
-0.011750616133213043,
0.013026700355112553,
0.03894634544849396,
-0.03470006585121155,
0.0001357321598334238,
0.016600729897618294,
-0.021514641121029854,
0.04737238213419914,
0.0075544980354607105,
-0.028699694201350212,
0.025787197053432465,
-0.03676537051796913,
-0.042906709015369415,
-0.04379892721772194,
-0.05879034847021103,
-0.04760657995939255,
-0.03453543409705162,
0.0037339176051318645,
0.037518683820962906,
0.08255676180124283,
-0.012488077394664288,
6.121723868177962e-34,
0.04449944570660591,
0.06492642313241959,
0.005475163459777832,
0.022290103137493134,
0.03260164335370064,
0.03956853225827217,
0.10673040896654129,
0.0022468739189207554,
0.0738663300871849,
-0.0716540664434433,
-0.013409730046987534,
-0.061819858849048615,
-0.009364655241370201,
-0.02224087156355381,
0.057868607342243195,
-0.027161242440342903,
0.0073384172283113,
-0.035136327147483826,
-0.08657780289649963,
0.012611452490091324,
-0.0010235427180305123,
0.027388816699385643,
0.034442704170942307,
-0.025253718718886375,
-0.10474710166454315,
0.041401207447052,
-0.02333206869661808,
0.06511043757200241,
0.03959327191114426,
0.021615898236632347,
-0.003466284368187189,
-0.0029772657435387373,
-0.007057040464133024,
-0.06261035054922104,
-0.08898240327835083,
0.03331694006919861,
-0.003982402849942446,
0.009139683097600937,
0.008619590662419796,
0.03476103022694588,
-0.002112198621034622,
0.019086195155978203,
0.021602997556328773,
0.015345723368227482,
0.03590720146894455,
-0.06427986919879913,
-0.05812142416834831,
0.0516507588326931,
-0.04900359734892845,
0.0257702749222517,
-0.013224653899669647,
0.032742083072662354,
-0.038251928985118866,
0.044772472232580185,
-0.08163909614086151,
-0.04593251273036003,
-0.06680544465780258,
0.015100080519914627,
0.10764891654253006,
-0.0074727945029735565,
-0.038381773978471756,
-0.0043761939741671085,
-0.06862347573041916,
-0.11658252030611038,
-0.04877958446741104,
0.05281694605946541,
-0.006500112358480692,
-0.04712117835879326,
0.05069546028971672,
-0.018143387511372566,
0.08826752007007599,
0.006458922289311886,
-0.006001787260174751,
0.023091891780495644,
0.01399166975170374,
-0.02306588925421238,
0.038680609315633774,
-0.033173609524965286,
0.016353705897927284,
-0.07406558841466904,
-0.06003657355904579,
-0.05147429183125496,
0.0423174686729908,
-0.023872334510087967,
0.016977809369564056,
0.12167032808065414,
0.013821310363709927,
-0.01101735420525074,
-0.017622068524360657,
0.13577181100845337,
-0.06883572787046432,
0.059892572462558746,
0.009313656948506832,
0.03637956082820892,
-0.00769271794706583,
-3.7187522394788175e-8,
-0.10081105679273605,
0.029401428997516632,
-0.001151876524090767,
0.027236279100179672,
-0.06410432606935501,
-0.04751378297805786,
0.11585012823343277,
-0.07296255230903625,
0.11290339380502701,
-0.057083941996097565,
0.07613305747509003,
0.023834042251110077,
0.08022404462099075,
0.05388168618083,
-0.08799301087856293,
0.04464774578809738,
-0.04385073482990265,
0.09031936526298523,
-0.024884264916181564,
-0.031261518597602844,
-0.02702619694173336,
0.05490363389253616,
0.01778830774128437,
-0.04524937644600868,
0.04284071549773216,
0.0015741593670099974,
0.007847948931157589,
0.07005006074905396,
0.06463060528039932,
-0.012755083851516247,
-0.0509648397564888,
0.04136635363101959,
-0.044255707412958145,
-0.08334171026945114,
-0.003011047840118408,
0.03662729635834694,
-0.023503614589571953,
-0.035262688994407654,
-0.01901017315685749,
-0.014493280090391636,
-0.0148959094658494,
0.06778768450021744,
-0.03539007529616356,
0.03602980449795723,
-0.04436422884464264,
0.022281305864453316,
0.061296090483665466,
-0.10038100928068161,
0.05719902366399765,
0.006219252943992615,
-0.052223868668079376,
-0.05522115156054497,
0.009112140163779259,
0.09987910836935043,
0.09343326836824417,
0.05620190128684044,
0.02523590251803398,
0.05159197002649307,
0.051589738577604294,
0.026828816160559654,
-0.062357913702726364,
0.018862994387745857,
0.018207019194960594,
-0.022372866049408913
] | 0 | 0 | 0 | 0 |
|
CyberHost: A One-stage Diffusion Framework for Audio-driven Talking Body Generation | https://openreview.net/forum?id=vaEPihQsAA | [
"Gaojie Lin",
"Jianwen Jiang",
"Chao Liang",
"Tianyun Zhong",
"Jiaqi Yang",
"Zerong Zheng",
"Yanbo Zheng"
] | Oral | Diffusion-based video generation technology has advanced significantly, catalyzing a proliferation of research in human animation. While breakthroughs have been made in driving human animation through various modalities for portraits, most of current solutions for human body animation still focus on video-driven methods, leaving audio-driven taking body generation relatively underexplored. In this paper, we introduce CyberHost, a one-stage audio-driven talking body generation framework that addresses common synthesis degradations in half-body animation, including hand integrity, identity consistency, and natural motion.
CyberHost's key designs are twofold. Firstly, the Region Attention Module (RAM) maintains a set of learnable, implicit, identity-agnostic latent features and combines them with identity-specific local visual features to enhance the synthesis of critical local regions. Secondly, the Human-Prior-Guided Conditions introduce more human structural priors into the model, reducing uncertainty in generated motion patterns and thereby improving the stability of the generated videos.
To our knowledge, CyberHost is the first one-stage audio-driven human diffusion model capable of zero-shot video generation for the human body. Extensive experiments demonstrate that CyberHost surpasses previous works in both quantitative and qualitative aspects. CyberHost can also be extended to video-driven and audio-video hybrid-driven scenarios, achieving similarly satisfactory results. | Audio-driven Human Animation.+Diffusion Model.+Generative Model.+Human Video Generation | We propose a one-stage audio-driven talking body generation framework, CyberHost, designed to produce human videos that match the input audio with high expressiveness and realism. | 4,230 | null | [
0.00657408032566309,
-0.06273941695690155,
0.043589383363723755,
-0.03761158511042595,
0.08147399127483368,
-0.00019113808230031282,
0.0056024217046797276,
-0.09960002452135086,
0.06084680184721947,
-0.09795080125331879,
0.0054278913885355,
-0.06134692579507828,
-0.06346379220485687,
0.012346459552645683,
0.020638063549995422,
-0.08929188549518585,
0.07751142978668213,
0.05096054822206497,
-0.001598991802893579,
0.05108267441391945,
0.0327257439494133,
-0.006674458738416433,
0.10395626723766327,
-0.007970603182911873,
0.010161259211599827,
0.016076097264885902,
0.020109057426452637,
-0.052988748997449875,
0.05233404412865639,
-0.0333968810737133,
0.04822402819991112,
0.0036345182452350855,
0.04788427799940109,
-0.013961262069642544,
-0.15533488988876343,
0.04296598210930824,
-0.0007734154351055622,
0.0652618259191513,
-0.11939257383346558,
-0.06410890817642212,
-0.026904860511422157,
0.0457727424800396,
0.026952872052788734,
0.03106464445590973,
0.0539952889084816,
0.01670149341225624,
-0.0459684357047081,
-0.010527282021939754,
0.020971838384866714,
-0.009947703219950199,
-0.04088485613465309,
-0.0869184136390686,
0.0419326014816761,
0.04926541447639465,
0.009240863844752312,
0.06008221581578255,
0.04389161244034767,
0.0057039158418774605,
0.029721755534410477,
0.006812892388552427,
-0.023896876722574234,
-0.04980321601033211,
0.0341288298368454,
-0.03814015910029411,
0.0848463922739029,
-0.06013799086213112,
0.10335119813680649,
-0.06078226491808891,
-0.023369748145341873,
0.034926462918519974,
-0.1266295462846756,
-0.06657937169075012,
-0.02218233235180378,
-0.001758275437168777,
-0.011762299574911594,
0.07509085536003113,
0.03531121462583542,
-0.07041638344526291,
0.002815788146108389,
0.0234494861215353,
0.04899675399065018,
-0.06751856207847595,
0.026705002412199974,
-0.045321203768253326,
-0.04893425852060318,
0.057136714458465576,
0.013787152245640755,
0.030080348253250122,
-0.07307010143995285,
0.04820747673511505,
-0.16692398488521576,
0.020474698394536972,
0.037163589149713516,
-0.10783622413873672,
0.04615165293216705,
-0.011781769804656506,
-0.05544900521636009,
-0.03852149844169617,
0.07378832995891571,
0.005028340965509415,
0.005674360319972038,
-0.061152148991823196,
0.05190715938806534,
0.005160839296877384,
0.09297505766153336,
-0.05019384250044823,
0.02528361603617668,
0.032041437923908234,
0.001732841250486672,
0.038064103573560715,
0.024547617882490158,
-0.04781884700059891,
0.000057407985877944157,
0.03159167245030403,
0.10177348554134369,
-0.03513488546013832,
-0.042629797011613846,
-0.02662116475403309,
0.04418621584773064,
-0.005063313990831375,
0.09324578940868378,
0.0030453642830252647,
-0.06621696054935455,
0.0007233503274619579,
0.031099652871489525,
-0.06508874148130417,
-0.007026239763945341,
5.7368334977626474e-34,
0.02012450061738491,
-0.0098947174847126,
0.09265610575675964,
0.109269879758358,
0.01715809665620327,
-0.03520192205905914,
-0.008065707981586456,
-0.03372902423143387,
-0.04325970634818077,
0.004937571473419666,
0.01785595528781414,
-0.05571199581027031,
-0.043151434510946274,
0.10419802367687225,
0.05065245181322098,
-0.005227858200669289,
-0.034193914383649826,
0.0061193350702524185,
0.04021759331226349,
0.04658402130007744,
0.004500456154346466,
-0.005366076715290546,
0.009113280102610588,
-0.025345467031002045,
0.01916322484612465,
0.02169780619442463,
0.02006904035806656,
-0.04595381021499634,
0.015612483955919743,
-0.015352758578956127,
-0.1883850395679474,
0.017083626240491867,
-0.018710598349571228,
-0.03127022460103035,
0.04466542229056358,
-0.007766999304294586,
0.023007085546851158,
-0.04718255251646042,
0.013645398430526257,
0.019963031634688377,
0.013050487264990807,
0.03669557720422745,
-0.03929330036044121,
-0.04220234602689743,
-0.0731409415602684,
-0.0017318911850452423,
0.09461402893066406,
0.1623917669057846,
-0.03969798982143402,
0.019117150455713272,
-0.015360328368842602,
0.0477529801428318,
0.022072596475481987,
-0.0628117024898529,
-0.07891996949911118,
-0.03258639574050903,
0.07263447344303131,
-0.03243372589349747,
-0.032115910202264786,
-0.0259024016559124,
0.014887410216033459,
0.06265483796596527,
-0.06455182284116745,
0.0033927345648407936,
0.03259354829788208,
-0.0007392101688310504,
-0.0006537418230436742,
0.07237032055854797,
0.018857767805457115,
-0.0159855168312788,
0.01103154569864273,
0.0012866383185610175,
0.006166813429445028,
-0.008241351693868637,
-0.014849659986793995,
0.0009892887901514769,
0.04560208320617676,
-0.0027718592900782824,
0.004794393200427294,
0.047777093946933746,
-0.003838839242234826,
0.05132333189249039,
-0.06471416354179382,
-0.014065477065742016,
0.024522891268134117,
-0.047080885618925095,
0.0000451402593171224,
-0.0363016240298748,
-0.09876664727926254,
-0.05572638660669327,
-0.060534775257110596,
-0.03800349310040474,
-0.0028236026410013437,
0.07547891139984131,
-0.050454214215278625,
-4.3045276104192686e-34,
0.01949550397694111,
0.06683409214019775,
-0.03185802698135376,
0.028542572632431984,
0.03841320425271988,
-0.0013110532891005278,
0.08728767186403275,
0.018441660329699516,
0.05718681961297989,
-0.057974059134721756,
0.00008639154839329422,
-0.06118606775999069,
0.012403723783791065,
-0.05616673082113266,
0.05550694838166237,
-0.03433459252119064,
0.02314269170165062,
-0.014175144955515862,
-0.055890701711177826,
0.05137695372104645,
0.06455745548009872,
0.034563928842544556,
0.0003383489092811942,
0.012073352932929993,
-0.09273175150156021,
0.04233529791235924,
-0.007541281171143055,
0.06736450642347336,
-0.020054088905453682,
-0.010583902709186077,
0.008808169513940811,
0.011340944096446037,
0.00589798716828227,
-0.04499232396483421,
-0.07892752438783646,
0.019032135605812073,
0.002466766629368067,
0.03133879974484444,
0.040639810264110565,
0.016719818115234375,
0.06488288938999176,
0.019846230745315552,
-0.018438799306750298,
0.03323177993297577,
0.009556137956678867,
-0.041522346436977386,
-0.10228797793388367,
0.04077493026852608,
-0.06896182149648666,
0.04779309406876564,
-0.025057444348931313,
0.06890077888965607,
0.0018394511425867677,
-0.07355896383523941,
-0.0893763080239296,
-0.07706113904714584,
-0.01630839891731739,
-0.019292645156383514,
0.16949577629566193,
-0.015756716951727867,
-0.017604650929570198,
0.03416643291711807,
-0.04283417388796806,
-0.11455120891332626,
-0.01977422833442688,
0.08279073238372803,
-0.018713178113102913,
0.020744824782013893,
-0.0028913056012243032,
0.01277196779847145,
0.039609599858522415,
0.01972968690097332,
0.024561380967497826,
0.02188999205827713,
-0.02922830358147621,
-0.021159706637263298,
-0.010295035317540169,
-0.030331896618008614,
0.0009798073442652822,
-0.06654918193817139,
0.010058773681521416,
-0.05844437703490257,
0.03475239872932434,
0.040591586381196976,
0.07050475478172302,
0.12526553869247437,
-0.019231503829360008,
0.0620361790060997,
-0.028389817103743553,
0.09874570369720459,
-0.06375203281641006,
0.07962258160114288,
-0.007713408209383488,
0.07756706327199936,
-0.013214988633990288,
-3.959136307685185e-8,
-0.12559449672698975,
-0.0054708137176930904,
0.044939182698726654,
0.009242549538612366,
-0.03343098983168602,
0.004106225911527872,
0.05067259073257446,
-0.03661399707198143,
0.06929060071706772,
-0.05025468021631241,
-0.017122907564044,
-0.011940041556954384,
0.10433821380138397,
0.0061912862583994865,
-0.05086645111441612,
0.05704395845532417,
-0.06385830789804459,
0.06930536031723022,
-0.06328199058771133,
-0.07591221481561661,
0.01420670934021473,
-0.017585502937436104,
-0.011113069951534271,
-0.031475190073251724,
0.05877685546875,
-0.05564388260245323,
-0.031241878867149353,
0.022087011486291885,
-0.006162299774587154,
-0.02370743826031685,
-0.05596059933304787,
0.027239611372351646,
0.021592942997813225,
-0.018792085349559784,
0.03821106255054474,
-0.0014391272561624646,
0.008408622816205025,
-0.02347782999277115,
-0.01807127334177494,
-0.015348576009273529,
-0.013576535508036613,
0.046053387224674225,
0.0125007638707757,
0.005567575339227915,
-0.043252091854810715,
0.014583989046514034,
0.013061712495982647,
-0.12457703053951263,
0.05937149375677109,
0.00302490359172225,
-0.016409048810601234,
-0.04129122197628021,
-0.01051377784460783,
0.081805020570755,
0.11381091177463531,
0.053791340440511703,
0.0619121678173542,
0.070094034075737,
0.10101879388093948,
0.07072286307811737,
0.004950141068547964,
0.03685399144887924,
-0.013301960192620754,
-0.004773131106048822
] | 0 | 0 | 0 | 0 |
|
Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model | https://openreview.net/forum?id=SI2hI0frk6 | [
"Chunting Zhou",
"LILI YU",
"Arun Babu",
"Kushal Tirumala",
"Michihiro Yasunaga",
"Leonid Shamis",
"Jacob Kahn",
"Xuezhe Ma",
"Luke Zettlemoyer",
"Omer Levy"
] | Oral | We introduce Transfusion, a recipe for training a multi-modal model over discrete and continuous data.
Transfusion combines the language modeling loss function (next token prediction) with diffusion to train a single transformer over mixed-modality sequences.
We pretrain multiple Transfusion models up to 7B parameters from scratch on a mixture of text and image data, establishing scaling laws with respect to a variety of uni- and cross-modal benchmarks.
Our experiments show that Transfusion scales significantly better than quantizing images and training a language model over discrete image tokens.
By introducing modality-specific encoding and decoding layers, we can further improve the performance of Transfusion models, and even compress each image to just 16 patches.
We further demonstrate that scaling our Transfusion recipe to 7B parameters and 2T multi-modal tokens produces a model that can generate images and text on a par with similar scale diffusion models and language models, reaping the benefits of both worlds. | multimodal foundation model, multimodal generation and understanding, diffusion, next token prediction | Transfusion is a recipe for training a multi-modal model over discrete and continuous data. | 4,134 | 2408.11039 | [
-0.03319622576236725,
-0.09518912434577942,
0.07746583968400955,
0.06083105504512787,
0.061525363475084305,
-0.024493765085935593,
-0.06280100345611572,
0.003179040737450123,
0.08700460940599442,
-0.06498916447162628,
-0.021886952221393585,
-0.06730484217405319,
-0.007749909069389105,
0.09976068884134293,
0.017252132296562195,
-0.0006554367137141526,
0.008209370076656342,
0.06852088868618011,
-0.15631559491157532,
-0.08767461776733398,
0.019706977531313896,
-0.0012859029229730368,
0.026574663817882538,
-0.032899159938097,
0.07133971899747849,
0.03759472072124481,
-0.034451160579919815,
-0.08602399379014969,
0.05643444508314133,
-0.023229382932186127,
0.006117044482380152,
-0.029384801164269447,
-0.07137241214513779,
0.030773980543017387,
-0.07043221592903137,
0.049251824617385864,
-0.009113297797739506,
-0.033254195004701614,
0.0021564639173448086,
-0.053493525832891464,
0.01873827911913395,
-0.01034429483115673,
0.011121219955384731,
0.05189850181341171,
0.05453208088874817,
-0.006877434439957142,
0.017053987830877304,
-0.041160695254802704,
0.019322814419865608,
-0.0069018821232020855,
-0.03126172348856926,
-0.004283741116523743,
-0.03604517877101898,
0.08870477229356766,
-0.079145647585392,
-0.04672855883836746,
0.05881921201944351,
-0.04193256050348282,
0.007880243472754955,
-0.03929021582007408,
-0.09192022681236267,
-0.057501308619976044,
-0.016684357076883316,
-0.010604484006762505,
0.020058047026395798,
-0.04130963236093521,
0.0541326068341732,
0.02454770915210247,
-0.0016979059437289834,
-0.03303920105099678,
-0.03794043883681297,
0.030821729451417923,
-0.03468438237905502,
0.06272339075803757,
-0.005791757255792618,
0.014416656456887722,
0.07733945548534393,
-0.017456185072660446,
-0.032022055238485336,
-0.03319912403821945,
0.09801040589809418,
-0.0012386483140289783,
0.1248982846736908,
-0.033090878278017044,
0.07369969040155411,
-0.02917114645242691,
-0.04327641800045967,
0.054303862154483795,
-0.058102916926145554,
0.04253571853041649,
-0.06918185949325562,
0.037555817514657974,
0.0601164884865284,
0.011762553825974464,
0.00006605214002775028,
0.029788155108690262,
-0.03224986419081688,
-0.05263249948620796,
0.00029415119206532836,
0.09749745577573776,
-0.0082444092258811,
0.024963920935988426,
0.00850735604763031,
0.006776029709726572,
0.06410710513591766,
-0.07269790023565292,
0.07334078848361969,
0.00012762298865709454,
0.013213906437158585,
-0.04201705381274223,
0.07517340034246445,
0.065901018679142,
0.024529559537768364,
-0.03626301884651184,
0.05543137714266777,
-0.028050241991877556,
-0.02755606360733509,
0.016522476449608803,
0.031839918345212936,
0.08812639117240906,
-0.04048901051282883,
0.013498492538928986,
-0.09605340659618378,
-0.01927298493683338,
-0.03732824698090553,
-0.04037443548440933,
-0.018748871982097626,
7.662123348741662e-34,
0.035393986850976944,
0.008070358075201511,
0.08351635187864304,
0.142085999250412,
0.046881262212991714,
-0.002341196173802018,
-0.06752435117959976,
-0.057344406843185425,
-0.04759698733687401,
-0.015695424750447273,
-0.04390322044491768,
-0.004273874219506979,
-0.06753280758857727,
0.1239721029996872,
0.008160004392266273,
-0.06929680705070496,
-0.03142061084508896,
-0.00016538357886020094,
0.03744002804160118,
0.0603640191257,
0.006606218405067921,
-0.036540474742650986,
-0.03365165367722511,
0.0190859604626894,
-0.02332524210214615,
0.03599700704216957,
0.0839616134762764,
-0.07028403133153915,
-0.01598249189555645,
0.009653367102146149,
-0.17646999657154083,
-0.009615801274776459,
0.024965200573205948,
-0.013803399167954922,
0.04067569226026535,
-0.021049538627266884,
-0.025710953399538994,
-0.025457706302404404,
-0.013309781439602375,
-0.02579278126358986,
0.00016804743790999055,
0.044583581387996674,
-0.056496601551771164,
0.010999350808560848,
-0.05902493745088577,
-0.0001463608641643077,
0.02525421977043152,
0.013664276339113712,
-0.08169681578874588,
-0.0480027049779892,
0.03672399744391441,
-0.003801462473347783,
-0.0711444765329361,
-0.06605903059244156,
0.019042057916522026,
-0.015745922923088074,
0.057796183973550797,
-0.02720659226179123,
0.06761325150728226,
-0.01826924830675125,
0.05596136674284935,
0.049815889447927475,
-0.02057059481739998,
0.07564445585012436,
0.021329456940293312,
-0.007745305076241493,
-0.03290163353085518,
0.0012807302409783006,
-0.03704296424984932,
0.007673152722418308,
-0.08014973998069763,
0.08815900981426239,
0.019296128302812576,
-0.0487050823867321,
0.0487435944378376,
0.01155416015535593,
0.09651974588632584,
-0.08665492385625839,
0.00843556597828865,
0.04874705523252487,
-0.03575338050723076,
-0.0055632153525948524,
0.010135699063539505,
-0.06562832742929459,
0.015930892899632454,
-0.02997550368309021,
-0.03306421637535095,
-0.06456199288368225,
-0.06280923634767532,
-0.07238287478685379,
-0.03472955524921417,
0.013884356245398521,
0.0006049053044989705,
-0.035034921020269394,
0.020048364996910095,
5.557934163032664e-35,
0.03379661962389946,
0.05343673378229141,
-0.08605767786502838,
0.055323582142591476,
-0.0057341428473591805,
-0.07586831599473953,
0.057134926319122314,
0.07065001875162125,
0.07188402116298676,
-0.011418282054364681,
0.019495589658617973,
0.03510352969169617,
0.02544992044568062,
-0.0577566996216774,
-0.003321768483147025,
-0.016613002866506577,
0.019998909905552864,
0.04747780039906502,
-0.027168033644557,
0.00870909821242094,
0.022481927648186684,
0.09373712539672852,
-0.038048822432756424,
0.06416872143745422,
-0.05657592788338661,
0.04783358797430992,
-0.11595478653907776,
0.05733689293265343,
0.03744839131832123,
-0.04943527281284332,
0.0008969519403763115,
-0.05179800093173981,
0.1062626764178276,
-0.010748207569122314,
-0.08976516127586365,
0.03249037265777588,
0.055773742496967316,
0.008269275538623333,
-0.0438031330704689,
0.06858651340007782,
0.052448470145463943,
-0.06059280037879944,
-0.0740320160984993,
0.06674230843782425,
-0.03217707574367523,
-0.0374615453183651,
-0.08471690118312836,
-0.07026947289705276,
0.030323417857289314,
0.06357720494270325,
-0.028493449091911316,
0.040061675012111664,
-0.10692226141691208,
0.03032023087143898,
0.007961240597069263,
-0.1251993477344513,
-0.08529675006866455,
-0.05089510604739189,
0.06277415156364441,
0.012311439961194992,
-0.03050171211361885,
-0.030227094888687134,
0.03729997202754021,
-0.07670080661773682,
-0.009573575109243393,
-0.06720946729183197,
-0.02538604848086834,
-0.0202846247702837,
-0.022417932748794556,
0.05788462609052658,
0.025424372404813766,
-0.040348440408706665,
-0.006880237255245447,
0.03515997529029846,
0.03393277898430824,
-0.07302119582891464,
0.037638477981090546,
-0.018617304041981697,
-0.021173661574721336,
0.0013017496094107628,
0.004859560169279575,
-0.02607913315296173,
0.013855458237230778,
0.06169617176055908,
0.1338788866996765,
0.08848168700933456,
0.06162438914179802,
0.014517455361783504,
0.04730745777487755,
0.06231274455785751,
-0.05329669266939163,
0.02736148051917553,
0.08111786097288132,
0.010994703508913517,
-0.02129790186882019,
-3.48841489028473e-8,
-0.037543244659900665,
0.01994280517101288,
-0.032207634299993515,
0.0011895589996129274,
-0.039440739899873734,
-0.02254282869398594,
0.0011788394767791033,
0.03618757426738739,
0.019760916009545326,
0.006652611307799816,
0.06941558420658112,
0.0047363406047225,
-0.07263225317001343,
-0.03559914976358414,
-0.010416457429528236,
0.07108906656503677,
0.01613500714302063,
0.01068197377026081,
0.05250626429915428,
-0.007968934252858162,
0.007684141863137484,
0.05017353594303131,
0.053186286240816116,
-0.03441426903009415,
0.06804617494344711,
-0.018549485132098198,
-0.02071678452193737,
0.021617762744426727,
0.028300313279032707,
-0.08409076184034348,
0.007577615324407816,
0.005819603335112333,
0.008663218468427658,
-0.035217393189668655,
-0.11437525600194931,
0.08631065487861633,
0.08466541022062302,
-0.01584540121257305,
-0.02546599507331848,
-0.01618948206305504,
0.04056309163570404,
0.022626347839832306,
-0.09218475967645645,
0.012810169719159603,
0.01591222546994686,
0.0262113306671381,
0.05567814037203789,
-0.02055044285953045,
0.010285858996212482,
-0.028501562774181366,
0.040396444499492645,
-0.008013022132217884,
0.0041991788893938065,
0.1280234307050705,
0.11480213701725006,
-0.008743426762521267,
0.018757285550236702,
0.06182657182216644,
0.1041644960641861,
0.06166697293519974,
0.0011942675337195396,
-0.004158805590122938,
-0.03933684527873993,
-0.07733418047428131
] | 0 | 0 | 0 | 0 |
|
MoE++: Accelerating Mixture-of-Experts Methods with Zero-Computation Experts | https://openreview.net/forum?id=t7P5BUKcYv | [
"Peng Jin",
"Bo Zhu",
"Li Yuan",
"Shuicheng YAN"
] | Oral | In this work, we aim to simultaneously enhance the effectiveness and efficiency of Mixture-of-Experts (MoE) methods. To achieve this, we propose MoE++, a general and heterogeneous MoE framework that integrates both Feed-Forward Network (FFN) and zero-computation experts. Specifically, we introduce three types of zero-computation experts: the zero expert, copy expert, and constant expert, which correspond to discard, skip, and replace operations, respectively. This design offers three key advantages: (i) **Low Computing Overhead**: Unlike the uniform mixing mechanism for all tokens within vanilla MoE, MoE++ allows each token to engage with a dynamic number of FFNs, be adjusted by constant vectors, or even skip the MoE layer entirely. (ii) **High Performance**: By enabling simple tokens to utilize fewer FFN experts, MoE++ allows more experts to focus on challenging tokens, thereby unlocking greater performance potential than vanilla MoE. (iii) **Deployment Friendly**: Given that zero-computation experts have negligible parameters, we can deploy all zero-computation experts on each GPU, eliminating the significant communication overhead and expert load imbalance associated with FFN experts distributed across different GPUs. Moreover, we leverage gating residuals, enabling each token to consider the pathway taken in the previous layer when selecting the appropriate experts. Extensive experimental results demonstrate that MoE++ achieves better performance while delivering 1.1$\sim$2.1$\times$ expert forward throughput compared to a vanilla MoE model of the same size, which lays a solid foundation for developing advanced and efficient MoE-related models. | Mixture of Experts, Large Language Models, Efficient Foundation Models | We propose MoE++, a general and heterogeneous mixture-of-experts framework that achieves better performance while delivering 1.1$\sim$2.1$\times$ expert forward throughput compared to a vanilla MoE model of the same size. | 4,125 | 2410.07348 | [
-0.052420422434806824,
-0.024389786645770073,
0.018427710980176926,
0.049967702478170395,
-0.06355657428503036,
-0.08996792882680893,
-0.034079570323228836,
-0.023230714723467827,
-0.05735665559768677,
-0.06702686846256256,
-0.06663456559181213,
-0.0430426299571991,
-0.04748227074742317,
-0.032579757273197174,
-0.031213123351335526,
0.019787991419434547,
0.09766858071088791,
0.06088855117559433,
-0.0707128494977951,
-0.056423064321279526,
-0.07268606871366501,
-0.11107660084962845,
-0.029210451990365982,
0.017251772806048393,
-0.02570902369916439,
0.03745274245738983,
-0.08158638328313828,
-0.10345476120710373,
0.0929739847779274,
0.001532270573079586,
0.05715829133987427,
0.01063547283411026,
0.0384681262075901,
0.027015211060643196,
-0.05852783843874931,
0.03827479109168053,
0.017553051933646202,
0.026394372805953026,
-0.019520418718457222,
-0.04094257205724716,
0.026858054101467133,
0.013081256300210953,
-0.03223836049437523,
0.05261342599987984,
0.023236656561493874,
0.05781403183937073,
0.028330089524388313,
-0.01787336729466915,
0.026399625465273857,
-0.04203074797987938,
-0.022481044754385948,
-0.04186895862221718,
-0.039599910378456116,
0.018215548247098923,
0.0009875298710539937,
0.041837066411972046,
0.050156380981206894,
0.013495034538209438,
-0.0013625076971948147,
0.06596525758504868,
-0.08858992159366608,
-0.08029262721538544,
-0.009002147242426872,
0.043669961392879486,
0.06872078776359558,
0.0100028607994318,
0.045923810452222824,
-0.00526386359706521,
0.027527417987585068,
0.03006628341972828,
-0.01093758549541235,
0.07635430991649628,
-0.06506448239088058,
0.004808470606803894,
-0.048162151128053665,
0.05721471086144447,
0.015019279904663563,
0.03752053156495094,
0.002089274814352393,
-0.027298826724290848,
0.047339800745248795,
-0.01730397343635559,
-0.0074536451138556,
-0.05707555636763573,
0.10922383517026901,
-0.033480558544397354,
-0.047376446425914764,
0.008810908533632755,
0.09321162104606628,
0.0037777884863317013,
-0.09127183258533478,
0.046223510056734085,
0.07933618873357773,
-0.04752655327320099,
0.04753461852669716,
0.06478668004274368,
-0.00023371944553218782,
-0.02849655970931053,
-0.08530005812644958,
0.12222506105899811,
-0.04621818661689758,
-0.014622078277170658,
0.07248887419700623,
0.017759114503860474,
0.015141618438065052,
0.07827892154455185,
-0.06800580024719238,
0.04869399219751358,
0.013306016102433205,
-0.001118483836762607,
0.040359899401664734,
0.06964296102523804,
0.05184045061469078,
-0.07301423698663712,
-0.06819116324186325,
-0.009806565940380096,
-0.06905442476272583,
-0.01897806115448475,
0.027247454971075058,
0.11576665192842484,
-0.07816821336746216,
-0.012970115058124065,
0.034570854157209396,
0.04426736384630203,
-0.014226004481315613,
0.05150463432073593,
-0.0827142596244812,
6.5475646046850865e-34,
0.004892450291663408,
0.021755781024694443,
-0.005005067214369774,
-0.05747209116816521,
0.004533002153038979,
0.03157142549753189,
0.06278571486473083,
0.0031919609755277634,
-0.023104360327124596,
-0.061180002987384796,
0.036686010658741,
0.05283128097653389,
-0.02425875887274742,
0.07447829842567444,
0.09199284017086029,
-0.06359301507472992,
0.03798986226320267,
0.05500958114862442,
0.03320033475756645,
-0.03180081397294998,
0.07560443878173828,
-0.019768930971622467,
-0.06698407977819443,
-0.014932000078260899,
0.013820597901940346,
0.04451942443847656,
0.019000841304659843,
-0.08859134465456009,
-0.015290005132555962,
0.035396791994571686,
-0.06523054093122482,
-0.05951914191246033,
-0.00274081202223897,
0.04001385346055031,
0.05479520186781883,
-0.02952483855187893,
-0.0017164662713184953,
-0.05466560646891594,
0.015502279624342918,
-0.000584460562095046,
-0.07475887984037399,
-0.003964925184845924,
-0.013293433003127575,
-0.07845905423164368,
-0.122186079621315,
-0.05382992699742317,
0.02053029276430607,
0.03838478401303291,
0.027164112776517868,
0.025177499279379845,
0.05030583590269089,
0.0032038588542491198,
-0.04661690071225166,
-0.01388995349407196,
0.01258057076483965,
-0.09115621447563171,
0.030029011890292168,
0.0009643524535931647,
0.08240779489278793,
0.06319436430931091,
-0.06950471550226212,
0.0626673698425293,
-0.0682365819811821,
0.07074352353811264,
-0.009493930265307426,
0.013226890005171299,
-0.016839394345879555,
0.050779033452272415,
0.06812512129545212,
-0.00240521552041173,
-0.053497232496738434,
0.06231677904725075,
0.02842537686228752,
-0.01173068955540657,
0.03841574862599373,
0.005443891044706106,
0.034057918936014175,
-0.07908640801906586,
-0.04090522229671478,
0.02163294143974781,
-0.0038294384721666574,
0.12059709429740906,
-0.03289654105901718,
0.01837829500436783,
0.014014404267072678,
-0.05597073212265968,
0.0016793509712442756,
0.04583613947033882,
-0.0830358937382698,
-0.08541207760572433,
-0.02564803883433342,
-0.08440884202718735,
0.018645809963345528,
0.040183693170547485,
-0.051042575389146805,
-7.894504347397324e-34,
-0.013126333244144917,
-0.004162894561886787,
-0.0149083212018013,
0.10466104000806808,
-0.024512477219104767,
-0.03512609004974365,
0.028981739655137062,
-0.09709009528160095,
-0.004719399381428957,
-0.050876569002866745,
-0.07961975038051605,
-0.019168414175510406,
0.06299398839473724,
-0.006302159745246172,
0.01447724737226963,
-0.019085422158241272,
-0.15374119579792023,
0.022206684574484825,
0.014213931746780872,
0.025282779708504677,
0.01933368481695652,
0.042536210268735886,
-0.10025990009307861,
0.06979447603225708,
-0.007543462328612804,
0.040839556604623795,
-0.06284678727388382,
0.09271172434091568,
0.035412099212408066,
0.01206591259688139,
-0.000708544161170721,
0.029224621132016182,
-0.017596757039427757,
-0.10196810215711594,
0.06676578521728516,
0.08101005107164383,
0.045877452939748764,
0.1585656851530075,
0.02165800705552101,
-0.02104169875383377,
0.09669119119644165,
-0.040921300649642944,
-0.03827971965074539,
0.03462143614888191,
-0.00582889886572957,
0.08273276686668396,
-0.07594458013772964,
-0.03721633553504944,
-0.015740348026156425,
-0.011887925677001476,
-0.05446719005703926,
0.021051298826932907,
-0.05909934639930725,
-0.015901541337370872,
-0.05114053562283516,
-0.053884394466876984,
0.026610927656292915,
0.067050501704216,
0.026096396148204803,
-0.0046568820253014565,
0.02940448932349682,
-0.09106122702360153,
0.09820999950170517,
-0.0028511385899037123,
0.0040734377689659595,
0.04330281913280487,
0.012765890918672085,
0.012892262078821659,
-0.026628781110048294,
0.03665754944086075,
0.0070602926425635815,
0.004074657801538706,
-0.0012696565827354789,
-0.006486136931926012,
-0.016173982992768288,
0.061546213924884796,
0.0476604700088501,
-0.07078694552183151,
0.028418418020009995,
0.047332506626844406,
-0.01826339215040207,
0.06810905784368515,
0.011154097504913807,
0.0025082684587687254,
0.07768483459949493,
0.010336276143789291,
0.03278857469558716,
0.015247822739183903,
0.017734071239829063,
-0.005821090191602707,
-0.03777085989713669,
0.07417786121368408,
0.025953877717256546,
0.014191114343702793,
-0.06175621226429939,
-4.2881765693891793e-8,
0.014807001687586308,
-0.02868383377790451,
0.009099009446799755,
0.06586755067110062,
0.08854358643293381,
-0.0026417854242026806,
-0.06225300952792168,
-0.036040760576725006,
-0.039757952094078064,
-0.04329301789402962,
0.08305152505636215,
-0.0396508127450943,
-0.09767021238803864,
-0.06233019754290581,
0.039237286895513535,
0.04186352714896202,
0.014054588973522186,
-0.07256507873535156,
-0.0551716648042202,
-0.0800606980919838,
-0.03452351689338684,
0.031763847917318344,
-0.013958464376628399,
-0.08572623878717422,
-0.022553585469722748,
-0.09276672452688217,
0.08541680127382278,
0.061525456607341766,
0.09072752296924591,
0.014280276373028755,
-0.04006750509142876,
0.0029707802459597588,
0.031010692939162254,
0.014651704579591751,
0.025258982554078102,
0.01825353316962719,
0.032743848860263824,
0.06030577793717384,
-0.03773835301399231,
-0.05582193285226822,
-0.06779089570045471,
0.00834646075963974,
0.0069480957463383675,
-0.01433063205331564,
0.024951769039034843,
-0.04338670149445534,
-0.056935377418994904,
-0.020228775218129158,
0.035226646810770035,
0.054055895656347275,
0.003504793159663677,
0.008945833891630173,
-0.09241835027933121,
0.10663428902626038,
0.02875078283250332,
0.004374288953840733,
-0.012381467968225479,
-0.05124184116721153,
0.04475126415491104,
0.08760828524827957,
0.019336825236678123,
-0.06160086393356323,
-0.01632002182304859,
-0.0011446673888713121
] | https://github.com/skyworkai/moe-plus-plus | 211 | 0 | 0 | 0 |
Compositional Entailment Learning for Hyperbolic Vision-Language Models | https://openreview.net/forum?id=3i13Gev2hV | [
"Avik Pal",
"Max van Spengler",
"Guido Maria D'Amely di Melendugno",
"Alessandro Flaborea",
"Fabio Galasso",
"Pascal Mettes"
] | Oral | Image-text representation learning forms a cornerstone in vision-language models, where pairs of images and textual descriptions are contrastively aligned in a shared embedding space. Since visual and textual concepts are naturally hierarchical, recent work has shown that hyperbolic space can serve as a high-potential manifold to learn vision-language representation with strong downstream performance. In this work, for the first time we show how to fully leverage the innate hierarchical nature of hyperbolic embeddings by looking beyond individual image-text pairs. We propose Compositional Entailment Learning for hyperbolic vision-language models. The idea is that an image is not only described by a sentence but is itself a composition of multiple object boxes, each with their own textual description. Such information can be obtained freely by extracting nouns from sentences and using openly available localized grounding models. We show how to hierarchically organize images, image boxes, and their textual descriptions through contrastive and entailment-based objectives. Empirical evaluation on a hyperbolic vision-language model trained with millions of image-text pairs shows that the proposed compositional learning approach outperforms conventional Euclidean CLIP learning, as well as recent hyperbolic alternatives, with better zero-shot and retrieval generalization and clearly stronger hierarchical performance. | Vision-Language Models, Hyperbolic Geometry, Representation Learning, CLIP | We explore the benefits brought in when using visual-semantic compositional hierarchies for learning hyperbolic representations through unsupervised contrastive training. | 4,111 | 2410.06912 | [
0.03150210157036781,
-0.06907740235328674,
0.030561033636331558,
-0.01958661712706089,
0.09670105576515198,
-0.013591548427939415,
-0.035270098596811295,
-0.038237519562244415,
0.10382423549890518,
-0.051346298307180405,
0.030105436220765114,
-0.0485326386988163,
0.0381954126060009,
0.12071983516216278,
0.013421441428363323,
0.0383932963013649,
0.07418788224458694,
0.057594675570726395,
-0.08722543716430664,
-0.06721879541873932,
0.061904191970825195,
0.024417180567979813,
0.049115050584077835,
-0.013943305239081383,
0.024317869916558266,
0.07908929139375687,
-0.031952161341905594,
-0.033330246806144714,
-0.01295777503401041,
-0.01566356047987938,
0.013956569135189056,
-0.031521692872047424,
-0.015545887872576714,
0.0627334713935852,
-0.014857702888548374,
0.17333774268627167,
-0.01760481484234333,
0.09768890589475632,
-0.026141639798879623,
-0.005782176740467548,
-0.07030941545963287,
-0.03603321686387062,
-0.00910305604338646,
-0.017953183501958847,
0.10226192325353622,
-0.014131653122603893,
0.009147911332547665,
-0.04341580346226692,
0.013464540243148804,
0.03726780787110329,
-0.09776344150304794,
-0.024506358429789543,
-0.04601294919848442,
0.05405845493078232,
0.003459502477198839,
0.012941870838403702,
0.013363657519221306,
-0.010801011696457863,
0.005488629452884197,
-0.0645781010389328,
0.008214286528527737,
-0.050968050956726074,
-0.06743317097425461,
0.015171147882938385,
-0.0016972478479146957,
-0.03497547656297684,
0.027293261140584946,
0.01550359558314085,
-0.10436266660690308,
0.034414611756801605,
-0.04895325377583504,
0.06834720075130463,
-0.030251583084464073,
-0.014582003466784954,
-0.060918550938367844,
0.07018040120601654,
0.05637864023447037,
-0.12045282125473022,
0.05103161931037903,
-0.12432682514190674,
0.051434971392154694,
0.08485402166843414,
0.09892336279153824,
0.021393753588199615,
0.0865773931145668,
0.052828773856163025,
0.012528548948466778,
0.05561336129903793,
-0.015457767993211746,
-0.05626611039042473,
-0.04368561878800392,
-0.08792023360729218,
0.07554033398628235,
-0.006715658586472273,
0.0765899270772934,
-0.009078335016965866,
-0.06934726238250732,
-0.0706266239285469,
-0.009286417625844479,
0.09341022372245789,
0.023858273401856422,
0.03140929713845253,
0.0325033962726593,
-0.03727220371365547,
-0.014501324854791164,
0.0012692210730165243,
-0.014864192344248295,
0.003045350778847933,
-0.008284366689622402,
-0.08701588958501816,
0.03265686705708504,
0.03644577041268349,
-0.014406159520149231,
0.033915840089321136,
0.02321593463420868,
-0.011698824353516102,
0.021263986825942993,
-0.09472769498825073,
0.03329947218298912,
-0.0018128957599401474,
-0.016374630853533745,
0.047613706439733505,
0.025901371613144875,
0.06130211055278778,
0.03705906867980957,
-0.08305436372756958,
0.026243338361382484,
-8.766633911986111e-34,
0.0625046044588089,
0.07984697073698044,
0.07242702692747116,
0.03681106120347977,
0.07752541452646255,
-0.004360843915492296,
-0.041790205985307693,
0.06479453295469284,
-0.035194188356399536,
-0.024090243503451347,
-0.021672168746590614,
-0.009930746629834175,
-0.0011855510529130697,
0.10401217639446259,
0.0010689969640225172,
-0.007437772583216429,
-0.03697342425584793,
0.041763897985219955,
-0.007011848967522383,
0.03556392341852188,
-0.04344785585999489,
0.014246365055441856,
-0.015829909592866898,
-0.051128752529621124,
-0.04130367934703827,
-0.013746737502515316,
0.08359503000974655,
-0.09425674378871918,
-0.07367502152919769,
0.002800796879455447,
-0.1098163053393364,
-0.028041759505867958,
0.013660130091011524,
0.029458824545145035,
0.03055788204073906,
-0.03510965779423714,
-0.02069859579205513,
0.0040642512030899525,
0.05035226792097092,
-0.043919067829847336,
-0.002691212808713317,
0.024590544402599335,
-0.00903589092195034,
-0.011896060779690742,
-0.06954913586378098,
0.009455460123717785,
-0.006260388996452093,
0.0679505318403244,
-0.08998320251703262,
-0.07625049352645874,
0.07272342592477798,
0.031635988503694534,
-0.01359780877828598,
-0.07240412384271622,
0.007215622812509537,
0.024106556549668312,
-0.0010505610844120383,
0.02837495505809784,
0.025703035295009613,
-0.016704708337783813,
0.03305627778172493,
0.06150282546877861,
-0.012098944745957851,
0.09226257354021072,
-0.02645556814968586,
-0.000871671421919018,
-0.01139597687870264,
0.023661069571971893,
0.048150788992643356,
0.021839337423443794,
-0.032025568187236786,
0.03203181177377701,
-0.030287226662039757,
-0.034262992441654205,
0.06272284686565399,
0.017312388867139816,
0.03643525391817093,
-0.09755820035934448,
-0.007761909626424313,
0.07465202361345291,
-0.069903165102005,
-0.06106777489185333,
0.039506763219833374,
-0.11771353334188461,
-0.07957487553358078,
-0.024178029969334602,
0.07666855305433273,
-0.1072157621383667,
0.006536728236824274,
-0.06670932471752167,
-0.11548632383346558,
-0.05906946212053299,
-0.07464343309402466,
0.007119466084986925,
0.09938670694828033,
-4.7400450531565605e-34,
0.005919867195188999,
0.04100998118519783,
-0.009199528023600578,
-0.009695605374872684,
-0.014139018021523952,
-0.0014562869910150766,
0.01595364883542061,
0.03222232684493065,
-0.05522865429520607,
-0.04547601938247681,
-0.055704131722450256,
0.053066231310367584,
-0.017844775691628456,
-0.0497053824365139,
0.033882737159729004,
-0.03710826486349106,
0.019765838980674744,
0.041620612144470215,
0.025289813056588173,
0.02863827534019947,
0.015114253386855125,
0.1023898720741272,
-0.0760052502155304,
0.07991715520620346,
-0.02869832143187523,
0.05397406592965126,
0.022263435646891594,
0.05246521160006523,
-0.02206462435424328,
0.026878975331783295,
-0.04068051651120186,
-0.058322321623563766,
-0.013062792830169201,
-0.022061621770262718,
-0.106272853910923,
0.09698550403118134,
0.006304759532213211,
-0.01293987687677145,
-0.0761352926492691,
0.05727541446685791,
0.015789968892931938,
-0.03640154376626015,
-0.07180453091859818,
0.01666228286921978,
-0.0807824358344078,
-0.05074697732925415,
-0.031986478716135025,
-0.03225156292319298,
-0.01684523932635784,
0.03479393199086189,
-0.09734894335269928,
-0.0605589896440506,
-0.052148494869470596,
0.0174233578145504,
-0.0629916563630104,
-0.04527567699551582,
-0.09700538218021393,
0.0040595694445073605,
0.019768087193369865,
-0.04585055634379387,
-0.009471191093325615,
0.01484934613108635,
-0.053505778312683105,
0.00593511201441288,
0.022483965381979942,
-0.09934861212968826,
-0.07049520313739777,
-0.03896750137209892,
0.00911271758377552,
0.03191836550831795,
0.0015178847825154662,
-0.05873735621571541,
-0.018488174304366112,
0.06524616479873657,
0.04037776589393616,
-0.020757466554641724,
0.05004285275936127,
0.028599325567483902,
-0.026850352063775063,
-0.050266217440366745,
0.04102116450667381,
-0.05430569499731064,
0.03154758736491203,
0.08743011951446533,
0.11672725528478622,
0.041644349694252014,
0.03475021570920944,
0.023402933031320572,
0.049805715680122375,
0.0169757679104805,
-0.049305811524391174,
-0.002575090155005455,
-0.014033585786819458,
0.09537751227617264,
-0.027104206383228302,
-3.277396132261856e-8,
-0.021483169868588448,
-0.0026526721194386482,
0.04160451889038086,
-0.03844042494893074,
-0.06983821094036102,
-0.04044622182846069,
0.005405177362263203,
0.028957560658454895,
-0.04114591330289841,
-0.08225986361503601,
0.039520926773548126,
0.031013764441013336,
-0.04993017017841339,
-0.0798441618680954,
-0.09829385578632355,
0.07823982834815979,
0.02891029417514801,
0.029066234827041626,
0.03635524958372116,
0.057705171406269073,
0.05903877317905426,
0.03694883733987808,
0.04583898186683655,
-0.0181292537599802,
-0.04189946874976158,
-0.03236083686351776,
-0.038310784846544266,
0.01074787974357605,
0.049040645360946655,
-0.014156284742057323,
0.038098063319921494,
0.08173703402280807,
0.018712567165493965,
-0.012791489250957966,
0.0021215288434177637,
0.022020390257239342,
-0.052484605461359024,
-0.0014930178876966238,
0.01964404061436653,
-0.07387170940637589,
0.08141179382801056,
0.03518809750676155,
0.006187103688716888,
0.009574799798429012,
0.09688020497560501,
0.03117981180548668,
0.0018630338599905372,
-0.08780061453580856,
0.0649670660495758,
0.030059637501835823,
-0.018811123445630074,
-0.015449426136910915,
0.004599050618708134,
0.06624061614274979,
0.010338549502193928,
-0.06086291745305061,
0.01180626917630434,
-0.030695682391524315,
0.07260189950466156,
0.03095652349293232,
0.011191188357770443,
0.04098793864250183,
-0.013476704247295856,
-0.06526827812194824
] | https://github.com/PalAvik/hycoclip | 60 | 0 | 0 | 0 |
Advantage Alignment Algorithms | https://openreview.net/forum?id=QFO1asgas2 | [
"Juan Agustin Duque",
"Milad Aghajohari",
"Tim Cooijmans",
"razvan ciuca",
"Tianyu Zhang",
"Gauthier Gidel",
"Aaron Courville"
] | Oral | Artificially intelligent agents are increasingly being integrated into human decision-making: from large language model (LLM) assistants to autonomous vehicles. These systems often optimize their individual objective, leading to conflicts, particularly in general-sum games where naive reinforcement learning agents empirically converge to Pareto-suboptimal Nash equilibria. To address this issue, opponent shaping has emerged as a paradigm for finding socially beneficial equilibria in general-sum games. In this work, we introduce Advantage Alignment, a family of algorithms derived from first principles that perform opponent shaping efficiently and intuitively. We achieve this by aligning the advantages of interacting agents, increasing the probability of mutually beneficial actions when their interaction has been positive. We prove that existing opponent shaping methods implicitly perform Advantage Alignment. Compared to these methods, Advantage Alignment simplifies the mathematical formulation of opponent shaping, reduces the computational burden and extends to continuous action domains. We demonstrate the effectiveness of our algorithms across a range of social dilemmas, achieving state-of-the-art cooperation and robustness against exploitation. | Multi-agent Reinforcement Learning, Opponent Shaping, Social Dilemmas, General-Sum Games | We introduce Advantage Alignment, a new family of algorithms for opponent shaping in general-sum games, designed to promote cooperation and avoid suboptimal outcomes. | 3,875 | 2406.14662 | [
-0.0661979541182518,
-0.04633864015340805,
-0.07316466420888901,
-0.0704231932759285,
-0.018084438517689705,
0.02448619157075882,
0.06816036999225616,
0.0029523149132728577,
0.0009462089510634542,
0.06433427333831787,
-0.047726936638355255,
-0.04013870656490326,
0.047812215983867645,
0.06507711857557297,
-0.010432345792651176,
0.04195259511470795,
0.16537965834140778,
0.08752718567848206,
-0.047214262187480927,
0.006742536555975676,
-0.029995277523994446,
-0.0027050767093896866,
0.04005828872323036,
-0.03115474432706833,
-0.01428200863301754,
-0.03417019173502922,
-0.005870550870895386,
-0.02378702349960804,
0.011472117155790329,
-0.03392576053738594,
0.05998808145523071,
-0.061971671879291534,
0.0900903195142746,
-0.023038726300001144,
-0.07482894510030746,
0.047934725880622864,
-0.15452377498149872,
-0.05296240746974945,
-0.03760617598891258,
-0.07531620562076569,
-0.013377499766647816,
-0.05165615305304527,
0.012827827595174313,
-0.05770963430404663,
0.005717300344258547,
0.06011568382382393,
-0.007245450746268034,
0.018433580175042152,
0.03982196003198624,
-0.03556084632873535,
-0.043145474046468735,
0.011999787762761116,
0.010571707040071487,
0.06784643232822418,
0.016282791271805763,
0.03531254082918167,
0.0001567465951666236,
0.07196913659572601,
0.00013523387315217406,
-0.028497187420725822,
0.0212448388338089,
-0.12889617681503296,
0.016247255727648735,
-0.0008927215822041035,
-0.045876938849687576,
-0.041259441524744034,
0.011945141479372978,
0.09218447655439377,
-0.09358055144548416,
0.062434375286102295,
0.04125054553151131,
0.03899264335632324,
0.04215746000409126,
-0.0052903457544744015,
0.036776281893253326,
0.07498550415039062,
0.012039321474730968,
-0.03800239413976669,
0.05579986795783043,
0.00933215580880642,
-0.017493611201643944,
0.0077146016992628574,
0.014774797484278679,
0.10533302277326584,
0.09096064418554306,
-0.07193878293037415,
-0.019307544454932213,
0.01681143045425415,
0.18350830674171448,
0.028473539277911186,
-0.07017830014228821,
0.017425457015633583,
0.09861859679222107,
-0.004955834709107876,
0.03198888897895813,
0.06428850442171097,
-0.05937249958515167,
-0.0827084630727768,
-0.051994286477565765,
0.1187608391046524,
-0.02440408617258072,
0.07705000042915344,
-0.037146519869565964,
-0.05616490915417671,
0.05574069544672966,
-0.005022757686674595,
-0.04735490679740906,
0.008670573122799397,
0.038749806582927704,
-0.06275109946727753,
0.002431392204016447,
-0.005821216385811567,
0.05111251771450043,
0.09475727379322052,
0.014522056095302105,
0.0022458552848547697,
0.06077047437429428,
0.050824861973524094,
-0.02194674126803875,
0.005872610490769148,
0.02342497743666172,
-0.012144679203629494,
-0.007463201880455017,
0.05927367880940437,
0.057930078357458115,
-0.0028850985690951347,
0.009977729059755802,
-6.698350224181854e-34,
0.038325682282447815,
0.02879968471825123,
0.08222053200006485,
0.05885724723339081,
0.018248779699206352,
-0.0310377087444067,
0.010213767178356647,
-0.011569352820515633,
-0.08787042647600174,
0.013965235091745853,
-0.03170191869139671,
0.06259015202522278,
0.010820357128977776,
0.07635966688394547,
0.01796688325703144,
-0.03352201357483864,
-0.11324217170476913,
0.038601264357566833,
-0.05714845657348633,
-0.03965463861823082,
0.08893992751836777,
0.007688477169722319,
0.006762044504284859,
-0.03519552946090698,
0.019530514255166054,
0.0039531695656478405,
-0.015354093164205551,
-0.04450669884681702,
0.03317643702030182,
-0.025864923372864723,
-0.08458344638347626,
0.031631819903850555,
-0.06918203830718994,
0.04134851694107056,
0.07233577221632004,
0.05406571924686432,
-0.10836824774742126,
-0.043146904557943344,
-0.07253340631723404,
0.06353872269392014,
-0.01703496277332306,
0.045610781759023666,
0.005752686411142349,
-0.03318857401609421,
-0.06892841309309006,
-0.053030453622341156,
-0.01767251454293728,
0.03309706225991249,
-0.08952982723712921,
-0.0734056830406189,
-0.018024444580078125,
-0.007139833644032478,
0.03399866446852684,
0.00294074066914618,
-0.06304922699928284,
0.016871903091669083,
0.007451292593032122,
0.06652095168828964,
-0.05740318074822426,
-0.027693353593349457,
-0.004501409363001585,
-0.059407610446214676,
0.04184846207499504,
0.04076746106147766,
-0.005531147588044405,
0.010519570671021938,
-0.05939739570021629,
0.06923488527536392,
0.06012261286377907,
-0.03724585101008415,
0.06677659600973129,
0.04679327458143234,
-0.014541840180754662,
-0.03163957595825195,
0.021549146622419357,
0.038096822798252106,
0.11041250079870224,
-0.06291715055704117,
0.03409048914909363,
0.000022975395040703006,
-0.04237707331776619,
-0.008379739709198475,
-0.007014332804828882,
-0.06765340268611908,
-0.04569614678621292,
-0.07891730964183807,
-0.010821687988936901,
-0.06533432751893997,
0.025202535092830658,
0.06155429035425186,
-0.10826689004898071,
-0.004814375191926956,
-0.005300447810441256,
0.09733579307794571,
0.018250932916998863,
-7.137089715301849e-34,
-0.038085732609033585,
-0.08480944484472275,
-0.012176714837551117,
-0.007403064053505659,
0.011249104514718056,
0.006371181923896074,
-0.0014352918369695544,
-0.003932522144168615,
-0.005730169825255871,
0.008656146004796028,
-0.08724452555179596,
0.0438232421875,
0.0916118323802948,
-0.07172627002000809,
0.009807253256440163,
-0.10212182998657227,
0.04007435962557793,
0.06585218757390976,
0.02893972396850586,
-0.011387678794562817,
0.021307094022631645,
0.06638610363006592,
-0.09238048642873764,
0.009809414856135845,
0.04592667147517204,
0.0018814844079315662,
0.012115813791751862,
-0.008051213808357716,
-0.0076264929957687855,
0.09393057227134705,
-0.03371676430106163,
-0.03416522964835167,
-0.11232004314661026,
-0.004318862222135067,
-0.06814733147621155,
0.09676777571439743,
-0.028223490342497826,
0.014448187313973904,
-0.011745529249310493,
0.07631123811006546,
-0.01963185705244541,
-0.10529020428657532,
-0.06688463687896729,
0.010470482520759106,
-0.0024151764810085297,
0.009629783220589161,
-0.014327279292047024,
-0.04920310154557228,
-0.018462633714079857,
-0.021692588925361633,
0.05746077373623848,
0.02644732967019081,
-0.03386987745761871,
0.023307273164391518,
-0.06316075474023819,
-0.07204858213663101,
0.024307064712047577,
0.0010002747876569629,
0.05413054674863815,
0.009607614949345589,
-0.005997727159410715,
0.058531004935503006,
-0.048400264233350754,
0.029608024284243584,
-0.06301353126764297,
0.011996588669717312,
-0.026974119246006012,
0.024284852668642998,
0.005545551888644695,
-0.06330473721027374,
0.03878514841198921,
0.036676932126283646,
-0.008972289972007275,
-0.0019991183653473854,
0.018540989607572556,
0.02299892157316208,
-0.009427163749933243,
-0.013155563734471798,
0.023345820605754852,
-0.05037786066532135,
0.03007388487458229,
0.07849431782960892,
0.09003420919179916,
0.05956604331731796,
-0.04267258942127228,
0.05966314300894737,
0.034117307513952255,
0.12133350968360901,
0.019438736140727997,
0.024070871993899345,
0.014430909417569637,
-0.0022374584805220366,
0.025007283315062523,
0.014752221293747425,
-0.10288888216018677,
-4.243504392320574e-8,
-0.07438217848539352,
0.015882251784205437,
0.0031875881832093,
0.043366581201553345,
-0.02506808564066887,
0.004962880630046129,
-0.06565175205469131,
0.00954088382422924,
-0.04202410578727722,
-0.019916415214538574,
0.053710050880908966,
0.07357583194971085,
0.0727335587143898,
-0.06410055607557297,
-0.024947121739387512,
0.08104097098112106,
-0.03132118657231331,
-0.06021419167518616,
-0.025330955162644386,
0.05245399847626686,
-0.0039094421081244946,
-0.003603777615353465,
-0.06203727051615715,
-0.046852920204401016,
-0.01836041547358036,
-0.05454092100262642,
-0.0856291875243187,
0.03218628093600273,
0.07987921684980392,
0.022982295602560043,
-0.013508165255188942,
-0.0004728996136691421,
0.03125796839594841,
-0.0007040766649879515,
0.05622101575136185,
0.0715474858880043,
-0.04919705167412758,
-0.039597153663635254,
-0.007557287812232971,
-0.03455357253551483,
0.008320346474647522,
0.1043672040104866,
-0.005142011679708958,
-0.03816577419638634,
0.010859270580112934,
0.08507943153381348,
-0.08976852148771286,
-0.08319691568613052,
-0.025268979370594025,
-0.03150413557887077,
0.04992136359214783,
-0.022050824016332626,
-0.007941659539937973,
0.013067810796201229,
0.040964338928461075,
-0.027607714757323265,
-0.018705066293478012,
-0.039110537618398666,
0.11200136691331863,
0.01870710216462612,
0.007420179434120655,
-0.0015551509568467736,
-0.0528329573571682,
-0.006678965408354998
] | 0 | 0 | 0 | 0 |
|
Scaling In-the-Wild Training for Diffusion-based Illumination Harmonization and Editing by Imposing Consistent Light Transport | https://openreview.net/forum?id=u1cQYxRI1H | [
"Lvmin Zhang",
"Anyi Rao",
"Maneesh Agrawala"
] | Oral | Diffusion-based image generators are becoming unique methods for illumination harmonization and editing. The current bottleneck in scaling up the training of diffusion-based illumination editing models is mainly in the difficulty of preserving the underlying image details and maintaining intrinsic properties, such as albedos, unchanged. Without appropriate constraints, directly training the latest large image models with complex, varied, or in-the-wild data is likely to produce a structure-guided random image generator, rather than achieving the intended goal of precise illumination manipulation. We propose Imposing Consistent Light (IC-Light) transport during training, rooted in the physical principle that the linear blending of an object's appearances under different illumination conditions is consistent with its appearance under mixed illumination. This consistency allows for stable and scalable illumination learning, uniform handling of various data sources, and facilitates a physically grounded model behavior that modifies only the illumination of images while keeping other intrinsic properties unchanged. Based on this method, we can scale up the training of diffusion-based illumination editing models to large data quantities (> 10 million), across all available data types (real light stages, rendered samples, in-the-wild synthetic augmentations, etc), and using strong backbones (SDXL, Flux, etc). We also demonstrate that this approach reduces uncertainties and mitigates artifacts such as mismatched materials or altered albedos. | diffusion model, illumination editing, image editing | Diffusion-based image illumination harmonization and editing model | 3,821 | null | [
-0.007782071828842163,
-0.08687491714954376,
0.09077583253383636,
0.111810602247715,
0.06002460792660713,
-0.07019916921854019,
-0.08999157696962357,
-0.06968017667531967,
0.021122077479958534,
0.012202853336930275,
0.0006495702546089888,
-0.04188133776187897,
0.006272981408983469,
0.014314287342131138,
0.00751422718167305,
-0.0035126018337905407,
0.07688729465007782,
0.11352425068616867,
-0.16238656640052795,
-0.015531018376350403,
-0.003620551200583577,
-0.08655628561973572,
-0.03769389167428017,
-0.05611851438879967,
0.06190161034464836,
-0.001567221712321043,
0.05636323615908623,
-0.07195696979761124,
0.06250947713851929,
-0.08080343902111053,
0.03090618923306465,
-0.020199865102767944,
-0.04898246377706528,
0.01597447134554386,
-0.05537240952253342,
0.019946279004216194,
0.021014709025621414,
0.035781893879175186,
-0.0021110130473971367,
-0.031864337623119354,
0.0442577563226223,
0.004968369845300913,
-0.013220868073403835,
0.027416473254561424,
0.02812814712524414,
-0.014334778301417828,
0.020818306133151054,
-0.07250970602035522,
-0.010305064730346203,
0.011130076833069324,
0.04245534539222717,
-0.0787716954946518,
-0.07248303294181824,
0.10029558092355728,
-0.02043309435248375,
0.0643574595451355,
0.065057672560215,
-0.031778134405612946,
0.05872257426381111,
-0.045281603932380676,
-0.023087402805685997,
0.04281505569815636,
-0.005413986276835203,
-0.01053357683122158,
-0.030625300481915474,
-0.024631604552268982,
-0.0013632056070491672,
-0.032868318259716034,
-0.030234429985284805,
-0.012678303755819798,
-0.021862102672457695,
0.003047283971682191,
-0.04538171365857124,
0.00916453916579485,
-0.01616654545068741,
0.02786991186439991,
-0.010078955441713333,
-0.029384994879364967,
-0.023191288113594055,
-0.07241035997867584,
0.08504938334226608,
0.009452023543417454,
0.046736203134059906,
0.03412433713674545,
0.07381132245063782,
-0.018158039078116417,
0.03032968007028103,
0.08405111730098724,
-0.01821071468293667,
0.03299912437796593,
-0.07685114443302155,
0.07925186306238174,
-0.09972633421421051,
-0.05663261562585831,
-0.06616070866584778,
-0.001156051643192768,
0.06408165395259857,
-0.07427827268838882,
0.024356158450245857,
0.10662853717803955,
0.033491604030132294,
-0.03265822306275368,
-0.04373754933476448,
0.05619377642869949,
0.03925417736172676,
-0.0634898915886879,
0.06629064679145813,
0.0491984486579895,
-0.021916108205914497,
-0.014338070526719093,
0.10785781592130661,
0.04126979783177376,
-0.05185024440288544,
-0.011081886477768421,
0.005819770507514477,
-0.013153309933841228,
0.005267074331641197,
0.016698185354471207,
-0.002851285273209214,
0.03680642321705818,
0.00131856685038656,
0.0716555193066597,
-0.016891149803996086,
0.07800392806529999,
0.019045136868953705,
-0.08440355211496353,
-0.01795915886759758,
8.199899616450877e-34,
0.054836563766002655,
0.02925090119242668,
0.009034997783601284,
0.01365289743989706,
0.12021335959434509,
-0.020836932584643364,
0.027982063591480255,
-0.017941921949386597,
-0.032178156077861786,
-0.04609895125031471,
0.046841517090797424,
0.03457477316260338,
-0.008631479926407337,
0.1247265636920929,
0.06098854914307594,
0.03437940031290054,
-0.0031539250630885363,
0.06488299369812012,
-0.020017627626657486,
0.04776252433657646,
0.004779140930622816,
-0.046198632568120956,
0.019519519060850143,
-0.03437017649412155,
-0.07481224089860916,
0.05359670892357826,
0.016159510239958763,
0.03144353628158569,
0.008931368589401245,
0.00944739580154419,
-0.07656177133321762,
0.05669495463371277,
0.021607238799333572,
0.05354853719472885,
0.004144605249166489,
-0.016950955614447594,
0.045314520597457886,
-0.09289305657148361,
0.0023810798302292824,
-0.030045298859477043,
-0.003137232270091772,
0.08278542757034302,
0.007095569744706154,
-0.0013050693087279797,
-0.10084635764360428,
0.08401432633399963,
0.00910415593534708,
0.050207071006298065,
-0.08847285062074661,
-0.02005390077829361,
0.027309691533446312,
-0.008516903035342693,
-0.03230581805109978,
-0.09467604756355286,
-0.054489146918058395,
-0.01693471521139145,
0.041708409786224365,
-0.07006610929965973,
0.0736120343208313,
-0.024067820981144905,
0.04480286315083504,
0.04590930789709091,
-0.05806518718600273,
0.10247839987277985,
-0.002555775223299861,
0.011974404565989971,
0.012075399979948997,
0.049518510699272156,
-0.026469726115465164,
-0.04574846476316452,
-0.0676160603761673,
0.05577157437801361,
-0.062104951590299606,
-0.07234187424182892,
0.06507347524166107,
-0.046063899993896484,
0.07866965234279633,
-0.03843149542808533,
0.003454801859334111,
0.08258332312107086,
-0.07439495623111725,
0.04197515547275543,
-0.04749705642461777,
-0.15382321178913116,
-0.04677474871277809,
-0.025110524147748947,
-0.06713659316301346,
0.01811152510344982,
-0.13552047312259674,
-0.09079980105161667,
0.0007091153529472649,
-0.02165105938911438,
-0.006068517453968525,
-0.019599393010139465,
0.001158292405307293,
-8.370476703724653e-34,
0.05065523460507393,
0.04273415356874466,
-0.014713839627802372,
0.07758259773254395,
0.020457476377487183,
-0.014202448539435863,
-0.01913783699274063,
0.02630421705543995,
0.06002848595380783,
-0.04572305455803871,
0.01736324280500412,
0.027434417977929115,
-0.045866623520851135,
-0.07143528014421463,
-0.04011628404259682,
-0.09735812991857529,
-0.015146513469517231,
0.006259065121412277,
-0.06088123843073845,
-0.011544718407094479,
-0.022904401645064354,
0.05632006749510765,
-0.01528384629637003,
0.03680218756198883,
-0.07940248399972916,
0.05451122298836708,
-0.034731023013591766,
0.10494115948677063,
-0.016175752505660057,
0.005517969373613596,
-0.04425622522830963,
0.01368990633636713,
0.021495122462511063,
-0.06874099373817444,
-0.0959942638874054,
0.09588342159986496,
0.06838209182024002,
0.011343042366206646,
-0.06744665652513504,
0.05554689094424248,
0.03169684484601021,
-0.06277143955230713,
-0.04601163789629936,
0.08093936741352081,
-0.01043345034122467,
-0.03362486511468887,
-0.06819271296262741,
-0.04563768580555916,
-0.01086938101798296,
0.0331418439745903,
0.011394456028938293,
0.018435027450323105,
-0.11842837184667587,
0.05284716561436653,
-0.019420385360717773,
-0.0589548721909523,
-0.12201175093650818,
0.012158158235251904,
0.0659419521689415,
-0.007846532389521599,
-0.04309849068522453,
-0.028190989047288895,
-0.01358860544860363,
-0.009486514143645763,
-0.05972549319267273,
0.0008722213096916676,
-0.061667513102293015,
-0.03946579620242119,
0.04050415754318237,
0.11555743217468262,
0.025617599487304688,
-0.05484255403280258,
-0.0024475674144923687,
0.04406651481986046,
0.04080752283334732,
-0.0835452675819397,
0.1064547523856163,
-0.05540525168180466,
-0.006253342609852552,
-0.008700843900442123,
-0.0012173407012596726,
-0.02269979752600193,
0.020312732085585594,
0.04631616547703743,
0.09722672402858734,
0.027658257633447647,
-0.019700638949871063,
-0.000006334404588415055,
0.0019993267487734556,
0.06513234972953796,
-0.047899603843688965,
0.04863518103957176,
0.05490029230713844,
-0.005977522116154432,
-0.0326540544629097,
-4.0307845949882903e-8,
-0.009839983657002449,
0.026489751413464546,
0.02111216075718403,
0.0267020296305418,
0.017849501222372055,
0.005434683058410883,
0.07025624811649323,
0.03603778034448624,
0.011006848886609077,
-0.050824180245399475,
0.06825399398803711,
-0.041543811559677124,
0.02163136750459671,
-0.012836645357310772,
0.007821436040103436,
0.08648265153169632,
-0.019618308171629906,
0.00984654575586319,
-0.004658171907067299,
0.04499474912881851,
-0.139090895652771,
0.046909064054489136,
0.033302634954452515,
-0.07202082127332687,
0.09444521367549896,
0.01177983544766903,
0.008607868105173111,
0.08021806925535202,
0.04711421951651573,
0.016026033088564873,
0.02506074868142605,
-0.014915784820914268,
0.04389895498752594,
0.048043981194496155,
-0.04000042378902435,
0.031697213649749756,
-0.0037865329068154097,
-0.0063917264342308044,
-0.036055680364370346,
-0.058666735887527466,
-0.007117034401744604,
0.04056941717863083,
-0.009965638630092144,
0.002412367146462202,
0.03233462944626808,
0.0462479442358017,
0.028614860028028488,
-0.05020787939429283,
-0.001568895997479558,
0.0322117805480957,
0.03161058574914932,
-0.0802539736032486,
-0.025623619556427002,
0.027457743883132935,
0.05914049595594406,
-0.03420612961053848,
0.008353878743946552,
0.012542211450636387,
0.030614379793405533,
0.0637570470571518,
0.05714666098356247,
-0.041478224098682404,
-0.03800813481211662,
-0.018533382564783096
] | 0 | 0 | 0 | 0 |
|
AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models | https://openreview.net/forum?id=HvSytvg3Jh | [
"Junfeng Fang",
"Houcheng Jiang",
"Kun Wang",
"Yunshan Ma",
"Jie Shi",
"Xiang Wang",
"Xiangnan He",
"Tat-Seng Chua"
] | Oral | Large language models (LLMs) often exhibit hallucinations, producing incorrect or outdated knowledge. Hence, model editing methods have emerged to enable targeted knowledge updates. To achieve this, a prevailing paradigm is the locating-then-editing approach, which first locates influential parameters and then edits them by introducing a perturbation. While effective, current studies have demonstrated that this perturbation inevitably disrupt the originally preserved knowledge within LLMs, especially in sequential editing scenarios.
To address this, we introduce AlphaEdit, a novel solution that projects perturbation onto the null space of the preserved knowledge before applying it to the parameters. We theoretically prove that this projection ensures the output of post-edited LLMs remains unchanged when queried about the preserved knowledge, thereby mitigating the issue of disruption.
Extensive experiments on various LLMs, including LLaMA3, GPT2-XL, and GPT-J, show that AlphaEdit boosts the performance of most locating-then-editing methods by an average of 36.7% with a single line of additional code for projection solely. | Model Editing, Null-Space, Large Language Model | We propose a novel model editing method named AlphaEdit to minimize the disruption to the preserved knowledge during editing. | 3,792 | 2410.02355 | [
-0.034345537424087524,
-0.0509110726416111,
-0.03306851163506508,
0.025230785831809044,
0.03385644033551216,
0.05144432932138443,
-0.0776713490486145,
-0.023903869092464447,
0.04742714762687683,
0.0456894226372242,
-0.05274282395839691,
0.07292810082435608,
0.03166251629590988,
-0.035542912781238556,
-0.03650083392858505,
0.015414727851748466,
0.025456205010414124,
0.06542237102985382,
-0.10636124759912491,
-0.07317283004522324,
-0.0025468531530350447,
0.030976902693510056,
0.052856557071208954,
0.02543145976960659,
0.060463763773441315,
0.044253937900066376,
-0.014907052740454674,
-0.030927028506994247,
0.03723512217402458,
-0.05311259999871254,
-0.048184167593717575,
0.058596525341272354,
0.03109094314277172,
0.04259376972913742,
-0.033077090978622437,
0.0857667401432991,
-0.03222770616412163,
0.06156540662050247,
-0.03342688828706741,
-0.08260199427604675,
-0.012644096277654171,
0.0494556687772274,
0.0010701615829020739,
0.029795287176966667,
0.028628580272197723,
-0.005530599970370531,
-0.012091971933841705,
0.023751750588417053,
-0.06697499752044678,
0.029730848968029022,
-0.06928938627243042,
-0.09394102543592453,
-0.036398909986019135,
0.016615217551589012,
-0.05305201932787895,
0.03365268558263779,
0.06357111781835556,
-0.001601449679583311,
-0.008986189030110836,
-0.0620134063065052,
0.003942050039768219,
-0.022547893226146698,
-0.0413089394569397,
0.010841354727745056,
-0.011546509340405464,
-0.01801089197397232,
0.013418633490800858,
0.06119447201490402,
-0.05312344431877136,
0.00894634798169136,
0.007154821418225765,
0.08709196746349335,
-0.09136568754911423,
0.04105105623602867,
-0.0728597640991211,
0.01378693524748087,
0.006051543168723583,
-0.027981024235486984,
0.05910078063607216,
-0.045493073761463165,
0.09682096540927887,
-0.013578305020928383,
-0.009851066395640373,
0.007629149593412876,
0.04573511704802513,
-0.013830150477588177,
0.035734884440898895,
0.020100213587284088,
0.10439719259738922,
0.094973623752594,
-0.010434670373797417,
-0.05714675411581993,
0.04994989186525345,
-0.023486284539103508,
0.03295722231268883,
-0.011421309784054756,
0.03981073573231697,
-0.0074315145611763,
-0.0025677327066659927,
0.10692871361970901,
-0.0053205108270049095,
0.09393178671598434,
-0.0500384196639061,
-0.05727308988571167,
-0.05288250744342804,
-0.03812924027442932,
0.00036633736453950405,
0.004913967102766037,
-0.0031922352500259876,
-0.03245849534869194,
0.023303445428609848,
0.023403527215123177,
0.008038471452891827,
-0.0002803719835355878,
-0.00950571708381176,
-0.02513110637664795,
0.05949699133634567,
0.019403543323278427,
-0.040982745587825775,
0.040264979004859924,
-0.006694462150335312,
0.01861312985420227,
0.07416772842407227,
0.050374094396829605,
-0.0167801845818758,
-0.026547346264123917,
-0.08313987404108047,
-9.291356223093949e-34,
0.1536664068698883,
0.04484763741493225,
-0.04600277915596962,
0.03667530417442322,
0.024612383916974068,
0.07241938263177872,
-0.03923403471708298,
0.04913624748587608,
-0.018103519454598427,
-0.08724725991487503,
0.05129385367035866,
-0.04174400866031647,
-0.02203228697180748,
0.026579804718494415,
0.06379188597202301,
0.046446822583675385,
0.05600225552916527,
0.11364975571632385,
-0.15464770793914795,
-0.019797615706920624,
0.03483045846223831,
0.02252591960132122,
-0.043841779232025146,
-0.07818423956632614,
-0.028906986117362976,
0.08116215467453003,
-0.01417350023984909,
-0.012663885951042175,
-0.03742391988635063,
-0.013203673996031284,
-0.11030280590057373,
0.0014878518413752317,
0.003435568418353796,
0.08646249771118164,
-0.009165924973785877,
-0.023565994575619698,
-0.008638495579361916,
-0.13401643931865692,
0.06758811324834824,
-0.03194103017449379,
0.012832618318498135,
0.04698143154382706,
0.018331438302993774,
-0.08027715235948563,
-0.044452447444200516,
-0.0486576147377491,
0.024971013888716698,
0.06853145360946655,
-0.01349879801273346,
-0.04275137931108475,
0.07821724563837051,
0.12151327729225159,
-0.06204771250486374,
0.02158173732459545,
-0.05622649937868118,
0.011745471507310867,
-0.000522494432516396,
-0.0472150556743145,
0.05262167751789093,
0.05655745789408684,
0.048027560114860535,
0.0880628451704979,
0.025827055796980858,
0.07709411531686783,
0.06225275248289108,
0.06240535527467728,
-0.05108949542045593,
0.02395363710820675,
0.11673220247030258,
-0.045343413949012756,
-0.07044319063425064,
-0.07068351656198502,
-0.01563093811273575,
0.05790521949529648,
-0.013939322903752327,
-0.07928109169006348,
-0.0328669510781765,
-0.05262330546975136,
-0.001806667074561119,
-0.05735208839178085,
0.021388530731201172,
-0.05951516702771187,
-0.002366908360272646,
-0.018125228583812714,
-0.023052627220749855,
-0.09662928432226181,
0.0474218986928463,
-0.0024981950409710407,
-0.06036142632365227,
-0.08624201267957687,
-0.001130159362219274,
-0.05459168180823326,
-0.04785666987299919,
0.023776888847351074,
0.023277288302779198,
-2.050052505106973e-34,
-0.045299433171749115,
-0.0939294770359993,
-0.0034865373745560646,
0.0870925560593605,
-0.07099534571170807,
-0.10952853411436081,
-0.03410382196307182,
0.014371149241924286,
0.004995162133127451,
-0.08121408522129059,
-0.04517032578587532,
-0.06141567975282669,
0.01612139493227005,
-0.03450155630707741,
0.06320632994174957,
-0.03649361804127693,
0.05398977920413017,
-0.037384867668151855,
-0.055855900049209595,
0.04512585327029228,
0.031197773292660713,
-0.020283330231904984,
-0.07185078412294388,
0.06536680459976196,
0.0006878773565404117,
0.0011012105969712138,
0.02788263000547886,
0.04006470367312431,
-0.01892906241118908,
0.026352228596806526,
-0.051629360765218735,
0.01698150299489498,
-0.07438968122005463,
-0.02513045258820057,
-0.03719617426395416,
0.01956830732524395,
0.06748802959918976,
0.007971459999680519,
-0.06257562339305878,
0.0662032887339592,
0.06088530272245407,
0.035611219704151154,
-0.04271909222006798,
0.021927857771515846,
0.01855292357504368,
0.052842557430267334,
-0.13230185210704803,
0.048896465450525284,
0.05375482141971588,
-0.028070706874132156,
-0.015337487682700157,
-0.03651341795921326,
-0.015558729879558086,
-0.007543501909822226,
0.013595914468169212,
-0.06832403689622879,
-0.07228011637926102,
-0.02771170623600483,
0.054658886045217514,
-0.03183089569211006,
-0.052743151783943176,
-0.005526080261915922,
0.005021994933485985,
-0.043973565101623535,
-0.0025469886604696512,
-0.0072690825909376144,
-0.04929215833544731,
0.06909819692373276,
0.02523278072476387,
-0.02187090553343296,
0.009967056103050709,
-0.03736012056469917,
-0.034900013357400894,
0.009004379622638226,
0.025277551263570786,
0.013437544927001,
0.047241855412721634,
-0.07787175476551056,
0.05663067847490311,
-0.09971798211336136,
-0.03338364511728287,
0.053112514317035675,
0.041015349328517914,
0.011036618612706661,
0.04952255263924599,
0.0610533207654953,
0.036386262625455856,
0.10811232030391693,
-0.0006785683217458427,
0.02416069805622101,
-0.05651995167136192,
0.03686284273862839,
-0.0020345577504485846,
0.03904465585947037,
-0.09396985918283463,
-4.0671224610377976e-8,
-0.04004021733999252,
0.019352268427610397,
-0.020990435034036636,
0.10773740708827972,
0.09276797622442245,
-0.09542801976203918,
0.012339414097368717,
0.09755156189203262,
0.020359668880701065,
-0.05867207050323486,
-0.0020939530804753304,
0.007990939542651176,
-0.018314749002456665,
-0.06991071999073029,
-0.0365980826318264,
0.06989166140556335,
-0.013290319591760635,
0.03061552904546261,
-0.06907106190919876,
0.014110840857028961,
0.025066377595067024,
0.08030698448419571,
0.0036628874950110912,
-0.04494297131896019,
0.06569483131170273,
-0.06896824389696121,
-0.01988537609577179,
0.0897442176938057,
0.06177613511681557,
0.06106141209602356,
0.04149215295910835,
-0.025390030816197395,
0.07746710628271103,
0.04649489000439644,
0.0061753191985189915,
0.042241018265485764,
0.0515252985060215,
0.012920434586703777,
0.019962036982178688,
-0.04636376351118088,
0.07650316506624222,
-0.025376256555318832,
-0.00827134307473898,
0.0708608627319336,
-0.007640146650373936,
-0.051071371883153915,
0.0014549923362210393,
-0.11382367461919785,
-0.009719341062009335,
-0.08818215131759644,
-0.007672912906855345,
0.03723157197237015,
-0.0013428909005597234,
0.003013501176610589,
0.03584439307451248,
0.07376222312450409,
-0.010660949163138866,
0.010739833116531372,
0.035519491881132126,
-0.04825722053647041,
0.02157757245004177,
-0.05630200728774071,
-0.02466757409274578,
0.04023897275328636
] | https://github.com/jianghoucheng/alphaedit | 174 | 0 | 0 | 0 |
DeepLTL: Learning to Efficiently Satisfy Complex LTL Specifications for Multi-Task RL | https://openreview.net/forum?id=9pW2J49flQ | [
"Mathias Jackermeier",
"Alessandro Abate"
] | Oral | Linear temporal logic (LTL) has recently been adopted as a powerful formalism for specifying complex, temporally extended tasks in multi-task reinforcement learning (RL). However, learning policies that efficiently satisfy arbitrary specifications not observed during training remains a challenging problem. Existing approaches suffer from several shortcomings: they are often only applicable to finite-horizon fragments of LTL, are restricted to suboptimal solutions, and do not adequately handle safety constraints. In this work, we propose a novel learning approach to address these concerns. Our method leverages the structure of Büchi automata, which explicitly represent the semantics of LTL specifications, to learn policies conditioned on sequences of truth assignments that lead to satisfying the desired formulae. Experiments in a variety of discrete and continuous domains demonstrate that our approach is able to zero-shot satisfy a wide range of finite- and infinite-horizon specifications, and outperforms existing methods in terms of both satisfaction probability and efficiency. Code available at: https://deep-ltl.github.io/ | reinforcement learning, linear temporal logic, ltl, generalization | null | 3,756 | null | [
-0.07681623101234436,
0.02300921455025673,
0.020348625257611275,
-0.034328117966651917,
0.048953041434288025,
0.025625044479966164,
-0.03778226301074028,
-0.05313735455274582,
-0.012159699574112892,
0.0050032902508974075,
-0.05855526402592659,
-0.09766149520874023,
-0.036386314779520035,
0.08388101309537888,
0.03946219012141228,
0.017088888213038445,
0.04203303903341293,
0.008957026526331902,
-0.10955223441123962,
-0.08208394795656204,
0.08070238679647446,
0.0376410111784935,
0.05088530480861664,
0.07913147658109665,
-0.012166551314294338,
0.0023414483293890953,
0.029742849990725517,
-0.08099484443664551,
0.043633587658405304,
-0.010958781465888023,
0.0320795513689518,
-0.045912690460681915,
-0.028453271836042404,
-0.041975535452365875,
0.03519890457391739,
0.07003994286060333,
-0.08841239660978317,
-0.05365465208888054,
-0.058626897633075714,
-0.1037750393152237,
-0.008482779376208782,
-0.06489778310060501,
0.009758517146110535,
0.025657931342720985,
0.05105185508728027,
-0.013931069523096085,
0.03957780450582504,
-0.052450891584157944,
0.019527604803442955,
0.017868679016828537,
0.023396814242005348,
0.025232942774891853,
0.04213697090744972,
0.025355586782097816,
0.03552306443452835,
-0.0052664815448224545,
-0.033939484506845474,
-0.02580753155052662,
-0.0510738343000412,
-0.03241020441055298,
-0.05167496204376221,
-0.08234749734401703,
-0.08107214421033859,
-0.053234465420246124,
-0.07967259734869003,
0.003932544030249119,
-0.11110921949148178,
0.04515813663601875,
0.016108661890029907,
0.0729847401380539,
-0.018109261989593506,
-0.040309153497219086,
-0.01716422848403454,
0.0535554513335228,
0.02219662442803383,
-0.01609354093670845,
0.030841989442706108,
-0.08074178546667099,
0.003913684748113155,
-0.05733173340559006,
0.017539964988827705,
-0.026510154828429222,
-0.025267796590924263,
0.07698386162519455,
0.06885223835706711,
-0.018771275877952576,
-0.06227812543511391,
0.08512182533740997,
0.09550167620182037,
-0.0032671308144927025,
0.021197382360696793,
0.004832819569855928,
0.04812973365187645,
0.031974636018276215,
-0.004496712703257799,
-0.0017923336708918214,
-0.03648272529244423,
-0.10356559604406357,
-0.003600197611376643,
0.12272647768259048,
-0.029753433540463448,
0.04137426242232323,
0.0036525896284729242,
-0.0012603523209691048,
0.00040737062226980925,
0.01261928305029869,
-0.006610284559428692,
0.06674228608608246,
-0.024541623890399933,
-0.13246509432792664,
0.03914501890540123,
0.005394138395786285,
0.08214429020881653,
0.024623677134513855,
-0.03950300067663193,
0.017284192144870758,
-0.04714169725775719,
-0.004739738069474697,
-0.013062377460300922,
0.0196614321321249,
-0.0467926487326622,
-0.01742599904537201,
0.0026240330189466476,
0.08023101836442947,
0.030934561043977737,
-0.10879445821046829,
0.001086374162696302,
1.3249681920919341e-33,
0.06526698172092438,
0.031854987144470215,
-0.024440892040729523,
0.00981550570577383,
0.005620243027806282,
0.02928532473742962,
0.07350827008485794,
0.0032573577482253313,
-0.03839825093746185,
0.04729757457971573,
0.011792344972491264,
0.01587822660803795,
-0.11344970762729645,
0.04347555711865425,
0.024989860132336617,
0.00005392523235059343,
-0.05426599457859993,
0.06228341907262802,
0.0348045639693737,
-0.03072267957031727,
0.060881953686475754,
-0.007303199730813503,
-0.0694461464881897,
0.0025526368990540504,
-0.009394707158207893,
0.057853590697050095,
0.0804414451122284,
-0.01193250808864832,
0.04268151521682739,
0.017760057002305984,
0.007789390627294779,
0.03769703581929207,
0.033173032104969025,
0.07894228398799896,
0.05086258798837662,
0.05526745319366455,
-0.06268555670976639,
0.0014071973273530602,
0.017680836841464043,
-0.012336409650743008,
0.007507548667490482,
-0.02529321238398552,
-0.030349798500537872,
0.05620855465531349,
-0.029212040826678276,
-0.12111382186412811,
-0.054178327322006226,
0.04689959064126015,
-0.08043933659791946,
-0.05629635602235794,
0.07226825505495071,
0.014675179496407509,
0.0031632056925445795,
-0.10741589218378067,
0.017526943236589432,
0.027542579919099808,
-0.012548519298434258,
0.1210535392165184,
0.01325863879173994,
0.09947539120912552,
0.017966406419873238,
-0.004324670415371656,
-0.03817557170987129,
0.06672048568725586,
-0.015542364679276943,
0.07110859453678131,
-0.07422436773777008,
0.06371863931417465,
0.12434219568967819,
-0.04652886465191841,
0.014074615202844143,
0.017849596217274666,
0.04186127707362175,
-0.048583053052425385,
0.11253414303064346,
-0.014385845512151718,
0.08587539941072464,
-0.09405211359262466,
-0.04477432370185852,
0.04703536629676819,
-0.0038051453884691,
0.05041256919503212,
-0.011182669550180435,
0.027476802468299866,
0.06331682205200195,
-0.0288083516061306,
-0.01154885534197092,
-0.0502653494477272,
0.010991333052515984,
-0.018120722845196724,
-0.10661683976650238,
-0.08343054354190826,
0.0296833124011755,
0.0065239546820521355,
0.010311360470950603,
-1.5992927577303544e-33,
0.07999396324157715,
-0.030488520860671997,
-0.070821613073349,
0.03219009190797806,
-0.03777981922030449,
-0.09393917769193649,
-0.07616850733757019,
-0.018631882965564728,
0.024301880970597267,
0.002387827727943659,
-0.04616519808769226,
0.012032750062644482,
0.010259091854095459,
-0.015348799526691437,
0.01313870307058096,
-0.04124181345105171,
-0.004438829608261585,
-0.04894113168120384,
-0.054350417107343674,
0.042100127786397934,
-0.017338747158646584,
0.07825952023267746,
-0.09487830102443695,
0.07726193219423294,
0.0032382295466959476,
0.035086847841739655,
-0.03220882639288902,
0.06075296178460121,
-0.016748623922467232,
0.05134446546435356,
-0.01746976561844349,
-0.11391470581293106,
0.0007454722071997821,
-0.015206080861389637,
0.0019087080145254731,
-0.020326055586338043,
0.08321019262075424,
0.05471615865826607,
-0.07258256524801254,
-0.010047505609691143,
0.01677578315138817,
-0.12732510268688202,
-0.06446991115808487,
-0.05454956740140915,
-0.01931365393102169,
0.0486266054213047,
-0.0327029824256897,
-0.017525048926472664,
-0.04457097873091698,
0.01942174695432186,
-0.025411419570446014,
-0.034212131053209305,
0.0022424531634896994,
-0.07108191400766373,
-0.02989836037158966,
-0.0699595957994461,
-0.025198515504598618,
-0.026800021529197693,
0.05178108438849449,
0.040621716529130936,
-0.016121748834848404,
0.03515901416540146,
0.02889799140393734,
0.0823718011379242,
0.02297038398683071,
-0.002647981746122241,
-0.013112632557749748,
-0.019556978717446327,
0.05823447182774544,
-0.02290421724319458,
0.024064498022198677,
0.010827595368027687,
-0.03348555043339729,
0.02868742309510708,
0.04373304545879364,
-0.024404849857091904,
-0.04030165076255798,
-0.037617847323417664,
0.0026609261985868216,
-0.02115919440984726,
-0.01647675782442093,
0.008564377203583717,
0.057188332080841064,
0.1017439216375351,
0.015334288589656353,
0.015308858826756477,
0.025474172085523605,
0.12358725816011429,
0.04465071111917496,
0.008825372904539108,
-0.011437293142080307,
0.03469131886959076,
-0.006026606075465679,
0.03927064314484596,
-0.0492841973900795,
-4.312461499011988e-8,
-0.012142794206738472,
0.06146657466888428,
-0.0004970236332155764,
0.007921330630779266,
0.058210186660289764,
0.017254361882805824,
-0.07390688359737396,
-0.03837181627750397,
0.08163513243198395,
-0.0005614880938082933,
0.06089462712407112,
0.05322487652301788,
-0.012345971539616585,
-0.10434511303901672,
-0.12026871740818024,
0.08934386819601059,
0.0794055163860321,
-0.0820283368229866,
0.01579371839761734,
0.023733600974082947,
0.031167343258857727,
0.05435305833816528,
0.004992375615984201,
-0.012804991565644741,
-0.009421800263226032,
-0.028197338804602623,
-0.053537409752607346,
0.11200808733701706,
0.09827227145433426,
0.09396058320999146,
0.06171504408121109,
-0.02852701023221016,
0.08352186530828476,
0.015562517568469048,
0.050313301384449005,
0.04972701147198677,
0.016103465110063553,
-0.028548765927553177,
0.018535524606704712,
-0.08138526976108551,
0.036124441772699356,
0.08863475173711777,
-0.0473325178027153,
0.003788130823522806,
-0.033535636961460114,
-0.06619113683700562,
-0.05683794990181923,
-0.11475210636854172,
0.02815098501741886,
-0.009298179298639297,
0.010118101723492146,
-0.0026852479204535484,
-0.014681915752589703,
0.07385502755641937,
0.009361729957163334,
0.0032295077107846737,
0.01675325259566307,
-0.08608879894018173,
0.0227506123483181,
0.04174419492483139,
0.0155211566016078,
0.04296569898724556,
-0.02294081635773182,
0.006230274215340614
] | 0 | 0 | 0 | 0 |
|
On the Role of Attention Heads in Large Language Model Safety | https://openreview.net/forum?id=h0Ak8A5yqw | [
"Zhenhong Zhou",
"Haiyang Yu",
"Xinghua Zhang",
"Rongwu Xu",
"Fei Huang",
"Kun Wang",
"Yang Liu",
"Junfeng Fang",
"Yongbin Li"
] | Oral | Large language models (LLMs) achieve state-of-the-art performance on multiple language tasks, yet their safety guardrails can be circumvented, leading to harmful generations. In light of this, recent research on safety mechanisms has emerged, revealing that when safety representations or component are suppressed, the safety capability of LLMs are compromised. However, existing research tends to overlook the safety impact of multi-head attention mechanisms, despite their crucial role in various model functionalities. Hence, in this paper, we aim to explore the connection between standard attention mechanisms and safety capability to fill this gap in the safety-related mechanistic interpretability. We propose an novel metric which tailored for multi-head attention, the Safety Head ImPortant Score (Ships), to assess the individual heads' contributions to model safety. Base on this, we generalize Ships to the dataset level and further introduce the Safety Attention Head AttRibution Algorithm (Sahara) to attribute the critical safety attention heads inside the model. Our findings show that special attention head has a significant impact on safety. Ablating a single safety head allows aligned model (e.g., Llama-2-7b-chat) to respond to **16$\times\uparrow$** more harmful queries, while only modifying **0.006\%** $\downarrow$ of the parameters, in contrast to the $\sim$ **5\%** modification required in previous studies. More importantly, we demonstrate that attention heads primarily function as feature extractors for safety and models fine-tuned from the same base model exhibit overlapping safety heads through comprehensive experiments. Together, our attribution approach and findings provide a novel perspective for unpacking the black box of safety mechanisms in large models. | interpretability, large language model, multi-head attention, safety, harmful content | We identify safety-critical attention heads in large language models, and when these heads are ablated, the model safety is significantly compromised. | 3,741 | 2410.13708 | [
0.028913913294672966,
-0.025908952578902245,
-0.05274828150868416,
-0.0243581123650074,
0.13618768751621246,
0.06258545070886612,
0.06673029810190201,
0.04702937975525856,
0.05167117342352867,
0.03173447027802467,
-0.02207445539534092,
-0.06973020732402802,
0.0903179943561554,
0.05380891636013985,
-0.04121271148324013,
-0.031745437532663345,
0.0658755898475647,
0.0069167776964604855,
-0.11451686173677444,
-0.022439181804656982,
0.06309878081083298,
0.08077656477689743,
0.09099981933832169,
0.03355024755001068,
-0.1087011843919754,
-0.0668725073337555,
-0.0819411426782608,
-0.01646413467824459,
-0.03892245516180992,
-0.02087143063545227,
-0.005047152284532785,
-0.060324277728796005,
0.041366320103406906,
0.07059527933597565,
-0.06321334838867188,
0.006852512247860432,
-0.039178650826215744,
0.012710384093225002,
-0.012956883758306503,
-0.03746435046195984,
-0.10802356898784637,
-0.015579038299620152,
0.008317075669765472,
0.02429588884115219,
0.06390748918056488,
-0.0704430565237999,
0.035907045006752014,
-0.05948058143258095,
-0.021847113966941833,
-0.0033789118751883507,
-0.07839538156986237,
-0.023519840091466904,
0.052626051008701324,
0.0005966289900243282,
0.034961991012096405,
-0.05059744045138359,
-0.006322994362562895,
0.026270950213074684,
-0.04477956146001816,
-0.014926912263035774,
-0.0836317241191864,
-0.05369383096694946,
-0.022084226831793785,
-0.05551619082689285,
-0.014123989269137383,
0.015749895945191383,
-0.06158025190234184,
0.0025727220345288515,
-0.046430330723524094,
0.06254421174526215,
-0.025179365649819374,
0.02979876846075058,
0.03899480774998665,
-0.027278216555714607,
0.058619216084480286,
0.05715254321694374,
0.020159272477030754,
-0.01695992238819599,
0.040833961218595505,
-0.0632636621594429,
-0.058101240545511246,
-0.01710418239235878,
0.0687268003821373,
-0.008761206641793251,
0.11453455686569214,
0.036784619092941284,
-0.013352343812584877,
0.044500626623630524,
0.009367312304675579,
0.08543137460947037,
-0.0680852010846138,
-0.08653736859560013,
0.09660273045301437,
0.043277837336063385,
0.06638621538877487,
0.047542963176965714,
-0.08573569357395172,
-0.03904726356267929,
-0.019616978242993355,
0.013852514326572418,
0.04707038402557373,
0.04517345502972603,
0.007398678455501795,
-0.009134707041084766,
0.029639586806297302,
-0.016023587435483932,
0.01348022185266018,
0.005410401150584221,
-0.030383501201868057,
-0.04153524339199066,
-0.013535740785300732,
0.035020001232624054,
0.014410888776183128,
-0.006900501903146505,
-0.03685304522514343,
-0.007044641766697168,
-0.08906111866235733,
-0.03577251359820366,
0.04087154567241669,
0.008898342959582806,
0.07066188007593155,
0.017127972096204758,
0.13079039752483368,
0.0030858456157147884,
0.026506338268518448,
0.0074457693845033646,
0.008450430817902088,
1.3196531209162942e-33,
0.0009531693067401648,
0.07545343041419983,
-0.06187177821993828,
0.031690578907728195,
-0.005091871600598097,
-0.04988551139831543,
-0.07073311507701874,
-0.0034286745358258486,
-0.0001137831641244702,
0.03566567972302437,
-0.03175337612628937,
0.017649177461862564,
-0.006106206215918064,
-0.005276513751596212,
0.04671359062194824,
0.021412456408143044,
0.01706974022090435,
0.02168221026659012,
-0.10767815262079239,
-0.05456836149096489,
0.0599893219769001,
-0.047079216688871384,
0.047132473438978195,
-0.04605600982904434,
0.061681825667619705,
0.06963963806629181,
0.02939620055258274,
-0.026020314544439316,
-0.08251076191663742,
0.06005121022462845,
-0.1335810273885727,
0.025960028171539307,
0.00692829629406333,
-0.03616395220160484,
0.019226158037781715,
-0.07248780876398087,
-0.0952693521976471,
-0.05095837637782097,
0.03327963128685951,
-0.015276962891221046,
-0.025893665850162506,
0.030974024906754494,
0.008282387629151344,
0.04780373349785805,
-0.02169433981180191,
0.02729763463139534,
-0.03384731709957123,
0.013449913822114468,
-0.022211872041225433,
-0.01933198608458042,
-0.007277504075318575,
0.008590160869061947,
0.07682714611291885,
-0.03437526151537895,
-0.07529419660568237,
0.06332539767026901,
0.04363298416137695,
0.05162198469042778,
-0.009959680959582329,
0.03457687795162201,
-0.03162973374128342,
0.0705472007393837,
0.025578109547495842,
0.030175242573022842,
0.11717987805604935,
0.026820799335837364,
-0.04053918272256851,
0.0419054739177227,
0.0780719518661499,
-0.014806280843913555,
-0.03331226482987404,
0.026172485202550888,
-0.02935333177447319,
-0.044578514993190765,
-0.08468227833509445,
-0.022967694327235222,
0.06880547851324081,
-0.0470968596637249,
0.06603182107210159,
-0.008418201468884945,
0.028572317212820053,
0.04047722741961479,
0.00651852460578084,
-0.005754206795245409,
-0.006842542439699173,
-0.026064708828926086,
-0.0020096153020858765,
-0.03602699190378189,
-0.04954444617033005,
0.04866126924753189,
0.050118573009967804,
-0.014692154712975025,
0.014420456252992153,
0.002068700734525919,
-0.10005193203687668,
-1.0751307349845768e-33,
-0.08024412393569946,
-0.005124425049871206,
-0.06757055222988129,
0.009299717843532562,
-0.05372690409421921,
-0.0756416991353035,
0.02235717512667179,
-0.018680933862924576,
-0.022452015429735184,
0.03784438967704773,
-0.05265055224299431,
-0.13053490221500397,
0.020580103620886803,
0.02427038736641407,
0.07389738410711288,
-0.06017247959971428,
0.037212032824754715,
0.08885962516069412,
-0.013924424536526203,
0.01983751356601715,
0.09741687029600143,
0.05703438073396683,
-0.12202954292297363,
0.03941383585333824,
-0.07586850225925446,
0.02671162225306034,
-0.059474460780620575,
-0.06853650510311127,
-0.02835233509540558,
-0.025456776842474937,
-0.0506855733692646,
0.05720354989171028,
-0.023132458329200745,
0.02419748343527317,
-0.07657991349697113,
-0.033214304596185684,
-0.019989339634776115,
-0.028795434162020683,
-0.04667845368385315,
0.07273101806640625,
0.0685989111661911,
0.03454811871051788,
-0.028182778507471085,
-0.04091500863432884,
-0.059880055487155914,
0.018051989376544952,
-0.03555825352668762,
-0.00789653044193983,
-0.0391109324991703,
-0.014746008440852165,
-0.06549515575170517,
-0.045112647116184235,
0.015396809205412865,
-0.007208775728940964,
-0.036088261753320694,
-0.015640171244740486,
0.025436894968152046,
-0.08163836598396301,
0.07405893504619598,
-0.03875379264354706,
-0.016808919608592987,
0.03687235713005066,
-0.022238261997699738,
0.049472875893116,
0.03814613074064255,
0.0038096897769719362,
-0.0520818866789341,
-0.05847082659602165,
0.033065371215343475,
-0.058090247213840485,
0.10588826984167099,
-0.015469731763005257,
-0.039634834975004196,
0.00513859186321497,
-0.03620712086558342,
-0.005062705837190151,
-0.055955469608306885,
-0.04081867262721062,
-0.030920328572392464,
0.010191449895501137,
0.02526962012052536,
0.00019661785336211324,
0.004689273424446583,
0.10381922870874405,
0.030128881335258484,
0.027412934228777885,
0.06636319309473038,
0.05167265236377716,
0.005709900055080652,
-0.041778601706027985,
-0.05063321068882942,
0.06519719958305359,
-0.04239267483353615,
0.05843290686607361,
-0.08478187769651413,
-4.1792372229565444e-8,
-0.06726283580064774,
0.0030630407854914665,
-0.025569193065166473,
0.06355339288711548,
0.00781993381679058,
-0.021518195047974586,
-0.09648118168115616,
0.031154746189713478,
-0.07241552323102951,
0.030704369768500328,
0.06232761964201927,
0.0020608077757060528,
-0.0192372128367424,
-0.02105114609003067,
-0.058114416897296906,
0.05963640287518501,
0.036997467279434204,
0.10684603452682495,
-0.03071608766913414,
-0.05896873027086258,
0.051034584641456604,
0.08379749953746796,
-0.0017694371053948998,
0.018565239384770393,
0.034058790653944016,
0.00685929274186492,
-0.06672482192516327,
0.10267536342144012,
0.033609114587306976,
0.010965518653392792,
0.04899822920560837,
-0.006992734968662262,
-0.056667573750019073,
0.02361576445400715,
0.01709570735692978,
0.0645153596997261,
0.008088666945695877,
0.0005631990497931838,
0.060334548354148865,
0.07345115393400192,
0.06511928886175156,
0.0761856883764267,
0.02377631701529026,
0.1155899316072464,
0.05156068876385689,
0.057623349130153656,
-0.0738169401884079,
-0.09168215095996857,
-0.006229456514120102,
-0.08021528273820877,
0.05756652355194092,
-0.06623237580060959,
-0.025245049968361855,
0.13507047295570374,
0.009961001574993134,
0.032536301761865616,
0.045731622725725174,
-0.027148675173521042,
0.012898239307105541,
0.04387468472123146,
0.10114254802465439,
0.05882979556918144,
0.010758712887763977,
-0.006254015024751425
] | https://github.com/ydyjya/safetyheadattribution | 21 | 0 | 0 | 0 |
Influence Functions for Scalable Data Attribution in Diffusion Models | https://openreview.net/forum?id=esYrEndGsr | [
"Bruno Kacper Mlodozeniec",
"Runa Eschenhagen",
"Juhan Bae",
"Alexander Immer",
"David Krueger",
"Richard E. Turner"
] | Oral | Diffusion models have led to significant advancements in generative modelling. Yet their widespread adoption poses challenges regarding data attribution and interpretability. In this paper, we aim to help address such challenges in diffusion models by extending influence functions. Influence function-based data attribution methods approximate how a model's output would have changed if some training data were removed. In supervised learning, this is usually used for predicting how the loss on a particular example would change. For diffusion models, we focus on predicting the change in the probability of generating a particular example via several proxy measurements. We show how to formulate influence functions for such quantities and how previously proposed methods can be interpreted as particular design choices in our framework. To ensure scalability of the Hessian computations in influence functions, we use a K-FAC approximation based on generalised Gauss-Newton matrices specifically tailored to diffusion models. We show that our recommended method outperforms previously proposed data attribution methods on common data attribution evaluations, such as the Linear Data-modelling Score (LDS) or retraining without top influences, without the need for method-specific hyperparameter tuning. | diffusion models, influence functions, Generalised Gauss Newton, GGN, data attribution, Hessian approximation, interpretability, curvature, Kronecker-Factored Approximate Curvature, K-FAC | We present a method for attributing the influence of training data on diffusion model’s output by adapting influence functions and a KFAC approximation for diffusion models, and we explore what measurements we want to attribute for in the first place | 3,597 | 2410.13850 | [
-0.06496269255876541,
-0.13771456480026245,
-0.009076908230781555,
0.04842138662934303,
0.0994720458984375,
-0.010610613971948624,
-0.024530544877052307,
-0.01381781604140997,
0.03975657373666763,
0.06670418381690979,
0.03562064468860626,
0.005837712436914444,
0.033691007643938065,
0.014862128533422947,
-0.07479602098464966,
-0.054441727697849274,
0.0918952152132988,
0.0254367608577013,
-0.1284252554178238,
-0.050882719457149506,
0.0029115795623511076,
0.025110196322202682,
0.0000459919756394811,
0.011534673161804676,
0.07033134996891022,
-0.027586549520492554,
0.011470980942249298,
-0.01295774057507515,
-0.039469823241233826,
-0.03174916282296181,
0.017143335193395615,
-0.012047539465129375,
-0.0386221706867218,
-0.019974516704678535,
-0.11426372826099396,
0.07615642249584198,
0.007786169648170471,
0.05531147122383118,
0.06833357363939285,
-0.00038824742659926414,
0.019452769309282303,
0.007199923507869244,
0.045564427971839905,
0.05908913165330887,
0.04852854833006859,
-0.04928681626915932,
0.06164480373263359,
-0.06350728124380112,
-0.05019247159361839,
0.08920731395483017,
-0.014009444043040276,
-0.043372463434934616,
-0.035808492451906204,
0.0010829060338437557,
-0.0028329542838037014,
-0.06627479195594788,
0.04700721427798271,
0.009604571387171745,
-0.01181513536721468,
0.019264696165919304,
-0.13059277832508087,
-0.10371672362089157,
-0.06498858332633972,
-0.0429886057972908,
-0.04351244494318962,
-0.019686998799443245,
-0.013648548163473606,
0.06879284232854843,
-0.11029237508773804,
0.08732598274946213,
0.07153335213661194,
0.06919348984956741,
-0.07735834270715714,
-0.033701565116643906,
0.0155209144577384,
-0.020307030528783798,
0.032151028513908386,
0.06941406428813934,
0.008238883689045906,
-0.06382373720407486,
0.027437396347522736,
0.056783128529787064,
0.04287487640976906,
-0.01198656391352415,
0.05147711560130119,
-0.05142587050795555,
0.08717367798089981,
0.045815326273441315,
0.08414717018604279,
0.02322649396955967,
-0.01711389794945717,
0.09561397135257721,
0.015635201707482338,
-0.05501789599657059,
-0.028653355315327644,
0.0030329080764204264,
-0.036638304591178894,
-0.08359760046005249,
0.07946500182151794,
0.0569661408662796,
-0.0623333603143692,
0.11744431406259537,
-0.0017188654746860266,
-0.04607033729553223,
0.0556442067027092,
-0.06677670031785965,
-0.031838804483413696,
0.06576161831617355,
0.036485906690359116,
-0.01243057195097208,
-0.05751405283808708,
0.08437132835388184,
-0.0016409088857471943,
-0.03694742172956467,
0.06387316435575485,
-0.11778134107589722,
0.011128457263112068,
-0.06970944255590439,
-0.04774976521730423,
0.06281442195177078,
-0.013686226680874825,
-0.0211295448243618,
0.001927836099639535,
0.030171668156981468,
0.005467076785862446,
-0.012227523140609264,
-0.09423775970935822,
3.215362391240835e-33,
-0.037709739059209824,
-0.03916635736823082,
0.026601992547512054,
0.0649646520614624,
0.06739489734172821,
-0.013510754331946373,
-0.020154591649770737,
-0.04279430955648422,
0.017528336495161057,
-0.07443809509277344,
-0.03151926025748253,
0.06607542186975479,
-0.054019514471292496,
0.017560403794050217,
0.045536596328020096,
0.08048516511917114,
-0.017793981358408928,
0.02267645299434662,
-0.016135506331920624,
-0.057225242257118225,
0.04337063804268837,
-0.09219688922166824,
-0.00173678167629987,
-0.014898001216351986,
0.002248854376375675,
-0.024980701506137848,
0.025542160496115685,
-0.012973141856491566,
-0.028731346130371094,
0.017137233167886734,
-0.026206832379102707,
-0.04865594953298569,
0.009751725941896439,
-0.04361187666654587,
-0.028869107365608215,
0.020346516743302345,
-0.0798722431063652,
-0.02634236589074135,
0.017196480184793472,
0.0025450829416513443,
-0.038710642606019974,
0.02869255281984806,
-0.002969341352581978,
0.020074419677257538,
-0.13471528887748718,
0.007865886203944683,
0.04478836804628372,
0.0011963326251134276,
0.02602994255721569,
-0.07555024325847626,
0.05032968521118164,
-0.03707515075802803,
0.07068679481744766,
0.004015748389065266,
-0.049611687660217285,
0.024217763915657997,
-0.02194310910999775,
-0.06834089756011963,
-0.0283683929592371,
-0.04911663755774498,
0.06482977420091629,
0.049798496067523956,
-0.008689776994287968,
0.06099763140082359,
0.032180335372686386,
0.0043450044468045235,
-0.029271163046360016,
0.02616155333817005,
0.03587432950735092,
0.058139145374298096,
0.029204348102211952,
0.0378376841545105,
-0.0786624327301979,
-0.07234116643667221,
0.020113546401262283,
-0.03147028014063835,
-0.03400450944900513,
-0.07662807404994965,
0.08161251991987228,
0.017256878316402435,
-0.042058344930410385,
-0.027498116716742516,
0.022964857518672943,
-0.039184361696243286,
-0.03734998032450676,
-0.04198363795876503,
-0.007762709632515907,
-0.06578546017408371,
-0.10775598883628845,
-0.0368911549448967,
-0.05001526698470116,
-0.05674634501338005,
-0.04472598806023598,
-0.013502990826964378,
0.010052863508462906,
-3.3835186948358867e-33,
-0.1338893175125122,
0.0684925764799118,
0.052332088351249695,
0.03351802006363869,
0.003903896315023303,
0.006918426603078842,
-0.05811046436429024,
0.05120265483856201,
0.09363320469856262,
-0.036895740777254105,
-0.010511263273656368,
-0.051205363124608994,
-0.0015177220338955522,
0.009522686712443829,
0.053536951541900635,
-0.02189747616648674,
-0.0041258675046265125,
-0.0019108847482129931,
-0.04621893912553787,
-0.04660504311323166,
0.0007586191059090197,
0.04014487937092781,
-0.03574226424098015,
0.0015365170547738671,
-0.00603160634636879,
0.00600002147257328,
-0.04476584494113922,
0.11711706221103668,
0.008689150214195251,
-0.02881619893014431,
-0.11497821658849716,
0.013981163501739502,
-0.012163530103862286,
-0.03755565732717514,
-0.08518005162477493,
0.0857694000005722,
0.034654807299375534,
0.05807812139391899,
-0.06724289059638977,
0.08131077885627747,
0.0129753602668643,
0.01799061708152294,
-0.00022897669987287372,
-0.006360197439789772,
0.013738013803958893,
0.0024449946358799934,
-0.017445843666791916,
0.04082516208291054,
0.08160251379013062,
-0.003841549623757601,
0.015793077647686005,
0.06229965388774872,
-0.05976991355419159,
0.12171930819749832,
-0.10276979953050613,
-0.003788662375882268,
-0.006218758411705494,
-0.023202020674943924,
0.056656647473573685,
-0.002821501577273011,
-0.08311022073030472,
-0.012712385505437851,
0.04498439282178879,
-0.03156329691410065,
-0.07498703896999359,
-0.03803657367825508,
-0.005757995881140232,
0.03029857575893402,
0.05686108395457268,
0.08264300227165222,
-0.029574399814009666,
-0.008581260219216347,
0.02632107585668564,
0.025273891165852547,
-0.01039784774184227,
-0.10115153342485428,
0.01319604367017746,
-0.03433028236031532,
-0.03714093938469887,
0.008114819414913654,
0.1427207589149475,
0.06355390697717667,
0.043153610080480576,
-0.014277435839176178,
0.052504703402519226,
-0.002420334843918681,
0.13280150294303894,
0.0262474212795496,
0.05235934630036354,
0.01170930452644825,
-0.053142186254262924,
-0.013192341662943363,
0.0016099122585728765,
0.04778989776968956,
-0.08337593078613281,
-4.1902609382304945e-8,
-0.05516660213470459,
0.05085169896483421,
0.06089654937386513,
0.08460391312837601,
0.0708048865199089,
0.035829417407512665,
0.007297139149159193,
0.09892195463180542,
-0.039143938571214676,
-0.030977966263890266,
0.04746735841035843,
0.06239529326558113,
-0.0046910070814192295,
-0.03686054050922394,
0.002504719654098153,
0.05491851270198822,
0.018677396699786186,
-0.018355829641222954,
-0.017145995050668716,
-0.037590142339468,
-0.035994116216897964,
0.026322077959775925,
0.05213203653693199,
0.004849947057664394,
0.09884746372699738,
-0.0680665597319603,
-0.06923232972621918,
0.0009329162421636283,
0.04435502737760544,
0.00599410617724061,
0.020637748762965202,
-0.007895799353718758,
0.008163818158209324,
0.03575081378221512,
-0.02340024709701538,
0.06386573612689972,
0.027585553005337715,
-0.04769875109195709,
-0.00524088554084301,
0.047097064554691315,
0.03072773665189743,
0.08367624878883362,
-0.012047910131514072,
0.07504501938819885,
0.027088718488812447,
0.04688934236764908,
0.004265846684575081,
-0.07097449153661728,
0.03539823368191719,
0.022704020142555237,
0.047359105199575424,
-0.04544765502214432,
-0.04996144771575928,
0.04229588061571121,
0.04913248494267464,
-0.07775762677192688,
-0.03639630600810051,
-0.011656155809760094,
0.03416762501001358,
0.03818294033408165,
0.04923707991838455,
-0.008153459057211876,
-0.025331007316708565,
0.0561804436147213
] | 0 | 0 | 0 | 0 |
|
Second-Order Min-Max Optimization with Lazy Hessians | https://openreview.net/forum?id=ijbA5swmoK | [
"Lesi Chen",
"Chengchang Liu",
"Jingzhao Zhang"
] | Oral | This paper studies second-order methods for convex-concave minimax optimization.
Monteiro & Svaiter (2012) proposed a method to solve the problem with an optimal iteration complexity of
$\mathcal{O}(\epsilon^{-3/2})$ to find an $\epsilon$-saddle point. However, it is unclear whether the
computational complexity, $\mathcal{O}((N+ d^2) d \epsilon^{-2/3})$, can be improved. In the above, we follow Doikov et al. (2023) and assume the complexity of obtaining a first-order oracle as $N$ and the complexity of obtaining a second-order oracle as $dN$.
In this paper, we show that the computation cost can be reduced by reusing Hessian across iterations. Our methods take the overall computational complexity of $\tilde{\mathcal{O}}( (N+d^2)(d+ d^{2/3}\epsilon^{-2/3}))$, which improves those of previous methods by a factor of $d^{1/3}$.
Furthermore, we generalize our method to strongly-convex-strongly-concave minimax problems and establish the complexity of $\tilde{\mathcal{O}}((N+d^2) (d + d^{2/3} \kappa^{2/3}) )$ when the condition number of the problem is $\kappa$, enjoying a similar speedup upon the state-of-the-art method.
Numerical experiments on both real and synthetic datasets also verify the efficiency of our method. | min-max optimization; second-order methods; computational complexity | We propose novel second-order methods for min-max optimization that are provably better than existing optimal methods | 3,596 | 2410.09568 | [
-0.06853403151035309,
-0.06910460442304611,
-0.025730736553668976,
-0.03332200273871422,
0.03814961761236191,
-0.12416665256023407,
-0.03258806839585304,
0.00890367105603218,
0.04236001893877983,
0.03714897111058235,
-0.020543035119771957,
0.10934267193078995,
0.011339661665260792,
0.0059479582123458385,
-0.04273436591029167,
0.026700355112552643,
0.02610340714454651,
0.06888739764690399,
0.022039689123630524,
-0.007627298589795828,
0.004312901291996241,
-0.0034938042517751455,
-0.035174064338207245,
-0.04175180569291115,
0.03608600050210953,
-0.035497430711984634,
0.0038553953636437654,
0.035722628235816956,
-0.014350158162415028,
0.04952143505215645,
0.008470028638839722,
-0.03215779736638069,
0.03274798020720482,
-0.07263776659965515,
0.010703470557928085,
0.09173579514026642,
0.04189639911055565,
-0.0006988789536990225,
-0.031144848093390465,
0.04622150585055351,
-0.05921733379364014,
0.037569817155599594,
0.03174006938934326,
0.030268685892224312,
0.050340671092271805,
-0.027668774127960205,
0.009260938502848148,
-0.06954213231801987,
0.05885092169046402,
-0.07011744379997253,
0.005099752452224493,
0.0020553171634674072,
-0.02097475901246071,
-0.04796791449189186,
-0.0379897765815258,
-0.022165218368172646,
-0.022610075771808624,
0.019523292779922485,
-0.016635363921523094,
0.054556380957365036,
0.027420789003372192,
-0.09882120788097382,
-0.013771318830549717,
-0.009070276282727718,
-0.004063678439706564,
-0.02580191381275654,
0.04401387646794319,
-0.08314329385757446,
-0.06552795320749283,
0.10264142602682114,
-0.044183216989040375,
0.04093734920024872,
-0.03422736003994942,
-0.019614892080426216,
0.056441374123096466,
-0.05322016030550003,
0.1290217489004135,
0.02475438453257084,
0.026199309155344963,
-0.1222706288099289,
0.03868121653795242,
0.06310324370861053,
-0.026630917564034462,
-0.016192061826586723,
0.00959579274058342,
-0.039034001529216766,
-0.007650545798242092,
0.012133411131799221,
0.14109142124652863,
-0.013042210601270199,
-0.02307923696935177,
0.0026835156604647636,
-0.1568853110074997,
-0.043347664177417755,
-0.03079485520720482,
-0.05990714579820633,
0.06813999265432358,
0.09364692866802216,
0.01734747551381588,
0.09015155583620071,
-0.006458972115069628,
0.013465802185237408,
-0.0062014576978981495,
-0.04565461724996567,
0.05460774526000023,
-0.07703644782304764,
0.0024766481947153807,
0.056337177753448486,
0.012857580557465553,
-0.0478811040520668,
0.05741416662931442,
0.01773848570883274,
0.04815724119544029,
0.002756084781140089,
0.023274080827832222,
0.04983429238200188,
0.01264461874961853,
-0.05264860764145851,
0.04285183176398277,
0.05539887771010399,
-0.026267964392900467,
-0.03471247851848602,
0.08130623400211334,
0.0008402739767916501,
0.00798881333321333,
0.06873847544193268,
-0.03284141793847084,
4.293427453998652e-34,
-0.010049077682197094,
0.1244397684931755,
0.06744817644357681,
-0.06438097357749939,
-0.01721588708460331,
0.07186431437730789,
0.06405582278966904,
0.0014450751477852464,
-0.04742132127285004,
0.024591438472270966,
-0.015288176946341991,
-0.07568740099668503,
-0.024736367166042328,
0.0032723371405154467,
0.11605974286794662,
0.029108459129929543,
-0.0012892669765278697,
-0.007022489327937365,
-0.0401710644364357,
-0.08788879960775375,
0.11221116036176682,
-0.059963397681713104,
0.012474820017814636,
-0.043934110552072525,
-0.03256496042013168,
-0.04983152449131012,
0.10401780903339386,
-0.028003288432955742,
-0.05895821377635002,
0.023894667625427246,
-0.027144253253936768,
-0.08676803857088089,
-0.00926588661968708,
0.07404579967260361,
0.032314836978912354,
-0.002226646291092038,
0.03366239741444588,
-0.018452877178788185,
0.003301723627373576,
-0.018245717510581017,
-0.025957657024264336,
0.07075826823711395,
0.048730865120887756,
-0.08095994591712952,
-0.10604475438594818,
-0.035987745970487595,
0.04406489059329033,
0.08720705658197403,
0.07925596088171005,
-0.08384936302900314,
-0.01255556009709835,
0.03603934869170189,
0.013293451629579067,
0.010712387040257454,
-0.026971327140927315,
-0.0001746417983667925,
-0.010516630485653877,
-0.02540191449224949,
0.03775594010949135,
0.051013823598623276,
0.010428444482386112,
-0.04034062847495079,
-0.00998846348375082,
-0.011059271171689034,
-0.01163802482187748,
0.056716836988925934,
-0.0036118680145591497,
-0.008950397372245789,
0.04429630562663078,
-0.0372992604970932,
-0.0018119486048817635,
-0.02502557262778282,
0.003483777865767479,
-0.09753970801830292,
0.0773405134677887,
-0.023008137941360474,
0.11206857115030289,
-0.06425876170396805,
0.04047601297497749,
-0.09577323496341705,
-0.07974227517843246,
0.06570214778184891,
-0.004373309202492237,
-0.025388572365045547,
-0.07842820137739182,
-0.03134152665734291,
0.021197298541665077,
-0.0017826728289946914,
-0.1001807153224945,
-0.04417717456817627,
-0.12117934972047806,
-0.017249947413802147,
-0.0016457702731713653,
-0.010332905687391758,
-0.019966525956988335,
-1.4167923006484017e-33,
-0.049655601382255554,
-0.09180855751037598,
0.02239922247827053,
-0.009943674318492413,
0.017752353101968765,
-0.08362804353237152,
-0.08174461871385574,
0.01743457280099392,
-0.06190077215433121,
-0.11997371166944504,
-0.015919873490929604,
-0.013959744945168495,
0.02123848907649517,
0.023692773655056953,
0.06494738906621933,
0.01676037907600403,
-0.09920772165060043,
0.03666292503476143,
0.03331200033426285,
0.03320618346333504,
0.10717056691646576,
0.048435941338539124,
0.0014592346269637346,
0.0513240285217762,
-0.023918434977531433,
-0.004481084179133177,
-0.015218258835375309,
0.08351729065179825,
-0.05006750300526619,
0.02719351276755333,
-0.033813368529081345,
0.045591358095407486,
-0.031164759770035744,
0.09303572028875351,
-0.026857485994696617,
0.08200277388095856,
-0.008975864388048649,
0.03506447374820709,
-0.03367069736123085,
0.12550224363803864,
0.01702074706554413,
0.003133100690320134,
0.010449239052832127,
0.00032079682569019496,
0.03851596266031265,
0.021297108381986618,
-0.04326837882399559,
0.013677550479769707,
-0.00957629643380642,
0.0653582215309143,
-0.004738962277770042,
-0.10438766330480576,
-0.04406506568193436,
0.0980873703956604,
-0.0395624078810215,
0.03906041756272316,
-0.0853879451751709,
0.0424046516418457,
0.03727344051003456,
-0.07207280397415161,
-0.0540129654109478,
0.02854214608669281,
-0.08454453200101852,
0.0831037238240242,
0.03423925116658211,
0.033378321677446365,
-0.028670968487858772,
-0.031265802681446075,
0.024269554764032364,
0.05377047508955002,
-0.11912281811237335,
-0.010885866358876228,
0.017595967277884483,
0.003969050012528896,
-0.041697658598423004,
-0.016992652788758278,
0.070048026740551,
0.023000003769993782,
-0.02724895440042019,
0.020795084536075592,
0.02092086710035801,
0.032443128526210785,
0.08160706609487534,
0.008674434386193752,
-0.00697736581787467,
0.027331754565238953,
0.08613277226686478,
0.026431558653712273,
-0.014667372219264507,
-0.053189363330602646,
-0.053644511848688126,
0.03593944385647774,
0.06773527711629868,
-0.029593665152788162,
-0.03853233531117439,
-4.373952222636035e-8,
-0.006436011753976345,
-0.07259073108434677,
0.030140038579702377,
-0.008550051599740982,
0.06888527423143387,
-0.0449858084321022,
-0.017369916662573814,
0.0429999865591526,
-0.06234850361943245,
-0.01590951718389988,
0.02916283905506134,
0.03891400620341301,
0.007539255544543266,
-0.04288391023874283,
-0.08253297954797745,
0.03422215208411217,
0.03606545552611351,
-0.05063994601368904,
-0.06284572184085846,
-0.050878994166851044,
-0.07909630984067917,
0.06699500232934952,
0.028014905750751495,
-0.01106267049908638,
0.03956179320812225,
-0.05650710314512253,
-0.003786307293921709,
-0.04849286004900932,
0.09511704742908478,
0.04544032737612724,
-0.06483876705169678,
0.030884793028235435,
-0.005985686089843512,
0.051030825823545456,
-0.04553568363189697,
0.04110340029001236,
-0.053911346942186356,
0.04124313220381737,
-0.05749005824327469,
-0.03428807109594345,
0.010440940037369728,
0.0710277333855629,
-0.0205233171582222,
0.01625404879450798,
0.09657242149114609,
0.004150021821260452,
-0.10206788778305054,
-0.044431429356336594,
0.026194654405117035,
-0.012752378359436989,
0.02768174186348915,
0.007304313592612743,
0.004695652052760124,
0.05406700074672699,
-0.028847048059105873,
-0.003656370798125863,
-0.08083414286375046,
0.009944770485162735,
0.00801384262740612,
0.014448298141360283,
0.018082326278090477,
0.009164953604340553,
0.06929288059473038,
-0.00332050072029233
] | 0 | 0 | 0 | 0 |
|
Composing Unbalanced Flows for Flexible Docking and Relaxation | https://openreview.net/forum?id=gHLWTzKiZV | [
"Gabriele Corso",
"Vignesh Ram Somnath",
"Noah Getz",
"Regina Barzilay",
"Tommi Jaakkola",
"Andreas Krause"
] | Oral | Diffusion models have emerged as a successful approach for molecular docking, but they often cannot model protein flexibility or generate nonphysical poses. We argue that both these challenges can be tackled by framing the problem as a transport between distributions. Still, existing paradigms lack the flexibility to define effective maps between such complex distributions. To address this limitation we propose Unbalanced Flow Matching, a generalization of Flow Matching (FM) that allows trading off sample efficiency with approximation accuracy and enables more accurate transport. Empirically, we apply Unbalanced FM on flexible docking and structure relaxation, demonstrating our ability to model protein flexibility and generate energetically favorable poses. On the PDBBind docking benchmark, our method FlexDock improves the docking performance while increasing the proportion of energetically favorable poses from 30% to 73%. | molecular docking, flow matching, structure relaxation, unbalanced transport | A new generalized flow matching paradigm and its applications to flexible docking and relaxation | 3,566 | null | [
-0.048352498561143875,
-0.061301469802856445,
0.018526878207921982,
-0.026991818100214005,
0.038172025233507156,
-0.06978583335876465,
-0.02363753877580166,
-0.008532254956662655,
0.0754161849617958,
-0.025703852996230125,
-0.06763815879821777,
-0.045347269624471664,
-0.02807178720831871,
0.10377084463834763,
-0.008448142558336258,
-0.05835561826825142,
0.008129887282848358,
0.08838821202516556,
-0.012551325373351574,
0.10716000944375992,
-0.09101882576942444,
0.02871456742286682,
-0.011746411211788654,
0.00681300600990653,
-0.038508340716362,
0.014416159130632877,
0.04029806703329086,
-0.09348804503679276,
0.007458245847374201,
-0.07814711332321167,
0.05655433237552643,
0.11406225711107254,
-0.08511339128017426,
-0.03393006697297096,
0.0072554019279778,
0.05573447793722153,
-0.05853377655148506,
-0.02361457236111164,
-0.05371265113353729,
-0.05899844691157341,
0.08682822436094284,
0.03974328190088272,
0.055452849715948105,
0.010260910727083683,
0.026969723403453827,
0.007886623032391071,
-0.04800025746226311,
-0.013481108471751213,
0.05598067864775658,
0.07790057361125946,
-0.008046100847423077,
-0.0083012031391263,
-0.0359516479074955,
0.07156258076429367,
-0.047381043434143066,
0.03265279904007912,
-0.013221535831689835,
-0.05888545140624046,
0.005369602236896753,
-0.035869237035512924,
0.05631216615438461,
-0.07572007179260254,
0.03192587569355965,
-0.005989149212837219,
0.01093415729701519,
-0.06434258073568344,
0.04765333607792854,
0.11003374308347702,
0.04377973452210426,
0.05633585900068283,
-0.012070278637111187,
0.08754883706569672,
-0.0945439264178276,
-0.036946408450603485,
0.06865577399730682,
0.008020373992621899,
-0.008486992679536343,
0.043545082211494446,
0.04980970174074173,
0.0014987278264015913,
-0.03131634369492531,
-0.015108448453247547,
0.04552329704165459,
0.011833949014544487,
0.01572864130139351,
-0.009692635387182236,
-0.029550716280937195,
0.017224900424480438,
0.02414531446993351,
-0.07846392691135406,
0.0020420649088919163,
0.06227729469537735,
-0.022904621437191963,
-0.055967945605516434,
0.01580350659787655,
0.014912745915353298,
0.006859390530735254,
0.006341060157865286,
0.02044176310300827,
0.03935495391488075,
0.01573462411761284,
0.062181055545806885,
0.02538723312318325,
-0.036960963159799576,
0.08396735787391663,
-0.10591284930706024,
0.07186217606067657,
0.0017414243193343282,
0.01656084880232811,
-0.0035443915985524654,
0.04781802371144295,
-0.015240185894072056,
0.11681751906871796,
0.025432273745536804,
-0.014394371770322323,
0.0003520044847391546,
-0.07693155854940414,
-0.028582166880369186,
-0.014977470971643925,
0.04482996091246605,
-0.0697263777256012,
0.0024055810645222664,
0.005543278995901346,
-0.04953155294060707,
-0.05081634223461151,
0.060975197702646255,
-0.06415310502052307,
1.3902277813587489e-34,
0.010234521701931953,
-0.04139314964413643,
0.07302021980285645,
0.03896602243185043,
0.06690750271081924,
-0.06414996832609177,
-0.04279562830924988,
-0.04129111021757126,
-0.029270391911268234,
0.016607431694865227,
-0.0808788612484932,
0.03190402686595917,
-0.00655282661318779,
0.044038187712430954,
-0.04310518503189087,
-0.07581253349781036,
-0.038109805434942245,
0.10182484984397888,
-0.0447145514190197,
0.022835208103060722,
-0.0027335546910762787,
-0.05050108954310417,
-0.01792539656162262,
-0.06072147935628891,
0.07194802910089493,
0.05965099111199379,
-0.07065370678901672,
0.019222315400838852,
-0.01455220952630043,
0.02804712951183319,
-0.07464605569839478,
-0.035461489111185074,
-0.07061102241277695,
0.011010320857167244,
0.09024100005626678,
-0.016690094023942947,
0.04574863612651825,
-0.03044198267161846,
-0.028557533398270607,
0.0030370510648936033,
-0.023253705352544785,
-0.012135317549109459,
-0.010901223868131638,
-0.05752425640821457,
-0.06368878483772278,
-0.015317265875637531,
0.040851231664419174,
0.05971004068851471,
-0.07161960750818253,
0.007521101273596287,
0.06577545404434204,
-0.033258989453315735,
-0.019214944913983345,
-0.027576714754104614,
-0.09414836764335632,
-0.08765078336000443,
-0.024753835052251816,
0.015927884727716446,
-0.045588474720716476,
0.14242608845233917,
-0.0619192011654377,
0.045753542333841324,
-0.10296528786420822,
-0.040454115718603134,
0.08258552104234695,
0.05096409469842911,
-0.08941221237182617,
0.05651659518480301,
0.040526267141103745,
0.07529197633266449,
-0.009346412494778633,
0.040239084511995316,
-0.021080518141388893,
-0.07222313433885574,
0.06153280287981033,
0.020369933918118477,
0.04788520559668541,
-0.017990242689847946,
0.03494894877076149,
0.056057464331388474,
-0.1276407539844513,
0.01820247620344162,
-0.06061987951397896,
0.008587844669818878,
-0.10381986945867538,
-0.13333304226398468,
0.04084702208638191,
-0.036307595670223236,
-0.11213313788175583,
-0.019000964239239693,
0.03300241753458977,
-0.02908344753086567,
0.010795723646879196,
-0.03928842395544052,
0.01290969643741846,
-2.1603062528679223e-33,
0.031329236924648285,
0.0009138968889601529,
0.06545639038085938,
-0.005312164314091206,
0.040391333401203156,
0.03155334293842316,
0.10102683305740356,
-0.030997617170214653,
0.052474815398454666,
-0.08779633045196533,
-0.03673234209418297,
0.03431008383631706,
-0.01158086210489273,
-0.029011031612753868,
0.04337587580084801,
0.012330095283687115,
-0.08345901966094971,
-0.027212437242269516,
0.03692084178328514,
0.04952302202582359,
0.010080709122121334,
0.001773958676494658,
0.035648416727781296,
0.04446178674697876,
-0.053757183253765106,
0.0697627067565918,
0.03157087042927742,
-0.0470513179898262,
-0.01757282204926014,
0.03164507821202278,
-0.045302052050828934,
-0.012785693630576134,
-0.10069490969181061,
0.02886378951370716,
-0.04935753718018532,
0.09648361057043076,
-0.03730706125497818,
0.09201223403215408,
0.03732331469655037,
-0.027125749737024307,
-0.024811530485749245,
-0.00009043379395734519,
-0.09492477029561996,
0.08701222389936447,
0.07308569550514221,
0.006590354721993208,
-0.069015271961689,
0.018137985840439796,
-0.009038912132382393,
-0.013586445711553097,
-0.03409156575798988,
0.03458145633339882,
-0.022754443809390068,
0.006165198516100645,
-0.035356320440769196,
-0.05200117453932762,
-0.06523019075393677,
-0.02654680237174034,
0.01972467079758644,
-0.0075153023935854435,
-0.044728901237249374,
-0.047346778213977814,
-0.029843514785170555,
0.022338109090924263,
-0.06106136739253998,
-0.02517325058579445,
0.026240594685077667,
-0.03824981674551964,
-0.0391981303691864,
0.04068450257182121,
0.052574753761291504,
0.002858982887119055,
0.08126906305551529,
0.08064919710159302,
0.05674032121896744,
-0.03811350464820862,
0.08893934637308121,
-0.07978998124599457,
0.07958287000656128,
-0.03129024803638458,
-0.046875957399606705,
0.0763683021068573,
0.04222814738750458,
-0.036321934312582016,
0.05266400799155235,
0.038082610815763474,
0.03301123157143593,
-0.04165587201714516,
0.0016939317574724555,
0.060822706669569016,
-0.012639437802135944,
-0.031077301129698753,
0.011142377741634846,
0.07338334619998932,
-0.010468481108546257,
-3.855898356164289e-8,
-0.005271183326840401,
0.02895510010421276,
0.07100480794906616,
-0.014494072645902634,
-0.03871695324778557,
0.10959628969430923,
-0.016757285222411156,
0.04142887890338898,
0.024976762011647224,
0.013025362975895405,
0.059425272047519684,
-0.022618163377046585,
-0.03061044029891491,
-0.02954997681081295,
0.00638549355790019,
0.10737118124961853,
-0.03425411507487297,
0.01155490055680275,
-0.06371407210826874,
-0.03314054012298584,
-0.07796306163072586,
0.06180236488580704,
0.03422525152564049,
0.003251026850193739,
0.0316399447619915,
-0.05761919915676117,
-0.002612900221720338,
-0.009097781963646412,
0.0838804841041565,
-0.06512311100959778,
-0.02536899782717228,
-0.019636794924736023,
-0.03294257074594498,
0.0793239176273346,
0.012038327753543854,
0.003249696921557188,
-0.0699777901172638,
-0.004136246163398027,
-0.05700043588876724,
0.05603870004415512,
-0.04357496276497841,
0.09140602499246597,
-0.06701955199241638,
-0.07850538939237595,
-0.014928575605154037,
0.06793270260095596,
0.03125490993261337,
0.07156436145305634,
-0.0012820796109735966,
0.057790275663137436,
-0.006440332625061274,
-0.04264650493860245,
0.013558482751250267,
0.017615627497434616,
-0.010659899562597275,
0.07695842534303665,
-0.10519251227378845,
-0.04234139621257782,
0.042617131024599075,
-0.008923610672354698,
-0.018521815538406372,
0.03040224127471447,
-0.027952853590250015,
-0.01620735041797161
] | 0 | 0 | 0 | 0 |
|
Learning Distributions of Complex Fluid Simulations with Diffusion Graph Networks | https://openreview.net/forum?id=uKZdlihDDn | [
"Mario Lino Valencia",
"Tobias Pfaff",
"Nils Thuerey"
] | Oral | Physical systems with complex unsteady dynamics, such as fluid flows, are often poorly represented by a single mean solution. For many practical applications, it is crucial to access the full distribution of possible states, from which relevant statistics (e.g., RMS and two-point correlations) can be derived. Here, we propose a graph-based latent diffusion model that enables direct sampling of states from their equilibrium distribution, given a mesh discretization of the system and its physical parameters. This allows for the efficient computation of flow statistics without running long and expensive numerical simulations. The graph-based structure enables operations on unstructured meshes, which is critical for representing complex geometries with spatially localized high gradients, while latent-space diffusion modeling with a multi-scale GNN allows for efficient learning and inference of entire distributions of solutions. A key finding of our work is that the proposed networks can accurately learn full distributions even when trained on incomplete data from relatively short simulations. We apply this method to a range of fluid dynamics tasks, such as predicting pressure distributions on 3D wing models in turbulent flow, demonstrating both accuracy and computational efficiency in challenging scenarios. The ability to directly sample accurate solutions, and capturing their diversity from short ground-truth simulations, is highly promising for complex scientific modeling tasks. | Graph Neural Networks, Diffusion Models, Physics Simulations | We propose an efficient graph-based latent diffusion model, which allows us to directly sample unsteady flow states from their equilibrium distribution given a mesh discretisation of the system and its physical parameters. | 3,559 | null | [
-0.022784044966101646,
-0.08193778246641159,
0.03027273900806904,
0.029804885387420654,
0.03908372297883034,
-0.05220054090023041,
-0.047805123031139374,
-0.05834703519940376,
0.05971504747867584,
-0.006792972330003977,
-0.0563441663980484,
-0.059334784746170044,
-0.032197486609220505,
0.10615704208612442,
-0.012036509811878204,
-0.043029412627220154,
-0.02026957832276821,
0.05180332809686661,
-0.05930265039205551,
0.009135362692177296,
0.01989879086613655,
0.08044867217540741,
-0.07918655127286911,
0.004839681554585695,
0.11999190598726273,
0.022953882813453674,
0.05218116194009781,
-0.0462183840572834,
-0.003820215119048953,
-0.0022063707001507282,
0.11085925251245499,
-0.03807913511991501,
-0.08090052008628845,
0.0037130212876945734,
-0.04385314881801605,
-0.03658491000533104,
-0.04893247038125992,
0.035490065813064575,
-0.05196528509259224,
0.053872816264629364,
0.026056408882141113,
0.08650621771812439,
0.05269705504179001,
0.038032133132219315,
0.11931765079498291,
0.0320538692176342,
-0.013075942173600197,
-0.06781205534934998,
0.03170204535126686,
0.09101694822311401,
-0.056315626949071884,
-0.03211820125579834,
-0.0468854196369648,
-0.013653689995408058,
0.04501473903656006,
-0.02619459666311741,
0.05125977098941803,
-0.05918710678815842,
0.00046835304237902164,
-0.05186957120895386,
0.04368823021650314,
-0.09944137185811996,
0.011405841447412968,
-0.002034792210906744,
-0.05303144454956055,
-0.009897079318761826,
0.017645463347434998,
0.10225892066955566,
0.01162461843341589,
0.024532265961170197,
0.006726941559463739,
0.05246046558022499,
-0.07755989581346512,
0.023346800357103348,
0.06100282073020935,
0.018162649124860764,
0.015995346009731293,
-0.008484184741973877,
-0.07688961178064346,
-0.04829899221658707,
0.07433798909187317,
0.060280393809080124,
0.043260909616947174,
-0.01365306880325079,
0.003271948080509901,
0.01881367713212967,
-0.03717254474759102,
0.06978801637887955,
0.010142343118786812,
-0.030771272256970406,
-0.02654273435473442,
0.10094663500785828,
-0.0666387602686882,
-0.039247166365385056,
0.04716220125555992,
0.06661991029977798,
0.08092537522315979,
-0.047359857708215714,
0.036900799721479416,
0.09307559579610825,
-0.03770183399319649,
-0.06244225800037384,
0.0675862655043602,
0.012809106148779392,
0.010689598508179188,
0.004605185240507126,
0.07615189999341965,
0.023434581235051155,
0.038173772394657135,
-0.06190890818834305,
-0.0368288978934288,
0.030171703547239304,
-0.04098131135106087,
0.043278660625219345,
0.03151930868625641,
-0.0687311515212059,
0.011546512134373188,
-0.10658329725265503,
0.003687227377668023,
0.10016389191150665,
-0.11559490114450455,
0.017635159194469452,
-0.03110167384147644,
0.01338367722928524,
0.02432767115533352,
0.029806451871991158,
-0.046734798699617386,
-1.7722246907966136e-33,
0.010754715651273727,
-0.008610328659415245,
0.03208820894360542,
0.10332009941339493,
0.0857519581913948,
0.04482506215572357,
-0.0019911997951567173,
-0.03120996430516243,
0.10019482672214508,
0.04162408038973808,
-0.1155548244714737,
0.041353873908519745,
-0.09560548514127731,
0.01286214217543602,
-0.00043671324965544045,
-0.015849079936742783,
-0.01340950932353735,
-0.015219567343592644,
0.019241029396653175,
0.004061681218445301,
0.09436433017253876,
-0.11310303211212158,
-0.021876279264688492,
-0.040903475135564804,
-0.03633647412061691,
0.031553104519844055,
0.04172280430793762,
-0.00007807044312357903,
0.03065512515604496,
0.007509974762797356,
-0.05365278944373131,
0.030196350067853928,
-0.03907597064971924,
0.019722238183021545,
0.01813673973083496,
0.033237162977457047,
0.03191978856921196,
-0.032507073134183884,
-0.01549648866057396,
-0.06740254163742065,
-0.03843000903725624,
-0.0080933952704072,
-0.016707811504602432,
-0.002391247544437647,
-0.09749399870634079,
-0.041281744837760925,
-0.029081571847200394,
-0.0308594461530447,
0.014781978912651539,
-0.08280360698699951,
0.015207753516733646,
0.006413836497813463,
0.018840555101633072,
-0.015529458411037922,
-0.017482243478298187,
0.02743469923734665,
0.06973745673894882,
-0.02382308803498745,
0.023243287578225136,
0.02228407748043537,
-0.012054714374244213,
0.004321509040892124,
-0.08031812310218811,
-0.003456695703789592,
0.04516473412513733,
0.05524785816669464,
-0.029778551310300827,
0.16003930568695068,
0.06765913218259811,
-0.050102002918720245,
0.00017452328756917268,
0.02903391793370247,
-0.052312836050987244,
-0.09719246625900269,
0.012544592842459679,
0.02825295925140381,
0.011183240450918674,
-0.03814342990517616,
-0.030405718833208084,
0.06407011300325394,
-0.05843612179160118,
-0.023021791130304337,
-0.0356583371758461,
0.008826544508337975,
-0.015554012730717659,
-0.03904449939727783,
0.029236115515232086,
-0.07038074731826782,
-0.04059695079922676,
-0.0668574795126915,
-0.09546797722578049,
-0.06977270543575287,
0.03281192108988762,
-0.009194349870085716,
-0.008256559260189533,
-4.843774623599454e-34,
-0.06772907078266144,
0.0931636169552803,
0.01873866468667984,
0.0498930849134922,
0.0457543209195137,
-0.011385263875126839,
-0.029244044795632362,
-0.0011051759356632829,
0.05601010099053383,
-0.06334978342056274,
-0.0727839469909668,
-0.04919806867837906,
-0.003937747795134783,
-0.041162244975566864,
0.1335388720035553,
-0.0015884886961430311,
-0.03458617255091667,
-0.0385863296687603,
0.013544874265789986,
-0.036024391651153564,
-0.0341571569442749,
-0.031800203025341034,
-0.09353441745042801,
-0.02209477499127388,
-0.007798593491315842,
0.008197234943509102,
-0.015962084755301476,
-0.01356737595051527,
-0.005844250321388245,
0.0758982002735138,
-0.033350519835948944,
-0.03958386927843094,
0.0004241073620505631,
0.002177545800805092,
-0.030078012496232986,
0.057362645864486694,
0.03649774566292763,
0.04546218365430832,
-0.005401939153671265,
-0.045124877244234085,
0.06767994165420532,
-0.1064431369304657,
-0.05652739480137825,
-0.003947526682168245,
-0.0062226466834545135,
0.05456821620464325,
0.014082655310630798,
-0.012994864024221897,
0.03844689577817917,
0.04114501550793648,
-0.005450427997857332,
0.016825878992676735,
-0.03305891901254654,
0.03227679431438446,
-0.025569545105099678,
-0.06663015484809875,
0.01201939582824707,
0.04107227921485901,
0.0935833528637886,
-0.05898585543036461,
-0.09982020407915115,
-0.06931650638580322,
-0.043480854481458664,
0.013245999813079834,
-0.07931504398584366,
-0.04151595011353493,
-0.050662409514188766,
-0.02341775968670845,
0.0290271807461977,
0.06696661561727524,
0.015842372551560402,
0.024540117010474205,
0.04085824638605118,
-0.021854642778635025,
0.05663105472922325,
-0.09567582607269287,
0.018338508903980255,
-0.014567947015166283,
-0.03727167472243309,
0.0055358209647238255,
0.06720926612615585,
0.0218978114426136,
0.05023669824004173,
0.04562099650502205,
0.12426276504993439,
-0.013662368059158325,
0.040106527507305145,
-0.012988338246941566,
0.03063311241567135,
0.03467392921447754,
0.018236879259347916,
0.04304123669862747,
-0.011942129582166672,
-0.013717669062316418,
-0.08185378462076187,
-4.0403303813718594e-8,
-0.13606172800064087,
0.03804351016879082,
0.08235147595405579,
-0.05616305395960808,
0.006477903574705124,
0.07122461497783661,
0.034040212631225586,
0.0792047381401062,
0.01026502251625061,
-0.017018292099237442,
0.0803726315498352,
-0.09145954251289368,
-0.012364511378109455,
-0.07201100885868073,
-0.04033908247947693,
0.04923288896679878,
-0.004576385952532291,
0.009201308712363243,
0.05082443729043007,
-0.0016392642864957452,
-0.054960738867521286,
0.02663489431142807,
0.01722101867198944,
0.04859315603971481,
0.02744195982813835,
-0.050319477915763855,
-0.06894118338823318,
-0.06137240305542946,
0.09495944529771805,
-0.012582341209053993,
-0.059725817292928696,
-0.03726879879832268,
0.044548459351062775,
0.03904717415571213,
-0.00172632304020226,
0.08144019544124603,
0.07972771674394608,
-0.029598047956824303,
-0.013161075301468372,
-0.05460859835147858,
-0.0788031816482544,
0.07773783057928085,
0.03037387691438198,
-0.02524002082645893,
0.05740208923816681,
0.07757741212844849,
0.029748139902949333,
0.015269502066075802,
0.05305721238255501,
-0.0195672195404768,
-0.02887929603457451,
0.011406111530959606,
-0.024017103016376495,
0.07493553310632706,
0.019539710134267807,
0.0423576645553112,
-0.09685500711202621,
-0.01719585992395878,
0.06283789873123169,
-0.050913188606500626,
-0.0424787662923336,
0.0242941714823246,
-0.07844831049442291,
0.011240644380450249
] | 0 | 0 | 0 | 0 |
|
Training Language Models to Self-Correct via Reinforcement Learning | https://openreview.net/forum?id=CjwERcAU7w | [
"Aviral Kumar",
"Vincent Zhuang",
"Rishabh Agarwal",
"Yi Su",
"John D Co-Reyes",
"Avi Singh",
"Kate Baumli",
"Shariq Iqbal",
"Colton Bishop",
"Rebecca Roelofs",
"Lei M Zhang",
"Kay McKinney",
"Disha Shrivastava",
"Cosmin Paduraru",
"George Tucker",
"Doina Precup",
"Feryal Behbahani",
"Aleksandra Faust"
] | Oral | Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks. | language models, reinforcement learning | null | 3,518 | 2409.12917 | [
-0.017976630479097366,
-0.031271085143089294,
0.07718449831008911,
0.08750918507575989,
-0.015229729004204273,
0.05324329435825348,
0.003296456765383482,
0.024271633476018906,
0.028135132044553757,
-0.03766132518649101,
-0.05740715190768242,
-0.04895678535103798,
0.014920404180884361,
0.009463418275117874,
0.027536870911717415,
-0.05615261197090149,
0.06546401232481003,
0.0705084428191185,
-0.032374057918787,
-0.10009094327688217,
-0.001036768197081983,
0.07645341753959656,
0.11192043870687485,
0.03018401563167572,
-0.0030810581520199776,
-0.0785410925745964,
-0.045065030455589294,
-0.01966015063226223,
0.007826329208910465,
-0.023368457332253456,
0.010717778466641903,
0.0011520475381985307,
-0.01450082566589117,
-0.0015178355388343334,
-0.09899035841226578,
0.0008345502428710461,
-0.04082497954368591,
0.015541437081992626,
0.000879022351000458,
-0.006184048485010862,
-0.014787489548325539,
-0.016945013776421547,
0.03864578902721405,
0.01724301651120186,
0.06408288329839706,
-0.02575165219604969,
-0.04113869369029999,
-0.018824966624379158,
-0.05505397170782089,
0.024679865688085556,
-0.03829028829932213,
-0.04783475026488304,
-0.009829453192651272,
0.0934908390045166,
-0.02928292192518711,
0.0599747896194458,
0.10998482257127762,
0.09596769511699677,
0.05867050960659981,
-0.053487032651901245,
0.002055336022749543,
-0.11765668541193008,
-0.04516381770372391,
-0.01816699281334877,
-0.05388752743601799,
0.008529948070645332,
-0.011939983814954758,
0.06526104360818863,
-0.006229293067008257,
0.03252435848116875,
-0.036481089890003204,
0.021027976647019386,
-0.0025693504139781,
0.04909422993659973,
0.01004168950021267,
0.026896249502897263,
-0.010045642964541912,
-0.02659192867577076,
0.06634488701820374,
0.057118628174066544,
-0.050158459693193436,
0.0007174163474701345,
0.018622756004333496,
0.0009485160699114203,
0.06642106175422668,
0.0012270965380594134,
0.05618669465184212,
-0.007470262236893177,
0.13318319618701935,
0.01936260238289833,
0.05118376389145851,
-0.008573519997298717,
0.028719991445541382,
-0.017612693831324577,
-0.07281050831079483,
0.013953677378594875,
0.022162580862641335,
-0.10676079243421555,
-0.021059876307845116,
0.057504113763570786,
0.08901260793209076,
0.12261746823787689,
-0.04541940987110138,
-0.028673406690359116,
-0.05114929750561714,
0.03201562911272049,
0.03752985969185829,
0.10258106887340546,
0.02487052045762539,
-0.11606892198324203,
0.0055298469960689545,
0.01958026923239231,
-0.005638997070491314,
0.018911248072981834,
0.02041628584265709,
0.04693794995546341,
-0.0022145467810332775,
-0.02788165956735611,
-0.04346228018403053,
0.06078324466943741,
-0.03396258503198624,
-0.06774911284446716,
0.0710967630147934,
-0.03435223922133446,
0.015731563791632652,
-0.060759685933589935,
-0.0642593502998352,
1.0599572151318727e-33,
0.12132415175437927,
0.06006522849202156,
0.011123339645564556,
-0.004399692174047232,
0.020409973338246346,
0.004121268168091774,
0.018583590164780617,
0.04448861628770828,
0.004553516861051321,
-0.05096350610256195,
0.05284075811505318,
-0.02535940520465374,
-0.03166815638542175,
0.06297619640827179,
0.014930197969079018,
0.04741028696298599,
-0.0605178102850914,
0.008590404875576496,
0.000598766200710088,
0.0000951275578700006,
0.08892663568258286,
-0.02361162379384041,
0.03971260413527489,
-0.1024860069155693,
0.06615041196346283,
0.06717070192098618,
0.08017978072166443,
-0.011581753380596638,
-0.023265210911631584,
-0.0007780282176099718,
-0.031038273125886917,
-0.005971564911305904,
0.04677524417638779,
0.056581657379865646,
0.02808196097612381,
-0.053359780460596085,
-0.02187526226043701,
-0.03284770995378494,
-0.005555863492190838,
-0.02510533295571804,
0.0024480847641825676,
0.010773432441055775,
0.05774911865592003,
0.01667834259569645,
-0.047048572450876236,
-0.10937613993883133,
-0.033949047327041626,
-0.03011218272149563,
-0.05976676195859909,
-0.050061602145433426,
-0.006384654901921749,
0.13589541614055634,
-0.05915073677897453,
-0.049952711910009384,
-0.06584735214710236,
0.05511019751429558,
-0.019952619448304176,
0.03703347593545914,
-0.055713798850774765,
0.008786527439951897,
0.11782315373420715,
0.03501938655972481,
-0.0014912873739376664,
0.08318020403385162,
0.039630062878131866,
0.02800157479941845,
0.036627527326345444,
0.00649225153028965,
0.037880800664424896,
-0.06287749856710434,
-0.0473635196685791,
-0.014915515668690205,
-0.012130817398428917,
-0.09037712961435318,
0.05876985564827919,
-0.0809636265039444,
0.04628581553697586,
-0.050272393971681595,
0.025978978723287582,
-0.036226749420166016,
0.0759945660829544,
-0.02403143234550953,
-0.01968185231089592,
-0.06088346615433693,
0.020947838202118874,
-0.13409914076328278,
0.012970821000635624,
-0.028804700821638107,
0.014168989844620228,
0.013757754117250443,
0.003773057134822011,
-0.03224341198801994,
-0.00619549723342061,
0.0008528088801540434,
-0.023616742342710495,
-6.785505324278153e-34,
0.0046295542269945145,
-0.048463333398103714,
-0.05478138104081154,
0.09009905904531479,
-0.05628195032477379,
-0.05493868142366409,
0.026730136945843697,
0.043156012892723083,
0.03431602194905281,
-0.05559860169887543,
-0.07131858915090561,
-0.042098429054021835,
-0.014969203621149063,
0.03544413298368454,
0.010137141682207584,
-0.05175304412841797,
-0.0010653299978002906,
-0.0033591454848647118,
0.0022063888609409332,
-0.0006933509139344096,
0.018663350492715836,
0.022932693362236023,
-0.04619893059134483,
0.10954033583402634,
-0.02246098779141903,
0.01966269128024578,
-0.028199931606650352,
0.049794215708971024,
-0.0074166348204016685,
0.08205665647983551,
-0.019998567178845406,
0.01768147572875023,
0.01077062264084816,
-0.04423605278134346,
-0.032990962266922,
0.04956764355301857,
0.058988671749830246,
-0.0361977182328701,
-0.053879670798778534,
0.12577450275421143,
0.05751611292362213,
-0.012572076171636581,
-0.09559805691242218,
-0.0056108031421899796,
-0.0314054973423481,
-0.052828963845968246,
-0.12184847146272659,
0.039057932794094086,
-0.007179111707955599,
0.04784331098198891,
-0.07567641884088516,
-0.007043907884508371,
-0.08069068193435669,
0.01840517297387123,
-0.03274784982204437,
-0.059354960918426514,
-0.02191847935318947,
-0.026785951107740402,
-0.04117843508720398,
0.008817018009722233,
-0.11150713264942169,
-0.02146405167877674,
0.0556073822081089,
-0.011191901750862598,
0.011024504899978638,
-0.0672592744231224,
-0.021609347313642502,
0.02029852569103241,
0.08517646789550781,
0.002597823040559888,
-0.00527195306494832,
0.024066291749477386,
0.043756794184446335,
-0.05663641169667244,
0.030385728925466537,
-0.04089099168777466,
-0.0017120861448347569,
-0.016467150300741196,
-0.04208159074187279,
-0.008352751843631268,
0.021441208198666573,
0.03547443076968193,
0.019973721355199814,
0.09558270871639252,
0.04342618212103844,
0.02390368841588497,
0.05150041729211807,
0.1712885946035385,
0.009416695684194565,
-0.0018946524942293763,
0.025231873616576195,
0.008611505851149559,
0.02146950550377369,
0.03494301810860634,
-0.07080002129077911,
-4.2869793048794236e-8,
-0.08589957654476166,
-0.015530459582805634,
0.06139439716935158,
0.15782156586647034,
0.0401567742228508,
-0.006580396555364132,
-0.021786512807011604,
0.03003959357738495,
-0.04196571931242943,
-0.054102249443531036,
-0.018897350877523422,
0.046420540660619736,
-0.045556724071502686,
-0.12542501091957092,
-0.01056833565235138,
0.0006316447979770601,
-0.03541341796517372,
0.05887078493833542,
-0.04488659277558327,
0.002220319351181388,
0.019217489287257195,
-0.020514795556664467,
-0.00012280119699425995,
-0.07836111634969711,
0.05319185182452202,
0.03983518108725548,
0.013094883412122726,
0.06202388182282448,
0.024764027446508408,
-0.04198438674211502,
0.05410192161798477,
-0.08341797441244125,
0.05449211597442627,
-0.03409012779593468,
0.03931066393852234,
0.04396648705005646,
0.1084272712469101,
-0.0459781289100647,
0.024165496230125427,
-0.003635329194366932,
0.009475532919168472,
0.09151556342840195,
-0.029691334813833237,
0.025475218892097473,
0.007200389634817839,
-0.05498098209500313,
-0.06891772896051407,
-0.08294305950403214,
0.04204774275422096,
-0.12659744918346405,
0.02191665954887867,
-0.040159665048122406,
-0.06287206709384918,
0.02220034971833229,
0.07161468267440796,
0.009377463720738888,
-0.005670771934092045,
-0.034913670271635056,
0.003945534583181143,
0.11950318515300751,
0.0031734283547848463,
0.010205228812992573,
-0.032747380435466766,
-0.0092852134257555
] | 0 | 0 | 0 | 0 |
|
AI as Humanity’s Salieri: Quantifying Linguistic Creativity of Language Models via Systematic Attribution of Machine Text against Web Text | https://openreview.net/forum?id=ilOEOIqolQ | [
"Ximing Lu",
"Melanie Sclar",
"Skyler Hallinan",
"Niloofar Mireshghallah",
"Jiacheng Liu",
"Seungju Han",
"Allyson Ettinger",
"Liwei Jiang",
"Khyathi Chandu",
"Nouha Dziri",
"Yejin Choi"
] | Oral | Creativity has long been considered one of the most difficult aspect of human intelligence for AI to mimic. However, the rise of Large Language Models (LLMs), like ChatGPT, has raised questions about whether AI can match or even surpass human creativity. We present CREATIVITY INDEX as the first step to quantify the linguistic creativity of a text by reconstructing it from existing text snippets on the web. CREATIVITY INDEX is motivated by the hypothesis that the seemingly remarkable creativity of LLMs may be attributable in large part to the creativity of human-written texts on the web. To compute CREATIVITY INDEX efficiently, we introduce DJ SEARCH, a novel dynamic programming algorithm that can search verbatim and near-verbatim matches of text snippets from a given document against the web. Experiments reveal that the CREATIVITY INDEX of professional human authors is on average 66.2% higher than that of LLMs, and that alignment reduces the CREATIVITY INDEX of LLMs by an average of 30.1%. In addition, we find that distinguished authors like Hemingway exhibit measurably higher CREATIVITY INDEX compared to other human writers. Finally, we demonstrate that CREATIVITY INDEX can be used as a surprisingly effective criterion for zero-shot machine text detection, surpassing the strongest existing zero-shot system, DetectGPT, by a significant margin of 30.2%, and even outperforming the strongest supervised system, GhostBuster, in five out of six domains. | Machine Creativity, Large Language Model, Science of LLM, Machine Text Detection | We present CREATIVITY INDEX, a metric that quantifies the creativity of a text by reconstructing it from existing web snippets, supported by a novel dynamic programming algorithm, DJ SEARCH, for efficient computation. | 3,478 | null | [
-0.09879282861948013,
-0.045103948563337326,
0.02962455525994301,
0.053274914622306824,
0.04110117256641388,
-0.02076238952577114,
0.03897576406598091,
-0.014074915088713169,
0.07419048994779587,
-0.031500864773988724,
-0.0840756893157959,
-0.00011962187272729352,
0.10308028012514114,
-0.04016406461596489,
-0.008857431821525097,
0.04559868574142456,
0.03705412149429321,
0.013002016581594944,
-0.023556997999548912,
-0.105807363986969,
0.08340635150671005,
0.07769894599914551,
0.07826803624629974,
-0.05634542927145958,
0.055199142545461655,
0.017553173005580902,
0.02119489200413227,
-0.022623369470238686,
0.03912637382745743,
0.01941613480448723,
-0.005166963674128056,
0.013808458112180233,
0.04075505584478378,
0.03348638862371445,
-0.05519723892211914,
0.054732318967580795,
-0.08371752500534058,
0.015008115209639072,
0.11950257420539856,
-0.0491151437163353,
-0.07806158810853958,
-0.029432740062475204,
-0.004108784254640341,
0.05401502549648285,
0.06591158360242844,
0.02060890384018421,
-0.09150545299053192,
-0.0029410840943455696,
-0.02043302170932293,
0.04798274114727974,
-0.1913139373064041,
-0.03093981370329857,
0.001977349631488323,
-0.013874378055334091,
-0.03841118514537811,
-0.016331175342202187,
-0.017337236553430557,
0.02188134379684925,
0.018834885209798813,
-0.06269975006580353,
0.03449966758489609,
-0.02456837333738804,
0.04650969058275223,
-0.01893388293683529,
0.026850339025259018,
-0.06027116999030113,
-0.014624477364122868,
0.047091249376535416,
-0.03859248757362366,
0.001920596812851727,
0.05679475516080856,
0.049855515360832214,
-0.0431138277053833,
0.12310221791267395,
0.015309345908463001,
0.08833876997232437,
-0.01914922520518303,
-0.012673545628786087,
-0.01639491505920887,
-0.01639486663043499,
-0.01530879084020853,
-0.023370658978819847,
0.034122247248888016,
0.03378339484333992,
0.02509816363453865,
-0.02201571874320507,
0.024178238585591316,
0.04994560033082962,
0.07297246903181076,
0.004798075184226036,
-0.06208260357379913,
-0.08650267869234085,
0.06860815733671188,
-0.10596997290849686,
0.0528235137462616,
0.0330260694026947,
-0.0012113525299355388,
-0.0466214157640934,
0.03098941221833229,
0.043152227997779846,
-0.024140479043126106,
0.05689062923192978,
0.0016582782845944166,
-0.06460271775722504,
-0.035265855491161346,
0.03968977555632591,
0.004441892262548208,
0.035613466054201126,
0.06723104417324066,
-0.052234843373298645,
-0.041783884167671204,
0.07511680573225021,
0.007342999335378408,
-0.027305293828248978,
0.09719521552324295,
-0.044957470148801804,
0.016391854733228683,
0.002612407086417079,
0.040179699659347534,
0.09438904374837875,
0.04745117574930191,
0.08611047267913818,
0.001799218007363379,
0.03406623378396034,
-0.041036464273929596,
0.014770551584661007,
-0.03094523213803768,
1.232117270852924e-33,
0.10791026800870895,
0.03849766030907631,
0.021384187042713165,
0.08190050721168518,
0.0428876169025898,
-0.06171746924519539,
0.018653180450201035,
0.03748951852321625,
-0.047093186527490616,
-0.041042450815439224,
-0.05688217282295227,
-0.024080779403448105,
0.0180425513535738,
0.11719965934753418,
-0.05997690558433533,
-0.022249475121498108,
0.0005353429587557912,
-0.00025830513914115727,
-0.04001201316714287,
-0.064241424202919,
0.008645005524158478,
-0.0033993725664913654,
0.09155948460102081,
-0.03686746582388878,
-0.03005952015519142,
0.0074079749174416065,
-0.009874178096652031,
-0.08285649120807648,
-0.008930782787501812,
0.011391919106245041,
-0.09103239327669144,
-0.04855848103761673,
-0.009470060467720032,
-0.039156727492809296,
-0.059367530047893524,
-0.028204672038555145,
-0.06443911790847778,
-0.028477367013692856,
0.04964546114206314,
0.060096729546785355,
-0.07586650550365448,
-0.004865522496402264,
0.02574089914560318,
-0.002126124920323491,
-0.04736745357513428,
0.04742998257279396,
-0.06914451718330383,
-0.009795267134904861,
0.04808438941836357,
0.012466135434806347,
0.04154830053448677,
0.05851228907704353,
0.027081716805696487,
0.08785000443458557,
0.03202512860298157,
-0.008901411667466164,
0.02376149781048298,
-0.01369949895888567,
0.022313278168439865,
0.025210030376911163,
0.03843602165579796,
0.05207841098308563,
-0.03676547110080719,
0.04246994107961655,
0.07860326021909714,
0.05840584635734558,
0.03640928491950035,
0.012595311738550663,
0.061993032693862915,
0.070030577480793,
0.07268290221691132,
-0.0065405662171542645,
-0.04969613999128342,
-0.04809124767780304,
-0.0998523011803627,
-0.006217330228537321,
0.01820516586303711,
-0.04033427685499191,
0.029239490628242493,
0.006336753256618977,
-0.052311159670352936,
-0.010821154341101646,
0.02450341358780861,
-0.10061023384332657,
-0.06672258675098419,
-0.07188298553228378,
0.012506753206253052,
-0.08519652485847473,
-0.02178994193673134,
-0.006620751228183508,
0.017395392060279846,
-0.029504908248782158,
-0.024768041446805,
-0.08022592216730118,
0.005836841184645891,
-1.4109769096896931e-33,
-0.13646334409713745,
-0.05320598930120468,
-0.004909604787826538,
0.06770677864551544,
-0.057093389332294464,
-0.06736332178115845,
0.004041232634335756,
0.03038877621293068,
-0.04623071104288101,
-0.010462737642228603,
-0.04660942032933235,
-0.05948564037680626,
0.008278978057205677,
-0.045563336461782455,
0.017469147220253944,
0.007446997333317995,
0.052144717425107956,
0.02896832674741745,
-0.08620389550924301,
0.1001354306936264,
0.040705494582653046,
0.058729853481054306,
-0.04513313248753548,
0.013537275604903698,
0.023898130282759666,
-0.011909795925021172,
0.028753196820616722,
0.023733016103506088,
-0.027335232123732567,
-0.049579866230487823,
0.049310874193906784,
0.0026377514004707336,
-0.10535167157649994,
-0.014229830354452133,
-0.03158550336956978,
0.043439045548439026,
-0.021893678233027458,
-0.039170779287815094,
-0.02941601537168026,
0.0913551077246666,
0.05971543490886688,
0.02250809781253338,
-0.01307699829339981,
0.009510723873972893,
-0.05764269083738327,
-0.048441264778375626,
-0.13330544531345367,
-0.058502256870269775,
0.003900405252352357,
-0.008038065396249294,
0.04111374169588089,
-0.03670642897486687,
-0.06429100036621094,
-0.04930755868554115,
-0.028117239475250244,
-0.07254515588283539,
-0.04519393667578697,
-0.0178223866969347,
-0.0052922372706234455,
-0.003242610255256295,
-0.07095908373594284,
0.04176666960120201,
0.045125700533390045,
0.0017032696632668376,
0.005931633524596691,
-0.10057671368122101,
-0.007509414106607437,
0.07216756045818329,
-0.05575123429298401,
-0.04960470646619797,
0.09010139852762222,
-0.06432914733886719,
-0.017034431919455528,
0.03115450032055378,
-0.03226702660322189,
0.033088985830545425,
0.027614878490567207,
0.012149588204920292,
-0.010037032887339592,
-0.10808321088552475,
0.0077682253904640675,
-0.014810190536081791,
0.002416422124952078,
0.01272304356098175,
0.022851748391985893,
0.06844323873519897,
-0.04278905317187309,
0.03868430480360985,
-0.022944020107388496,
0.04246903955936432,
-0.008251059800386429,
-0.00169118563644588,
-0.05196154862642288,
0.07041344046592712,
0.032488711178302765,
-3.7152098286696855e-8,
-0.03136352077126503,
0.03629397228360176,
-0.068771593272686,
0.11039240658283234,
0.01851559430360794,
0.028221720829606056,
-0.020579710602760315,
0.006064190063625574,
-0.0905713215470314,
-0.05827604979276657,
0.08486136794090271,
-0.04059920832514763,
0.003779227379709482,
0.0219110157340765,
-0.01911085844039917,
-0.0078076692298054695,
-0.03852338716387749,
0.046943385154008865,
0.025789540261030197,
-0.021005256101489067,
0.0987771674990654,
0.10477183759212494,
-0.06576065719127655,
-0.007640100084245205,
0.00895901583135128,
0.01820831373333931,
-0.06941543519496918,
0.01969081535935402,
-0.05890995264053345,
0.05977775529026985,
0.032661326229572296,
0.0011150523787364364,
-0.05494929477572441,
0.017463093623518944,
0.07941263169050217,
-0.014225145801901817,
-0.03903844207525253,
-0.07926352322101593,
-0.11400175839662552,
0.00915944017469883,
0.07271170616149902,
0.051264867186546326,
0.030405424535274506,
0.02616044133901596,
0.06423473358154297,
-0.03469923883676529,
-0.00765015184879303,
-0.09703357517719269,
0.07489649206399918,
-0.009184770286083221,
-0.0017759395996108651,
-0.07733232527971268,
0.013134793378412724,
-0.013020891696214676,
0.07372795790433884,
-0.03512256592512131,
-0.0652998834848404,
0.08463034778833389,
0.03356151655316353,
0.013195852749049664,
0.09045641124248505,
0.05040884390473366,
0.06365660578012466,
-0.014273944310843945
] | 0 | 0 | 0 | 0 |
|
Comparing noisy neural population dynamics using optimal transport distances | https://openreview.net/forum?id=cNmu0hZ4CL | [
"Amin Nejatbakhsh",
"Victor Geadah",
"Alex H Williams",
"David Lipshutz"
] | Oral | Biological and artificial neural systems form high-dimensional neural representations that underpin their computational capabilities. Methods for quantifying geometric similarity in neural representations have become a popular tool for identifying computational principles that are potentially shared across neural systems. These methods generally assume that neural responses are deterministic and static. However, responses of biological systems, and some artificial systems, are noisy and dynamically unfold over time. Furthermore, these characteristics can have substantial influence on a system’s computational capabilities. Here, we demonstrate that existing metrics can fail to capture key differences between neural systems with noisy dynamic responses. We then propose a metric for comparing the geometry of noisy neural trajectories, which can be derived as an optimal transport distance between Gaussian processes. We use the metric to compare models of neural responses in different regions of the motor system and to compare the dynamics of latent diffusion models for text-to-image synthesis. | Representational similarity, shape metrics, optimal transport, Wasserstein distance | We propose using optimal transport distances on stochastic processes to compare noisy neural trajectories. | 3,439 | 2412.14421 | [
-0.027801277115941048,
-0.17639754712581635,
0.060196202248334885,
-0.013543648645281792,
-0.04278673231601715,
0.004802390467375517,
0.031115662306547165,
-0.03857646882534027,
0.12602028250694275,
-0.07279998064041138,
0.0027314156759530306,
-0.01778959110379219,
0.021809447556734085,
0.08323770761489868,
-0.112635038793087,
-0.0373050831258297,
0.06122938171029091,
0.02612028643488884,
-0.09008095413446426,
0.0020742155611515045,
0.0005950117483735085,
0.011312560178339481,
0.025153664872050285,
0.030720766633749008,
0.04204821586608887,
0.08085759729146957,
0.014888287521898746,
-0.05416690185666084,
-0.018212588503956795,
0.01898735761642456,
0.024870740249753,
-0.042539745569229126,
-0.06457722932100296,
-0.04459868744015694,
-0.06872782856225967,
0.06992147117853165,
-0.09178727120161057,
-0.008498608134686947,
0.03271458297967911,
-0.02162962034344673,
0.03607887029647827,
0.05424942076206207,
0.04218357428908348,
0.09786955267190933,
0.05025705695152283,
0.039225023239851,
-0.039665572345256805,
-0.08721094578504562,
-0.017517952248454094,
-0.012393850833177567,
-0.0969855859875679,
-0.016092991456389427,
-0.041448257863521576,
0.039734095335006714,
0.05181008577346802,
0.07964348047971725,
0.044027816504240036,
0.014133697375655174,
-0.026357872411608696,
0.022616393864154816,
-0.011201356537640095,
-0.04031846299767494,
0.03096100687980652,
-0.024500465020537376,
0.03965933993458748,
0.0480930358171463,
0.0207256730645895,
0.016481684520840645,
-0.04807467386126518,
0.06702500581741333,
-0.07452628761529922,
0.13173730671405792,
-0.11535388976335526,
0.02580738067626953,
-0.012294242158532143,
-0.006830665748566389,
0.00495761027559638,
0.03244514763355255,
0.006618243642151356,
-0.1107025071978569,
0.02355087175965309,
-0.026160461828112602,
0.06244071200489998,
0.028105471283197403,
0.08417797833681107,
0.046519581228494644,
0.02934102900326252,
0.06397851556539536,
0.007204301655292511,
-0.03356563672423363,
-0.019266633316874504,
-0.025579968467354774,
-0.052148908376693726,
-0.038840651512145996,
-0.03366578742861748,
0.027391789481043816,
-0.032615114003419876,
-0.015057994984090328,
0.09046781063079834,
0.018704045563936234,
-0.014831659384071827,
-0.03681301698088646,
0.011393411085009575,
0.05026267468929291,
0.053203947842121124,
0.02049647457897663,
0.10302553325891495,
0.0039228894747793674,
-0.009218763560056686,
-0.02921230159699917,
0.044845979660749435,
-0.021721256896853447,
-0.009931854903697968,
0.1803789734840393,
0.022988274693489075,
-0.09034410864114761,
0.05603969469666481,
-0.014854544773697853,
0.02868705801665783,
0.09152591973543167,
-0.01719931699335575,
-0.04713393747806549,
-0.0941348597407341,
-0.026137400418519974,
0.004601711872965097,
-0.03611103445291519,
-0.05114208906888962,
1.716162059128037e-33,
-0.06951244920492172,
0.0032507102005183697,
0.06258518248796463,
0.035679277032613754,
0.04038543999195099,
-0.0404982715845108,
-0.056193675845861435,
-0.08500850200653076,
0.07177885621786118,
-0.021284665912389755,
-0.15793055295944214,
0.03297770395874977,
-0.06718817353248596,
0.06520514190196991,
0.0609486885368824,
0.0010000672191381454,
-0.04227818176150322,
0.049735624343156815,
0.02230285108089447,
-0.10431923717260361,
0.0442817285656929,
-0.011107590049505234,
-0.04799162968993187,
-0.05588836967945099,
-0.03336150571703911,
-0.0632401779294014,
-0.0664973258972168,
-0.049415696412324905,
-0.015514243394136429,
-0.016077321022748947,
-0.11594343185424805,
-0.0014101382112130523,
-0.0424930676817894,
0.011518385261297226,
0.015598473139107227,
-0.025526605546474457,
0.06494943797588348,
-0.014704179018735886,
-0.02283652313053608,
0.0018288014689460397,
0.01070941798388958,
0.05301232263445854,
0.006171579472720623,
-0.03664574399590492,
-0.03631985932588577,
0.057017531245946884,
0.020446280017495155,
0.009563918225467205,
0.004437846131622791,
-0.09662105143070221,
0.011741633526980877,
0.005280633922666311,
0.03739623352885246,
-0.09730639308691025,
0.01335290726274252,
-0.06957037001848221,
-0.032779645174741745,
-0.01237382274121046,
0.0036166219506412745,
0.08938117325305939,
0.058612074702978134,
0.029480017721652985,
0.06874334812164307,
-0.014209708198904991,
0.11213689297437668,
0.037825532257556915,
-0.02528625726699829,
0.02659313753247261,
0.04620329663157463,
0.0001635713124414906,
0.03482697531580925,
0.011040763929486275,
-0.03880392387509346,
-0.08281072974205017,
0.03378676623106003,
-0.005044858902692795,
0.04490182548761368,
-0.08477400243282318,
-0.05613446235656738,
-0.0030365330167114735,
-0.09836629778146744,
-0.06941837072372437,
-0.08092477917671204,
-0.05223478004336357,
-0.02665247581899166,
-0.038948774337768555,
0.05124557763338089,
-0.07185579091310501,
-0.10118568688631058,
0.0027866980526596308,
-0.020811645314097404,
-0.049890656024217606,
-0.049273017793893814,
-0.022558465600013733,
-0.035036977380514145,
-3.225689109112809e-33,
-0.04564165323972702,
0.0038238943088799715,
0.08903287351131439,
0.07596981525421143,
-0.044939879328012466,
0.0589117594063282,
0.023397177457809448,
0.061214037239551544,
0.040308788418769836,
0.010766560211777687,
-0.05635638162493706,
0.015967991203069687,
0.019878797233104706,
-0.03955630585551262,
0.013262378983199596,
-0.06425851583480835,
0.022537752985954285,
-0.017969394102692604,
0.03259192779660225,
-0.07544008642435074,
0.03043527714908123,
0.05768722668290138,
-0.042299047112464905,
-0.02788909524679184,
-0.04311510547995567,
0.06807152181863785,
-0.03153664246201515,
0.04466725513339043,
-0.04771043732762337,
0.02334720827639103,
-0.08432692289352417,
-0.02801463007926941,
-0.05516916885972023,
-0.02924174815416336,
-0.052812691777944565,
0.07069458067417145,
0.01344139315187931,
0.021901389583945274,
-0.002481837058439851,
0.0781160295009613,
0.04297348111867905,
-0.04433053731918335,
0.030712326988577843,
0.027063196524977684,
0.022870343178510666,
-0.02434534579515457,
-0.069771409034729,
-0.04299283027648926,
-0.009779088199138641,
0.021504178643226624,
0.08525678515434265,
0.0961851179599762,
-0.06665333360433578,
-0.019099272787570953,
-0.01489546149969101,
0.018749037757515907,
-0.03837558999657631,
-0.06974747031927109,
0.09519791603088379,
-0.030212756246328354,
-0.0630902647972107,
-0.07601239532232285,
-0.03339334949851036,
-0.03645776957273483,
-0.041251290589571,
0.015306184999644756,
-0.033085353672504425,
0.004356634337455034,
0.05691857635974884,
0.04883701726794243,
0.030154284089803696,
0.020931996405124664,
0.027236390858888626,
0.06020013242959976,
0.032808851450681686,
-0.03384988009929657,
0.09259332716464996,
-0.010416985489428043,
0.017862074077129364,
-0.012927290983498096,
0.06105205789208412,
0.07133302837610245,
0.0037540756165981293,
0.007145352195948362,
-0.010960224084556103,
0.05605001747608185,
-0.012554077431559563,
-0.03150642290711403,
0.047260455787181854,
0.02497035823762417,
0.052949320524930954,
-0.025320300832390785,
0.023886417970061302,
-0.010409904643893242,
-0.08481045812368393,
-4.1871544453897513e-8,
-0.07727430015802383,
0.014959574677050114,
0.10370676964521408,
0.004510806407779455,
0.02582198567688465,
0.022921765223145485,
0.036493755877017975,
0.06785843521356583,
-0.04013845697045326,
-0.018921013921499252,
0.0971008688211441,
-0.013039813376963139,
0.023257426917552948,
0.021458569914102554,
-0.0191271360963583,
0.08755385130643845,
-0.046748947352170944,
0.004123082384467125,
0.01400105468928814,
-0.004859264008700848,
0.008189552463591099,
0.043540261685848236,
-0.03086441196501255,
0.04321020096540451,
0.045093730092048645,
-0.061559539288282394,
-0.056717224419116974,
0.011555933393537998,
0.017012719064950943,
-0.029044315218925476,
0.07610227167606354,
0.050539691001176834,
0.03898214176297188,
0.014767834916710854,
-0.018850920721888542,
-0.0039980155415833,
-0.022755060344934464,
-0.03459582105278969,
-0.06750279664993286,
0.004376974422484636,
-0.04492378979921341,
0.03318130224943161,
-0.0017026496352627873,
0.012382913380861282,
0.08308657258749008,
0.01576085202395916,
0.09420537948608398,
-0.0767495259642601,
-0.007565319072455168,
0.0018524897750467062,
0.0033919985871762037,
-0.02343990094959736,
-0.03029840998351574,
0.05951802432537079,
0.01840651035308838,
-0.04300824925303459,
-0.014286471530795097,
-0.05361586809158325,
0.019509000703692436,
0.035502973943948746,
0.0031697971280664206,
0.13118763267993927,
-0.01350834034383297,
-0.04420782998204231
] | 0 | 0 | 0 | 0 |
|
Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport | https://openreview.net/forum?id=gQlxd3Mtru | [
"Zhenyi Zhang",
"Tiejun Li",
"Peijie Zhou"
] | Oral | Reconstructing dynamics using samples from sparsely time-resolved snapshots is an important problem in both natural sciences and machine learning. Here, we introduce a new deep learning approach for solving regularized unbalanced optimal transport (RUOT) and inferring continuous unbalanced stochastic dynamics from observed snapshots. Based on the RUOT form, our method models these dynamics without requiring prior knowledge of growth and death processes or additional information, allowing them to be learned directly from data. Theoretically, we explore the connections between the RUOT and Schrödinger bridge problem and discuss the key challenges and potential solutions. The effectiveness of our method is demonstrated with a synthetic gene regulatory network, high-dimensional Gaussian Mixture Model, and single-cell RNA-seq data from blood development. Compared with other methods, our approach accurately identifies growth and transition patterns, eliminates false transitions, and constructs the Waddington developmental landscape. Our code is available at: [https://github.com/zhenyiizhang/DeepRUOT](https://github.com/zhenyiizhang/DeepRUOT). | optimal transport, Schrödinger bridge, trajectory inference, single-cell | null | 3,337 | 2410.00844 | [
-0.10701880604028702,
-0.08018507808446884,
0.06355518847703934,
0.038316912949085236,
0.015255860053002834,
0.03927348554134369,
-0.053038109093904495,
-0.034687772393226624,
0.033881936222314835,
-0.02464202605187893,
-0.04940313845872879,
-0.020875109359622,
-0.08641834557056427,
0.05535503104329109,
-0.0794883593916893,
0.05026126652956009,
-0.05829959362745285,
0.05093551427125931,
-0.09745492786169052,
0.001131830271333456,
-0.005549089517444372,
0.08576931804418564,
-0.02006787620484829,
0.03810940310359001,
0.041969798505306244,
-0.019549556076526642,
0.06470867246389389,
-0.022763317450881004,
-0.015233572572469711,
-0.0584266223013401,
0.07536947727203369,
0.004530779551714659,
-0.0035264466423541307,
-0.0378759391605854,
-0.025102097541093826,
0.0813571959733963,
-0.07727596908807755,
0.00348467193543911,
-0.0015536299906671047,
0.017610883340239525,
0.12243340164422989,
0.048278070986270905,
-0.06124259531497955,
0.06554698199033737,
0.037252191454172134,
-0.00238973181694746,
-0.008080019615590572,
-0.06971964985132217,
0.04154906049370766,
-0.0005623332690447569,
0.012951107695698738,
-0.022103650495409966,
-0.012456403113901615,
0.11777389049530029,
0.06272632628679276,
0.01837668940424919,
-0.011498264968395233,
-0.07704157382249832,
-0.00691516138613224,
-0.08905980736017227,
0.01540436316281557,
-0.07635939866304398,
0.0007797417347319424,
-0.07250602543354034,
0.0335160456597805,
-0.0001677667023614049,
0.02261742576956749,
0.0911078006029129,
0.002528063952922821,
0.024581853300333023,
0.008700220845639706,
0.11501585692167282,
-0.07113026827573776,
0.03305961564183235,
0.04863293096423149,
-0.017193689942359924,
0.03655116260051727,
0.03488899767398834,
0.033897679299116135,
0.014880409464240074,
-0.008245504461228848,
-0.03663726523518562,
0.023247847333550453,
0.011082148179411888,
0.01780211180448532,
0.03775043413043022,
-0.06405939161777496,
0.08443786948919296,
-0.019431861117482185,
-0.038894448429346085,
-0.044393450021743774,
0.052721377462148666,
0.050020694732666016,
-0.012710459530353546,
0.028630128130316734,
0.07872913777828217,
0.028740325942635536,
-0.055046048015356064,
0.07424850016832352,
0.03424295410513878,
-0.059633512049913406,
0.013730442151427269,
0.07379478961229324,
0.042188383638858795,
0.06896667927503586,
-0.06329742074012756,
0.038649749010801315,
0.028754595667123795,
-0.012828594073653221,
0.02553211897611618,
0.11354193836450577,
0.014733262360095978,
0.07039280980825424,
0.06898636370897293,
0.030998507514595985,
-0.007030004635453224,
-0.00009585598309058696,
-0.006285766139626503,
-0.08922965824604034,
0.10919936001300812,
-0.04189509525895119,
-0.01192691270262003,
-0.08539704233407974,
0.0489223413169384,
-0.027889253571629524,
0.046129487454891205,
-0.010851808823645115,
3.096214651183473e-33,
0.044128112494945526,
-0.07166595011949539,
0.048226092010736465,
0.006624401547014713,
0.012937719002366066,
-0.02564341016113758,
-0.07586902379989624,
-0.027440907433629036,
-0.03755425661802292,
-0.04105743020772934,
-0.09130704402923584,
0.006706953980028629,
-0.07147662341594696,
-0.011299415491521358,
-0.06071572005748749,
-0.01997513696551323,
-0.13577081263065338,
0.01011575572192669,
0.08815757185220718,
-0.06834256649017334,
0.004729945212602615,
-0.06569361686706543,
-0.057238467037677765,
-0.06674493104219437,
-0.08638666570186615,
-0.035746555775403976,
0.058031003922224045,
-0.03208998218178749,
0.049127064645290375,
-0.0035168647300451994,
-0.11215641349554062,
0.01356634683907032,
-0.034434884786605835,
0.01583980582654476,
0.06645361334085464,
-0.023641860112547874,
0.02389582246541977,
0.014680059626698494,
-0.04516949504613876,
0.02584969997406006,
-0.028204305097460747,
0.007456013932824135,
-0.06867360323667526,
-0.06625359505414963,
-0.001931175123900175,
-0.07120674848556519,
0.042402010411024094,
-0.035051751881837845,
-0.02956007979810238,
0.00628826254978776,
-0.001930215978063643,
-0.0754891037940979,
0.003945059143006802,
-0.07025416195392609,
-0.08337049931287766,
-0.01609380915760994,
0.01634661853313446,
-0.03595731407403946,
-0.01928597129881382,
0.08221276849508286,
0.03350253030657768,
0.02107914164662361,
0.046214509755373,
0.04703006520867348,
0.05178377777338028,
0.1060306578874588,
-0.06381874531507492,
-0.009293402545154095,
0.03238246589899063,
-0.014847164042294025,
0.04119294881820679,
-0.02140933647751808,
0.0061949193477630615,
-0.06338238716125488,
0.04042964428663254,
0.02950013056397438,
0.07212257385253906,
0.051205072551965714,
-0.040010012686252594,
0.08084389567375183,
-0.013245350681245327,
0.024943333119153976,
-0.09085973352193832,
0.0535120815038681,
-0.008830766193568707,
-0.0693584531545639,
0.006622072774916887,
-0.10880595445632935,
-0.09639839828014374,
0.0003178214537911117,
-0.037906426936388016,
-0.06700774282217026,
-0.01607450284063816,
-0.004891610238701105,
0.04291059821844101,
-2.7028859762036812e-33,
0.03228403627872467,
0.045216068625450134,
-0.014762298204004765,
0.012805163860321045,
0.00446081580594182,
0.0003284239792264998,
-0.01592961698770523,
0.03255032002925873,
0.03559323027729988,
0.004166931379586458,
0.06055055558681488,
0.06367442011833191,
0.024796191602945328,
-0.0044779907912015915,
0.03527910262346268,
-0.02201102115213871,
0.008309833705425262,
0.002272437559440732,
-0.0024026099126785994,
0.01832866482436657,
-0.02625594474375248,
0.03966878354549408,
-0.09764710068702698,
0.028496555984020233,
-0.028380099684000015,
0.03880339488387108,
-0.02028229832649231,
0.10197100043296814,
-0.028806693851947784,
0.00245921709574759,
-0.06401806324720383,
-0.08219596743583679,
-0.028385000303387642,
0.02348206378519535,
-0.13000813126564026,
0.11522017419338226,
0.05197031423449516,
0.08630789816379547,
0.049100372940301895,
-0.06386524438858032,
0.04552637040615082,
-0.010595114901661873,
-0.10664810985326767,
0.01599659025669098,
0.025842299684882164,
-0.0244737658649683,
-0.0028265812434256077,
0.030463524162769318,
0.053692374378442764,
0.02588474377989769,
0.01562049612402916,
0.07897831499576569,
-0.0615476556122303,
0.028600018471479416,
-0.025945890694856644,
-0.011670341715216637,
0.049350764602422714,
0.003232651622965932,
0.006291561294347048,
0.0029298290610313416,
-0.11210174858570099,
-0.06500095874071121,
-0.0383547767996788,
-0.03718584403395653,
-0.0368315652012825,
-0.07315727323293686,
0.022278904914855957,
-0.04232103005051613,
-0.003929318394511938,
-0.025969067588448524,
0.010623625479638577,
0.03875340148806572,
0.08597894012928009,
0.025587143376469612,
0.06807520985603333,
-0.06564196199178696,
0.010427356697618961,
-0.013331018388271332,
-0.003108158940449357,
0.004663470201194286,
0.03183092176914215,
-0.003386223688721657,
0.03955642506480217,
0.034400824457407,
0.09955628216266632,
-0.01854170858860016,
0.0015353340422734618,
-0.05222718417644501,
0.06883927434682846,
0.01680952124297619,
-0.07013719528913498,
-0.02578086592257023,
-0.038083769381046295,
0.0024485629983246326,
-0.033844027668237686,
-4.4280341171543114e-8,
0.006942157167941332,
0.03194084390997887,
0.057279206812381744,
-0.027261096984148026,
0.09752655774354935,
-0.015173472464084625,
-0.013162532821297646,
0.08419828116893768,
0.021559715270996094,
0.04648789018392563,
0.07412740588188171,
0.021226737648248672,
-0.05416306480765343,
-0.09002489596605301,
-0.01612134650349617,
0.11278147995471954,
-0.020923705771565437,
0.001687174546532333,
-0.013390129432082176,
-0.0368804931640625,
-0.0029606700409203768,
0.0075052292086184025,
-0.000914841250050813,
0.05350008234381676,
0.09006283432245255,
-0.07767898589372635,
-0.014318868517875671,
-0.03347645699977875,
0.06454978138208389,
-0.08381272852420807,
0.0013672256609424949,
0.010812031105160713,
0.08999986201524734,
0.04519670829176903,
-0.08562254905700684,
0.04791644215583801,
0.0550890676677227,
0.043951719999313354,
-0.05420364812016487,
-0.032212480902671814,
-0.014312813058495522,
0.1072661504149437,
-0.07073058187961578,
-0.04619808495044708,
-0.019473567605018616,
0.056985851377248764,
0.0067209517583251,
0.010445394553244114,
0.09148593246936798,
0.015176133252680302,
0.04365842416882515,
-0.02775573544204235,
0.07985193282365799,
0.03703853860497475,
0.023822296410799026,
0.053000640124082565,
-0.05660624802112579,
-0.062183499336242676,
0.0141398711130023,
0.056310854852199554,
0.026844937354326248,
0.05933409556746483,
-0.06140464171767235,
-0.010397413745522499
] | https://github.com/zhenyiizhang/deepruot | 13 | 0 | 0 | 0 |
Prioritized Generative Replay | https://openreview.net/forum?id=5IkDAfabuo | [
"Renhao Wang",
"Kevin Frans",
"Pieter Abbeel",
"Sergey Levine",
"Alexei A Efros"
] | Oral | Sample-efficient online reinforcement learning often uses replay buffers to store experience for reuse when updating the value function.
However, uniform replay is inefficient, since certain classes of transitions can be more relevant to learning. While prioritization of more useful samples is helpful, this strategy can also lead to overfitting, as useful samples are likely to be more rare.
In this work, we instead propose a prioritized, parametric version of an agent's memory, using generative models to capture online experience. This paradigm enables (1) densification of past experience, with new generations that benefit from the generative model's generalization capacity and (2) guidance via a family of ``relevance functions'' that push these generations towards more useful parts of an agent's acquired history. We show this recipe can be instantiated using conditional diffusion models and simple relevance functions such as curiosity- or value-based metrics.
Our approach consistently improves performance and sample efficiency in both state- and pixel-based domains. We expose the mechanisms underlying these gains, showing how guidance promotes diversity in our generated transitions and reduces overfitting. We also showcase how our approach can train policies with even higher update-to-data ratios than before, opening up avenues to better scale online RL agents. | online learning, model-based reinforcement learning, generative modeling, synthetic data, continual learning | We construct a conditional generative model of an agent's online memory, allowing us to replay high-priority data at large quantities to accelerate training of online RL agents. | 3,226 | 2410.18082 | [
-0.05044179409742355,
-0.0824681743979454,
0.04505253583192825,
0.012356148101389408,
0.03616897389292717,
0.08964398503303528,
0.021773822605609894,
0.004021867178380489,
0.05453987047076225,
-0.07731641829013824,
-0.0410187691450119,
-0.02964867278933525,
0.03015773370862007,
-0.0064040059223771095,
-0.0023392413277179003,
-0.006759202107787132,
0.10579006373882294,
0.04804063215851784,
-0.08620459586381912,
-0.043923527002334595,
0.050177592784166336,
0.04054476320743561,
0.08517395704984665,
-0.02917362004518509,
-0.02312728948891163,
-0.007085734978318214,
0.031036943197250366,
-0.023811977356672287,
0.0895710438489914,
-0.03981707990169525,
0.044282179325819016,
0.053506165742874146,
-0.06972142308950424,
-0.01054157130420208,
-0.0401206836104393,
0.08332152664661407,
-0.13057728111743927,
0.05525124445557594,
0.007249807473272085,
-0.013019213452935219,
0.006418190430849791,
-0.004416859243065119,
0.008614441379904747,
0.021152852103114128,
0.028053538873791695,
-0.031436167657375336,
-0.007473039906471968,
-0.01774180307984352,
0.01697089895606041,
0.04167948663234711,
-0.008416539058089256,
-0.019332341849803925,
-0.06414610892534256,
0.0388302281498909,
0.05644780769944191,
0.05895514041185379,
0.05506523698568344,
0.026542240753769875,
-0.04523969441652298,
-0.05471543222665787,
-0.08625543117523193,
-0.12554068863391876,
0.022439785301685333,
-0.062272071838378906,
-0.06812729686498642,
-0.07307490706443787,
0.017239999026060104,
0.04415525868535042,
0.006276475731283426,
0.04177776724100113,
-0.029394103214144707,
0.034790266305208206,
-0.026800574734807014,
0.00187642069067806,
-0.014391101896762848,
0.06314253807067871,
0.029600387439131737,
-0.0593612976372242,
-0.03479645028710365,
-0.029185930266976357,
0.0031343221198767424,
-0.05561905354261398,
-0.004259541630744934,
0.03971632942557335,
0.0867369994521141,
-0.01320225466042757,
0.025650421157479286,
-0.029067115858197212,
0.11424580216407776,
0.024219388142228127,
-0.0050227344036102295,
0.02973587065935135,
0.007291337009519339,
-0.00509515218436718,
0.030683893710374832,
0.038352157920598984,
0.03662041202187538,
-0.10444607585668564,
0.01717302016913891,
0.07511737942695618,
0.017848225310444832,
0.09835076332092285,
-0.0001623164862394333,
0.005130942445248365,
0.008735968731343746,
-0.05604051798582077,
0.02489345706999302,
0.11941711604595184,
0.02079082280397415,
-0.07257938385009766,
-0.01479372475296259,
-0.015358994714915752,
0.06384564936161041,
0.006073057185858488,
0.01305581908673048,
-0.026145486161112785,
0.02945205196738243,
-0.00646726880222559,
-0.04698025807738304,
0.04972018674015999,
0.07306273281574249,
-0.002146477112546563,
0.008790422230958939,
0.013353529386222363,
0.017620142549276352,
-0.07540380954742432,
-0.02428528480231762,
1.6632975067776664e-33,
0.09659723937511444,
0.02329222857952118,
-0.01403597928583622,
0.04736451059579849,
0.003702769288793206,
0.0670558363199234,
0.04248783364892006,
-0.002847424941137433,
-0.030157579109072685,
-0.08519815653562546,
-0.014964721165597439,
0.04450717568397522,
-0.10901106894016266,
0.08678795397281647,
0.03375081717967987,
-0.016194026917219162,
-0.15506437420845032,
0.08134203404188156,
0.03303246945142746,
-0.024526508525013924,
0.023791581392288208,
0.0162657480686903,
-0.008453819900751114,
-0.03808839991688728,
0.008191614411771297,
0.016563965007662773,
0.024930160492658615,
-0.002197942463681102,
-0.0016117674531415105,
-0.042093340307474136,
-0.04871929809451103,
-0.01753586158156395,
-0.019139224663376808,
0.0019518804037943482,
-0.010367210023105145,
0.049416542053222656,
-0.05408013239502907,
-0.04322563484311104,
-0.0024849807377904654,
-0.03476789966225624,
0.0047133867628872395,
0.017511870712041855,
-0.0006206168327480555,
-0.04010569304227829,
-0.129983589053154,
-0.1808461993932724,
0.03857487440109253,
-0.016840225085616112,
-0.03400712087750435,
-0.06586810201406479,
0.06986545026302338,
0.011165576986968517,
-0.014701561070978642,
-0.06245550885796547,
-0.0448075532913208,
-0.03373320400714874,
-0.022653590887784958,
0.031378552317619324,
-0.06775759160518646,
0.014772109687328339,
0.0742541030049324,
0.029044251888990402,
0.03229868784546852,
0.08224637806415558,
0.005215695593506098,
0.03584584966301918,
0.04712036997079849,
0.10862788558006287,
0.04386744648218155,
0.011305087246000767,
0.010726818814873695,
-0.0026179051492363214,
-0.029034189879894257,
-0.15091824531555176,
0.08428559452295303,
-0.010012849234044552,
0.04562857374548912,
-0.055794645100831985,
0.05363723635673523,
-0.06551381200551987,
-0.013765070587396622,
-0.07102245092391968,
-0.07634338736534119,
-0.004127309657633305,
0.038562141358852386,
-0.02391185238957405,
0.046642936766147614,
-0.030551118776202202,
-0.04476136714220047,
0.014538062736392021,
-0.09516394138336182,
-0.035040754824876785,
-0.024731703102588654,
0.015789039433002472,
0.008723458275198936,
-1.8195941743988748e-33,
-0.011703240685164928,
-0.03220946341753006,
0.05708262324333191,
0.03640695661306381,
-0.022725678980350494,
-0.06407661736011505,
-0.05516934022307396,
-0.007264374289661646,
-0.06127594783902168,
-0.09982280433177948,
-0.0740482360124588,
0.037092845886945724,
0.06943833827972412,
0.012270260602235794,
-0.08200202137231827,
0.006296121049672365,
0.0028966276440769434,
-0.02660958468914032,
-0.05319628864526749,
-0.023115111514925957,
0.02999548427760601,
0.0946200042963028,
-0.08989489823579788,
0.045850008726119995,
-0.018438320606946945,
0.0809221863746643,
-0.008361006155610085,
0.08075742423534393,
0.04116065055131912,
-0.014639606699347496,
-0.012998146936297417,
-0.06598035246133804,
0.021399615332484245,
0.013265252113342285,
-0.030930718407034874,
0.07231321185827255,
0.12543970346450806,
-0.0028521709609776735,
-0.11863932013511658,
0.17574134469032288,
0.08424333482980728,
-0.05693333223462105,
-0.07214025408029556,
0.025761937722563744,
-0.013337429612874985,
0.020615560933947563,
-0.029099417850375175,
0.052754875272512436,
0.03372567519545555,
0.002762298099696636,
-0.0063252560794353485,
0.0714908093214035,
-0.11157893389463425,
-0.022115370258688927,
-0.06764383614063263,
-0.08080898225307465,
-0.04883219674229622,
-0.036340273916721344,
0.06664927303791046,
-0.01988738402724266,
-0.006199878174811602,
-0.01386314444243908,
-0.06267951428890228,
-0.031190255656838417,
-0.04002981260418892,
-0.006159862503409386,
-0.007937328889966011,
-0.009688695892691612,
-0.047763049602508545,
0.013304361142218113,
0.0218943003565073,
0.006031447555869818,
0.010920966975390911,
-0.06450959295034409,
0.06427530199289322,
-0.02436448074877262,
0.03614412993192673,
0.017857177183032036,
0.019769517704844475,
-0.057487569749355316,
-0.06283238530158997,
0.04695836454629898,
0.01908932812511921,
0.010089561343193054,
0.0213872492313385,
0.0943230465054512,
-0.000009992576451622881,
0.14247660338878632,
0.04493048042058945,
-0.04078838974237442,
-0.0329803042113781,
-0.04922158271074295,
-0.010136003606021404,
0.006144920829683542,
-0.06221197172999382,
-4.743066739365531e-8,
-0.07085105031728745,
0.08931349962949753,
0.0417037196457386,
0.08668331056833267,
0.08304353803396225,
-0.004910966847091913,
-0.025356734171509743,
0.04985111206769943,
0.00449158763512969,
-0.05505490303039551,
0.05299505963921547,
0.014065775088965893,
0.013250139541924,
-0.03055562637746334,
0.049123793840408325,
0.0237896591424942,
-0.005776465870440006,
-0.03909504786133766,
-0.010510357096791267,
0.012355360202491283,
0.013696372509002686,
-0.03745441138744354,
-0.012676321901381016,
-0.017262540757656097,
0.012770958244800568,
-0.00796485971659422,
0.0532035231590271,
0.02932337485253811,
0.0773649662733078,
0.012620142661035061,
0.04062552750110626,
-0.0007374476990662515,
0.08723387122154236,
-0.01855834573507309,
-0.004919734317809343,
0.029414013028144836,
0.023102955892682076,
-0.06244521960616112,
0.027353085577487946,
-0.01099297683686018,
-0.04201271012425423,
0.005480475723743439,
-0.007505039218813181,
0.02459530718624592,
0.0015602406347170472,
-0.009132688865065575,
-0.02247629128396511,
-0.06654535979032516,
0.037462782114744186,
-0.019806277006864548,
-0.025455836206674576,
-0.08019411563873291,
0.04363756626844406,
0.03513123467564583,
0.11816145479679108,
-0.01709582842886448,
0.041388168931007385,
0.026833996176719666,
0.07704612612724304,
0.06366654485464096,
0.03712696209549904,
-0.017053373157978058,
-0.06003476306796074,
0.01893192157149315
] | 0 | 0 | 0 | 0 |
|
The Geometry of Categorical and Hierarchical Concepts in Large Language Models | https://openreview.net/forum?id=bVTM2QKYuA | [
"Kiho Park",
"Yo Joong Choe",
"Yibo Jiang",
"Victor Veitch"
] | Oral | The linear representation hypothesis is the informal idea that semantic concepts are encoded as linear directions in the representation spaces of large language models (LLMs). Previous work has shown how to make this notion precise for representing binary concepts that have natural contrasts (e.g., {male, female}) as _directions_ in representation space. However, many natural concepts do not have natural contrasts (e.g., whether the output is about an animal). In this work, we show how to extend the formalization of the linear representation hypothesis to represent features (e.g., is_animal) as _vectors_. This allows us to immediately formalize the representation of categorical concepts as polytopes in the representation space. Further, we use the formalization to prove a relationship between the hierarchical structure of concepts and the geometry of their representations. We validate these theoretical results on the Gemma and LLaMA-3 large language models, estimating representations for 900+ hierarchically related concepts using data from WordNet. | categorical concepts, hierarchical concepts, linear representation hypothesis, causal inner product, interpretability | We extend the linear representation hypothesis to general concepts and show that hierarchical relationships are encoded as orthogonality. | 3,176 | 2406.01506 | [
0.015396418049931526,
-0.09153055399656296,
0.04677187278866768,
0.02842932567000389,
0.0015525452326983213,
0.022438915446400642,
-0.06247709318995476,
-0.03203596919775009,
0.022270027548074722,
-0.030148649588227272,
-0.05790512263774872,
-0.06134188920259476,
0.07953421771526337,
0.04062017798423767,
-0.003086785087361932,
0.0467955656349659,
0.1337105631828308,
0.07589493691921234,
-0.03492885082960129,
-0.005402008071541786,
0.11045074462890625,
0.0052301050163805485,
0.018558621406555176,
-0.012537805363535881,
-0.006896767299622297,
0.047819167375564575,
-0.04530942067503929,
-0.012955380603671074,
0.002988185966387391,
0.009637529030442238,
-0.03187345713376999,
0.09481736272573471,
0.03821967914700508,
0.08485372364521027,
-0.025932677090168,
0.06641560792922974,
-0.05358484387397766,
0.05355853587388992,
-0.019282478839159012,
-0.025526942685246468,
-0.014380848035216331,
0.03101293370127678,
0.04342825710773468,
0.02885611541569233,
0.12099765986204147,
0.014035274274647236,
-0.021523533388972282,
-0.06100616231560707,
-0.06497752666473389,
-0.0051278006285429,
-0.06532130390405655,
-0.045569293200969696,
-0.039209477603435516,
0.04963595047593117,
-0.09395884722471237,
-0.06673678010702133,
0.006168793886899948,
-0.03661837428808212,
-0.02997010573744774,
-0.049220651388168335,
-0.033233851194381714,
-0.002912365598604083,
-0.026508284732699394,
0.016144851222634315,
-0.02297879010438919,
-0.00023978804529178888,
-0.037969157099723816,
0.0796416625380516,
-0.013425386510789394,
0.09336400777101517,
-0.0002788925776258111,
0.0407257154583931,
-0.09687912464141846,
0.05344867333769798,
-0.04168827459216118,
-0.014585267752408981,
0.10945631563663483,
0.015399092808365822,
0.04903149977326393,
-0.041907306760549545,
-0.04089892655611038,
0.02137484960258007,
0.04403236508369446,
0.0029114549979567528,
0.1072142943739891,
0.09298919141292572,
-0.022639881819486618,
0.016042083501815796,
-0.13823477923870087,
0.0010944110108539462,
-0.07556911557912827,
-0.10947263240814209,
0.097514308989048,
0.040543802082538605,
0.047345347702503204,
-0.023945175111293793,
0.050428569316864014,
-0.04162582755088806,
0.07267393171787262,
0.07069391012191772,
0.014595761895179749,
0.019706720486283302,
-0.007721212692558765,
-0.11292514950037003,
0.011277355253696442,
-0.058674056082963943,
0.0019784788601100445,
-0.027610106393694878,
0.014738678932189941,
-0.06903120130300522,
-0.02795925922691822,
-0.02667125128209591,
-0.0486580990254879,
0.060554541647434235,
0.02332250401377678,
-0.046289004385471344,
0.024132994934916496,
-0.06663361936807632,
0.10619425773620605,
-0.046541258692741394,
0.03362837806344032,
0.08044102042913437,
0.055015187710523605,
0.02902848646044731,
0.03840365633368492,
-0.016606971621513367,
-0.049025584012269974,
-2.653223360262355e-33,
0.03035074658691883,
-0.034290608018636703,
0.001735221710987389,
0.08041992783546448,
0.07923616468906403,
0.0066883559338748455,
0.028172176331281662,
0.0014990901108831167,
0.006956682074815035,
0.009244720451533794,
-0.0721394345164299,
0.06704611331224442,
-0.03419685363769531,
0.035610925406217575,
-0.0179342869669199,
0.051970694214105606,
0.018445728346705437,
0.002896319143474102,
-0.07458391040563583,
-0.041933655738830566,
-0.07812819629907608,
0.09350787848234177,
-0.028947120532393456,
-0.018033236265182495,
-0.02656441554427147,
-0.005203685723245144,
0.013418491929769516,
-0.11004149913787842,
-0.07560823112726212,
0.0006620160420425236,
-0.10811183601617813,
0.007884953171014786,
0.0026333495043218136,
0.021286869421601295,
-0.02007710188627243,
-0.05216525122523308,
0.010632321238517761,
-0.07023397833108902,
0.05025605857372284,
0.012109130620956421,
-0.06065830960869789,
0.002590079326182604,
-0.046672333031892776,
-0.07160650938749313,
-0.00300741009414196,
-0.0007647160091437399,
0.020418252795934677,
0.020053990185260773,
-0.0363127775490284,
-0.014796203002333641,
0.057078491896390915,
0.04395885393023491,
-0.005712502636015415,
0.00016552515444345772,
-0.027277354151010513,
0.06416476517915726,
-0.02166767604649067,
0.049091558903455734,
-0.026113145053386688,
0.04912351816892624,
-0.011353849433362484,
0.022958584129810333,
0.12542115151882172,
-0.009846293367445469,
-0.013917696662247181,
-0.007739145774394274,
-0.034187592566013336,
0.05937119945883751,
0.05025481805205345,
0.031574446707963943,
-0.05284249782562256,
0.04274097830057144,
-0.051147595047950745,
-0.06718326359987259,
0.0028074183501303196,
0.021489689126610756,
0.00912938080728054,
-0.1045013815164566,
-0.06928809732198715,
0.09256339073181152,
-0.010638362728059292,
-0.03715277463197708,
0.022997602820396423,
-0.07269241660833359,
-0.06745890527963638,
-0.02921256050467491,
0.08437356352806091,
-0.09813113510608673,
0.04895659163594246,
-0.04903842508792877,
-0.05217813700437546,
-0.042424965649843216,
-0.011516950093209743,
-0.027739280834794044,
0.06890479475259781,
-4.80128739047944e-34,
-0.05747365951538086,
0.004002473782747984,
-0.05258180573582649,
-0.04477273300290108,
-0.10597018152475357,
-0.10476653277873993,
0.05366436764597893,
0.04401920735836029,
-0.06285738199949265,
0.019712217152118683,
-0.011667629703879356,
-0.04929603263735771,
0.025888711214065552,
-0.02018417976796627,
0.05778926983475685,
-0.04960007593035698,
-0.0357687808573246,
0.031732138246297836,
0.05430815741419792,
0.019094908609986305,
0.03703288361430168,
0.10025209188461304,
-0.11300970613956451,
0.06961052864789963,
-0.008185924030840397,
0.008682113140821457,
-0.003732085693627596,
-0.001847440842539072,
-0.0035585996229201555,
0.010999927297234535,
-0.013041190803050995,
-0.05460306257009506,
-0.01864493638277054,
0.007660615257918835,
-0.05673634260892868,
-0.03104926086962223,
-0.019796378910541534,
-0.03978016600012779,
0.011650160886347294,
0.05879887565970421,
0.047942135483026505,
-0.04467454552650452,
-0.06191441789269447,
-0.035197511315345764,
-0.03627149015665054,
-0.0039695557206869125,
-0.05366864427924156,
-0.02208438515663147,
0.025875210762023926,
-0.02192925475537777,
0.016698598861694336,
-0.014428810216486454,
0.004172793123871088,
0.04169413819909096,
0.030569007620215416,
-0.07262706011533737,
0.0058353557251393795,
-0.06371927261352539,
0.05286634713411331,
0.008224249817430973,
-0.025920959189534187,
-0.0470668263733387,
0.0005113664665259421,
0.0015911542577669024,
-0.03406777232885361,
-0.0676746517419815,
-0.10511317104101181,
-0.07236102223396301,
0.030542943626642227,
-0.01892901584506035,
0.0026130129117518663,
-0.004640610888600349,
-0.09698285907506943,
0.09892748296260834,
-0.02506317012012005,
-0.010252590291202068,
0.04426097869873047,
-0.021563569083809853,
-0.02861345000565052,
-0.042719703167676926,
0.0022340072318911552,
-0.05115116760134697,
0.09760961681604385,
0.044853102415800095,
0.07728041708469391,
0.018277794122695923,
-0.00653880788013339,
0.1338890790939331,
0.04576504975557327,
0.02087346650660038,
0.003382662544026971,
0.008013393729925156,
-0.03740366920828819,
0.06789648532867432,
-0.05463465675711632,
-4.2287293666731784e-8,
-0.09436837583780289,
0.0020471708849072456,
-0.001246289350092411,
0.03611854463815689,
-0.010254058055579662,
-0.006734273862093687,
0.013525978662073612,
0.01916298270225525,
-0.0702284425497055,
0.03160584717988968,
0.04166850820183754,
0.05557555332779884,
-0.054601430892944336,
-0.05700969696044922,
0.03404969349503517,
0.05781489238142967,
0.04328536614775658,
-0.01694364845752716,
0.061871182173490524,
0.07318432629108429,
0.06888649612665176,
0.05896451324224472,
-0.05334411561489105,
0.05134015530347824,
0.031152233481407166,
-0.031158674508333206,
-0.033447764813899994,
0.1730886697769165,
0.04099058732390404,
0.0029395075980573893,
-0.012124432250857353,
0.06849201768636703,
-0.05606666952371597,
-0.04636906087398529,
-0.0033622237388044596,
0.07847937941551208,
-0.0006294014165177941,
-0.04345197230577469,
0.06947428733110428,
-0.03684403374791145,
0.027124525979161263,
0.027818845584988594,
-0.04221012815833092,
0.019698305055499077,
0.12454105913639069,
0.022665970027446747,
0.0072539071552455425,
-0.053056132048368454,
-0.011300339363515377,
0.0061580948531627655,
-0.0586482435464859,
0.05809110030531883,
-0.003973528277128935,
0.03861502185463905,
-0.009225547313690186,
-0.0006462433957494795,
-0.08506543189287186,
-0.006057619117200375,
-0.003466269001364708,
-0.026620646938681602,
0.07445916533470154,
0.060357432812452316,
0.02214748039841652,
-0.029900657013058662
] | https://github.com/kihopark/llm_categorical_hierarchical_representations | 91 | 0 | 0 | 0 |
Generator Matching: Generative modeling with arbitrary Markov processes | https://openreview.net/forum?id=RuP17cJtZo | [
"Peter Holderrieth",
"Marton Havasi",
"Jason Yim",
"Neta Shaul",
"Itai Gat",
"Tommi Jaakkola",
"Brian Karrer",
"Ricky T. Q. Chen",
"Yaron Lipman"
] | Oral | We introduce Generator Matching, a modality-agnostic framework for generative modeling using arbitrary Markov processes. Generators characterize the infinitesimal evolution of a Markov process, which we leverage for generative modeling in a similar vein to flow matching: we construct conditional generators which generate single data points, then learn to approximate the marginal generator which generates the full data distribution. We show that Generator Matching unifies various generative modeling methods, including diffusion models, flow matching and discrete diffusion models. Furthermore, it expands the design space to new and unexplored Markov processes such as jump processes. Finally, Generator Matching enables the construction of superpositions of Markov generative models and enables the construction of multimodal models in a rigorous manner. We empirically validate our method on image and multimodal generation, e.g. showing that superposition with a jump process improves performance. | Flow matching, Markov process, Diffusion model, Generative Modeling | The core principles of flow matching can be vastly generalized to practically all continuous-time Markov processes using Markov generators, unifying all previous methods and opening the door to new generative models agnostic to data modality. | 3,162 | 2410.20587 | [
-0.1239093616604805,
-0.08167548477649689,
0.05239056795835495,
-0.03850816562771797,
0.04242464527487755,
-0.03382603079080582,
-0.026751568540930748,
-0.05825582891702652,
0.08921726793050766,
-0.0846395492553711,
-0.013702272437512875,
-0.08121257275342941,
0.06509773433208466,
0.0030380054377019405,
0.002648272318765521,
-0.0023772777058184147,
-0.004748580977320671,
0.08065935969352722,
-0.043979186564683914,
-0.047430939972400665,
0.04858139157295227,
0.015956155955791473,
0.00236311717890203,
-0.07163781672716141,
0.06526844203472137,
0.08754709362983704,
0.030886100605130196,
-0.047086652368307114,
0.08477886766195297,
-0.026257654651999474,
0.07288989424705505,
0.06834527850151062,
-0.1037236675620079,
0.017857013270258904,
0.028009139001369476,
0.11764335632324219,
-0.09162843227386475,
0.10106856375932693,
0.013280668295919895,
-0.022234762087464333,
0.032565515488386154,
0.025699255988001823,
0.008463875390589237,
-0.007993128150701523,
0.028847699984908104,
0.004473461303859949,
-0.014835613779723644,
-0.04888545349240303,
-0.014334632083773613,
-0.004324422217905521,
-0.04075322672724724,
0.021707911044359207,
-0.01225736178457737,
0.02930621989071369,
0.040588125586509705,
0.049867648631334305,
0.08976244181394577,
-0.07172566652297974,
0.008424484170973301,
-0.013823085464537144,
-0.06202709302306175,
-0.09190896153450012,
-0.05753163620829582,
-0.05192340165376663,
0.023409904912114143,
-0.06290251016616821,
-0.016154015436768532,
0.08310039341449738,
0.029339885339140892,
0.028668593615293503,
-0.07047499716281891,
0.03078332543373108,
-0.024829303845763206,
0.04470968246459961,
-0.042018529027700424,
0.04879426956176758,
0.04619866609573364,
-0.08699256926774979,
-0.01471853256225586,
-0.10872667282819748,
0.03551435098052025,
-0.04671074450016022,
0.09031084179878235,
-0.03267042711377144,
0.07209248095750809,
0.04237949103116989,
0.019152140244841576,
0.06328672915697098,
0.03410019353032112,
0.02401154674589634,
-0.10923031717538834,
0.07868164777755737,
-0.0025940658524632454,
-0.03217870369553566,
0.03781392052769661,
0.09000851958990097,
0.05370089411735535,
-0.10705725103616714,
0.07667304575443268,
0.0634775459766388,
0.06992436200380325,
0.029055433347821236,
0.03331605717539787,
-0.011156696826219559,
0.08115652203559875,
-0.1149468868970871,
0.0030542106833308935,
0.06274720281362534,
0.020343000069260597,
-0.035237304866313934,
0.049826715141534805,
-0.013093510642647743,
0.07445815950632095,
0.007738989777863026,
-0.077131487429142,
-0.0669793039560318,
-0.04067407175898552,
-0.020561130717396736,
-0.030139392241835594,
0.025650210678577423,
-0.014889094978570938,
0.07084101438522339,
-0.05915073677897453,
0.0417313389480114,
-0.03744344413280487,
-0.050278209149837494,
-0.0012290170416235924,
3.983566683944416e-34,
0.028747044503688812,
-0.06517749279737473,
0.02832694724202156,
0.07484368979930878,
0.0919487476348877,
0.02577418088912964,
-0.04302884638309479,
-0.046167537569999695,
-0.02741244062781334,
-0.03551434725522995,
-0.06944438070058823,
0.0023922284599393606,
-0.1280486136674881,
0.11659007519483566,
-0.0038658466655761003,
0.002218573121353984,
-0.09199270606040955,
0.023678092285990715,
0.046502240002155304,
0.03836081922054291,
0.028333930298686028,
-0.011720077134668827,
-0.05537434294819832,
0.0822862833738327,
0.001900172675959766,
0.027837561443448067,
0.0295440424233675,
-0.04706890508532524,
-0.010786252096295357,
-0.032577138394117355,
-0.13056625425815582,
0.0030211941339075565,
0.0176340714097023,
-0.011414957232773304,
0.033195141702890396,
-0.0053427694365382195,
0.008191910572350025,
-0.042254768311977386,
-0.036983951926231384,
-0.037532053887844086,
-0.054549701511859894,
0.003102112328633666,
-0.023883290588855743,
-0.008657178841531277,
-0.15899714827537537,
-0.047935448586940765,
0.049360595643520355,
0.0329112745821476,
-0.01597188040614128,
-0.021760258823633194,
0.08469546586275101,
0.013105171732604504,
0.02169843763113022,
-0.07732666283845901,
-0.05801992863416672,
-0.015380949713289738,
0.027717087417840958,
0.02792781963944435,
0.0776100605726242,
0.052266545593738556,
-0.025836410000920296,
0.10013682395219803,
0.0005039729876443744,
0.018697073683142662,
0.0644611269235611,
-0.030779501423239708,
-0.008898071944713593,
0.028866983950138092,
0.1159244105219841,
0.030060317367315292,
-0.010764352977275848,
0.0225558802485466,
-0.026516791433095932,
-0.0855470821261406,
0.07615289837121964,
-0.002617400139570236,
-0.04250186309218407,
-0.002496246015653014,
-0.036200203001499176,
0.07306551933288574,
-0.0731230154633522,
0.0010725700994953513,
-0.006417225115001202,
-0.07150325179100037,
0.06060776114463806,
-0.005794119089841843,
-0.015342939645051956,
-0.1055750623345375,
-0.09920761734247208,
-0.05515555664896965,
0.032684892416000366,
0.0013161603128537536,
-0.0016459343023598194,
0.04187823459506035,
0.0242853295058012,
-2.5766883924482544e-33,
-0.007836133241653442,
0.08619186282157898,
0.08052192628383636,
0.03899206966161728,
0.023190150037407875,
-0.007290130015462637,
0.02252170816063881,
-0.02468976378440857,
-0.005587012507021427,
-0.07696613669395447,
-0.06049920246005058,
0.0402773953974247,
0.03450116887688637,
-0.08841083943843842,
-0.04690961539745331,
-0.06422349810600281,
0.04850033298134804,
-0.0036275924649089575,
-0.03841153159737587,
0.007543686777353287,
0.01743919402360916,
0.10947848111391068,
-0.06763522326946259,
0.029212094843387604,
-0.0501023568212986,
0.052619464695453644,
-0.036743197590112686,
0.028498992323875427,
-0.01350706908851862,
0.004514315165579319,
0.018120046705007553,
-0.02065373957157135,
0.01603716053068638,
-0.007051508873701096,
-0.06897150725126266,
0.048323534429073334,
0.05449310690164566,
0.07732916623353958,
-0.0699722021818161,
0.0522092804312706,
0.01789381541311741,
-0.028032969683408737,
-0.027313200756907463,
0.08413606137037277,
-0.026957610622048378,
-0.04147018864750862,
-0.02818482182919979,
0.01337039191275835,
0.027433525770902634,
-0.010773343034088612,
-0.00681662792339921,
0.06689712405204773,
-0.11374403536319733,
-0.02005823329091072,
-0.02917144075036049,
-0.10836463421583176,
-0.05303821712732315,
-0.05143171176314354,
0.027187863364815712,
-0.011151215992867947,
-0.004763277713209391,
-0.11443088948726654,
0.00226486730389297,
-0.04523807391524315,
-0.0006844675517641008,
0.008799298666417599,
-0.05040909722447395,
-0.053065307438373566,
-0.006698145531117916,
0.03760648891329765,
0.055893708020448685,
0.01409953460097313,
0.0004220358678139746,
0.03628242015838623,
0.08310278505086899,
-0.04639417678117752,
0.10683583468198776,
0.03204760700464249,
0.0038796812295913696,
-0.05797482654452324,
-0.03520357236266136,
0.036672595888376236,
0.02205602265894413,
0.033953502774238586,
-0.021331407129764557,
0.04017196595668793,
-0.0014840044314041734,
-0.04039238765835762,
0.03952779620885849,
-0.012952434830367565,
-0.022918924689292908,
0.02997465431690216,
0.015779132023453712,
0.02789216861128807,
-0.06687243282794952,
-3.7899901883520215e-8,
-0.03506084531545639,
0.01988869346678257,
0.05539534613490105,
-0.017268715426325798,
0.05807209014892578,
0.02810061350464821,
0.024967001751065254,
0.047627370804548264,
0.06072033941745758,
-0.09766826778650284,
0.05818507820367813,
-0.027931062504649162,
0.008080576546490192,
-0.024482864886522293,
0.02649078145623207,
-0.00029922605608589947,
-0.003913392312824726,
0.04238814488053322,
-0.0014105670852586627,
-0.023430056869983673,
-0.07538363337516785,
0.004815594293177128,
-0.057143934071063995,
0.07454070448875427,
0.0478852204978466,
-0.025623928755521774,
0.002595864934846759,
-0.03599102795124054,
0.06797473877668381,
-0.025119202211499214,
-0.02354486472904682,
-0.01259686890989542,
0.006992799695581198,
0.011007498949766159,
0.03849133476614952,
0.038186412304639816,
-0.017062297090888023,
-0.017480740323662758,
-0.01943051628768444,
-0.10588347166776657,
-0.005861469078809023,
0.03088223747909069,
-0.04894149675965309,
0.03539714217185974,
-0.022449292242527008,
0.02370982989668846,
0.043027374893426895,
-0.09120055288076401,
0.025297880172729492,
0.05747402086853981,
0.025301290675997734,
-0.07190550863742828,
0.05108654499053955,
-0.011541830375790596,
0.04072890803217888,
-0.01866903342306614,
0.0439925342798233,
-0.009488385170698166,
0.08816521614789963,
0.032457929104566574,
-0.0073732477612793446,
0.030176864936947823,
-0.009441225789487362,
0.0014886613935232162
] | 0 | 0 | 0 | 0 |
|
No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images | https://openreview.net/forum?id=P4o9akekdf | [
"Botao Ye",
"Sifei Liu",
"Haofei Xu",
"Xueting Li",
"Marc Pollefeys",
"Ming-Hsuan Yang",
"Songyou Peng"
] | Oral | We introduce NoPoSplat, a feed-forward model capable of reconstructing 3D scenes parameterized by 3D Gaussians from unposed sparse multi-view images. Our model, trained exclusively with photometric loss, achieves real-time 3D Gaussian reconstruction during inference. To eliminate the need for accurate pose input during reconstruction, we anchor one input view's local camera coordinates as the canonical space and train the network to predict Gaussian primitives for all views within this space. This approach obviates the need to transform Gaussian primitives from local coordinates into a global coordinate system, thus avoiding errors associated with per-frame Gaussians and pose estimation. To resolve scale ambiguity, we design and compare various intrinsic embedding methods, ultimately opting to convert camera intrinsics into a token embedding and concatenate it with image tokens as input to the model, enabling accurate scene scale prediction. We utilize the reconstructed 3D Gaussians for novel view synthesis and pose estimation tasks and propose a two-stage coarse-to-fine pipeline for accurate pose estimation. Experimental results demonstrate that our pose-free approach can achieve superior novel view synthesis quality compared to pose-required methods, particularly in scenarios with limited input image overlap. For pose estimation, our method, trained without ground truth depth or explicit matching loss, significantly outperforms the state-of-the-art methods with substantial improvements. This work makes significant advances in pose-free generalizable 3D reconstruction and demonstrates its applicability to real-world scenarios. Code and trained models are available at https://noposplat.github.io/. | 3D Gaussian Splatting, Pose Free, Pose Estimation, Novel View Synthesis, 3D Reconstruction | NoPoSplat is a novel feed-forward model that reconstructs scenes from unposed images by predicting Gaussians in a canonical space, demonstrating superior performance in both novel view synthesis and pose estimation. | 3,116 | 2410.24207 | [
-0.02851003035902977,
-0.0925145149230957,
0.023567551746964455,
-0.022609353065490723,
0.09119702875614166,
-0.0065258522517979145,
-0.020629355683922768,
-0.039942704141139984,
0.018414173275232315,
-0.059157293289899826,
0.043098222464323044,
-0.061981044709682465,
0.05293440818786621,
0.0759463682770729,
-0.03738338500261307,
-0.05145882070064545,
0.08911070972681046,
0.13916762173175812,
-0.05215258523821831,
0.05296466872096062,
-0.046102412045001984,
-0.03961598500609398,
0.02884729579091072,
-0.014737091958522797,
0.06561451405286789,
-0.01944921165704727,
0.05995699763298035,
-0.07862313091754913,
-0.021638428792357445,
-0.01671532541513443,
0.0338829904794693,
-0.007102838717401028,
-0.05558723583817482,
-0.04154631495475769,
-0.02022235468029976,
0.05999292433261871,
0.023045143112540245,
0.010625002905726433,
-0.010366322472691536,
-0.0338967926800251,
0.06484303623437881,
0.05533165484666824,
-0.01573222316801548,
0.07228269428014755,
0.09334010630846024,
0.09433387219905853,
0.07554232329130173,
-0.0723382905125618,
0.021645596250891685,
-0.009565908461809158,
-0.024083415046334267,
-0.0014767873799428344,
0.030408291146159172,
-0.014680681750178337,
-0.015165943652391434,
-0.0409356914460659,
-0.023229006677865982,
-0.1051565632224083,
0.04194742068648338,
0.0446564182639122,
0.042441025376319885,
-0.021216293796896935,
0.006265763193368912,
0.04203638434410095,
-0.03830411657691002,
0.004107092507183552,
0.029268914833664894,
-0.024887120351195335,
-0.04410453140735626,
0.08918062597513199,
-0.06982095539569855,
0.018611053004860878,
0.028871243819594383,
0.004955936688929796,
0.00593984592705965,
0.0363992415368557,
0.03538262099027634,
-0.07659226655960083,
0.00555505882948637,
-0.02340218797326088,
0.07147997617721558,
0.016962213441729546,
-0.02852666936814785,
0.01773647405207157,
0.055848393589258194,
0.012816994450986385,
-0.009212346747517586,
0.10969768464565277,
0.008943690918385983,
-0.02231558784842491,
-0.09809628874063492,
-0.02752099558711052,
-0.04751795157790184,
-0.07445669174194336,
0.0545785017311573,
-0.031528547406196594,
-0.0009409462800249457,
-0.03412025794386864,
-0.018235495314002037,
0.05069225654006004,
0.012315456755459309,
0.0750819742679596,
0.07302635908126831,
-0.014644897542893887,
0.021196046844124794,
0.025594061240553856,
-0.02513526752591133,
0.026751989498734474,
0.04601210728287697,
-0.04039434343576431,
0.02347988821566105,
-0.019942138344049454,
0.013400556519627571,
0.011641652323305607,
0.026177478954195976,
0.011899968609213829,
0.00796584878116846,
0.0002995124086737633,
-0.01830863207578659,
-0.019485559314489365,
0.000797622837126255,
-0.029015041887760162,
-0.03732672333717346,
0.029775278642773628,
-0.03166569396853447,
0.021958835422992706,
-0.022328149527311325,
1.7616465273361593e-33,
0.02567257359623909,
0.10443156212568283,
0.08525405079126358,
0.03490187227725983,
0.04796040430665016,
0.055253662168979645,
-0.02697211503982544,
0.03871341794729233,
-0.03609778732061386,
0.033858656883239746,
0.003085874719545245,
-0.08651852607727051,
0.03209031745791435,
0.10604788362979889,
-0.025795333087444305,
0.06811349093914032,
-0.0380045510828495,
0.010243494063615799,
0.02662346325814724,
0.08796840906143188,
-0.032255031168460846,
-0.02061452344059944,
-0.05181046202778816,
-0.020609315484762192,
-0.06102987751364708,
-0.0462479405105114,
0.05170026421546936,
-0.026053955778479576,
-0.08204817026853561,
-0.0425577238202095,
-0.06467384099960327,
-0.00776022020727396,
0.010404650121927261,
0.017164526507258415,
0.05318516120314598,
-0.10816925764083862,
0.052321091294288635,
-0.02075747586786747,
-0.012039551511406898,
-0.02672368474304676,
-0.0064435382373631,
0.086203433573246,
-0.04764334857463837,
-0.019585052505135536,
-0.09581570327281952,
-0.014449521899223328,
0.04092052951455116,
0.09312774240970612,
-0.05582607537508011,
0.002835653955116868,
0.103133924305439,
-0.00703719025477767,
-0.07099144905805588,
-0.08434662967920303,
-0.08433408290147781,
-0.04422897472977638,
-0.014348256401717663,
-0.026214206591248512,
0.031490929424762726,
-0.04595677927136421,
-0.005518456920981407,
-0.011619262397289276,
-0.025138363242149353,
0.02930174209177494,
-0.044297464191913605,
0.0010890401899814606,
0.004319565836340189,
0.05095042660832405,
0.08783946931362152,
0.07361654937267303,
-0.023221053183078766,
-0.030684344470500946,
-0.020146923139691353,
-0.012942187488079071,
0.05890515074133873,
0.06679446995258331,
-0.01763145439326763,
-0.07507569342851639,
0.036072175949811935,
0.1526261568069458,
-0.034079793840646744,
-0.022968165576457977,
0.0028669522143900394,
0.04538686200976372,
-0.07855325192213058,
-0.0402413047850132,
0.011428400874137878,
0.0024763362016528845,
0.011938581243157387,
-0.05227801203727722,
-0.023734761402010918,
-0.099746473133564,
-0.09579838067293167,
-0.057144612073898315,
-0.02276476100087166,
-7.11714717813366e-34,
0.11485495418310165,
0.04217063635587692,
-0.05248429626226425,
-0.030276194214820862,
-0.006163980346173048,
-0.10835152864456177,
0.018270496279001236,
0.02702314965426922,
0.05893446505069733,
-0.04220426455140114,
-0.03806737810373306,
-0.009323887526988983,
0.02094290778040886,
0.029765989631414413,
0.03977251052856445,
0.052833832800388336,
-0.06595254689455032,
-0.0023854044266045094,
-0.08552294224500656,
-0.04142060875892639,
0.03957758843898773,
0.023286940529942513,
0.02419973351061344,
0.048963554203510284,
-0.01687777042388916,
0.18788355588912964,
0.0007335268310271204,
0.08049462735652924,
-0.05984165519475937,
-0.01737169548869133,
-0.020529575645923615,
-0.03455139696598053,
0.03267867863178253,
-0.0704633891582489,
-0.05172742158174515,
0.07963816076517105,
0.019854875281453133,
0.05437029153108597,
-0.054663460701704025,
-0.010427182540297508,
-0.0026268246583640575,
0.003985664341598749,
-0.026059376075863838,
0.0933334231376648,
0.012338676489889622,
0.042525023221969604,
0.04926717281341553,
0.002705489518120885,
0.017760029062628746,
0.12024470418691635,
-0.10854709148406982,
0.13454702496528625,
-0.0804094597697258,
-0.005974461790174246,
-0.015555726364254951,
-0.08422011882066727,
-0.09659582376480103,
-0.05689188838005066,
0.09194529056549072,
-0.029981758445501328,
-0.000556496437638998,
-0.04090004414319992,
-0.0685652419924736,
-0.00017604818276595324,
-0.03647112473845482,
-0.009687240235507488,
0.011800795793533325,
0.004905487410724163,
-0.020987417548894882,
0.05324451997876167,
-0.02035163715481758,
-0.009208754636347294,
-0.0009335234062746167,
-0.005156459286808968,
-0.010379414074122906,
0.010705849155783653,
0.018927643075585365,
0.010671796277165413,
0.07250913232564926,
-0.006829211488366127,
0.08273949474096298,
-0.07210678607225418,
0.02922947145998478,
0.06764426827430725,
0.09799225628376007,
0.10033377259969711,
0.001078032306395471,
0.014296081848442554,
0.06645273417234421,
0.012410636059939861,
-0.07259391993284225,
0.014616439118981361,
-0.036697763949632645,
0.042764030396938324,
0.013966808095574379,
-3.6976373962716025e-8,
-0.09094998985528946,
0.09175100922584534,
0.008794334717094898,
-0.01617240533232689,
-0.0744965597987175,
-0.055537305772304535,
0.004989197012037039,
0.05388626083731651,
0.023496437817811966,
-0.06549840420484543,
0.0472727045416832,
-0.09751206636428833,
-0.010511282831430435,
-0.035023033618927,
0.03529331833124161,
0.040520329028367996,
0.05168605595827103,
-0.021867826581001282,
-0.04469722509384155,
0.02166147343814373,
-0.055654559284448624,
0.020347312092781067,
0.01988019421696663,
0.007853842340409756,
0.004914363846182823,
-0.05598606914281845,
-0.03863631188869476,
0.06522496789693832,
0.11926640570163727,
-0.010164650157094002,
-0.025871820747852325,
-0.007234958466142416,
0.035685352981090546,
-0.03733181208372116,
-0.05843265727162361,
-0.006836824119091034,
-0.03455886244773865,
-0.02633659355342388,
0.04915250837802887,
-0.09337255358695984,
-0.07365050166845322,
0.04917382076382637,
0.000005301382316247327,
0.013420459814369678,
-0.036185093224048615,
0.0576421394944191,
0.03757239505648613,
-0.017888536676764488,
0.03958900272846222,
-0.02460402250289917,
-0.026604704558849335,
-0.02692681923508644,
-0.00420468021184206,
0.11321030557155609,
0.041374944150447845,
-0.007438631262630224,
0.03519108518958092,
-0.05948486924171448,
0.0563335120677948,
0.023938389495015144,
-0.06789582967758179,
-0.0193374864757061,
-0.06859484314918518,
-0.022757600992918015
] | https://github.com/cvg/NoPoSplat | 744 | 0 | 0 | 0 |
Variational Diffusion Posterior Sampling with Midpoint Guidance | https://openreview.net/forum?id=6EUtjXAvmj | [
"Badr MOUFAD",
"Yazid Janati",
"Lisa Bedin",
"Alain Oliviero Durmus",
"randal douc",
"Eric Moulines",
"Jimmy Olsson"
] | Oral | Diffusion models have recently shown considerable potential in solving Bayesian inverse problems when used as priors. However, sampling from the resulting denoising posterior distributions remains a challenge as it involves intractable terms. To tackle this issue, state-of-the-art approaches formulate the problem as that of sampling from a surrogate diffusion model targeting the posterior and decompose its scores into two terms: the prior score and an intractable guidance term. While the former is replaced by the pre-trained score of the considered diffusion model, the guidance term has to be estimated. In this paper, we propose a novel approach that utilises a decomposition of the transitions which, in contrast to previous methods, allows a trade-off between the complexity of the intractable guidance term and that of the prior transitions. We validate the proposed approach through extensive experiments on linear and nonlinear inverse problems, including challenging cases with latent diffusion models as priors, and demonstrate its effectiveness in reconstructing electrocardiogram (ECG) from partial measurements for accurate cardiac diagnosis. | Diffusion models, Inverse problems, posterior sampling | null | 3,058 | 2410.09945 | [
-0.0204713623970747,
-0.06354806572198868,
0.08635801076889038,
0.05638383328914642,
0.0466940775513649,
-0.05554063990712166,
-0.015195099636912346,
-0.023426614701747894,
0.0390077643096447,
-0.02187807485461235,
0.020182490348815918,
-0.004345900844782591,
-0.0066343070939183235,
0.04886193946003914,
-0.046188417822122574,
-0.050368331372737885,
0.028444424271583557,
0.027996348217129707,
-0.04648251831531525,
0.03183199092745781,
-0.01252429373562336,
0.06977581232786179,
-0.06351301074028015,
-0.023829789832234383,
0.04980173334479332,
-0.01224728673696518,
0.06627896428108215,
-0.05796840414404869,
-0.04318477213382721,
-0.022935321554541588,
0.0722612738609314,
-0.017335185781121254,
0.02203129231929779,
0.028407979756593704,
-0.02929810993373394,
0.04317015781998634,
0.03149518370628357,
-0.0014898402150720358,
0.007211083080619574,
0.06154290959239006,
0.06010694429278374,
0.06789135187864304,
-0.006481717340648174,
0.10326797515153885,
0.15345197916030884,
-0.07294868677854538,
-0.03217579796910286,
-0.06949124485254288,
0.03875690698623657,
0.06343501061201096,
0.05534490570425987,
0.016060272231698036,
-0.043822310864925385,
0.027809349820017815,
0.03157608211040497,
0.022268477827310562,
0.01635284535586834,
-0.020504439249634743,
0.03684919700026512,
-0.03920101746916771,
-0.07366826385259628,
-0.05392739549279213,
-0.05182891711592674,
-0.04495181515812874,
-0.016984494403004646,
-0.00955670140683651,
0.04190245270729065,
0.023487554863095284,
0.04111235961318016,
0.11404350399971008,
-0.042746651917696,
0.03617091849446297,
-0.056294165551662445,
-0.04899987578392029,
-0.03492504358291626,
-0.013201084919273853,
-0.016781054437160492,
0.03439786657691002,
0.0025455355644226074,
-0.06973692029714584,
0.04348912090063095,
-0.03605616092681885,
0.023821530863642693,
-0.04643012583255768,
0.07146061956882477,
0.023368626832962036,
-0.002113115508109331,
0.0236516073346138,
0.036182135343551636,
0.01116036158055067,
-0.018758924677968025,
-0.06313934177160263,
-0.1093607097864151,
-0.03841014951467514,
-0.014329103752970695,
-0.007146092597395182,
0.011170958168804646,
-0.034223444759845734,
0.08931027352809906,
0.08752719312906265,
0.019211743026971817,
0.02543278969824314,
-0.036974940448999405,
0.0178962554782629,
-0.012511193752288818,
-0.09084127843379974,
0.06953199952840805,
0.07223419845104218,
0.0037994245067238808,
0.05643695592880249,
0.03222840651869774,
-0.03215064853429794,
0.06268062442541122,
0.022222520783543587,
0.08389269560575485,
-0.07258109748363495,
0.08927998691797256,
0.002809608355164528,
0.031155558302998543,
0.05878553166985512,
0.004120304714888334,
-0.0284525528550148,
-0.01985601708292961,
-0.024142414331436157,
0.035131145268678665,
0.029834909364581108,
-0.027576230466365814,
1.0409782159893737e-33,
-0.0011271425755694509,
-0.048062898218631744,
0.0701548308134079,
0.057600054889917374,
-0.0246218703687191,
0.002301407279446721,
-0.05014888197183609,
-0.06897080689668655,
-0.012706383131444454,
0.01792505569756031,
-0.03411705791950226,
0.04003880172967911,
-0.08535351604223251,
0.02169644460082054,
0.011247958056628704,
0.007922648452222347,
-0.03621074557304382,
0.06371922791004181,
-0.006963563151657581,
-0.03564073517918587,
0.019338350743055344,
-0.03858083114027977,
-0.06947088986635208,
-0.07762480527162552,
-0.027307361364364624,
-0.02420995756983757,
0.012095765210688114,
-0.04133861884474754,
-0.0008745254599489272,
0.028190134093165398,
-0.15802273154258728,
-0.030413338914513588,
-0.038345374166965485,
0.04124758020043373,
-0.035524882376194,
0.0773465558886528,
0.10740584880113602,
-0.0016921398928388953,
-0.008902687579393387,
-0.02491391822695732,
0.05129306763410568,
0.02121102437376976,
0.0525897741317749,
-0.04657324403524399,
-0.08361369371414185,
-0.056627966463565826,
0.0312860868871212,
-0.029353350400924683,
0.03312354534864426,
-0.0492943674325943,
-0.07265176624059677,
-0.07413355261087418,
0.06914573907852173,
-0.11339616775512695,
-0.05861172452569008,
0.010938305407762527,
-0.018604284152388573,
-0.005554200615733862,
0.0033048661425709724,
0.013714781031012535,
0.09820865094661713,
0.05239001661539078,
-0.0493372343480587,
0.026844199746847153,
0.04846428707242012,
0.04366256669163704,
-0.07175437361001968,
0.005949126090854406,
-0.0042859455570578575,
0.02084452472627163,
-0.04019257798790932,
0.004139284137636423,
-0.014545541256666183,
-0.05947377532720566,
0.04242442920804024,
0.03725219517946243,
-0.01675429195165634,
0.031999651342630386,
0.01247432455420494,
-0.13767379522323608,
-0.05672202259302139,
-0.0195122379809618,
-0.04605277627706528,
-0.0022062917705625296,
0.015502109192311764,
0.011517121456563473,
0.032413091510534286,
-0.044265735894441605,
-0.14876176416873932,
-0.0626152902841568,
-0.0547073557972908,
0.041760027408599854,
-0.050513025373220444,
0.04017509147524834,
0.022534793242812157,
-1.4025001260523259e-33,
-0.05572313070297241,
0.05292556807398796,
0.059127796441316605,
-0.02111527882516384,
-0.018818911164999008,
-0.04532698169350624,
0.037658222019672394,
0.06338290870189667,
0.08152581006288528,
-0.11832039058208466,
-0.04430798068642616,
-0.0413503535091877,
0.0211807768791914,
0.027187170460820198,
0.022808212786912918,
0.039638444781303406,
-0.059025101363658905,
0.005544843152165413,
0.01750623993575573,
-0.004471944645047188,
-0.04767917841672897,
0.004668307024985552,
0.0080998744815588,
0.05225006490945816,
-0.0697973445057869,
0.07011294364929199,
-0.027143262326717377,
0.09490147233009338,
0.019389808177947998,
-0.10253570973873138,
-0.07060161232948303,
-0.023718900978565216,
0.006570643279701471,
-0.03933093324303627,
-0.14672687649726868,
0.16579975187778473,
-0.01636841893196106,
0.06412992626428604,
-0.007764458656311035,
0.027126412838697433,
0.026799209415912628,
0.005632823798805475,
-0.018914321437478065,
-0.03788122162222862,
0.04869134724140167,
-0.07898285239934921,
0.06073492765426636,
0.017463291063904762,
0.02124132215976715,
0.01700872741639614,
-0.02365216799080372,
0.03379204869270325,
-0.06120757386088371,
0.087195485830307,
-0.13215544819831848,
-0.024153131991624832,
-0.07059383392333984,
-0.03000510111451149,
0.06166988983750343,
-0.010847213678061962,
-0.015396605245769024,
0.005230578128248453,
0.0415022149682045,
-0.04032742977142334,
-0.0057758851908147335,
0.03009173274040222,
-0.003184331813827157,
0.03321385756134987,
-0.009848157875239849,
0.03198152780532837,
-0.04400494694709778,
0.04450100287795067,
0.0033350414596498013,
-0.004082272294908762,
0.0717427134513855,
-0.09604372084140778,
-0.027739128097891808,
-0.048045992851257324,
-0.04174411669373512,
-0.05689170956611633,
0.08428696542978287,
-0.013539883308112621,
-0.040319137275218964,
-0.036220572888851166,
0.014537177979946136,
0.0106694670394063,
0.04282807558774948,
-0.05240960419178009,
-0.0012064382899552584,
-0.046261467039585114,
-0.009219215251505375,
-0.003786167362704873,
0.05657874792814255,
0.014430496841669083,
-0.10962596535682678,
-4.399239372787633e-8,
-0.10149377584457397,
0.08110751211643219,
0.062017813324928284,
0.03918154165148735,
0.06548375636339188,
0.0030093493405729532,
0.022238116711378098,
0.047264546155929565,
0.006705181673169136,
-0.07105579972267151,
0.059978682547807693,
-0.007239685859531164,
0.024576760828495026,
-0.09335153549909592,
-0.032205257564783096,
0.12859658896923065,
-0.0021902795415371656,
0.06837530434131622,
-0.012668807990849018,
-0.00812505278736353,
0.028188439086079597,
0.04594343900680542,
-0.0068059079349040985,
-0.035492945462465286,
0.12976382672786713,
0.004250227473676205,
-0.06764192879199982,
-0.011866354383528233,
0.049490612000226974,
0.044548001140356064,
-0.0008423493127338588,
0.01962745189666748,
0.029011759907007217,
0.11194775253534317,
-0.06260953843593597,
0.029435284435749054,
0.03148418664932251,
-0.024023042991757393,
-0.0981818288564682,
-0.031248388811945915,
-0.060295455157756805,
-0.04229637607932091,
-0.06976018100976944,
0.023403046652674675,
0.008749359287321568,
0.016082055866718292,
0.07058335095643997,
-0.00279808696359396,
0.046930573880672455,
0.015920842066407204,
0.0713672786951065,
-0.04245934262871742,
0.033379536122083664,
0.0064946687780320644,
0.05045849457383156,
-0.004324804991483688,
0.04386737197637558,
0.023043978959321976,
0.03524994105100632,
0.05171821266412735,
-0.0015500267036259174,
0.04355153068900108,
-0.022275026887655258,
0.08764799684286118
] | https://github.com/yazidjanati/mgps | 9 | 0 | 0 | 0 |
Towards Understanding Why FixMatch Generalizes Better Than Supervised Learning | https://openreview.net/forum?id=25kAzqzTrz | [
"Jingyang Li",
"Jiachun Pan",
"Vincent Y. F. Tan",
"Kim-chuan Toh",
"Pan Zhou"
] | Oral | Semi-supervised learning (SSL), exemplified by FixMatch (Sohn et al., 2020), has shown significant generalization advantages over supervised learning (SL), particularly in the context of deep neural networks (DNNs). However, it is still unclear, from a theoretical standpoint, why FixMatch-like SSL algorithms generalize better than SL on DNNs. In this work, we present the first theoretical justification for the enhanced test accuracy observed in FixMatch-like SSL applied to DNNs by taking convolutional neural networks (CNNs) on classification tasks as an example. Our theoretical analysis reveals that the semantic feature learning processes in FixMatch and SL are rather different. In particular, FixMatch learns all the discriminative features of each semantic class, while SL only randomly captures a subset of features due to the well-known lottery ticket hypothesis. Furthermore, we show that our analysis framework can be applied to other FixMatch-like SSL methods, e.g., FlexMatch, FreeMatch, Dash, and SoftMatch. Inspired by our theoretical analysis, we develop an improved variant of FixMatch, termed Semantic-Aware FixMatch (SA-FixMatch). Experimental results corroborate our theoretical findings and the enhanced generalization capability of SA-FixMatch. | deep semi-supervised learning, generalization error, feature learning | null | 2,984 | 2410.11206 | [
-0.09914189577102661,
-0.050590332597494125,
0.06386592239141464,
0.0389854721724987,
0.049703776836395264,
0.02944020926952362,
0.021811580285429955,
-0.07122904062271118,
-0.003243356943130493,
-0.04830094426870346,
-0.04868647828698158,
-0.02854202501475811,
0.005951968487352133,
0.05165239796042442,
-0.03746693581342697,
-0.08327198028564453,
0.08943195641040802,
0.07837270200252533,
-0.14719265699386597,
0.027388207614421844,
0.002306899754330516,
0.009320435114204884,
-0.004364767577499151,
-0.010160203091800213,
0.039917465299367905,
-0.010349586606025696,
-0.04760266840457916,
-0.025265995413064957,
-0.006028028205037117,
-0.010972663760185242,
0.009934470057487488,
0.09252184629440308,
-0.058330435305833817,
0.06241723150014877,
-0.025234706699848175,
-0.003967981319874525,
-0.028298886492848396,
0.1028764545917511,
0.019111817702651024,
-0.06317867338657379,
0.014066189527511597,
-0.0945245549082756,
-0.011593244969844818,
0.037709757685661316,
0.004151089116930962,
0.03023543953895569,
0.005929071456193924,
-0.0008777906186878681,
0.0003862953162752092,
-0.016639403998851776,
-0.03091004677116871,
0.042435213923454285,
-0.04763920605182648,
0.11870907247066498,
-0.09373795241117477,
0.016552019864320755,
0.053700607270002365,
0.0389983095228672,
-0.05253472179174423,
0.03274635970592499,
-0.02428826317191124,
-0.11325202137231827,
-0.013091848231852055,
-0.026934627443552017,
-0.011773769743740559,
-0.04147293046116829,
-0.00978870503604412,
-0.020301716402173042,
-0.0011351462453603745,
0.02528359554708004,
0.05136281996965408,
0.12114212661981583,
0.007923450320959091,
0.0784807950258255,
0.06746597588062286,
-0.021926430985331535,
0.048798952251672745,
-0.018818998709321022,
0.06539187580347061,
-0.0267349760979414,
0.015462270006537437,
-0.001973201986402273,
0.03966419771313667,
0.029825763776898384,
0.22032591700553894,
0.0107124587520957,
-0.021053694188594818,
-0.00438053160905838,
-0.0020050371531397104,
0.0033588637597858906,
-0.01907922327518463,
-0.05779118463397026,
0.016841571778059006,
-0.09576772153377533,
0.03288908675312996,
0.011949536390602589,
-0.0033742005471140146,
-0.037087228149175644,
-0.03687814250588417,
0.038956418633461,
-0.01064891368150711,
0.01743849366903305,
0.011336412280797958,
-0.12120935320854187,
0.05629779025912285,
0.05875778943300247,
-0.013734117150306702,
0.009651088155806065,
0.0847494900226593,
-0.15588577091693878,
0.006598072126507759,
0.05217772722244263,
0.05700475722551346,
0.0013728311751037836,
-0.037517085671424866,
0.0003044398035854101,
-0.10627731680870056,
-0.023789387196302414,
0.045111894607543945,
-0.011387139558792114,
-0.08555429428815842,
0.07252586632966995,
0.01601329818367958,
0.057484738528728485,
-0.0200816560536623,
-0.005796134937554598,
-0.06213401257991791,
3.7252319638840486e-33,
0.05302118882536888,
-0.01411966048181057,
-0.06852567195892334,
-0.04049965739250183,
0.014247599989175797,
-0.027765942737460136,
-0.016575152054429054,
0.09555987268686295,
-0.030816854909062386,
0.04661157354712486,
-0.021924050524830818,
0.015536155551671982,
-0.024076413363218307,
0.026908058673143387,
-0.023050660267472267,
0.05302875488996506,
-0.08110854774713516,
-0.03342097997665405,
-0.056333862245082855,
-0.04868816211819649,
0.043399661779403687,
0.007840272970497608,
0.07139665633440018,
-0.032958440482616425,
-0.031246468424797058,
-0.05936828628182411,
0.0042404308915138245,
0.0017976872622966766,
0.021663052961230278,
0.009204576723277569,
-0.06850176304578781,
0.02113563008606434,
0.06375965476036072,
0.10597901046276093,
0.051754966378211975,
0.004013980273157358,
0.01609046384692192,
-0.02165205404162407,
0.015495005995035172,
-0.028033772483468056,
-0.09226436913013458,
0.05308661609888077,
-0.056041840463876724,
-0.024560891091823578,
0.001389993354678154,
0.02072412148118019,
-0.018911411985754967,
-0.028730086982250214,
0.013669176027178764,
-0.02062392793595791,
0.04491626098752022,
0.00107968645170331,
-0.061745066195726395,
-0.042012251913547516,
-0.037489112466573715,
0.0041732401587069035,
0.06305040419101715,
0.031265802681446075,
-0.06354520469903946,
0.042378827929496765,
-0.07761449366807938,
-0.012903694994747639,
-0.007397179491817951,
0.008225720375776291,
-0.03940679877996445,
0.015604881569743156,
0.05593319609761238,
0.0642307698726654,
0.005433196667581797,
-0.012328263372182846,
0.014260602183640003,
0.06406602263450623,
-0.008392228744924068,
-0.029879311099648476,
0.06606322526931763,
-0.008257388137280941,
0.0738210678100586,
0.0064225574024021626,
-0.012105044908821583,
0.04312463477253914,
0.043963391333818436,
-0.03588533028960228,
-0.0032508710864931345,
-0.03174121305346489,
-0.11653534322977066,
-0.004062609281390905,
0.006138484459370375,
-0.0660439282655716,
0.02693098410964012,
-0.08950366079807281,
-0.04809541255235672,
-0.04040619358420372,
-0.004522965755313635,
0.008353904820978642,
-0.002201526425778866,
-3.428071032758005e-33,
-0.07529225945472717,
-0.005301984027028084,
-0.05886026471853256,
0.1137285903096199,
-0.039307914674282074,
-0.03619474917650223,
-0.004126826301217079,
0.02509334310889244,
-0.005877086892724037,
0.02883579023182392,
0.0340191125869751,
-0.05648144334554672,
-0.012969198636710644,
0.01817925274372101,
-0.0032973787747323513,
-0.03810359910130501,
-0.09335003048181534,
0.020400702953338623,
0.013665358535945415,
0.09943202883005142,
0.03581802174448967,
0.15554937720298767,
-0.06429237127304077,
0.11793947219848633,
-0.0086100148037076,
-0.011443857103586197,
-0.05874231830239296,
0.0278839822858572,
0.02131233923137188,
-0.003186842892318964,
-0.015687033534049988,
-0.005442970432341099,
-0.05365286394953728,
-0.019722068682312965,
0.026206588372588158,
0.018988344818353653,
0.013211756013333797,
-0.03456445410847664,
-0.011918424628674984,
0.04337102174758911,
0.041768547147512436,
-0.023814380168914795,
-0.06722215563058853,
0.053075917065143585,
0.040834005922079086,
-0.05072449892759323,
-0.05013851448893547,
0.029696810990571976,
0.010366029106080532,
-0.002496439265087247,
0.011470336467027664,
-0.0025860315654426813,
-0.058144982904195786,
0.01182537991553545,
-0.0353277213871479,
-0.0199881661683321,
-0.02306678704917431,
-0.013678919523954391,
-0.03492424637079239,
0.09276409447193146,
-0.09787826240062714,
-0.026287047192454338,
0.057089343667030334,
-0.0025686861481517553,
0.018128491938114166,
0.024367820471525192,
-0.08263687789440155,
0.04230004921555519,
0.015704941004514694,
0.033643923699855804,
-0.000299152365187183,
-0.03182215616106987,
0.006080233491957188,
0.03213917091488838,
-0.043278422206640244,
-0.0754581019282341,
0.05922573804855347,
-0.02092735469341278,
-0.04566393792629242,
0.019601717591285706,
0.04799932986497879,
-0.07233570516109467,
0.02346179448068142,
0.047790464013814926,
0.05260774865746498,
0.08210326731204987,
0.03989645093679428,
0.07053065299987793,
0.03267422690987587,
-0.02209239825606346,
-0.030991213396191597,
0.08095569163560867,
-0.011008361354470253,
0.030456801876425743,
0.01656874641776085,
-4.317897506211921e-8,
-0.08554695546627045,
0.09292924404144287,
-0.04465765133500099,
0.019202755764126778,
0.06004495546221733,
-0.04983743280172348,
-0.009672857820987701,
0.11063981801271439,
-0.05094977840781212,
-0.033297426998615265,
-0.013697127811610699,
0.06802409142255783,
-0.07625053077936172,
-0.06208192557096481,
-0.040270641446113586,
0.04230185225605965,
-0.042975544929504395,
-0.007646085228770971,
0.049869418144226074,
0.012821875512599945,
0.1315649002790451,
-0.008297929540276527,
0.02042970433831215,
-0.02533184178173542,
-0.04118737205862999,
-0.05888478830456734,
-0.04375568777322769,
0.0851454809308052,
0.005471188575029373,
0.006216925568878651,
-0.021862035617232323,
-0.03147487714886665,
0.03354742005467415,
-0.025615233927965164,
0.1400957703590393,
0.12374583631753922,
0.0033549091313034296,
-0.08373274654150009,
0.043684955686330795,
0.06155075877904892,
0.021851588040590286,
0.14353233575820923,
0.027374323457479477,
0.0009056424023583531,
-0.04143130034208298,
-0.037994999438524246,
-0.01652190275490284,
-0.08061020821332932,
0.06284387409687042,
0.01690296083688736,
0.0007680398412048817,
-0.018716074526309967,
-0.023351341485977173,
0.04101267084479332,
0.001276110066100955,
0.008841807022690773,
-0.003344483906403184,
-0.09132799506187439,
0.016466854140162468,
0.11085338145494461,
0.03260732814669609,
0.060077108442783356,
-0.019559260457754135,
0.006950829643756151
] | 0 | 0 | 0 | 0 |
|
NeuralPlane: Structured 3D Reconstruction in Planar Primitives with Neural Fields | https://openreview.net/forum?id=5UKrnKuspb | [
"Hanqiao Ye",
"Yuzhou Liu",
"Yangdong Liu",
"Shuhan Shen"
] | Oral | 3D maps assembled from planar primitives are compact and expressive in representing man-made environments. In this paper, we present **NeuralPlane**, a novel approach that explores **neural** fields for multi-view 3D **plane** reconstruction. Our method is centered upon the core idea of distilling geometric and semantic cues from inconsistent 2D plane observations into a unified 3D neural representation, which unlocks the full leverage of plane attributes. It is accomplished through several key designs, including: 1) a monocular module that generates geometrically smooth and semantically meaningful segments known as 2D plane observations, 2) a plane-guided training procedure that implicitly learns accurate 3D geometry from the multi-view plane observations, and 3) a self-supervised feature field termed *Neural Coplanarity Field* that enables the modeling of scene semantics alongside the geometry. Without relying on prior plane annotations, our method achieves high-fidelity reconstruction comprising planar primitives that are not only crisp but also well-aligned with the semantic content. Comprehensive experiments on ScanNetv2 and ScanNet++ demonstrate the superiority of our method in both geometry and semantics. | 3D Reconstruction, 3D Scene Understanding, Scene Abstraction, Neural Rendering | NeuralPlane rebuilds indoor scenes as arrangements of planar primitives from multi-view images. | 2,933 | null | [
-0.000860394851770252,
-0.07376120239496231,
0.009059257805347443,
0.003339008893817663,
0.04867983236908913,
-0.018087094649672508,
0.016076786443591118,
-0.09400386363267899,
-0.03489910438656807,
-0.022570425644516945,
-0.06432896107435226,
0.0019243824062868953,
-0.04044245928525925,
0.08979973942041397,
-0.010835142806172371,
0.00466646533459425,
0.0402081273496151,
0.07197312265634537,
-0.06982381641864777,
0.08069473505020142,
-0.0048673502169549465,
0.05979685112833977,
0.022024724632501602,
0.018780682235956192,
0.03433891013264656,
0.037675678730010986,
0.0741107240319252,
-0.028859630227088928,
0.007874676957726479,
0.010348782874643803,
0.1105131283402443,
0.043765947222709656,
-0.10820916295051575,
0.046687230467796326,
0.019719600677490234,
-0.02243557758629322,
0.030956517904996872,
0.06714573502540588,
-0.03656108304858208,
-0.05755849927663803,
-0.017287831753492355,
0.057931553572416306,
-0.019192714244127274,
0.09830041974782944,
0.0643547922372818,
0.10622233897447586,
0.00013625696010421962,
-0.10743464529514313,
0.06360390037298203,
-0.003403382608667016,
-0.026423033326864243,
-0.04071304202079773,
-0.014846645295619965,
-0.026539310812950134,
0.04443502426147461,
0.10403918474912643,
-0.003051694016903639,
-0.04150117561221123,
-0.02779732085764408,
-0.0035896177869290113,
0.04756750911474228,
0.011251157149672508,
0.00871084351092577,
-0.015911635011434555,
-0.05138558894395828,
-0.040373366326093674,
-0.02991677075624466,
-0.009586678817868233,
0.016312718391418457,
-0.010715490207076073,
-0.01365569420158863,
0.010725158266723156,
0.01422185730189085,
-0.0652657002210617,
0.011127113364636898,
0.02848823182284832,
0.025237303227186203,
-0.059381548315286636,
0.0034947495441883802,
-0.10635898262262344,
-0.032598014920949936,
0.029370740056037903,
0.004510356578975916,
0.09974121302366257,
-0.015500904992222786,
0.049009643495082855,
-0.009122077375650406,
0.10240995138883591,
-0.0447031632065773,
-0.008850939571857452,
-0.025531670078635216,
-0.004177127033472061,
-0.027922168374061584,
-0.06435032933950424,
-0.022199278697371483,
-0.042617883533239365,
0.042787741869688034,
-0.1471189260482788,
-0.02065776288509369,
0.06972941011190414,
0.023089643567800522,
0.01678377203643322,
0.025321701541543007,
-0.04080362990498543,
-0.009549589827656746,
-0.03294369578361511,
0.0668659508228302,
-0.0055995602160692215,
0.0007600969984196126,
-0.008390508592128754,
0.001456796657294035,
0.02127307467162609,
-0.01791582815349102,
0.0011604891624301672,
0.03607011213898659,
0.009302285499870777,
-0.06523258984088898,
-0.0031773089431226254,
0.08827933669090271,
-0.009837320074439049,
-0.08272416144609451,
0.006663214415311813,
-0.00312969577498734,
0.10694795101881027,
-0.021392663940787315,
-0.02841460332274437,
0.012069073505699635,
-2.945859556492694e-34,
0.012183436192572117,
0.08232390135526657,
0.02800222858786583,
0.047148387879133224,
0.05099746957421303,
0.034219659864902496,
0.0359230637550354,
0.032833028584718704,
0.04486182704567909,
0.04830249398946762,
-0.06169653311371803,
-0.1082255020737648,
0.02809038758277893,
0.10677830874919891,
0.011663215234875679,
-0.02771657332777977,
-0.06131216138601303,
-0.00008135750249493867,
-0.05959188938140869,
0.0918206125497818,
-0.01982695795595646,
-0.013937057927250862,
0.014353282749652863,
0.028970951214432716,
0.05029263347387314,
-0.02094019204378128,
-0.006845028139650822,
0.05607982724905014,
-0.03703194484114647,
0.0049115754663944244,
-0.043821509927511215,
0.036236707121133804,
0.04201261326670647,
0.019074833020567894,
-0.06141967698931694,
-0.04041968286037445,
-0.026736440137028694,
-0.029206611216068268,
-0.028404388576745987,
-0.016792573034763336,
0.00699496129527688,
0.04810171201825142,
-0.016365202143788338,
-0.01779688335955143,
-0.04802630841732025,
0.0019085032399743795,
0.024997742846608162,
0.10941699892282486,
-0.06617993116378784,
-0.04231619834899902,
0.06651385873556137,
0.042090874165296555,
0.016108674928545952,
-0.1270962357521057,
-0.060242436826229095,
-0.03811368718743324,
-0.00045159441651776433,
0.0010807630605995655,
0.07785002887248993,
0.013248474337160587,
0.009831753559410572,
0.04152262583374977,
-0.050523411482572556,
0.041491150856018066,
-0.04989120736718178,
0.04303162917494774,
0.057964593172073364,
0.07386446744203568,
0.08079244941473007,
0.02161574736237526,
-0.06888290494680405,
0.010816329158842564,
-0.019155271351337433,
-0.041740816086530685,
0.05418742820620537,
-0.001115776365622878,
0.004311359021812677,
-0.06984474509954453,
0.09471643716096878,
0.07384291291236877,
-0.0216231532394886,
-0.022849585860967636,
-0.010183333419263363,
-0.06291254609823227,
-0.02825641818344593,
-0.020054001361131668,
0.05188552290201187,
-0.05690687149763107,
0.07241296768188477,
-0.014305329881608486,
-0.07346699386835098,
-0.1395515650510788,
-0.03729686886072159,
-0.06815245747566223,
-0.0161858182400465,
-3.91584489319011e-34,
0.037180256098508835,
-0.022086745128035545,
-0.10185719281435013,
-0.03456389158964157,
-0.07662393897771835,
-0.06997635215520859,
-0.06844919919967651,
-0.0006370582268573344,
0.009540458209812641,
-0.0746532678604126,
-0.011901858262717724,
0.014000597409904003,
0.021467359736561775,
-0.03571619838476181,
0.02792140655219555,
0.0030586058273911476,
-0.1010417640209198,
-0.03891449794173241,
0.01590852625668049,
0.014405290596187115,
0.09330148249864578,
0.024161165580153465,
-0.07798414677381516,
-0.02898995205760002,
-0.08150915801525116,
0.06941317021846771,
-0.012454519979655743,
0.051600005477666855,
0.021504653617739677,
0.0546664334833622,
0.029366767033934593,
-0.06494660675525665,
0.035978540778160095,
-0.05709543079137802,
-0.05066714435815811,
0.011687801219522953,
0.0594351701438427,
-0.023357056081295013,
-0.041650548577308655,
-0.014540284872055054,
-0.03200628235936165,
-0.014612793922424316,
0.01178386528044939,
0.0488530695438385,
-0.05796148627996445,
0.028679300099611282,
0.00038388787652365863,
0.02871014177799225,
-0.05047640576958656,
0.1019415631890297,
-0.09364102780818939,
0.06845998018980026,
-0.06340110301971436,
-0.03795439749956131,
0.06693810969591141,
-0.0676976889371872,
-0.0505978949368,
-0.052179984748363495,
0.08534536510705948,
-0.022281311452388763,
-0.06865408271551132,
-0.09493255615234375,
-0.05800807103514671,
0.0828867182135582,
0.008058139123022556,
-0.02373991347849369,
-0.04887112230062485,
-0.03830597922205925,
0.04257391765713692,
0.04745770990848541,
-0.0030825904104858637,
0.006434364710003138,
-0.0130221638828516,
0.011174126528203487,
-0.006758539006114006,
-0.053307563066482544,
0.047739844769239426,
0.023725850507616997,
0.09096173942089081,
-0.01900179497897625,
0.09449270367622375,
-0.07908595353364944,
0.0464143231511116,
0.14149954915046692,
0.11245161294937134,
0.0003000945143867284,
-0.014588405378162861,
0.014008359983563423,
0.005594716407358646,
0.013461657799780369,
0.01396971195936203,
0.011960944160819054,
-0.05272887274622917,
0.031177151948213577,
-0.00007316766277654096,
-3.698630379744827e-8,
-0.09749271720647812,
0.07143770158290863,
0.082982636988163,
0.008032177574932575,
0.004217674024403095,
-0.0950695276260376,
0.052245475351810455,
0.0691358894109726,
-0.029840867966413498,
-0.06063263490796089,
-0.010758674703538418,
-0.031815268099308014,
-0.05773661658167839,
-0.028609350323677063,
0.020587915554642677,
0.09225388616323471,
-0.004325778689235449,
-0.022587111219763756,
0.02314188703894615,
0.02819719724357128,
-0.03342913091182709,
0.043521467596292496,
0.022678786888718605,
-0.031194524839520454,
0.07320471107959747,
0.016604376956820488,
-0.047505419701337814,
0.11954659223556519,
0.10535682737827301,
-0.051221814006567,
0.015041193924844265,
-0.005673524923622608,
0.044201355427503586,
-0.04373154416680336,
0.01732822135090828,
0.03811716288328171,
0.035149089992046356,
-0.020404981449246407,
0.027012968435883522,
-0.10996447503566742,
-0.02610689215362072,
0.09176862984895706,
0.03963034600019455,
0.001065200543962419,
0.052557528018951416,
0.0016128573333844543,
0.031083738431334496,
-0.05262114852666855,
-0.027807964012026787,
-0.055858127772808075,
-0.055510032922029495,
0.052983809262514114,
0.020701859146356583,
0.10385633260011673,
-0.010466750711202621,
0.053953658789396286,
-0.047726262360811234,
-0.057456035166978836,
-0.0027274044696241617,
0.06153695657849312,
-0.04555509611964226,
0.05056372284889221,
-0.05785655975341797,
-0.04806401580572128
] | 0 | 0 | 0 | 0 |
|
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models | https://openreview.net/forum?id=I4e82CIDxv | [
"Samuel Marks",
"Can Rager",
"Eric J Michaud",
"Yonatan Belinkov",
"David Bau",
"Aaron Mueller"
] | Oral | We introduce methods for discovering and applying **sparse feature circuits**. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms in neural networks. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors. | Interpretability, mechanistic interpretability, circuits, spurious correlations, generalization, dictionary learning | We automatically discover circuits of interpretable components and apply them to remove sensitivity to spurious correlates | 2,718 | 2403.19647 | [
-0.021420760080218315,
-0.05556326359510422,
0.0629282295703888,
0.08019492030143738,
0.06701295077800751,
-0.010708976536989212,
-0.05544448643922806,
-0.04766501858830452,
-0.0272881668061018,
-0.04137486219406128,
-0.04912832751870155,
-0.06339650601148605,
0.012550192885100842,
0.027259856462478638,
-0.0396689809858799,
0.016127459704875946,
0.06234431639313698,
0.047478873282670975,
-0.04968618229031563,
-0.06602689623832703,
0.014375808648765087,
0.015538965351879597,
-0.03261091560125351,
0.044542521238327026,
0.026883471757173538,
0.02403167635202408,
-0.026356130838394165,
-0.02755557931959629,
0.010388965718448162,
0.0041635120287537575,
-0.007541139610111713,
0.013833406381309032,
-0.03603225573897362,
-0.022851040586829185,
0.0068246363662183285,
0.11420448869466782,
0.06876268982887268,
0.04481053724884987,
-0.013187598437070847,
-0.06927977502346039,
0.059978410601615906,
-0.045210812240839005,
0.030169110745191574,
0.12115549296140671,
0.09045910090208054,
0.012813814915716648,
-0.006458953954279423,
-0.028590001165866852,
-0.013367105275392532,
-0.05900561437010765,
-0.026285171508789062,
-0.009835291653871536,
0.07594433426856995,
0.04846138879656792,
0.015842704102396965,
-0.005659447982907295,
0.035397667437791824,
0.06890425086021423,
-0.02358882687985897,
-0.019972847774624825,
-0.07377881556749344,
-0.11180609464645386,
-0.0034633807372301817,
0.018869781866669655,
-0.021637970581650734,
0.04030822217464447,
0.005492967087775469,
0.06388060748577118,
-0.03693297505378723,
0.028360407799482346,
0.06312142312526703,
0.06909438222646713,
-0.015908893197774887,
0.02980783022940159,
0.09699296951293945,
0.021694188937544823,
0.02629128098487854,
0.0033090943470597267,
0.04336622729897499,
-0.03602801263332367,
-0.07003893703222275,
-0.045112527906894684,
-0.014272520318627357,
-0.0171100664883852,
0.11319033801555634,
0.028653249144554138,
0.03205089643597603,
0.01504589058458805,
-0.02675304375588894,
0.08116692304611206,
-0.03190258890390396,
-0.11506309360265732,
0.08205950260162354,
-0.07699552178382874,
-0.05487266927957535,
0.05889343097805977,
0.002890300238505006,
-0.05919244512915611,
-0.03944732993841171,
0.10154982656240463,
-0.0396098718047142,
0.022327737882733345,
0.032507531344890594,
-0.013077696785330772,
0.008130217902362347,
-0.014338096603751183,
0.012732074595987797,
0.07278311252593994,
0.023773403838276863,
-0.08733520656824112,
-0.036676131188869476,
0.018145723268389702,
0.0024432383943349123,
0.05183441564440727,
0.06909366697072983,
-0.058258961886167526,
0.0409499816596508,
-0.06490328907966614,
0.01588488183915615,
0.021312283352017403,
-0.035450927913188934,
0.037506550550460815,
-0.04064864665269852,
0.01852318085730076,
0.017763346433639526,
-0.0009298524237237871,
-0.06556688249111176,
5.709350348004926e-34,
0.13100074231624603,
0.05166669189929962,
0.0220296923071146,
-0.006699774879962206,
0.05077727138996124,
0.004627353046089411,
-0.02576633170247078,
0.016916345804929733,
0.010232701897621155,
0.033034712076187134,
-0.05118190124630928,
0.033874183893203735,
-0.02469746023416519,
0.09846189618110657,
0.01709005795419216,
-0.0023508553858846426,
-0.05481097102165222,
0.019178267568349838,
0.024799134582281113,
-0.09936363250017166,
-0.06212865188717842,
-0.04443121701478958,
-0.05631030723452568,
-0.01438069622963667,
0.006484052632004023,
-0.06367693096399307,
0.047767091542482376,
-0.0676928237080574,
-0.07356694340705872,
0.018082614988088608,
-0.10698631405830383,
0.012611819431185722,
0.07312838733196259,
0.014415622688829899,
-0.0348353385925293,
-0.03195023536682129,
-0.026364101096987724,
-0.06715430319309235,
0.06759580969810486,
0.0003710785531438887,
0.013120286166667938,
-0.043614715337753296,
0.004492302890866995,
-0.05623616650700569,
-0.1195727288722992,
0.008199190720915794,
-0.004428833723068237,
0.03268925100564957,
-0.05071108415722847,
-0.02652590349316597,
0.05298382043838501,
0.04622889682650566,
0.019747352227568626,
-0.06914529949426651,
0.011889070272445679,
-0.009299768134951591,
-0.03215451166033745,
-0.01215591561049223,
0.004283658694475889,
0.05815539136528969,
0.04550044983625412,
0.06648615747690201,
-0.01837203837931156,
0.06368450075387955,
0.008386433124542236,
0.052900105714797974,
-0.04789469763636589,
0.05420290306210518,
0.039158374071121216,
0.004265666473656893,
-0.11397553980350494,
0.02830740436911583,
0.0010887583484873176,
-0.055179413408041,
0.06311054527759552,
0.022944563999772072,
0.05444470793008804,
-0.15731990337371826,
0.016746478155255318,
0.014347238466143608,
-0.0573466420173645,
-0.153194397687912,
0.006762382574379444,
0.042050112038850784,
-0.025521164759993553,
-0.010550891980528831,
0.037229884415864944,
-0.047424815595149994,
-0.06441455334424973,
-0.05244216322898865,
0.010892210528254509,
-0.02408195100724697,
0.0346449576318264,
0.03539329394698143,
0.02775099314749241,
-1.8201371058521608e-33,
-0.047467537224292755,
0.07361331582069397,
0.03753794729709625,
-0.0382721945643425,
-0.044303491711616516,
-0.021479537710547447,
-0.033301543444395065,
0.004454315640032291,
0.023333024233579636,
-0.025401998311281204,
-0.00009754560596775264,
-0.04648597165942192,
0.03764588013291359,
-0.023047277703881264,
-0.00637127086520195,
0.0498134009540081,
-0.039031896740198135,
0.004339155741035938,
0.06502331793308258,
0.04253283143043518,
0.004748734179884195,
0.06613245606422424,
-0.08062943816184998,
-0.0028851134702563286,
-0.0072774565778672695,
-0.006894703488796949,
-0.03739096224308014,
0.034004438668489456,
-0.013410354033112526,
-0.028652112931013107,
-0.10043764114379883,
-0.001110857934691012,
-0.06535005569458008,
-0.10708259046077728,
-0.04262888431549072,
0.06935670971870422,
0.019852256402373314,
0.009255608543753624,
-0.022706886753439903,
0.009961911477148533,
0.05665408819913864,
-0.01729988120496273,
-0.11603827029466629,
0.051777228713035583,
0.008255424909293652,
0.04291613772511482,
-0.03931953012943268,
0.08801227062940598,
-0.028627458959817886,
0.04048316180706024,
-0.06542528420686722,
0.03872567042708397,
0.019172631204128265,
-0.007567507214844227,
-0.08287585526704788,
-0.051831744611263275,
0.06535014510154724,
-0.014897190034389496,
0.12838414311408997,
-0.036512065678834915,
-0.08710242807865143,
-0.028505120426416397,
-0.034483738243579865,
0.003538914956152439,
0.006295075174421072,
-0.06893674284219742,
0.052302610129117966,
0.008839638903737068,
0.0166921466588974,
-0.027077583596110344,
0.08996016532182693,
0.01309154648333788,
-0.03134956210851669,
-0.06802556663751602,
0.002117933239787817,
-0.03530564159154892,
-0.031547363847494125,
-0.08324480801820755,
-0.0510227233171463,
-0.031256526708602905,
0.0942397490143776,
0.08997009694576263,
0.05209707096219063,
-0.04344788193702698,
0.034958820790052414,
0.09773606806993484,
0.06023908033967018,
0.10991886258125305,
0.059598807245492935,
0.03399853780865669,
-0.10430750250816345,
0.05985294282436371,
-0.04445217177271843,
0.10434676706790924,
-0.06891089677810669,
-3.5647104823510745e-8,
-0.08869876712560654,
0.028580687940120697,
-0.07180823385715485,
0.009864355437457561,
0.06734249740839005,
-0.06883136928081512,
-0.011093711480498314,
0.08361990004777908,
-0.023252345621585846,
-0.04600668326020241,
0.017665011808276176,
0.03442854806780815,
-0.04096103832125664,
-0.05322665348649025,
0.01430878322571516,
0.0765547975897789,
-0.011097901500761509,
0.012495637871325016,
-0.057570163160562515,
-0.02706754393875599,
0.07629472762346268,
-0.028548793867230415,
0.0006605609087273479,
0.06743471324443817,
0.052744921296834946,
-0.10750846564769745,
-0.030525581911206245,
0.08689119666814804,
0.06655334681272507,
-0.002502978779375553,
-0.02122899703681469,
0.03514597564935684,
0.005443184170871973,
-0.03559945151209831,
0.02979767695069313,
0.1387336403131485,
0.029041165485978127,
-0.04332133010029793,
-0.024501031264662743,
-0.03651738911867142,
-0.045220136642456055,
0.017067570239305496,
-0.006628277711570263,
0.020770970731973648,
-0.00814321730285883,
-0.049377020448446274,
-0.018499072641134262,
-0.07505642622709274,
0.05615341290831566,
-0.029579883441329002,
-0.020390339195728302,
0.02432614006102085,
0.011087125167250633,
0.0383584126830101,
0.008022827096283436,
0.08600030839443207,
0.011633075773715973,
-0.02081773802638054,
-0.03353523090481758,
0.0024840044789016247,
0.044476523995399475,
0.07233230769634247,
0.029440591111779213,
-0.05493246391415596
] | https://github.com/saprmarks/feature-circuits | 162 | 0 | 0 | 0 |
Retrieval Head Mechanistically Explains Long-Context Factuality | https://openreview.net/forum?id=EytBpUGB1Z | [
"Wenhao Wu",
"Yizhong Wang",
"Guangxuan Xiao",
"Hao Peng",
"Yao Fu"
] | Oral | Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache. | Large language models, long context, interpretability, attention | We study retrieval head, a special type of attention head that mechanistically explains long-context factuality | 2,659 | 2404.15574 | [
-0.00524255633354187,
-0.005770675837993622,
0.037024132907390594,
0.025465454906225204,
0.08860775828361511,
0.021663380786776543,
0.03273030370473862,
0.014855090528726578,
0.10622381418943405,
-0.04930877685546875,
-0.021564964205026627,
-0.0215847659856081,
0.06196686998009682,
0.02692767232656479,
0.04328811168670654,
-0.03709504008293152,
0.09601103514432907,
0.0624842643737793,
-0.15211018919944763,
-0.03606840595602989,
0.05996648594737053,
0.034910403192043304,
0.06202424317598343,
-0.020867440849542618,
-0.02813229151070118,
-0.028009112924337387,
-0.06778222322463989,
-0.07003346085548401,
0.09452616423368454,
-0.04247872531414032,
0.04758153483271599,
0.09705428779125214,
-0.014642071910202503,
0.03579443320631981,
-0.07300477474927902,
0.08267137408256531,
-0.06518083810806274,
0.02957312762737274,
0.045510344207286835,
-0.07439208775758743,
-0.006178108509629965,
0.0036940418649464846,
-0.007193969562649727,
0.0615837387740612,
0.06209035962820053,
-0.01697695441544056,
0.026942970231175423,
0.04213741794228554,
-0.01913134753704071,
-0.031569112092256546,
-0.12901084125041962,
-0.055398497730493546,
0.03507387638092041,
0.010018742643296719,
-0.024844011291861534,
0.03024541586637497,
0.018715929239988327,
0.03914725407958031,
-0.06556259095668793,
-0.015930917114019394,
-0.06414807587862015,
-0.0772324874997139,
-0.08804270625114441,
-0.04464733973145485,
-0.00014102492423262447,
-0.04776906222105026,
0.0027304873801767826,
-0.0003917106078006327,
-0.012823308818042278,
0.021341541782021523,
0.04669242724776268,
0.04157056659460068,
-0.03396587818861008,
0.035991422832012177,
0.0947098657488823,
0.014766550622880459,
0.010596809908747673,
-0.03882165625691414,
0.022127632051706314,
-0.03815019503235817,
0.039102572947740555,
0.012254077941179276,
0.0845421701669693,
-0.024077771231532097,
0.03883162885904312,
-0.00907664280384779,
-0.00046033330727368593,
-0.018004443496465683,
0.008937472477555275,
0.010794584639370441,
-0.021137528121471405,
-0.16707868874073029,
0.02596926875412464,
-0.053030405193567276,
0.08334609866142273,
0.008667983114719391,
0.009704585187137127,
0.011021474376320839,
0.012675185687839985,
0.039809588342905045,
0.05268724262714386,
0.01248498447239399,
-0.05661044269800186,
-0.021341972053050995,
-0.02372784912586212,
-0.059046123176813126,
-0.03794066235423088,
0.0687280148267746,
-0.021420108154416084,
-0.06303685903549194,
-0.019121460616588593,
0.07267650216817856,
0.0019308766350150108,
-0.020911119878292084,
-0.016421791166067123,
-0.0378418006002903,
-0.02788032405078411,
-0.03628139942884445,
0.0006422838196158409,
-0.0038589576724916697,
0.027231544256210327,
0.02842140570282936,
0.022569386288523674,
-0.044575296342372894,
-0.0018197264289483428,
-0.05929999426007271,
-0.03225032240152359,
3.159466165491297e-33,
0.022258395329117775,
0.025034505873918533,
-0.013762369751930237,
0.08755459636449814,
-0.012579082511365414,
0.008970669470727444,
-0.0238952599465847,
0.030551142990589142,
0.004872403107583523,
-0.00868797767907381,
-0.009794337674975395,
0.12197042256593704,
-0.04054438695311546,
0.025812357664108276,
-0.033343054354190826,
0.010302109643816948,
-0.033687081187963486,
0.06257245689630508,
-0.028256000950932503,
0.0029574583750218153,
0.06343300640583038,
0.010041346773505211,
0.030145741999149323,
-0.07171911001205444,
0.00706604216247797,
0.005898664705455303,
0.03760611265897751,
-0.05018562823534012,
-0.10354412347078323,
0.028624722734093666,
-0.14482173323631287,
0.034146808087825775,
-0.024336503818631172,
-0.00011522771819727495,
0.003959796857088804,
-0.06405366957187653,
-0.01737474836409092,
-0.043190326541662216,
0.042581140995025635,
-0.017369268462061882,
0.009669706225395203,
0.04579964652657509,
-0.010053061880171299,
0.0071291569620370865,
-0.09389176964759827,
-0.08250804245471954,
-0.01396807748824358,
0.05516333878040314,
-0.09636873006820679,
0.02296868897974491,
0.09152457863092422,
0.05757465958595276,
-0.04190022498369217,
-0.06299483776092529,
-0.008526558056473732,
0.04642126336693764,
-0.02244585007429123,
0.021309612318873405,
0.003391732694581151,
0.07646789401769638,
0.08666018396615982,
0.0421893410384655,
-0.005259458441287279,
0.09913676232099533,
0.10824714601039886,
0.018975352868437767,
-0.020775601267814636,
0.03019452840089798,
0.06770806014537811,
0.03453019633889198,
-0.00905511062592268,
-0.022516408935189247,
-0.013695111498236656,
-0.02657872624695301,
0.05207750201225281,
-0.048866283148527145,
-0.007745280396193266,
-0.05035094916820526,
0.018061498180031776,
-0.0008487271261401474,
0.03472168371081352,
-0.055071230977773666,
0.03659960627555847,
-0.003937725909054279,
-0.04412372037768364,
-0.014165707863867283,
0.03616909310221672,
-0.1839917004108429,
-0.04576623812317848,
-0.07104435563087463,
0.043565914034843445,
-0.004920050036162138,
-0.04365291818976402,
-0.02287069894373417,
-0.04903782904148102,
-3.663603733806313e-33,
0.025831041857600212,
-0.11610520631074905,
-0.0013334857067093253,
-0.0024496742989867926,
-0.04979279264807701,
-0.0465492382645607,
0.02184385433793068,
0.022534798830747604,
-0.12862199544906616,
-0.06852608174085617,
-0.035959623754024506,
-0.05183074250817299,
0.047327276319265366,
0.017750920727849007,
0.023505229502916336,
-0.006866385228931904,
0.05219988897442818,
-0.04031643643975258,
-0.039378244429826736,
0.06060534715652466,
0.0500023327767849,
0.012224837206304073,
-0.13632291555404663,
-0.01589885912835598,
-0.03412207216024399,
-0.01004723645746708,
-0.05953209474682808,
0.0003157946339342743,
0.015593042597174644,
0.008696692995727062,
-0.06826178729534149,
0.011357391253113747,
-0.03272940590977669,
-0.07360068708658218,
-0.08205967396497726,
0.05485425144433975,
-0.025232909247279167,
0.0052276779897511005,
-0.050646912306547165,
0.13476482033729553,
0.04920608550310135,
-0.015250307507812977,
0.02629406377673149,
0.010249657556414604,
-0.030758315697312355,
-0.06718633323907852,
-0.11911080777645111,
0.022443773224949837,
0.05753263086080551,
0.01474617887288332,
0.06039466708898544,
0.0019476739689707756,
-0.017626861110329628,
-0.028469350188970566,
-0.026538722217082977,
-0.03722532466053963,
-0.06114351004362106,
-0.09394048899412155,
0.0579395554959774,
-0.02073189616203308,
-0.03460855036973953,
-0.025680234655737877,
-0.05171624571084976,
0.025161653757095337,
0.07248321920633316,
-0.03399462625384331,
0.03785502910614014,
-0.04853011295199394,
0.009448055177927017,
-0.039430730044841766,
0.0475691594183445,
-0.05673312768340111,
-0.033866237848997116,
-0.05924784019589424,
0.0028581952210515738,
0.11567956209182739,
-0.021120693534612656,
-0.03733082488179207,
-0.00998118706047535,
-0.036325689405202866,
-0.08601099997758865,
0.004022097215056419,
-0.026419680565595627,
0.021734900772571564,
0.04321494325995445,
0.06670544296503067,
0.008775569498538971,
0.07085985690355301,
0.062040023505687714,
0.002529538469389081,
0.03433713689446449,
0.0013687275350093842,
-0.01110909041017294,
0.08773595094680786,
-0.051069967448711395,
-4.891538196716283e-8,
-0.09437619149684906,
0.03798184543848038,
-0.043804433196783066,
0.07745786756277084,
0.03772023692727089,
0.005172767676413059,
-0.035739466547966,
0.12974610924720764,
-0.03552272915840149,
0.00957533624023199,
0.022311661392450333,
0.005948895122855902,
-0.016969049349427223,
-0.011207850649952888,
0.033987030386924744,
0.09380828589200974,
0.09090888500213623,
0.03218673914670944,
-0.027896864339709282,
-0.05496887117624283,
0.06429111957550049,
0.05784698948264122,
0.0016830430831760168,
-0.029087623581290245,
0.01962924748659134,
0.009698748588562012,
-0.05291219428181648,
0.13998518884181976,
0.08048785477876663,
-0.03358929604291916,
0.0271495021879673,
0.033632997423410416,
-0.06816132366657257,
-0.04345395788550377,
-0.022912312299013138,
0.041361253708601,
0.0322798490524292,
-0.06476542353630066,
0.011836331337690353,
-0.002027126494795084,
0.07407254725694656,
0.05346233397722244,
-0.03262506052851677,
0.11215025186538696,
0.030497804284095764,
0.04256945848464966,
0.028340039774775505,
-0.07641281932592392,
0.005669334903359413,
-0.00567966140806675,
0.0018202324863523245,
-0.04436493292450905,
-0.011889463290572166,
0.06880833953619003,
0.04375184699892998,
0.054128438234329224,
0.027222836390137672,
0.0017762043280526996,
0.008846021257340908,
0.022333944216370583,
0.05962682515382767,
0.06701526045799255,
-0.05105702579021454,
0.02839023247361183
] | https://github.com/nightdessert/retrieval_head | 185 | 0 | 0 | 0 |
High-Dynamic Radar Sequence Prediction for Weather Nowcasting Using Spatiotemporal Coherent Gaussian Representation | https://openreview.net/forum?id=Cjz9Xhm7sI | [
"Ziye Wang",
"Yiran Qin",
"Lin Zeng",
"Ruimao Zhang"
] | Oral | Weather nowcasting is an essential task that involves predicting future radar echo sequences based on current observations, offering significant benefits for disaster management, transportation, and urban planning. Current prediction methods are limited by training and storage efficiency, mainly focusing on 2D spatial predictions at specific altitudes. Meanwhile, 3D volumetric predictions at each timestamp remain largely unexplored. To address such a challenge, we introduce a comprehensive framework for 3D radar sequence prediction in weather nowcasting, using the newly proposed SpatioTemporal Coherent Gaussian Splatting (STC-GS) for dynamic radar representation and GauMamba for efficient and accurate forecasting. Specifically, rather than relying on a 4D Gaussian for dynamic scene reconstruction, STC-GS optimizes 3D scenes at each frame by employing a group of Gaussians while effectively capturing their movements across consecutive frames. It ensures consistent tracking of each Gaussian over time, making it particularly effective for prediction tasks. With the temporally correlated Gaussian groups established, we utilize them to train GauMamba, which integrates a memory mechanism into the Mamba framework. This allows the model to learn the temporal evolution of Gaussian groups while efficiently handling a large volume of Gaussian tokens. As a result, it achieves both efficiency and accuracy in forecasting a wide range of dynamic meteorological radar signals. The experimental results demonstrate that our STC-GS can efficiently represent 3D radar sequences with over $16\times$ higher spatial resolution compared with the existing 3D representation methods, while GauMamba outperforms state-of-the-art methods in forecasting a broad spectrum of high-dynamic weather conditions. | 3D Gaussian, Dynamic Reconstruction, Radar Prediction, Weather Nowcasting | null | 2,603 | 2502.14895 | [
-0.009182075038552284,
-0.11779333651065826,
-0.002963405568152666,
0.044294800609350204,
0.09329714626073837,
0.01346471719443798,
0.02631758712232113,
-0.09179530292749405,
-0.05215800926089287,
-0.07124395668506622,
-0.06274032592773438,
-0.01294314581900835,
-0.011723008938133717,
0.06841810792684555,
-0.049420908093452454,
-0.04054822027683258,
-0.011767193675041199,
0.01930350810289383,
-0.0020708562806248665,
-0.013042359612882137,
0.011338342912495136,
0.014092900790274143,
-0.04020978510379791,
0.03054678812623024,
0.07636372745037079,
0.0775514468550682,
0.03353011608123779,
-0.07654811441898346,
-0.012147543020546436,
0.06103270873427391,
0.08112967759370804,
0.007598509080708027,
-0.011076579801738262,
-0.0402136854827404,
-0.059315431863069534,
-0.031949885189533234,
-0.042237289249897,
0.06842073798179626,
-0.020777782425284386,
0.052720438688993454,
0.06599507480859756,
0.07485432922840118,
-0.0434873141348362,
0.004594025667756796,
-0.004977291915565729,
0.0686788409948349,
0.06992850452661514,
-0.07857351750135422,
0.014593045227229595,
0.008595097810029984,
-0.06462643295526505,
0.05254741385579109,
-0.06069648265838623,
0.023924071341753006,
-0.037121910601854324,
0.06038132682442665,
-0.08147121965885162,
0.002088614972308278,
0.031228668987751007,
-0.006087892223149538,
-0.0539577454328537,
-0.052023183554410934,
0.012739179655909538,
-0.057579658925533295,
0.004805545788258314,
-0.010229691863059998,
0.01510702446103096,
0.04697771742939949,
-0.02897193282842636,
0.08830618858337402,
-0.047872137278318405,
0.11673793196678162,
0.022859420627355576,
-0.03028157167136669,
-0.031550537794828415,
0.03643058240413666,
0.022391924634575844,
-0.04574722424149513,
-0.009551250375807285,
-0.006345394998788834,
0.07336578518152237,
0.017396967858076096,
0.005364663433283567,
-0.022003943100571632,
0.008391269482672215,
0.028450997546315193,
-0.023339731618762016,
0.13005253672599792,
0.012504257261753082,
-0.03311513364315033,
0.020497919991612434,
-0.06259508430957794,
0.012973516248166561,
-0.03626846522092819,
-0.021457331255078316,
0.032029636204242706,
0.024405786767601967,
-0.08501767367124557,
0.06858181953430176,
0.018143635243177414,
0.027710847556591034,
-0.05414784327149391,
0.020940544083714485,
0.06938035786151886,
0.044690847396850586,
-0.07726569473743439,
0.04675886407494545,
-0.020635858178138733,
-0.005163495894521475,
0.02126252092421055,
0.005936163943260908,
0.03715436905622482,
0.044498663395643234,
-0.06424927711486816,
0.002711614826694131,
0.06148281693458557,
-0.1173926442861557,
-0.009360161609947681,
-0.06355907768011093,
0.04243491217494011,
0.025202294811606407,
-0.09388924390077591,
-0.00891337264329195,
0.06874461472034454,
0.04324259236454964,
-0.009862540289759636,
-0.06787353754043579,
1.4302215889225771e-33,
-0.007821089588105679,
0.02980343997478485,
0.051517631858587265,
0.03460663929581642,
0.009670181199908257,
0.030117515474557877,
-0.01591232232749462,
0.026459166780114174,
0.006135490257292986,
0.062440238893032074,
-0.04129817336797714,
-0.01298933383077383,
-0.05636562034487724,
0.11568432301282883,
-0.0022790960501879454,
0.03287769481539726,
-0.06155504658818245,
0.012866741046309471,
-0.03285592794418335,
0.05879432335495949,
-0.0008654762641526759,
0.01962330751121044,
-0.020955346524715424,
-0.029312385246157646,
0.009945975616574287,
0.026893174275755882,
0.043975431472063065,
-0.06111476197838783,
0.023814095184206963,
0.0073455567471683025,
-0.022189361974596977,
0.04595615714788437,
-0.05623398721218109,
-0.00016629480523988605,
0.057097792625427246,
-0.04121439903974533,
-0.07790978997945786,
-0.05829593911767006,
0.00518071511760354,
0.024970250204205513,
0.0007621258846484125,
-0.005725475028157234,
-0.023739317432045937,
-0.02637363225221634,
-0.037831779569387436,
-0.04855936020612717,
0.10315145552158356,
0.0650055855512619,
-0.018999509513378143,
0.015269000083208084,
0.10163023322820663,
0.009424101561307907,
-0.1024593859910965,
0.01269320398569107,
-0.029646530747413635,
-0.013834558427333832,
0.04738464951515198,
-0.009569765068590641,
0.0885375589132309,
-0.013781406916677952,
-0.061728376895189285,
-0.044338077306747437,
0.02847224846482277,
-0.030173439532518387,
-0.029465703293681145,
-0.0060919406823813915,
0.024809736758470535,
0.11031552404165268,
0.040371328592300415,
-0.0018435543170198798,
0.06190691515803337,
-0.04048546403646469,
0.04967139661312103,
0.03311717510223389,
0.03054666332900524,
-0.013569570146501064,
0.030799048021435738,
0.04439018294215202,
0.009665600024163723,
0.03746815025806427,
-0.05800351873040199,
0.014392725192010403,
-0.08730992674827576,
0.006744595244526863,
-0.06668723374605179,
-0.04199867323040962,
0.01594867929816246,
0.030428022146224976,
-0.08920074999332428,
-0.041464224457740784,
-0.08266134560108185,
-0.00463074678555131,
0.024057844653725624,
-0.030891625210642815,
-0.09712382405996323,
-2.413592592033493e-33,
0.08668019622564316,
0.05950562283396721,
-0.0616234727203846,
-0.005476851016283035,
-0.010836495086550713,
-0.021540185436606407,
-0.014650795608758926,
0.03268321603536606,
-0.016301289200782776,
-0.057217083871364594,
-0.047954946756362915,
-0.012173549272119999,
0.01888805627822876,
0.07041580229997635,
0.018683647736907005,
-0.03282599896192551,
-0.01872089132666588,
0.028152160346508026,
-0.07233430445194244,
-0.07315293699502945,
0.02914346382021904,
-0.08357452601194382,
-0.01619798131287098,
-0.0037461586762219667,
-0.02322901226580143,
0.03431335836648941,
0.09313176572322845,
0.07827530056238174,
-0.04989732801914215,
0.015366067178547382,
-0.1169709637761116,
-0.03607529029250145,
-0.041082605719566345,
-0.024864114820957184,
-0.037598371505737305,
0.052424974739551544,
0.11483974009752274,
-0.032103899866342545,
-0.06925243139266968,
-0.07208404690027237,
0.02654634788632393,
0.012418110854923725,
0.10376136004924774,
0.03765573725104332,
0.016286049038171768,
-0.00043684730189852417,
0.019257456064224243,
0.07829540967941284,
0.04187890514731407,
0.04669470340013504,
-0.043095044791698456,
0.018477683886885643,
-0.13668298721313477,
-0.014602822251617908,
0.03106316551566124,
-0.03653920069336891,
-0.10028640180826187,
-0.12161730229854584,
-0.046637870371341705,
-0.037267643958330154,
-0.032334499061107635,
0.021447131410241127,
0.008393327705562115,
-0.01768413372337818,
0.01660647988319397,
-0.0006065333145670593,
0.023368943482637405,
-0.04520636796951294,
-0.018374433740973473,
0.1062387004494667,
-0.02716582641005516,
-0.07119535654783249,
0.022235272452235222,
0.042547453194856644,
-0.049997732043266296,
0.014214648865163326,
-0.054289329797029495,
-0.002568138763308525,
-0.022141141816973686,
0.029733236879110336,
0.02397892251610756,
-0.049102287739515305,
0.01635424979031086,
0.0577392578125,
0.03080083057284355,
0.12437466531991959,
0.032827090471982956,
-0.022143909707665443,
0.08015616983175278,
-0.008317206054925919,
-0.04057442396879196,
0.018175214529037476,
0.020406421273946762,
0.04030917212367058,
-0.023112760856747627,
-3.844379392603514e-8,
-0.06588677316904068,
0.017366953194141388,
0.007682470139116049,
-0.052569132298231125,
-0.01667756959795952,
-0.07308491319417953,
0.008703259751200676,
0.025695186108350754,
0.03497137129306793,
-0.11357161402702332,
0.12270072847604752,
-0.16178490221500397,
-0.05053995922207832,
0.019170712679624557,
0.06136221066117287,
-0.010180280543863773,
-0.003669378813356161,
-0.08059437572956085,
-0.03452790156006813,
-0.012717911042273045,
-0.012458721175789833,
0.06507620215415955,
-0.028789300471544266,
0.052400339394807816,
0.058253541588783264,
0.025573309510946274,
-0.0032100360840559006,
0.09020794928073883,
0.05053987354040146,
0.04996204376220703,
0.02867097593843937,
-0.027029206976294518,
0.01676405780017376,
0.00554353091865778,
-0.05466911941766739,
-0.05766825005412102,
0.03620878607034683,
0.03737671673297882,
0.05310074985027313,
-0.050545476377010345,
-0.010719082318246365,
0.1016245037317276,
-0.04936700686812401,
0.0054906136356294155,
0.01724930666387081,
0.05492584407329559,
-0.0027835520450025797,
-0.08850251883268356,
0.0430724173784256,
0.0022252167109400034,
-0.0005674298154190183,
0.0026144180446863174,
-0.0315263606607914,
0.10849399864673615,
0.13275514543056488,
0.026462463662028313,
-0.0017241192981600761,
-0.069327712059021,
0.06930582970380783,
-0.024325307458639145,
-0.10164626687765121,
0.03491371497511864,
-0.1208072081208229,
0.05963055416941643
] | 0 | 0 | 0 | 0 |
|
Differential Transformer | https://openreview.net/forum?id=OvoCm1gGhN | [
"Tianzhu Ye",
"Li Dong",
"Yuqing Xia",
"Yutao Sun",
"Yi Zhu",
"Gao Huang",
"Furu Wei"
] | Oral | Transformer tends to overallocate attention to irrelevant context. In this work, we introduce Diff Transformer, which amplifies attention to the relevant context while canceling noise. Specifically, the differential attention mechanism calculates attention scores as the difference between two separate softmax attention maps. The subtraction cancels noise, promoting the emergence of sparse attention patterns. Experimental results on language modeling show that Diff Transformer outperforms Transformer in various settings of scaling up model size and training tokens. More intriguingly, it offers notable advantages in practical applications, such as long-context modeling, key information retrieval, hallucination mitigation, in-context learning, and reduction of activation outliers. By being less distracted by irrelevant context, Diff Transformer can mitigate hallucination in question answering and text summarization. For in-context learning, Diff Transformer not only enhances accuracy but is also more robust to order permutation, which was considered as a chronic robustness issue. The results position Diff Transformer as a highly effective and promising architecture for large language models. | sequence modeling, language models, model architecture, Transformer | null | 2,557 | 2410.05258 | [
-0.0822243019938469,
-0.0254822988063097,
0.012173532508313656,
0.013659697957336903,
0.0712650716304779,
-0.03573989495635033,
0.007014350034296513,
0.06080203875899315,
0.06838119775056839,
-0.06824120879173279,
-0.010614021681249142,
0.014490578323602676,
0.016038764268159866,
0.044973403215408325,
-0.047393616288900375,
0.013865544460713863,
0.0354386642575264,
0.13794264197349548,
-0.14089900255203247,
-0.060046322643756866,
-0.01392502710223198,
0.025308577343821526,
0.06605696678161621,
-0.008679772727191448,
0.041338808834552765,
0.03325014188885689,
-0.03745963051915169,
-0.042435307055711746,
0.016634609550237656,
-0.035392504185438156,
0.024865811690688133,
0.04355991631746292,
-0.00032796847517602146,
0.057346537709236145,
-0.08403436094522476,
0.06860002875328064,
-0.03780251741409302,
0.048783864825963974,
0.06101321801543236,
-0.008304604329168797,
0.011576053686439991,
0.012158887460827827,
0.0175740085542202,
-0.000013793804100714624,
0.07510030269622803,
-0.07679896801710129,
0.0028211285825818777,
0.010273616760969162,
0.03830275684595108,
-0.019670754671096802,
-0.021525947377085686,
-0.0029141728300601244,
-0.020236626267433167,
0.09175243973731995,
-0.036365993320941925,
0.0133505342528224,
0.009954860433936119,
0.05224251747131348,
-0.027215994894504547,
-0.03906761109828949,
-0.08723590523004532,
-0.06647694855928421,
0.018646514043211937,
-0.05237918347120285,
-0.10583760589361191,
-0.052264269441366196,
0.045598238706588745,
0.010540890507400036,
-0.0606832429766655,
-0.026730267331004143,
-0.018099576234817505,
0.07818106561899185,
0.00538845919072628,
0.07249484211206436,
0.07529807835817337,
-0.008863856084644794,
0.0055297622457146645,
-0.02941790595650673,
0.08513893932104111,
-0.013163617812097073,
0.06640418618917465,
-0.028862399980425835,
0.09585285186767578,
-0.0509573295712471,
0.11128494888544083,
-0.0448790080845356,
-0.0036423075944185257,
-0.002319381106644869,
-0.08319190889596939,
-0.0103383120149374,
-0.080675408244133,
-0.0988970622420311,
0.0806041806936264,
-0.04956956207752228,
0.027528943493962288,
0.05468522384762764,
-0.042451128363609314,
0.010334550403058529,
-0.03419424965977669,
0.08763141930103302,
0.014743137173354626,
-0.024967383593320847,
-0.05128483474254608,
-0.044244349002838135,
-0.062257055193185806,
-0.03149773180484772,
0.03258448466658592,
0.02935980074107647,
-0.05928764119744301,
-0.06254930794239044,
0.02017459087073803,
0.0859454944729805,
0.018655406311154366,
0.001566669438034296,
0.021952442824840546,
-0.041473839432001114,
0.06952925026416779,
0.007421455346047878,
0.03208538144826889,
0.038981907069683075,
0.051228225231170654,
0.021542668342590332,
-0.046044591814279556,
0.018019629642367363,
0.017028987407684326,
-0.0033742361702024937,
0.008917494677007198,
2.6946020471082458e-33,
0.010282292030751705,
0.007126952055841684,
-0.00028720407863147557,
0.03961845114827156,
0.07020550966262817,
-0.012889178469777107,
-0.014833093620836735,
0.053803566843271255,
0.00893893837928772,
-0.029136326164007187,
0.016636062413454056,
0.1000145897269249,
-0.04207461327314377,
0.01044546440243721,
-0.020679805427789688,
-0.037617165595293045,
-0.0123745147138834,
0.01939738728106022,
-0.049325451254844666,
-0.039422981441020966,
0.02461021952331066,
0.0028322741854935884,
0.034471455961465836,
-0.026032062247395515,
-0.02141481265425682,
0.011114402674138546,
0.06490596383810043,
-0.07075922191143036,
-0.005842114798724651,
-0.013192238286137581,
-0.13935238122940063,
-0.02870965003967285,
-0.026399383321404457,
-0.0013451643753796816,
0.04045204445719719,
-0.058549392968416214,
-0.007598778232932091,
-0.049738287925720215,
0.037071749567985535,
-0.05996015667915344,
-0.020557792857289314,
0.042067088186740875,
-0.04249557480216026,
-0.017740551382303238,
-0.07357893139123917,
0.06304295361042023,
0.046100225299596786,
0.00284005468711257,
0.021384071558713913,
-0.011575371026992798,
0.04287855699658394,
0.0042482297867536545,
-0.06533966958522797,
-0.052999868988990784,
0.011353212408721447,
0.028639670461416245,
0.08342274278402328,
-0.016208479180932045,
0.018962526693940163,
0.05039161816239357,
-0.009243285283446312,
0.041967276483774185,
0.005495607852935791,
0.1256018877029419,
0.06640222668647766,
-0.04225660488009453,
-0.03751472011208534,
0.07431341707706451,
0.0010164000559598207,
-0.0067880768328905106,
-0.015051123686134815,
0.02939109317958355,
-0.001716517610475421,
0.04280070960521698,
0.0383010171353817,
0.0326995886862278,
0.055922139436006546,
-0.007292613852769136,
-0.021705923601984978,
-0.044921889901161194,
0.011042995378375053,
-0.0270728450268507,
0.03209328278899193,
-0.08905075490474701,
-0.0882476195693016,
0.03663152456283569,
0.04878571256995201,
-0.041862886399030685,
0.015213606879115105,
-0.06248345226049423,
-0.020806852728128433,
0.0033754766918718815,
-0.017314767464995384,
-0.021450629457831383,
0.005783575586974621,
-1.4433009839563905e-33,
-0.05432986840605736,
-0.011982529424130917,
-0.057419609278440475,
0.028930088505148888,
-0.022343236953020096,
0.032574813812971115,
0.023779505863785744,
0.0691155418753624,
0.02578132227063179,
-0.014189734123647213,
-0.027257857844233513,
-0.015895476564764977,
0.0789017379283905,
-0.06547056138515472,
0.07105450332164764,
-0.016815470531582832,
0.034792426973581314,
0.026487179100513458,
-0.10529105365276337,
0.047070350497961044,
0.010125875473022461,
0.12368699163198471,
-0.08519352227449417,
0.004821901675313711,
-0.09962647408246994,
0.02133876271545887,
-0.12201083451509476,
0.06133168563246727,
0.04031972587108612,
-0.03794541954994202,
-0.03520864248275757,
0.04453844577074051,
0.027352796867489815,
-0.0636642798781395,
-0.10064671188592911,
0.006477373652160168,
-0.05036638304591179,
-0.056648608297109604,
-0.05130781978368759,
0.08669086545705795,
0.035641107708215714,
-0.039156265556812286,
-0.03623972460627556,
0.044105689972639084,
-0.042590707540512085,
0.0025614004116505384,
-0.06605491787195206,
-0.09222213178873062,
-0.035343918949365616,
0.07506353408098221,
-0.02619645558297634,
0.024330712854862213,
-0.06555826216936111,
0.026462813839316368,
-0.05606662854552269,
-0.03307636082172394,
-0.04145581275224686,
-0.025863071903586388,
0.03148999437689781,
0.0015635498566552997,
-0.020300213247537613,
0.008957663550972939,
0.014430144801735878,
-0.08718614280223846,
-0.006861830595880747,
-0.012968962080776691,
0.009505270048975945,
-0.03994714468717575,
0.0545223169028759,
0.05032097175717354,
0.10467837005853653,
0.004810665268450975,
0.031483422964811325,
-0.01010450441390276,
0.021947486326098442,
0.03193065896630287,
0.0350562259554863,
-0.08910152316093445,
-0.03213438764214516,
-0.01590873673558235,
-0.047946639358997345,
-0.027851292863488197,
-0.034703806042671204,
0.0172121599316597,
0.04030527547001839,
0.067599818110466,
0.08783692866563797,
0.0597757026553154,
0.05245077610015869,
0.033504560589790344,
0.01669178530573845,
-0.014057941734790802,
-0.014929882250726223,
0.059013813734054565,
-0.048812396824359894,
-4.187253921372758e-8,
-0.13249778747558594,
0.03074846789240837,
-0.12841090559959412,
0.012332144193351269,
-0.02573506534099579,
-0.11619050055742264,
-0.013025285676121712,
0.11975234001874924,
-0.051673851907253265,
-0.019555574283003807,
0.03634423017501831,
-0.04847610741853714,
-0.02555953338742256,
-0.05194292590022087,
0.02077256143093109,
0.15132629871368408,
0.08614552766084671,
0.05382727086544037,
-0.0056477561593055725,
-0.03332855924963951,
0.04525166377425194,
0.09857483953237534,
-0.014000226743519306,
0.0025593533646315336,
0.00570193026214838,
0.011850343085825443,
-0.024488598108291626,
0.0829354003071785,
0.03304681181907654,
-0.07667409628629684,
0.08152412623167038,
0.04977636784315109,
-0.028516793623566628,
-0.020182576030492783,
-0.08134142309427261,
0.10225537419319153,
0.06943347305059433,
-0.006496047601103783,
0.0020299802999943495,
0.05117085948586464,
0.053867973387241364,
0.002664936473593116,
-0.0467924103140831,
0.10513444989919662,
-0.09972453117370605,
0.031295597553253174,
-0.023059222847223282,
-0.1285599172115326,
0.021384725347161293,
-0.05525016039609909,
0.10429193079471588,
-0.03861646726727486,
0.011414142325520515,
-0.0280691497027874,
0.02881464920938015,
0.03110387548804283,
0.033694785088300705,
-0.009815499186515808,
-0.01723571866750717,
0.04046138748526573,
0.010103752836585045,
0.010052081197500229,
-0.031099749729037285,
0.06256379187107086
] | https://github.com/microsoft/unilm/blob/master/Diff-Transformer/ | 21,148 | 0 | 0 | 0 |
Open-Vocabulary Customization from CLIP via Data-Free Knowledge Distillation | https://openreview.net/forum?id=1aF2D2CPHi | [
"Yongxian Wei",
"Zixuan Hu",
"Li Shen",
"Zhenyi Wang",
"Chun Yuan",
"Dacheng Tao"
] | Oral | Vision-language models such as CLIP have demonstrated strong zero-shot performance, but their considerable size and inefficient inference limit customizable deployment for users. While knowledge distillation is a solution, it still requires the original data, which is not always available due to copyrights and privacy concerns. For many users seeking open-vocabulary customization, Data-Free Knowledge Distillation (DFKD) emerges as a promising direction. Upon rethinking DFKD, we find that existing methods fail on CLIP due to their heavy reliance on BatchNorm layers, which are unexpectedly unusable in CLIP. Based on our findings, we adopt image-text matching to achieve DFKD for CLIP, enabling customization based on arbitrary class texts. This involves (i) inversing a surrogate dataset from CLIP based on text prompts; and (ii) distilling a student model from CLIP using the surrogate dataset. Specifically, we introduce style dictionary diversification to enhance the diversity of synthetic images. To prevent uncontrollable semantics introduced by diversification, we propose a class consistency maintaining strategy to ensure the consistency of synthetic images. Based on synthetic images with various styles, we further propose meta knowledge distillation to train the student model with good generalization ability. Moreover, we introduce a simple yet effective method to enable customization based on few example images. Comprehensive experiments showcase the superiority of our approach across twelve customized tasks, achieving a 9.33\% improvement compared to existing DFKD methods. | Data-Free Learning, CLIP Model, Customization | Could we distill models from CLIP without data to meet customized tasks? | 2,525 | null | [
-0.041182827204465866,
-0.080133356153965,
0.029049335047602654,
0.018209945410490036,
0.08235961198806763,
-0.05633848160505295,
-0.0419275127351284,
-0.010622123256325722,
0.004818597808480263,
-0.018016476184129715,
0.025368299335241318,
0.0018709172727540135,
0.03366437926888466,
0.010222304612398148,
0.03772156313061714,
0.01833581179380417,
0.08034154772758484,
0.1449194997549057,
-0.14066624641418457,
0.0012151586124673486,
0.020910609513521194,
-0.05706080421805382,
0.088437519967556,
-0.005755510181188583,
0.05260098725557327,
0.039116837084293365,
0.01802431046962738,
-0.01618756726384163,
0.026706773787736893,
-0.06619594991207123,
0.015604013577103615,
0.011867023073136806,
-0.010373164899647236,
0.05684341490268707,
0.03459020331501961,
0.03797033429145813,
-0.023732248693704605,
0.09893112629652023,
-0.015974724665284157,
-0.01407942920923233,
-0.028263330459594727,
-0.004343253560364246,
-0.016572939231991768,
0.06879478693008423,
0.02712961658835411,
0.048446595668792725,
-0.04650723561644554,
-0.004975487012416124,
0.025983307510614395,
0.05578818544745445,
-0.0670294314622879,
-0.05750099942088127,
-0.08429726958274841,
0.022834938019514084,
0.01632716879248619,
0.013973441906273365,
0.047874223440885544,
-0.006429322063922882,
-0.01289284136146307,
-0.00087971385801211,
-0.041098110377788544,
-0.08428782969713211,
-0.03504921868443489,
0.03799368068575859,
-0.036763597279787064,
-0.04353496804833412,
0.023671548813581467,
0.05269549414515495,
0.06922398507595062,
-0.06066398322582245,
-0.0013133768225088716,
0.07733935862779617,
0.051251381635665894,
0.06575623154640198,
-0.014134606346487999,
0.05809725821018219,
0.029194267466664314,
-0.07121271640062332,
0.003275159979239106,
-0.1382446140050888,
0.06994417309761047,
-0.026922740042209625,
0.0831475555896759,
-0.021655967459082603,
0.09955765306949615,
0.019213788211345673,
-0.03205094486474991,
-0.0022999944631010294,
0.02672589384019375,
-0.005903993733227253,
-0.09162579476833344,
-0.04557808116078377,
0.05615838244557381,
-0.030997799709439278,
0.03662401810288429,
0.02321453019976616,
-0.009026125073432922,
-0.07174833118915558,
0.000449660059530288,
0.07893315702676773,
-0.031963448971509933,
0.04248635470867157,
-0.0027775720227509737,
-0.04303464666008949,
-0.023334680125117302,
0.01635611429810524,
0.05112249776721001,
-0.017563477158546448,
0.01369564887136221,
-0.04585131257772446,
0.1159372627735138,
0.04437513276934624,
0.0008535768720321357,
-0.06519919633865356,
0.04133973643183708,
-0.03248998522758484,
-0.00529522867873311,
-0.06582332402467728,
-0.003472615499049425,
0.036965660750865936,
-0.055951934307813644,
0.03128991648554802,
-0.01859305053949356,
-0.048799771815538406,
-0.048022061586380005,
0.009202724322676659,
-0.005970613565295935,
2.956701838840099e-33,
0.073443703353405,
0.05716536566615105,
-0.0171962920576334,
0.053517360240221024,
0.07850244641304016,
-0.07863277196884155,
-0.002904508961364627,
0.03216938674449921,
-0.17173421382904053,
-0.06620505452156067,
0.07008214294910431,
-0.012906764633953571,
-0.07793933898210526,
0.08380176872015,
0.029306435957551003,
0.007443365640938282,
-0.03331927955150604,
0.07417955994606018,
-0.024663642048835754,
0.044262103736400604,
0.05483956262469292,
-0.03047201968729496,
-0.018832482397556305,
0.01057469379156828,
-0.01975775510072708,
-0.064651258289814,
0.02048536017537117,
-0.012668740935623646,
0.02882097288966179,
0.013395119458436966,
-0.06361255049705505,
-0.009058759547770023,
0.09335707873106003,
0.04221794009208679,
0.026513585820794106,
-0.07449439913034439,
0.00342893460765481,
-0.06849173456430435,
0.07001623511314392,
-0.014049766585230827,
-0.042953427881002426,
0.03592878580093384,
-0.008465565741062164,
0.010305752977728844,
-0.09263916313648224,
-0.01420989353209734,
0.052646566182374954,
0.0817168727517128,
-0.05463610589504242,
-0.01637747511267662,
0.10306841880083084,
0.016431141644716263,
-0.08769267797470093,
-0.047672804445028305,
-0.05735001713037491,
-0.0715397298336029,
0.032402586191892624,
-0.04129785671830177,
0.07289082556962967,
0.031037740409374237,
-0.03181120753288269,
0.12021506577730179,
0.00207338179461658,
0.07191692292690277,
0.009935802780091763,
-0.005410470068454742,
0.02084159664809704,
-0.007521266583353281,
-0.00018718636420089751,
0.008939889259636402,
-0.07396066933870316,
0.026124216616153717,
-0.09764906764030457,
-0.0609312541782856,
0.05287492647767067,
-0.005049678031355143,
0.018896767869591713,
-0.039770644158124924,
0.03651750832796097,
0.013378611765801907,
0.0398198701441288,
-0.043973129242658615,
-0.015279862098395824,
-0.04882020130753517,
-0.10553387552499771,
-0.009343755431473255,
0.040823645889759064,
-0.11881749331951141,
0.030222157016396523,
-0.0018273286987096071,
-0.06130984425544739,
-0.03671090304851532,
-0.08544474095106125,
-0.00820138305425644,
0.050631120800971985,
-1.674684373617288e-33,
0.016906840726733208,
0.04175413027405739,
-0.028793631121516228,
0.08216820657253265,
0.012648951262235641,
0.01369915995746851,
0.005394496954977512,
0.05411937087774277,
-0.01965545490384102,
-0.17037269473075867,
0.00101191783323884,
0.035842735320329666,
-0.037056662142276764,
-0.031298477202653885,
-0.06484140455722809,
-0.043108053505420685,
-0.012607475742697716,
0.03153010085225105,
-0.07420207560062408,
0.04215128347277641,
-0.0014106073649600148,
0.0280871894210577,
-0.08248890936374664,
0.07583890855312347,
-0.03335980698466301,
0.07128513604402542,
-0.10655421763658524,
0.021169526502490044,
0.0414721705019474,
0.02798735350370407,
0.036473292857408524,
0.011838189326226711,
-0.020983532071113586,
-0.06626860797405243,
-0.09463119506835938,
0.08508072048425674,
0.01707516796886921,
0.011369915679097176,
-0.024147536605596542,
0.052231159061193466,
0.01588868908584118,
-0.05560214817523956,
-0.08222869038581848,
-0.05698549747467041,
-0.08584965765476227,
-0.09859880805015564,
-0.053268127143383026,
-0.023512259125709534,
0.05479058623313904,
0.04532922804355621,
-0.029763517901301384,
-0.0022292938083410263,
-0.04105387628078461,
0.032782021909952164,
0.0053989882580935955,
-0.09423048049211502,
-0.057036858052015305,
-0.0143672414124012,
-0.0015824318397790194,
0.030185718089342117,
0.023665795102715492,
-0.08088728040456772,
-0.047082848846912384,
-0.03119843453168869,
0.04497675597667694,
-0.05617932975292206,
-0.05227947235107422,
0.052241455763578415,
-0.03618626296520233,
0.0168046485632658,
0.008079130202531815,
-0.05254935845732689,
0.013838618993759155,
-0.0022142825182527304,
0.014030802994966507,
-0.00031185720581561327,
0.06465182453393936,
0.01739032194018364,
0.0021273032762110233,
0.07837830483913422,
0.027377454563975334,
-0.02725822851061821,
0.0560205914080143,
0.06666045635938644,
0.09772726893424988,
0.1289772391319275,
0.017485879361629486,
0.04210508614778519,
0.03675973787903786,
0.04594733193516731,
-0.05148816853761673,
-0.007920239120721817,
0.0060355388559401035,
0.08478029072284698,
0.013815134763717651,
-4.159136324233259e-8,
-0.05680041387677193,
0.040605202317237854,
0.02343973144888878,
0.10360924899578094,
-0.05459299311041832,
-0.031576987355947495,
-0.0059112077578902245,
0.03411261737346649,
0.02773585170507431,
-0.007675073575228453,
0.10416775941848755,
0.023994147777557373,
-0.011139399372041225,
-0.0819745808839798,
-0.04191216453909874,
0.10437769442796707,
-0.014113119803369045,
0.07283446937799454,
0.01520139817148447,
0.014131434261798859,
-0.016534490510821342,
0.01411769911646843,
0.03971005976200104,
-0.017262160778045654,
0.02158856764435768,
-0.024279508739709854,
-0.019948815926909447,
0.0656508356332779,
0.09458062797784805,
0.03091447241604328,
-0.004141872748732567,
-0.04490749537944794,
0.03045465797185898,
-0.05251788720488548,
-0.00008319559856317937,
0.04983742907643318,
-0.020905625075101852,
0.010599710047245026,
0.012820766307413578,
-0.009421263821423054,
0.05260052904486656,
-0.01809900999069214,
-0.015594358555972576,
0.05371378734707832,
-0.038862138986587524,
-0.01869802549481392,
0.04282187670469284,
-0.06670364737510681,
0.015643909573554993,
0.002088383538648486,
-0.02592441812157631,
-0.01840786635875702,
0.004452962893992662,
0.10511980950832367,
0.010226539336144924,
0.025474729016423225,
-0.028247110545635223,
0.08463679254055023,
0.05092553421854973,
0.07054214924573898,
-0.00727918790653348,
-0.068806953728199,
-0.028354870155453682,
0.0196070596575737
] | 0 | 0 | 0 | 0 |
|
Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreement | https://openreview.net/forum?id=UHPnqSTBPO | [
"Jaehun Jung",
"Faeze Brahman",
"Yejin Choi"
] | Oral | We present a principled approach to provide LLM-based evaluation with a rigorous guarantee of human agreement. We first propose that a reliable evaluation method should not uncritically rely on model preferences for pairwise evaluation, but rather assess the confidence of judge models and selectively decide when to trust its judgement. We then show that under this *selective evaluation* framework, human agreement can be provably guaranteed---such that the model evaluation aligns with that of humans to a user-specified agreement level. As part of our framework, we also introduce *Simulated Annotators*, a novel confidence estimation method that significantly improves judge calibration and thus enables high coverage of evaluated instances. Finally, we propose *Cascaded Selective Evaluation*, where we use cheaper models as initial judges and escalate to stronger models only when necessary---again, while still providing a provable guarantee of human agreement. Experimental results show that Cascaded Selective Evaluation guarantees strong alignment with humans, far beyond what LLM judges could achieve without selective evaluation. For example, on a subset of Chatbot Arena where GPT-4 almost never achieves 80% human agreement, our method, even while employing substantially cost-effective models such as Mistral-7B, *guarantees* over 80% human agreement with almost 80% test coverage. | Large Language Model, LLM, LLM Judge, Evaluation, Alignment | We propose Cascaded Selective Evaluation, an LLM-as-Judge framework that dynamically selects when to trust different judge models to reduce evaluation overhead, while providing a provable guarantee of human-judge agreement. | 2,430 | 2407.18370 | [
-0.08380675315856934,
-0.01806502416729927,
-0.027878424152731895,
0.0018596822628751397,
0.06811787188053131,
-0.03261634334921837,
0.0470079630613327,
0.004214040003716946,
0.062272198498249054,
-0.023430464789271355,
-0.06729546189308167,
-0.11185707896947861,
0.06603679805994034,
0.03396078571677208,
0.040418609976768494,
-0.039269816130399704,
0.11102280020713806,
0.06375791877508163,
-0.06391461193561554,
-0.015818003565073013,
-0.04137801378965378,
0.060528527945280075,
0.049139756709337234,
-0.010997647419571877,
-0.05017901584506035,
-0.1257076859474182,
-0.024383047595620155,
-0.008347230963408947,
-0.019846417009830475,
-0.03344390168786049,
-0.009246764704585075,
-0.02271786704659462,
0.015055865980684757,
0.02741851471364498,
-0.06802990287542343,
0.005400190595537424,
-0.061237119138240814,
-0.007225967478007078,
-0.04209739342331886,
-0.0495675727725029,
-0.08708138763904572,
-0.030212171375751495,
0.008559897541999817,
0.04163617268204689,
0.017271967604756355,
0.024028168991208076,
-0.016243379563093185,
0.021451180800795555,
-0.05700337514281273,
0.06202473118901253,
-0.08621688187122345,
0.003953063394874334,
0.07308386266231537,
0.011799594387412071,
-0.07495778053998947,
0.07175621390342712,
0.03479921445250511,
-0.05176275223493576,
0.015538507141172886,
-0.022347738966345787,
-0.04689953476190567,
-0.08856628090143204,
-0.054004084318876266,
0.025937018916010857,
0.021196093410253525,
-0.0035896776244044304,
0.04746291786432266,
0.04328981041908264,
-0.035103268921375275,
0.03764164820313454,
0.030121568590402603,
0.04952061548829079,
0.026834625750780106,
0.02724357508122921,
-0.03304227441549301,
0.07474945485591888,
-0.009658625349402428,
-0.0064702508971095085,
0.10756263136863708,
-0.012686328031122684,
-0.03403599560260773,
0.00017391775327268988,
0.037894949316978455,
-0.04517212510108948,
0.05193759873509407,
-0.02707407996058464,
0.01806737668812275,
0.03935428708791733,
-0.033134542405605316,
0.046721071004867554,
-0.06897301971912384,
0.009814832359552383,
0.06465422362089157,
0.014963239431381226,
0.001461386913433671,
0.0367434024810791,
-0.005256011616438627,
-0.07013825327157974,
0.01832948438823223,
0.12376118451356888,
-0.040556617081165314,
0.07602714747190475,
-0.03603459149599075,
-0.13716217875480652,
0.06863735616207123,
0.0059248339384794235,
-0.001738852821290493,
-0.00438012694939971,
0.05338431894779205,
-0.029130782932043076,
-0.01376614160835743,
-0.0010136981727555394,
0.06098730117082596,
0.0804397314786911,
0.10630075633525848,
-0.008689258247613907,
0.024881595745682716,
0.0418185256421566,
0.01616867072880268,
-0.027831871062517166,
0.04716815426945686,
-0.00016722831060178578,
0.08207405358552933,
0.00134233885910362,
0.07472038269042969,
0.04865929111838341,
-0.008999924175441265,
2.9164554836618362e-33,
0.07501385360956192,
-0.001106251496821642,
0.05428534746170044,
0.0419083870947361,
-0.01593129336833954,
0.05499055236577988,
-0.06944870948791504,
0.02571873553097248,
-0.11288464814424515,
-0.00006646225665463135,
-0.06955885142087936,
0.017435993999242783,
-0.03800885006785393,
0.019231706857681274,
0.031649213284254074,
0.07726041972637177,
-0.053288549184799194,
0.06106993928551674,
-0.08921221643686295,
0.043267250061035156,
0.08679905533790588,
-0.007713561877608299,
0.028989238664507866,
-0.004787390120327473,
0.0001670013298280537,
0.023618750274181366,
0.04100717976689339,
-0.007587259169667959,
0.0036046633031219244,
0.0005882922559976578,
-0.10454460978507996,
-0.02428959123790264,
0.038947101682424545,
0.03977680951356888,
0.04123738780617714,
-0.0034864135086536407,
-0.05645251274108887,
-0.0115905636921525,
-0.012293349020183086,
0.031613364815711975,
-0.02334209904074669,
-0.014615638181567192,
0.01546308770775795,
-0.02433418482542038,
-0.043829549103975296,
-0.050319015979766846,
-0.07511548697948456,
0.012815750204026699,
-0.02281980775296688,
-0.04725392535328865,
0.0023706555366516113,
0.07211946696043015,
0.0032555547077208757,
-0.01496537309139967,
-0.08285550028085709,
-0.03852154687047005,
-0.026802564039826393,
0.06651883572340012,
-0.04349521920084953,
-0.02908484637737274,
-0.0296095572412014,
0.051141608506441116,
-0.030001619830727577,
-0.009432157501578331,
-0.01146596111357212,
-0.023539727553725243,
-0.04464881122112274,
0.0024488545022904873,
0.009120233356952667,
-0.04137585312128067,
0.008577796630561352,
0.015732131898403168,
0.004941074643284082,
-0.009829477407038212,
-0.021848039701581,
-0.042163264006376266,
0.07972348481416702,
-0.007322588004171848,
0.07956843823194504,
-0.058020006865262985,
-0.0023820498026907444,
0.025812063366174698,
-0.06509170681238174,
-0.030458815395832062,
-0.0032987589947879314,
-0.1411360204219818,
0.03113461658358574,
-0.0039878846146166325,
-0.04845227673649788,
0.034797750413417816,
-0.03627748414874077,
0.002819277113303542,
-0.026428554207086563,
0.09273885935544968,
0.0015009966446086764,
-2.9005525144631492e-33,
-0.02296147681772709,
-0.005986273754388094,
0.05151392146945,
0.09293247759342194,
0.009713365696370602,
-0.029943834990262985,
0.020542221143841743,
-0.007219322491437197,
0.006015006452798843,
-0.056272704154253006,
-0.05767159163951874,
-0.04244229197502136,
0.08957493305206299,
-0.060652874410152435,
0.07520382106304169,
-0.035946138203144073,
-0.025806553661823273,
-0.01896917074918747,
0.02314842864871025,
0.042826615273952484,
0.10085943341255188,
0.054613884538412094,
-0.12955623865127563,
0.08568291366100311,
-0.010967345908284187,
0.0045220074243843555,
0.006045466288924217,
-0.07585341483354568,
0.037815988063812256,
0.0005545143503695726,
-0.027956904843449593,
-0.022538546472787857,
-0.10412346571683884,
-0.018609458580613136,
0.02442619577050209,
-0.029982388019561768,
0.10485921800136566,
-0.007794341538101435,
0.023360993713140488,
0.11580672860145569,
-0.014084597118198872,
-0.015745442360639572,
-0.16121885180473328,
-0.054455384612083435,
0.007228110451251268,
-0.012413606978952885,
-0.031423479318618774,
-0.03361065685749054,
-0.06457304954528809,
-0.00821395218372345,
-0.020448438823223114,
0.006030605640262365,
-0.0692669078707695,
-0.04299100860953331,
-0.06655346602201462,
-0.12073652446269989,
-0.04152993485331535,
-0.05126871541142464,
0.04676457867026329,
0.04509088397026062,
-0.07348724454641342,
0.018068598583340645,
0.04072732478380203,
0.043482180684804916,
0.0072877793572843075,
-0.008888063952326775,
0.016904493793845177,
0.07696767896413803,
0.00834578275680542,
0.03865468129515648,
0.05033930018544197,
-0.021443940699100494,
-0.04614124447107315,
0.01221170648932457,
0.11012740433216095,
0.04781857505440712,
0.01633766107261181,
-0.05553477257490158,
0.020661139860749245,
-0.07687686383724213,
0.07533839344978333,
0.010608535259962082,
0.09747766703367233,
-0.01678687147796154,
0.0706934854388237,
0.07195151597261429,
0.10023970901966095,
0.0814652070403099,
0.020354852080345154,
0.06462959945201874,
-0.041993312537670135,
-0.0015185594093054533,
0.020507296547293663,
-0.010034359991550446,
-0.09064298123121262,
-4.3709594166330135e-8,
-0.04054960981011391,
0.014819063246250153,
0.01989314891397953,
0.07751227170228958,
-0.030344601720571518,
-0.02290765941143036,
-0.02868705615401268,
0.02170437015593052,
-0.03931540250778198,
-0.012446407228708267,
0.06998193264007568,
0.03490361571311951,
-0.019160207360982895,
-0.05203160271048546,
0.004922351334244013,
0.08135378360748291,
-0.049170058220624924,
0.07936593145132065,
-0.054234519600868225,
0.009265069849789143,
-0.005658809561282396,
0.03666107356548309,
-0.04975041002035141,
-0.042780231684446335,
0.03295750916004181,
-0.09293215721845627,
-0.03154359757900238,
0.10418996214866638,
-0.03908362239599228,
0.060717374086380005,
-0.032464973628520966,
-0.024132316932082176,
-0.04284018650650978,
0.009963623248040676,
0.07550303637981415,
0.039188530296087265,
-0.052015725523233414,
-0.00219104066491127,
0.07907696813344955,
-0.00895314197987318,
-0.013998335227370262,
0.1354907602071762,
-0.09424683451652527,
0.011083791963756084,
0.0673699900507927,
-0.019197601824998856,
-0.06174061819911003,
-0.08395542204380035,
-0.03609497845172882,
-0.044157449156045914,
-0.013120892457664013,
-0.0864172875881195,
-0.025953255593776703,
0.05702155828475952,
0.06749148666858673,
0.07624713331460953,
0.03483005240559578,
-0.030030878260731697,
0.06774674355983734,
0.023149343207478523,
0.04736902192234993,
0.0024942297022789717,
-0.04071318730711937,
-0.05652950331568718
] | 0 | 0 | 0 | 0 |
|
Your Mixture-of-Experts LLM Is Secretly an Embedding Model for Free | https://openreview.net/forum?id=eFGQ97z5Cd | [
"Ziyue Li",
"Tianyi Zhou"
] | Oral | While large language models (LLMs) excel on generation tasks, their decoder-only architecture often limits their potential as embedding models if no further representation finetuning is applied. Does this contradict their claim of generalists? To answer the question, we take a closer look at Mixture-of-Experts (MoE) LLMs. Our study shows that the expert routers in MoE LLMs can serve as an off-the-shelf embedding model with promising performance on a diverse class of embedding-focused tasks, without requiring any finetuning. Moreover, our extensive analysis shows that the MoE routing weights (RW) is complementary to the hidden state (HS) of LLMs, a widely-used embedding. Compared to HS, we find that RW is more robust to the choice of prompts and focuses on high-level semantics. Motivated by the analysis, we propose MoEE combining RW and HS, which achieves better performance than using either separately. Our exploration of their combination and prompting strategy shed several novel insights, e.g., a weighted sum of RW and HS similarities outperforms the similarity on their concatenation. Our experiments are conducted on 6 embedding tasks with 20 datasets from the Massive Text Embedding Benchmark (MTEB). The results demonstrate the significant improvement brought by MoEE to LLM-based embedding without further finetuning. | Mixture of Experts | null | 2,416 | 2410.10814 | [
-0.042870938777923584,
-0.11159065365791321,
0.06472746282815933,
0.04907292127609253,
0.04836060479283333,
0.01256014034152031,
-0.047827135771512985,
0.041663650423288345,
0.04272031411528587,
-0.08146785944700241,
-0.13947106897830963,
-0.06116533651947975,
0.08230018615722656,
0.0013554863398894668,
0.0018777856603264809,
0.03917085379362106,
0.17102661728858948,
0.050395771861076355,
-0.08585416525602341,
-0.07212034612894058,
-0.026181135326623917,
0.000192728140973486,
0.05096609517931938,
-0.028024224564433098,
0.02697211503982544,
0.015413552522659302,
-0.015056753531098366,
-0.014643522910773754,
0.0504341796040535,
0.006822514813393354,
0.11196218430995941,
0.048135966062545776,
0.001063983771018684,
0.0706746056675911,
-0.04244515672326088,
0.083489790558815,
-0.0918315127491951,
0.03032604232430458,
-0.037850961089134216,
-0.07042283564805984,
-0.045011039823293686,
0.04635293036699295,
0.006943786051124334,
0.08502756804227829,
0.04210278019309044,
-0.046845074743032455,
-0.020605655387043953,
-0.003075266722589731,
-0.06648191064596176,
0.04355411231517792,
-0.06683257222175598,
-0.05423009768128395,
0.0030582197941839695,
0.017132800072431564,
-0.05252287909388542,
0.01005231961607933,
0.033904995769262314,
0.018851706758141518,
-0.06232116371393204,
-0.035318054258823395,
-0.12994711101055145,
-0.10757656395435333,
-0.001874928129836917,
-0.01878535933792591,
0.015552018769085407,
-0.030931958928704262,
-0.021085720509290695,
0.05425232648849487,
0.029164709150791168,
0.03408929333090782,
-0.028823794797062874,
0.006040873471647501,
-0.030190547928214073,
0.04102686792612076,
-0.030680468305945396,
0.031087292358279228,
0.01454070582985878,
-0.025579730048775673,
0.04323294013738632,
-0.042690176516771317,
-0.0019850439857691526,
-0.007530858274549246,
0.004412040114402771,
0.012267077341675758,
0.15093958377838135,
-0.017981110140681267,
0.029887136071920395,
0.021592671051621437,
0.02727469615638256,
-0.0004845824441872537,
-0.06745897233486176,
-0.10878866165876389,
0.11073196679353714,
0.03957865387201309,
0.0871422216296196,
0.06958655267953873,
0.007132614962756634,
-0.0773845985531807,
-0.038640495389699936,
0.07803229242563248,
-0.04502712935209274,
0.049344465136528015,
0.061675265431404114,
-0.08587127923965454,
-0.06997320055961609,
0.0002442875993438065,
-0.026258395984768867,
0.08044656366109848,
-0.005721195600926876,
-0.06617862731218338,
-0.018388524651527405,
0.06908126920461655,
0.0577140711247921,
0.028173772618174553,
-0.009151746518909931,
-0.03646549955010414,
0.012760423123836517,
-0.06780789792537689,
0.03850456327199936,
0.05258382484316826,
-0.06407644599676132,
0.05020923167467117,
0.044603850692510605,
0.019578993320465088,
0.007779617793858051,
-0.051671821624040604,
-0.025329694151878357,
3.6338604206526475e-33,
0.036933887749910355,
0.0404045432806015,
-0.028095824643969536,
0.06789818406105042,
0.05915958434343338,
0.02378402277827263,
0.05295867100358009,
0.06812145560979843,
-0.02674952894449234,
0.0043724398128688335,
-0.042646195739507675,
0.047287214547395706,
-0.015438101254403591,
0.08865790069103241,
0.05773930996656418,
-0.010801438242197037,
-0.033909764140844345,
0.014989657327532768,
-0.07220486551523209,
-0.0502149723470211,
-0.001973524922505021,
0.044544193893671036,
0.011256900615990162,
-0.010434643365442753,
0.012693229131400585,
-0.016325553879141808,
0.09791898727416992,
-0.11260218918323517,
-0.03057992272078991,
-0.004831242375075817,
-0.1481844186782837,
-0.06652510166168213,
0.005773109383881092,
0.025407185778021812,
0.023203743621706963,
-0.008806034922599792,
-0.002750868909060955,
-0.07236920297145844,
0.024394607171416283,
-0.014934664592146873,
0.022509194910526276,
0.017565937712788582,
0.00928159523755312,
0.01222104113548994,
-0.0934862345457077,
-0.028561916202306747,
-0.04572298377752304,
0.02249971404671669,
-0.049248769879341125,
0.0003164893132634461,
0.04850836843252182,
0.022493524476885796,
-0.007994125597178936,
-0.034832559525966644,
0.07221841812133789,
-0.05413096398115158,
-0.02679205872118473,
0.10956055670976639,
0.07967766374349594,
-0.0012051467783749104,
-0.021769296377897263,
0.0708007961511612,
0.024712510406970978,
0.07663179188966751,
0.021727122366428375,
0.013599982485175133,
-0.030175035819411278,
0.13137799501419067,
0.06186455860733986,
-0.0027904019225388765,
-0.013459328562021255,
0.029557811096310616,
-0.01436423510313034,
-0.002396190771833062,
-0.01431928388774395,
0.0030568144284188747,
-0.021708352491259575,
-0.0909268707036972,
0.035872362554073334,
0.017616374418139458,
0.02377769723534584,
-0.0076723298989236355,
-0.033646780997514725,
-0.012861710973083973,
-0.004542479291558266,
-0.06477762013673782,
0.023209434002637863,
-0.06879758089780807,
0.0018387719755992293,
0.04459391161799431,
-0.0058516450226306915,
-0.04896514490246773,
-0.048357050865888596,
0.04374523460865021,
0.002349712187424302,
-2.9295539804788505e-33,
-0.04745469242334366,
-0.0002083288854919374,
-0.08967344462871552,
0.07856249064207077,
-0.031021825969219208,
-0.09594690054655075,
0.011126364581286907,
-0.025331076234579086,
-0.03667627274990082,
-0.025510258972644806,
-0.07430154830217361,
-0.017980143427848816,
0.050017114728689194,
-0.04058201611042023,
0.0695318952202797,
-0.05827329307794571,
-0.041369255632162094,
0.026105687022209167,
0.06235283985733986,
0.037674665451049805,
-0.01085734087973833,
0.025307752192020416,
-0.13983379304409027,
0.034783218055963516,
0.02777986414730549,
0.05372925102710724,
-0.04217321053147316,
0.07322030514478683,
-0.04270261898636818,
0.058773018419742584,
-0.012184876948595047,
0.010409774258732796,
-0.04205344244837761,
-0.030865751206874847,
-0.06400483846664429,
0.0927305668592453,
0.04246022179722786,
0.04243023321032524,
-0.03205215930938721,
-0.001583048957400024,
0.05170310288667679,
-0.05693758651614189,
-0.06854276359081268,
0.03261594846844673,
-0.03827790170907974,
0.014866217970848083,
-0.11160579323768616,
-0.0527118556201458,
0.05308584123849869,
-0.034125056117773056,
-0.049328576773405075,
0.05514351651072502,
-0.0496877022087574,
-0.019131582230329514,
-0.029180286452174187,
-0.0646100863814354,
-0.03145119175314903,
-0.06382326036691666,
0.006854038219898939,
-0.03324628248810768,
0.0004267073527444154,
-0.04900114983320236,
0.07750118523836136,
-0.0928448960185051,
0.040683187544345856,
-0.03578498214483261,
-0.01026281900703907,
0.010575255379080772,
-0.009506839327514172,
0.0011664159828796983,
0.03452788665890694,
-0.06933685392141342,
0.02411220595240593,
0.08465808629989624,
0.08386893570423126,
0.03099062107503414,
0.03062513656914234,
-0.03323352709412575,
-0.013937952928245068,
-0.0818900540471077,
0.022578563541173935,
-0.0074515510350465775,
0.015069435350596905,
-0.0001272721419809386,
0.11397895216941833,
0.02550654299557209,
-0.029042314738035202,
0.07519444823265076,
-0.032005634158849716,
0.04182088375091553,
0.051894787698984146,
0.025052374228835106,
-0.006188406143337488,
0.07276241481304169,
-0.026123950257897377,
-5.671261149586826e-8,
-0.03945275768637657,
0.0319102443754673,
0.009109425358474255,
0.022017857059836388,
0.05489804968237877,
-0.06245254725217819,
-0.0585523396730423,
0.04479324817657471,
-0.008245027624070644,
0.02359955385327339,
0.0924263522028923,
-0.04054112732410431,
-0.11473636329174042,
-0.023148708045482635,
0.05447704717516899,
0.06293581426143646,
-0.0010848112870007753,
-0.02066422626376152,
-0.01950477994978428,
-0.013908972963690758,
0.05559918284416199,
0.062341928482055664,
-0.024655114859342575,
0.029983287677168846,
0.03159142658114433,
-0.036845773458480835,
-0.03228922560811043,
0.06615942716598511,
0.044018011540174484,
0.033030878752470016,
-0.03339485824108124,
0.04103371128439903,
-0.0102956248447299,
0.012312332168221474,
0.009140484035015106,
0.12081318348646164,
-0.0013147678691893816,
-0.055550284683704376,
0.003975113853812218,
-0.045850712805986404,
0.03145192191004753,
-0.007015906274318695,
-0.02189197763800621,
0.017351793125271797,
0.08693060278892517,
-0.004526602104306221,
-0.04486493766307831,
-0.09928294271230698,
0.006169918458908796,
-0.020129067823290825,
-0.005552010610699654,
-0.04817729815840721,
-0.02324596792459488,
0.08015263080596924,
0.055894043296575546,
0.04385517165064812,
-0.02584066055715084,
-0.08073101937770844,
0.042401671409606934,
0.009307761676609516,
0.025117550045251846,
0.03416173532605171,
-0.0019557459745556116,
-0.017462076619267464
] | https://github.com/tianyi-lab/moe-embedding | 65 | 0 | 0 | 0 |
REEF: Representation Encoding Fingerprints for Large Language Models | https://openreview.net/forum?id=SnDmPkOJ0T | [
"Jie Zhang",
"Dongrui Liu",
"Chen Qian",
"Linfeng Zhang",
"Yong Liu",
"Yu Qiao",
"Jing Shao"
] | Oral | Protecting the intellectual property of open-source Large Language Models (LLMs) is very important, because training LLMs costs extensive computational resources and data. Therefore, model owners and third parties need to identify whether a suspect model is a subsequent development of the victim model. To this end, we propose a training-free REEF to identify the relationship between the suspect and victim models from the perspective of LLMs' feature representations. Specifically, REEF computes and compares the centered kernel alignment similarity between the representations of a suspect model and a victim model on the same samples. This training-free REEF does not impair the model's general capabilities and is robust to sequential fine-tuning, pruning, model merging, and permutations. In this way, REEF provides a simple and effective way for third parties and models' owners to protect LLMs' intellectual property together. Our code is publicly accessible at https://github.com/AI45Lab/REEF. | Large Language Model, Fingerprint, Representation, Intellectual Property | null | 2,401 | 2410.14273 | [
-0.05628053471446037,
-0.09566366672515869,
-0.016017971560359,
-0.05170273408293724,
0.14207050204277039,
0.03532904386520386,
-0.07740306854248047,
-0.034203626215457916,
0.08686590939760208,
-0.06768520921468735,
0.030831346288323402,
-0.029714416712522507,
0.04516499489545822,
-0.007544247899204493,
-0.025846345350146294,
-0.004227004013955593,
-0.014262251555919647,
0.1192859336733818,
-0.05187948793172836,
-0.07807501405477524,
-0.028982248157262802,
0.04097200185060501,
-0.008129614405333996,
0.019747713580727577,
-0.006807663943618536,
0.017742237076163292,
0.03938807174563408,
0.022691458463668823,
0.0010251380736008286,
-0.08685098588466644,
0.004284583497792482,
0.05678614228963852,
0.03129786625504494,
0.028939388692378998,
-0.001713448902592063,
0.03768903762102127,
-0.09418945014476776,
-0.0015796272782608867,
-0.0021360847167670727,
-0.10111326724290848,
-0.02899034321308136,
-0.0159606896340847,
0.017707478255033493,
0.05520278587937355,
0.026817910373210907,
0.0347270630300045,
-0.012041804380714893,
0.0077256495133042336,
-0.0203451719135046,
-0.03012455627322197,
-0.1064906194806099,
-0.06673230230808258,
-0.00912543572485447,
0.04518810287117958,
-0.01694829761981964,
-0.10000940412282944,
0.036429595202207565,
-0.02157364785671234,
-0.019087351858615875,
-0.03986405208706856,
-0.00982141587883234,
-0.05510042607784271,
-0.06578877568244934,
0.016173070296645164,
-0.010067235678434372,
0.00042818457586690784,
0.035532839596271515,
0.09417743235826492,
-0.027351494878530502,
0.006485948339104652,
-0.013313757255673409,
0.053060729056596756,
0.009459949098527431,
0.08788697421550751,
-0.02384631335735321,
0.11191429942846298,
0.02420060895383358,
0.015200238674879074,
0.11140335351228714,
-0.0736164078116417,
-0.04513269290328026,
-0.0008260115864686668,
0.08223497867584229,
0.01942169852554798,
0.08351632952690125,
0.008322152309119701,
0.03402629867196083,
0.06853000074625015,
0.04525609686970711,
0.012816358357667923,
-0.040338125079870224,
-0.026286209002137184,
0.11949527263641357,
-0.06589987128973007,
0.0007323989411816001,
0.03550109639763832,
-0.012096462771296501,
-0.0008887836593203247,
-0.02171848528087139,
0.09850174933671951,
-0.06700079888105392,
0.03054419718682766,
0.07240235805511475,
-0.09188897907733917,
0.08557916432619095,
0.003130890429019928,
0.012259525246918201,
-0.02252878062427044,
0.07186727225780487,
-0.08923149108886719,
0.0357048325240612,
0.029080813750624657,
-0.03864327818155289,
0.03045123815536499,
0.08962666243314743,
-0.011923948302865028,
-0.016238141804933548,
-0.030469899997115135,
0.0003914798144251108,
0.005261179991066456,
-0.04148421809077263,
-0.011264918372035027,
-0.019265921786427498,
0.024008726701140404,
-0.02726338990032673,
-0.038143888115882874,
-0.11785002052783966,
1.7296516890629067e-35,
0.08878090232610703,
0.10544693470001221,
-0.028787774965167046,
-0.012745973654091358,
0.07693377882242203,
-0.075210802257061,
-0.011087971739470959,
0.029314938932657242,
-0.04878181591629982,
0.03085578791797161,
-0.05806003138422966,
0.004230031743645668,
-0.050782181322574615,
0.05340993031859398,
-0.04422873258590698,
0.024473773315548897,
-0.017688272520899773,
0.06211791932582855,
-0.08716975897550583,
-0.03190716356039047,
0.060987237840890884,
0.0001460028433939442,
0.0033618139568716288,
-0.021295420825481415,
0.010748093016445637,
-0.008040969260036945,
-0.015890365466475487,
-0.10821158438920975,
-0.009167126379907131,
0.02720433846116066,
-0.06519561260938644,
-0.025454219430685043,
0.02568335086107254,
0.062357865273952484,
0.07694091647863388,
-0.04859260842204094,
-0.02413107268512249,
-0.0848732441663742,
0.0317634642124176,
-0.02764134854078293,
0.034918636083602905,
0.014714285731315613,
0.02424726076424122,
-0.048614464700222015,
-0.010003577917814255,
-0.0542510449886322,
-0.056269265711307526,
-0.03553862124681473,
0.023829473182559013,
-0.044441282749176025,
0.11656046658754349,
-0.006588248070329428,
-0.08335412293672562,
-0.006718095857650042,
-0.07751831412315369,
0.009474636055529118,
-0.023769741877913475,
0.02519150823354721,
-0.003440745174884796,
0.07652357965707779,
-0.0270023625344038,
0.03116261586546898,
-0.02134348452091217,
0.04237447679042816,
0.06166147068142891,
-0.05213902145624161,
-0.027205076068639755,
-0.0034481524489820004,
0.0762963816523552,
0.005353657063096762,
-0.02971753291785717,
-0.030838923528790474,
-0.03195925056934357,
-0.03864164277911186,
-0.06677037477493286,
-0.04111657291650772,
0.07161808758974075,
-0.0747857466340065,
0.00925759319216013,
0.05247313156723976,
-0.02296147309243679,
0.021799635142087936,
0.029421355575323105,
-0.10576417297124863,
-0.09735459834337234,
-0.060178544372320175,
0.00799062754958868,
-0.05725756287574768,
-0.019669435918331146,
0.04824058338999748,
-0.025107121095061302,
-0.060659125447273254,
-0.017771847546100616,
0.038229361176490784,
0.003166212234646082,
-3.6481770234907014e-34,
-0.06256869435310364,
0.01635870337486267,
0.03621361404657364,
0.010023331269621849,
-0.09170366078615189,
-0.07107038050889969,
0.016807816922664642,
0.06717091053724289,
-0.042714569717645645,
-0.07641997933387756,
-0.0697515457868576,
-0.06727565079927444,
0.10635749995708466,
-0.03379210829734802,
0.07519989460706711,
-0.024015016853809357,
0.006035209633409977,
0.014000079594552517,
0.03706968203186989,
0.07109428942203522,
0.010302751325070858,
0.0884186327457428,
-0.06227978318929672,
0.07024328410625458,
-0.0025190243031829596,
0.026841316372156143,
0.07017502188682556,
-0.0064884996972978115,
0.06167485937476158,
0.001519550452940166,
-0.052479133009910583,
0.0202970989048481,
-0.08498615771532059,
-0.03523971885442734,
-0.1117909625172615,
-0.024403301998972893,
-0.0025977736804634333,
0.024096418172121048,
0.04089881852269173,
0.10871030390262604,
0.048150595277547836,
-0.005812434945255518,
-0.11391361057758331,
-0.015928488224744797,
-0.08062826097011566,
-0.031339310109615326,
-0.02566690556704998,
0.02183588221669197,
0.05793832987546921,
-0.12995906174182892,
0.020723458379507065,
0.0360521636903286,
-0.007107112091034651,
-0.0333566814661026,
-0.03922642394900322,
-0.05694730579853058,
0.01589675061404705,
-0.024089403450489044,
0.03940963372588158,
0.002072848379611969,
-0.04119803383946419,
-0.01590200886130333,
0.019560260698199272,
-0.03463159129023552,
0.04677680879831314,
-0.0006064196932129562,
-0.02544338069856167,
0.017049288377165794,
-0.02013569511473179,
0.0007291498477570713,
0.06783856451511383,
-0.007253836840391159,
-0.095675528049469,
0.04871169105172157,
-0.02485847845673561,
-0.05184992775321007,
-0.03063458949327469,
-0.07516924291849136,
-0.02596937119960785,
-0.027342915534973145,
0.035838786512613297,
-0.013396363705396652,
0.04074903205037117,
0.06525883823633194,
0.09567680954933167,
0.06325271725654602,
0.07764371484518051,
0.0414130799472332,
0.02653834968805313,
0.04217734560370445,
-0.04495877027511597,
0.0640382245182991,
-0.0020986066665500402,
0.08238697797060013,
-0.018924130126833916,
-4.263635133838761e-8,
-0.0441872663795948,
-0.010171535424888134,
0.00958157517015934,
0.08617522567510605,
0.00006176876195240766,
-0.02827921323478222,
-0.004964893218129873,
0.07721368223428726,
-0.036416277289390564,
0.03986356407403946,
0.047440219670534134,
0.023185910657048225,
-0.05285268649458885,
-0.05017523467540741,
0.00548706017434597,
0.07468628138303757,
-0.023952998220920563,
0.05411505326628685,
-0.03616386651992798,
0.02721124142408371,
0.03632177785038948,
0.06813343614339828,
0.014590454287827015,
0.008908690884709358,
0.037160858511924744,
-0.044891566038131714,
-0.020444152876734734,
0.08657132089138031,
0.045713040977716446,
0.078420490026474,
-0.02987007424235344,
0.02288763038814068,
-0.01025601290166378,
-0.03290571644902229,
0.07132776826620102,
0.10893849283456802,
-0.023839352652430534,
-0.0016896590823307633,
-0.01725620962679386,
0.013416588306427002,
0.0635777935385704,
0.07855325192213058,
-0.09859797358512878,
0.013068233616650105,
0.04427957907319069,
-0.0029911454766988754,
0.012029838748276234,
-0.13092704117298126,
0.04706981033086777,
-0.0011836906196549535,
0.03732617571949959,
-0.006528754252940416,
-0.012150737456977367,
0.06749438494443893,
-0.01130669191479683,
0.044838905334472656,
0.004352176561951637,
0.01761145517230034,
0.07408315688371658,
0.05205943435430527,
0.010097097605466843,
-0.029446717351675034,
0.07400991767644882,
-0.019114520400762558
] | https://github.com/tmylla/reef | 41 | 0 | 0 | 0 |
Flat Reward in Policy Parameter Space Implies Robust Reinforcement Learning | https://openreview.net/forum?id=4OaO3GjP7k | [
"Hyun Kyu Lee",
"Sung Whan Yoon"
] | Oral | Investigating flat minima on loss surfaces in parameter space is well-documented in the supervised learning context, highlighting its advantages for model generalization. However, limited attention has been paid to the reinforcement learning (RL) context, where the impact of flatter reward landscapes in policy parameter space remains largely unexplored. Beyond merely extrapolating from supervised learning, which suggests a link between flat reward landscapes and enhanced generalization, we aim to formally connect the flatness of the reward surface to the robustness of RL models. In policy models where a deep neural network determines actions, flatter reward landscapes in response to parameter perturbations lead to consistent rewards even when actions are perturbed. Moreover, robustness to action perturbations further enhances robustness against other variations, such as changes in state transition probabilities and reward functions. We extensively simulate various RL environments, confirming the consistent benefits of flatter reward landscapes in enhancing the robustness of RL under diverse conditions, including action selection, transition dynamics, and reward functions. The code for these experiments is available at https://github.com/HK-05/flatreward-RRL. | Reinforcement learning, Flat Minima, Robust Reinforcement learning | null | 2,326 | null | [
-0.08428554236888885,
-0.0786646381020546,
0.053606003522872925,
0.026942851021885872,
0.04999112710356712,
0.10245374590158463,
-0.023488182574510574,
-0.0003497490251902491,
0.0014428457943722606,
0.003299824893474579,
-0.07132035493850708,
-0.057901281863451004,
0.030103670433163643,
0.06649334728717804,
0.03774332255125046,
0.021494729444384575,
0.035903140902519226,
0.0430586002767086,
-0.06295033544301987,
-0.009307749569416046,
0.06423527747392654,
0.023940254002809525,
0.08459696173667908,
0.015296784229576588,
-0.05537017807364464,
-0.013627284206449986,
0.03749895840883255,
-0.009089775383472443,
-0.001347108162008226,
-0.06370542198419571,
0.08508391678333282,
-0.03791631758213043,
-0.005423582158982754,
-0.09439976513385773,
-0.02724728360772133,
0.08444233983755112,
-0.11447082459926605,
-0.020986270159482956,
-0.05493823438882828,
-0.04030005633831024,
0.023624299094080925,
0.0422576479613781,
0.028577301651239395,
0.08615339547395706,
0.09796398133039474,
0.02285609021782875,
0.03072371706366539,
-0.05382769927382469,
-0.018834644928574562,
-0.04169850796461105,
0.06468841433525085,
-0.03171838819980621,
0.023638322949409485,
-0.0008119532139971852,
-0.011077551171183586,
0.02441413328051567,
-0.002643667161464691,
0.02923072874546051,
-0.013983733020722866,
-0.06080405041575432,
0.008391018025577068,
-0.14293941855430603,
-0.016655590385198593,
-0.027712684124708176,
-0.0270686075091362,
-0.03736652061343193,
-0.013195300474762917,
-0.005653248634189367,
-0.07385975867509842,
-0.021338004618883133,
-0.017286131158471107,
0.04044710472226143,
-0.044124163687229156,
-0.10160966217517853,
0.06955116987228394,
0.0013406315119937062,
-0.012199996039271355,
0.013720531947910786,
-0.025760166347026825,
-0.003785383654758334,
-0.008417919278144836,
-0.023921502754092216,
-0.00878237560391426,
0.02262209728360176,
0.02173091284930706,
-0.05269722640514374,
-0.03981796279549599,
0.012889645993709564,
0.08201063424348831,
-0.007818984799087048,
-0.023694003000855446,
-0.003242139471694827,
-0.019833192229270935,
-0.013673679903149605,
-0.056145504117012024,
0.040884457528591156,
-0.0122519014403224,
-0.12493081390857697,
-0.10873586684465408,
0.08848785609006882,
0.05562227964401245,
0.0125560462474823,
0.018728388473391533,
0.0430208295583725,
0.09461495280265808,
-0.04326969385147095,
0.00885106436908245,
0.00806794036179781,
0.02614045888185501,
-0.07632559537887573,
0.022616488859057426,
-0.0065839351154863834,
0.08464265614748001,
0.08215402811765671,
-0.024251844733953476,
-0.0165493655949831,
-0.0065911151468753815,
-0.06995663791894913,
-0.059590667486190796,
0.024024151265621185,
0.0011937743984162807,
-0.0909857526421547,
0.044348493218421936,
-0.008992761373519897,
-0.015246065333485603,
-0.021756602451205254,
-0.041876643896102905,
4.910964870501998e-33,
0.07926075160503387,
0.008560349233448505,
-0.021481910720467567,
0.02985207736492157,
0.004677779506891966,
-0.0013861882034689188,
0.0784917026758194,
-0.010259415954351425,
-0.0543767511844635,
-0.0006783271674066782,
-0.03457944467663765,
0.07168068736791611,
-0.06959746032953262,
0.0954560860991478,
0.011312217451632023,
0.0149052944034338,
-0.01523516047745943,
0.05944947898387909,
0.03473204746842384,
-0.054404888302087784,
-0.012477231211960316,
0.04734490439295769,
0.03407168388366699,
-0.047829046845436096,
0.046847153455019,
0.02472303807735443,
0.006671269424259663,
-0.028312543407082558,
-0.053122278302907944,
-0.007467314600944519,
0.014515933580696583,
0.05269638076424599,
-0.001221403363160789,
0.04824436455965042,
0.030014511197805405,
0.08306563645601273,
-0.026939785107970238,
-0.013125224970281124,
0.013061681762337685,
-0.02075846865773201,
0.002589368959888816,
0.03532150760293007,
0.038938459008932114,
0.08458098024129868,
-0.03654327616095543,
-0.08796315640211105,
0.05972227081656456,
0.02051016502082348,
-0.10355114191770554,
0.0077202171087265015,
-0.021757271140813828,
0.05362379178404808,
0.050943851470947266,
-0.10681036114692688,
-0.03902968764305115,
0.00808405689895153,
0.005249380134046078,
0.04135013371706009,
-0.08187531679868698,
0.01065877266228199,
0.06711052358150482,
-0.0185343399643898,
-0.014343035407364368,
-0.0030282342340797186,
-0.004112995695322752,
0.10543162375688553,
0.01152378786355257,
0.0385226309299469,
0.016316115856170654,
0.03598204255104065,
0.051022350788116455,
0.08753710985183716,
-0.01593288592994213,
-0.11576981097459793,
0.10874336957931519,
0.0024335538037121296,
0.0631355345249176,
-0.11905068904161453,
0.02500862255692482,
-0.06272999942302704,
-0.020954109728336334,
-0.022039910778403282,
-0.05004214122891426,
0.02462136372923851,
-0.03533223643898964,
-0.022569924592971802,
0.04961273446679115,
-0.09752466529607773,
-0.002613817108795047,
0.003710429649800062,
-0.035905394703149796,
-0.05359179899096489,
0.031015094369649887,
0.06709899008274078,
0.03379480168223381,
-4.3432308537555885e-33,
0.006895930971950293,
-0.0316026508808136,
-0.08960907906293869,
0.05397139489650726,
-0.02301686815917492,
0.02700432762503624,
-0.028588540852069855,
0.04067239537835121,
-0.051505304872989655,
-0.03716350346803665,
-0.04166233539581299,
0.06805749982595444,
0.039127517491579056,
0.03746713325381279,
0.004943406209349632,
-0.05712452530860901,
-0.0029379650950431824,
0.048552896827459335,
-0.002935719210654497,
0.02192145772278309,
0.0743166133761406,
0.08939751982688904,
-0.13600404560565948,
0.05017619952559471,
-0.011997265741229057,
0.04515409842133522,
-0.07127238810062408,
0.047496210783720016,
0.03538038581609726,
0.03476596996188164,
-0.06808627396821976,
-0.03899220749735832,
-0.07511057704687119,
-0.02208094671368599,
-0.05073912814259529,
0.07731052488088608,
0.06425300985574722,
0.013740524649620056,
-0.11656694114208221,
0.06564594060182571,
-0.02402932569384575,
-0.06755444407463074,
-0.03868560865521431,
0.04553801566362381,
0.016828475520014763,
-0.05039900541305542,
-0.005753075238317251,
-0.018453454598784447,
-0.041646070778369904,
-0.016104748472571373,
-0.05208226665854454,
0.022265704348683357,
-0.03687024116516113,
0.038100309669971466,
-0.08217372000217438,
-0.01937316358089447,
-0.03652656450867653,
-0.029668696224689484,
0.08733972162008286,
0.009711056016385555,
-0.11884968727827072,
0.039554860442876816,
-0.07435774803161621,
0.021590018644928932,
-0.02799352817237377,
0.029600592330098152,
-0.02536662481725216,
-0.022776568308472633,
0.045843228697776794,
0.004885502625256777,
0.0452423170208931,
0.009783984161913395,
0.0834776982665062,
-0.03962681069970131,
0.033509641885757446,
-0.0031028417870402336,
0.014928348362445831,
-0.045918598771095276,
0.005920717027038336,
-0.022890815511345863,
0.0538020059466362,
0.03544200584292412,
0.05457853525876999,
0.017218006774783134,
0.01684364303946495,
0.044158514589071274,
0.010502927005290985,
0.09639975428581238,
-0.0013734044041484594,
-0.027251560240983963,
0.03588682413101196,
-0.017359459772706032,
0.020394986495375633,
0.036094631999731064,
-0.048810143023729324,
-4.321524826877976e-8,
-0.07786142081022263,
0.07890132069587708,
0.014852236025035381,
0.023625025525689125,
0.03926980495452881,
0.016751592978835106,
0.02942194975912571,
0.013545019552111626,
-0.024156710132956505,
-0.01773814484477043,
0.028314918279647827,
0.05242760479450226,
0.053846850991249084,
-0.11886126548051834,
-0.022206200286746025,
0.12997214496135712,
0.019190456718206406,
0.042464759200811386,
0.0022669683676213026,
0.04028324410319328,
0.009549599140882492,
0.008536787703633308,
-0.032803263515233994,
-0.10002434998750687,
0.02784314937889576,
-0.08524567633867264,
-0.05538294091820717,
0.11541395634412766,
0.09056945145130157,
0.006342207081615925,
0.0670069232583046,
-0.01834067516028881,
0.0707978904247284,
0.062376637011766434,
-0.018813231959939003,
0.05861663445830345,
0.04976819455623627,
-0.03606762737035751,
0.017499694600701332,
0.006273171398788691,
-0.019410457462072372,
0.07030736654996872,
-0.02349630370736122,
-0.021576259285211563,
-0.08735757321119308,
0.06325075030326843,
-0.03785143792629242,
-0.1074429377913475,
0.07999823242425919,
-0.07634596526622772,
0.0017924070125445724,
-0.028187181800603867,
-0.05787816271185875,
0.07917256653308868,
0.04818522557616234,
-0.01371452771127224,
0.008963806554675102,
-0.07958444207906723,
-0.031979039311409,
0.10464683175086975,
0.01656404882669449,
-0.015327226370573044,
-0.10759058594703674,
-0.0201419610530138
] | 0 | 0 | 0 | 0 |
|
LLM-SR: Scientific Equation Discovery via Programming with Large Language Models | https://openreview.net/forum?id=m2nmp8P5in | [
"Parshin Shojaee",
"Kazem Meidani",
"Shashank Gupta",
"Amir Barati Farimani",
"Chandan K. Reddy"
] | Oral | Mathematical equations have been unreasonably effective in describing complex natural phenomena across various scientific disciplines. However, discovering such insightful equations from data presents significant challenges due to the necessity of navigating extremely large combinatorial hypothesis spaces. Current methods of equation discovery, commonly known as symbolic regression techniques, largely focus on extracting equations from data alone, often neglecting the domain-specific prior knowledge that scientists typically depend on. They also employ limited representations such as expression trees, constraining the search space and expressiveness of equations. To bridge this gap, we introduce LLM-SR, a novel approach that leverages the extensive scientific knowledge and robust code generation capabilities of Large Language Models (LLMs) to discover scientific equations from data. Specifically, LLM-SR treats equations as programs with mathematical operators and combines LLMs' scientific priors with evolutionary search over equation programs. The LLM iteratively proposes new equation skeleton hypotheses, drawing from its domain knowledge, which are then optimized against data to estimate parameters. We evaluate LLM-SR on four benchmark problems across diverse scientific domains (e.g., physics, biology), which we carefully designed to simulate the discovery process and prevent LLM recitation. Our results demonstrate that LLM-SR discovers physically accurate equations that significantly outperform state-of-the-art symbolic regression baselines, particularly in out-of-domain test settings. We also show that LLM-SR's incorporation of scientific priors enables more efficient equation space exploration than the baselines. | Symbolic Regression, Equation Discovery, Large Language Models, Evolutionary Search | We introduce LLM-SR, an approach that harnesses Large Language Models (LLMs) to discover governing equations from data in an efficient, knowledge-guided manner. | 2,272 | null | [
-0.11292421072721481,
0.012586317025125027,
0.07334477454423904,
0.09288782626390457,
0.037688128650188446,
-0.02776372618973255,
-0.07304996252059937,
0.04660962149500847,
0.00215742620639503,
0.047470007091760635,
-0.02502899244427681,
-0.07833336293697357,
-0.0004954519681632519,
0.06985154747962952,
0.0365983247756958,
0.11760832369327545,
-0.04792952537536621,
0.042679887264966965,
-0.04021717235445976,
-0.08096206933259964,
0.021734943613409996,
0.04090490564703941,
-0.06113617122173309,
0.012480613775551319,
0.09754320979118347,
-0.02041032910346985,
0.0400078110396862,
0.007334788329899311,
0.006718483287841082,
0.09109277278184891,
0.02109234780073166,
0.03889904543757439,
0.009051626548171043,
-0.0064829750917851925,
-0.027484364807605743,
0.05578374117612839,
-0.1037091538310051,
-0.014267843216657639,
-0.033389050513505936,
-0.007381141185760498,
-0.02062777802348137,
-0.0054342313669621944,
0.0775359496474266,
-0.0028620585799217224,
0.07415478676557541,
-0.03909480944275856,
0.020615549758076668,
-0.022277552634477615,
-0.024162940680980682,
0.014894860796630383,
-0.06153004989027977,
-0.03888753429055214,
-0.05654941126704216,
0.00785098411142826,
0.08204537630081177,
-0.06880880147218704,
0.04674791544675827,
-0.06657898426055908,
0.01794172078371048,
-0.050505660474300385,
0.013628440909087658,
-0.08767952024936676,
0.0015319008380174637,
-0.04082344099879265,
0.028076771646738052,
0.006490177940577269,
0.019339658319950104,
0.1022818312048912,
0.019247395917773247,
0.08764728158712387,
-0.04684169590473175,
0.10303641110658646,
-0.06560023128986359,
0.07441391795873642,
0.055098071694374084,
0.04839364439249039,
0.07537391036748886,
0.04214973747730255,
0.12201332300901413,
-0.08292565494775772,
-0.013187818229198456,
-0.008542591705918312,
-0.07999369502067566,
0.02749805524945259,
0.039556119590997696,
-0.022817391902208328,
0.024022920057177544,
0.08033766597509384,
0.03600070998072624,
0.07444730401039124,
-0.003936107736080885,
-0.13828690350055695,
0.025736572220921516,
0.0074787745252251625,
-0.03361133858561516,
0.05530779808759689,
-0.02352042868733406,
-0.055506158620119095,
0.020714350044727325,
0.05055784434080124,
0.02092130109667778,
0.1521308571100235,
0.06633549183607101,
-0.07336077094078064,
-0.05712083354592323,
-0.03448871523141861,
-0.0074536618776619434,
0.017027201130986214,
0.026588493958115578,
-0.0524224229156971,
0.0020037787035107613,
0.02876483090221882,
-0.011406255885958672,
0.10426761955022812,
-0.0014747256645932794,
-0.005034203641116619,
0.09756588935852051,
-0.007671594154089689,
-0.011551184579730034,
-0.02024848200380802,
-0.004949730355292559,
0.031768474727869034,
-0.006213540211319923,
0.034575145691633224,
0.06608119606971741,
0.02537602372467518,
-0.1480049043893814,
2.3773082038304783e-33,
0.07889082282781601,
0.03900279849767685,
0.010402881540358067,
0.05933617055416107,
0.020811019465327263,
-0.02324766479432583,
0.0064462278969585896,
0.026787085458636284,
-0.061221733689308167,
0.09928792715072632,
-0.08070256561040878,
0.034672658890485764,
-0.0548907108604908,
0.0032844168599694967,
0.007017879281193018,
-0.0065968893468379974,
0.00944798905402422,
0.0149586396291852,
-0.08037049323320389,
-0.04875878244638443,
0.017669200897216797,
-0.07867582142353058,
-0.04043104499578476,
-0.04480681195855141,
0.08793231099843979,
0.03395506367087364,
0.04271114990115166,
-0.05690782144665718,
-0.014372596517205238,
-0.010740424506366253,
-0.06978611648082733,
-0.02145259827375412,
-0.027417706325650215,
0.0639905110001564,
0.0556112639605999,
-0.04559747874736786,
0.0014181575970724225,
-0.0957934632897377,
-0.012036311440169811,
0.03933506831526756,
-0.025580884888768196,
0.039676178246736526,
0.01813315600156784,
0.019702700898051262,
-0.02196001261472702,
-0.0522913821041584,
-0.029965506866574287,
0.003909842576831579,
0.05878002569079399,
0.010796580463647842,
-0.008331588469445705,
-0.016221780329942703,
0.01801133342087269,
-0.04258057102560997,
-0.0122615285217762,
0.003294976893812418,
-0.07770640403032303,
-0.016148529946804047,
0.030516739934682846,
0.051638759672641754,
-0.012597267515957355,
0.027253994718194008,
0.032650113105773926,
0.011649238876998425,
0.048419613391160965,
0.08196035027503967,
-0.06966690719127655,
0.027672117576003075,
0.07865944504737854,
0.05355113372206688,
0.011558420956134796,
0.026268448680639267,
-0.030373917892575264,
-0.0775163471698761,
0.09011176228523254,
-0.04075706750154495,
0.08720578253269196,
-0.10816679894924164,
-0.004221834242343903,
-0.04291936382651329,
-0.0381045788526535,
0.017253924161195755,
0.008314005099236965,
-0.036737773567438126,
-0.02620076946914196,
-0.07370219379663467,
-0.01650509424507618,
-0.08376830071210861,
-0.05943368747830391,
-0.12363201379776001,
-0.08476903289556503,
-0.015923738479614258,
-0.024536048993468285,
-0.0017415801994502544,
0.004049586597830057,
-3.6017760370309225e-33,
-0.09220872074365616,
-0.056919049471616745,
0.040315721184015274,
-0.028789689764380455,
-0.04621199145913124,
-0.09051545709371567,
-0.04866680130362511,
-0.022431906312704086,
-0.025112435221672058,
-0.13361197710037231,
0.011085006408393383,
-0.02803250402212143,
-0.015472001396119595,
-0.004407268483191729,
0.08716686069965363,
0.003340686671435833,
-0.015044119209051132,
-0.0020335197914391756,
-0.0014075104845687747,
0.09943364560604095,
-0.06893870234489441,
0.09550020843744278,
-0.0477232001721859,
0.010296099819242954,
-0.04841747507452965,
-0.014322159811854362,
0.057745590806007385,
0.025868169963359833,
-0.04381784796714783,
0.036025017499923706,
-0.04899374768137932,
-0.04574846103787422,
-0.06039950251579285,
-0.06643004715442657,
-0.07977418601512909,
0.026071785017848015,
0.08179519325494766,
0.04994713142514229,
-0.01070819515734911,
-0.011097392067313194,
0.05242994427680969,
0.0025980446953326464,
-0.013326974585652351,
0.006747706327587366,
-0.008208800107240677,
-0.01906507834792137,
-0.03600262477993965,
0.037624210119247437,
0.08907775580883026,
0.059470731765031815,
0.021403374150395393,
-0.05846758559346199,
0.03374195098876953,
-0.05773061513900757,
-0.04804249107837677,
-0.05274127796292305,
-0.09379366785287857,
-0.025109095498919487,
-0.0606364943087101,
-0.025150876492261887,
-0.050772856920957565,
-0.022790275514125824,
0.06258472800254822,
0.04664336144924164,
-0.05682532489299774,
-0.061583057045936584,
-0.060992058366537094,
0.022912941873073578,
-0.06983610987663269,
-0.04390536993741989,
0.017776207998394966,
-0.08535487204790115,
-0.033960551023483276,
0.02678685449063778,
0.02510196715593338,
0.0047934213653206825,
0.021828651428222656,
-0.031054511666297913,
0.028182754293084145,
0.007305464241653681,
0.07617060095071793,
0.06100129336118698,
0.08507143706083298,
-0.009200310334563255,
-0.010339800268411636,
-0.04476144164800644,
0.06567387282848358,
0.08116860687732697,
0.022088881582021713,
0.010074291378259659,
-0.1036933958530426,
0.031878016889095306,
-0.03277228772640228,
0.01037113182246685,
0.01117142103612423,
-4.109742590685528e-8,
-0.016815725713968277,
0.05192742496728897,
0.04566624388098717,
-0.006213483400642872,
0.08231131732463837,
0.032714925706386566,
-0.06546571105718613,
0.04774190112948418,
-0.03925255686044693,
0.022755613550543785,
0.061544131487607956,
0.0678774043917656,
-0.05932243913412094,
-0.050504088401794434,
-0.0452895313501358,
0.0640585646033287,
-0.029093222692608833,
-0.01599857583642006,
-0.07149208337068558,
0.0030227038078010082,
-0.015147920697927475,
0.04557105526328087,
-0.0009026611223816872,
0.008958451449871063,
0.1128987967967987,
-0.09290895611047745,
0.016563480719923973,
0.053990401327610016,
0.07351219654083252,
0.017286358401179314,
-0.024186959490180016,
-0.02312498539686203,
0.00673185708001256,
-0.0019223890267312527,
0.058900684118270874,
-0.006067174021154642,
0.014957370236515999,
-0.04153265431523323,
-0.007538854144513607,
-0.06943034380674362,
0.012707945890724659,
0.03477977588772774,
-0.01467290148139,
0.03506621718406677,
0.03592585027217865,
-0.012152008712291718,
0.020359251648187637,
0.009851891547441483,
0.0022209545131772757,
0.023524051532149315,
0.03006081096827984,
0.013329312205314636,
-0.07332758605480194,
-0.021009325981140137,
0.007922813296318054,
0.04831046238541603,
-0.07892187684774399,
0.049423012882471085,
0.03374193608760834,
-0.006222376599907875,
0.032898809760808945,
-0.03871695324778557,
0.048622921109199524,
0.04050270467996597
] | 0 | 0 | 0 | 0 |
|
Backtracking Improves Generation Safety | https://openreview.net/forum?id=Bo62NeU6VF | [
"Yiming Zhang",
"Jianfeng Chi",
"Hailey Nguyen",
"Kartikeya Upasani",
"Daniel M. Bikel",
"Jason E Weston",
"Eric Michael Smith"
] | Oral | Text generation has a fundamental limitation almost by definition: there is no taking back tokens that have been generated, even when they are clearly problematic.
In the context of language model safety, when a partial unsafe generation is produced, language models by their nature tend to happily keep on generating similarly unsafe additional text.
This is in fact how safety alignment of frontier models gets circumvented in the wild, despite great efforts in improving their safety.
Deviating from the paradigm of approaching safety alignment as prevention (decreasing the probability of harmful responses), we propose backtracking, a technique that allows language models to "undo" and recover from their own unsafe generation through the introduction of a special [RESET] token.
Our method can be incorporated into either SFT or DPO training to optimize helpfulness and harmlessness.
We show that models trained to backtrack are consistently safer than baseline models: backtracking Llama-3-8B is four times more safe than the baseline model (6.1\% $\to$ 1.5\%) in our evaluations without regression in helpfulness.
Our method additionally provides protection against four adversarial attacks including an adaptive attack, despite not being trained to do so. | AI safety, Generation algorithm, Backtracking | We introduce a backtracking technique that trains language models to recover from unsafe generations and substantially improves generation safety. | 2,265 | 2409.14586 | [
-0.0751231387257576,
-0.05979549139738083,
-0.04348417744040489,
0.04930814355611801,
0.06848183274269104,
0.06971425563097,
-0.02790897712111473,
-0.015271306969225407,
0.01832752861082554,
-0.018751922994852066,
-0.030169885605573654,
-0.04259880259633064,
0.08504371345043182,
-0.05141398683190346,
-0.07341448962688446,
-0.03783009573817253,
0.019403213635087013,
0.13861212134361267,
-0.07550717890262604,
-0.02766723372042179,
-0.02976902388036251,
0.1029123067855835,
0.03789361193776131,
0.050315454602241516,
-0.019140146672725677,
-0.009332267567515373,
-0.012182611972093582,
0.005249222740530968,
0.01843889430165291,
-0.012312411330640316,
-0.02759956754744053,
-0.02103322744369507,
-0.04597560316324234,
0.009298870339989662,
-0.04139009118080139,
0.04521239176392555,
-0.06108181178569794,
0.026822824031114578,
0.06726737320423126,
-0.05618467926979065,
-0.09078524261713028,
-0.03565046191215515,
-0.021268164739012718,
0.019847849383950233,
0.07003947347402573,
-0.07912196964025497,
-0.011383099481463432,
-0.005239993799477816,
-0.01866922527551651,
-0.03044949099421501,
0.02063123695552349,
-0.017841624096035957,
-0.007242207415401936,
0.00968088861554861,
-0.022655460983514786,
-0.054970256984233856,
0.05999890714883804,
0.05644052103161812,
-0.024323759600520134,
-0.0071982089430093765,
0.0007298762793652713,
-0.07163836061954498,
-0.08995350450277328,
-0.05669856816530228,
-0.012921779416501522,
-0.04173203185200691,
-0.03178417310118675,
0.0551508404314518,
-0.005028481595218182,
0.10065837949514389,
-0.017329558730125427,
0.05485495552420616,
-0.004014450591057539,
0.06180299073457718,
-0.01752324588596821,
0.16153454780578613,
0.02520664408802986,
0.01238700095564127,
0.02689390629529953,
-0.05618487671017647,
0.1102568581700325,
-0.02042323723435402,
0.10723590105772018,
0.06871385127305984,
0.04319847747683525,
0.03704450652003288,
-0.02233036607503891,
0.0012172335991635919,
0.1356744021177292,
0.051462262868881226,
-0.07313001155853271,
-0.033168453723192215,
0.14840251207351685,
0.05752301216125488,
0.015931205824017525,
0.004488945007324219,
-0.015237121842801571,
-0.06702326238155365,
-0.01973443292081356,
0.08604896813631058,
0.010638345964252949,
0.06069105491042137,
0.042022380977869034,
-0.05852018669247627,
0.06714518368244171,
-0.03400036320090294,
0.03689158335328102,
-0.03832167759537697,
0.011133585125207901,
-0.07227659225463867,
0.02691195346415043,
0.07531478255987167,
0.011159216985106468,
0.028057880699634552,
0.00835595466196537,
-0.028234297409653664,
-0.03529888018965721,
-0.05922772362828255,
-0.027605634182691574,
0.03257440775632858,
-0.024958763271570206,
0.04815784469246864,
0.08851225674152374,
0.031669732183218,
0.018992416560649872,
-0.09203509241342545,
0.011046958155930042,
3.7211416108851716e-33,
0.03482234105467796,
0.08903713524341583,
-0.009177443571388721,
0.009198885411024094,
0.036663252860307693,
0.00597678916528821,
-0.03788492828607559,
0.0024800440296530724,
-0.013675034046173096,
-0.0010883808135986328,
-0.018164662644267082,
-0.012775097042322159,
-0.011035824194550514,
0.01242795865982771,
0.08371662348508835,
0.006422580685466528,
-0.026342526078224182,
0.03911365196108818,
-0.050227563828229904,
-0.027410035952925682,
0.04926960915327072,
-0.03166806325316429,
-0.018584709614515305,
-0.08137214183807373,
0.011477689258754253,
0.06415911763906479,
0.06375815719366074,
-0.014973359182476997,
-0.06923616677522659,
0.05096173658967018,
-0.14195528626441956,
0.013869555667042732,
0.03715873882174492,
0.017012469470500946,
0.03549377992749214,
-0.030271267518401146,
-0.07030749320983887,
-0.06429517269134521,
0.008586493320763111,
-0.0028592401649802923,
-0.030771058052778244,
0.04127166420221329,
-0.011150861158967018,
0.010150466114282608,
0.06732456386089325,
-0.017735855653882027,
-0.03651929646730423,
-0.014564506709575653,
-0.041731346398591995,
-0.05599333718419075,
-0.006358681246638298,
0.06458736956119537,
0.019052991643548012,
-0.046915844082832336,
-0.13995996117591858,
-0.019853882491588593,
-0.00807931087911129,
0.04260673373937607,
-0.03232980892062187,
0.06555067747831345,
0.018151745200157166,
0.04137369990348816,
-0.032732848078012466,
0.018221525475382805,
0.0028817488346248865,
-0.032035648822784424,
-0.006768747698515654,
0.01629762165248394,
0.028796186670660973,
-0.027396010234951973,
-0.03277837485074997,
-0.021206798031926155,
-0.05089675262570381,
-0.029023081064224243,
-0.058408260345458984,
-0.014640321023762226,
0.073134645819664,
0.01483853068202734,
0.052735112607479095,
-0.07386279106140137,
0.050453588366508484,
-0.04422159865498543,
0.014944781549274921,
-0.0341070182621479,
-0.03646833449602127,
-0.06403293460607529,
0.02659367024898529,
-0.037048548460006714,
-0.015885567292571068,
0.03882396221160889,
-0.03700605407357216,
-0.09858783334493637,
0.04364967346191406,
0.022357728332281113,
-0.01192049216479063,
-3.545048516370162e-33,
-0.06578699499368668,
-0.009825467132031918,
-0.07707153260707855,
0.046598177403211594,
-0.0946921706199646,
-0.06055845320224762,
0.0005243641790002584,
0.05992713198065758,
-0.03616614267230034,
-0.04628114402294159,
-0.0626736581325531,
-0.024827435612678528,
0.055462390184402466,
-0.02603498287498951,
0.04312668740749359,
-0.0667668953537941,
-0.04167312756180763,
0.053032286465168,
-0.03374716266989708,
0.037247925996780396,
0.10544560849666595,
0.03349687159061432,
-0.1142878383398056,
0.05725511908531189,
-0.0016058868495747447,
0.02578110620379448,
-0.006901572924107313,
0.07122717052698135,
0.0016854229616001248,
0.021647049114108086,
0.011526844464242458,
0.03457862511277199,
0.03882952779531479,
-0.0016287673497572541,
-0.06399419158697128,
-0.010541521944105625,
0.04996125027537346,
-0.004424732644110918,
0.0005807211273349822,
0.05590353533625603,
0.0663517564535141,
0.05953489616513252,
-0.09173943102359772,
0.016929352656006813,
-0.04590339586138725,
0.006421384401619434,
-0.07879092544317245,
-0.012175645679235458,
0.03258557990193367,
-0.04680052027106285,
0.027591321617364883,
-0.06917356699705124,
0.00040638676728121936,
0.010017188265919685,
-0.019755735993385315,
-0.07878856360912323,
0.05358410254120827,
-0.04059721902012825,
-0.028667783364653587,
0.03938819840550423,
-0.08301116526126862,
0.0033803025726228952,
0.017694536596536636,
0.010622531175613403,
0.06642735004425049,
-0.03318250551819801,
-0.041817497462034225,
0.020556313917040825,
0.006176630966365337,
-0.04682707041501999,
0.08064952492713928,
0.026767564937472343,
-0.07451900839805603,
0.03039497323334217,
0.013683575205504894,
-0.03782599791884422,
-0.025018172338604927,
-0.09466075897216797,
-0.002312184777110815,
-0.0387282520532608,
0.022638030350208282,
-0.005207570735365152,
0.040926672518253326,
0.0964420884847641,
0.042821817100048065,
-0.018690723925828934,
0.048573728650808334,
0.04536401107907295,
-0.018059642985463142,
0.048219095915555954,
-0.07238354533910751,
0.0010698168771341443,
-0.014388040639460087,
0.054100874811410904,
-0.027202723547816277,
-4.367777250990912e-8,
-0.026908107101917267,
0.0070579880848526955,
-0.06052885204553604,
0.1285158097743988,
0.016423234716057777,
-0.03310815244913101,
-0.11787354946136475,
-0.0011708593228831887,
-0.0013991391751915216,
-0.0806991308927536,
0.008510957472026348,
0.003659662092104554,
-0.05169852823019028,
-0.03991690278053284,
-0.06537200510501862,
0.05895709618926048,
0.007791342679411173,
0.05316853150725365,
-0.05635339021682739,
-0.004821193404495716,
0.00015434381202794611,
0.10021843016147614,
0.008555318228900433,
0.008461198769509792,
0.08550940454006195,
0.010700112208724022,
-0.0347212478518486,
0.07243654876947403,
0.1132017970085144,
0.029886789619922638,
0.04140138626098633,
-0.06190170347690582,
-0.03440277278423309,
0.04109136387705803,
-0.06215120851993561,
0.14798946678638458,
-0.009167417883872986,
-0.05511708930134773,
0.048341210931539536,
0.027277642861008644,
0.10864199697971344,
0.051349516957998276,
-0.005530487280339003,
0.04730935022234917,
-0.018289407715201378,
-0.03428177535533905,
-0.020136170089244843,
-0.09882498532533646,
0.064209945499897,
-0.11292232573032379,
-0.03343085199594498,
-0.05222512036561966,
0.018329402431845665,
0.04213770106434822,
-0.0043890345841646194,
0.024233317002654076,
-0.05025908350944519,
-0.028509482741355896,
0.02038436383008957,
0.1453346610069275,
0.03250771388411522,
0.014420529827475548,
0.02681019902229309,
-0.02344696596264839
] | 0 | 0 | 0 | 0 |
|
Rethinking the generalization of drug target affinity prediction algorithms via similarity aware evaluation | https://openreview.net/forum?id=j7cyANIAxV | [
"Chenbin Zhang",
"Zhiqiang Hu",
"Jiang Chuchu",
"Wen Chen",
"JIE XU",
"Shaoting Zhang"
] | Oral | Drug-target binding affinity prediction is a fundamental task for drug discovery. It has been extensively explored in literature and promising results are reported. However, in this paper, we demonstrate that the results may be misleading and cannot be well generalized to real practice. The core observation is that the canonical randomized split of a test set in conventional evaluation leaves the test set dominated by samples with high similarity to the training set. The performance of models is severely degraded on samples with lower similarity to the training set but the drawback is highly overlooked in current evaluation. As a result, the performance can hardly be trusted when the model meets low-similarity samples in real practice. To address this problem, we propose a framework of similarity aware evaluation in which a novel split methodology is proposed to adapt to any desired distribution. This is achieved by a formulation of optimization problems which are approximately and efficiently solved by gradient descent. We perform extensive experiments across five representative methods in four datasets for two typical target evaluations and compare them with various counterpart methods. Results demonstrate that the proposed split methodology can significantly better fit desired distributions and guide the development of models. | Drug-Target Affinity Prediction, Similarity-Aware Evaluation | null | 2,093 | null | [
-0.043632134795188904,
-0.08079209178686142,
-0.048799172043800354,
0.003317692084237933,
0.05154227092862129,
-0.026280326768755913,
0.02964816614985466,
0.04081828519701958,
0.022040367126464844,
-0.10011032223701477,
-0.02566424384713173,
-0.08037316054105759,
0.055627789348363876,
0.07072107493877411,
0.006156220566481352,
0.007583437021821737,
0.06657742708921432,
0.04447263851761818,
-0.06177249550819397,
-0.013171335682272911,
0.0038285204209387302,
0.0244057085365057,
0.02637941762804985,
0.08083441853523254,
0.013549397699534893,
0.017733607441186905,
0.030614513903856277,
-0.07826219499111176,
-0.05221946910023689,
0.021199604496359825,
0.12163548171520233,
0.011454800143837929,
-0.023855721578001976,
-0.04915330559015274,
-0.055547747761011124,
0.050505951046943665,
-0.09029015153646469,
0.027315644547343254,
0.025762971490621567,
-0.03312824293971062,
0.03373925760388374,
-0.030616050586104393,
0.015323430299758911,
0.06588852405548096,
0.0007823462365195155,
-0.0583280511200428,
-0.027825862169265747,
0.020117953419685364,
0.04719563201069832,
0.06556159257888794,
-0.03384049981832504,
-0.051987458020448685,
-0.06848490983247757,
0.059262465685606,
-0.0405239574611187,
-0.006600634660571814,
0.002042653039097786,
0.009249616414308548,
-0.04056842252612114,
0.03516720235347748,
-0.016535695642232895,
-0.1158185601234436,
-0.007457883562892675,
0.0042166817001998425,
0.013173412531614304,
-0.06331244856119156,
0.06326545774936676,
0.022853264585137367,
0.02729308046400547,
-0.04188172146677971,
-0.04659804701805115,
0.11975804716348648,
-0.060008659958839417,
0.07605352997779846,
0.02846006117761135,
0.08167075365781784,
0.03537612408399582,
0.04978296533226967,
-0.0182584710419178,
-0.06422978639602661,
0.022444909438490868,
-0.009059186093509197,
0.043947938829660416,
0.041316188871860504,
0.09805306047201157,
0.027217136695981026,
0.020931405946612358,
0.025874651968479156,
0.04017702117562294,
-0.008818755857646465,
0.003137771273031831,
0.02773614227771759,
-0.014832981862127781,
-0.022422900423407555,
0.0022653790656477213,
0.00396225368604064,
-0.030794290825724602,
-0.04265624284744263,
-0.03334854170680046,
0.06606490910053253,
-0.0492546483874321,
0.0670749619603157,
0.0186774842441082,
-0.04048636555671692,
0.03249351680278778,
-0.028152601793408394,
0.049149565398693085,
-0.03316614404320717,
0.05285385251045227,
-0.08346372097730637,
0.026029186323285103,
0.07853265106678009,
0.01717747375369072,
0.05516944080591202,
0.051044363528490067,
0.0019327508052811027,
0.05906303972005844,
-0.033709101378917694,
-0.043687328696250916,
0.09210997819900513,
-0.10610295832157135,
-0.0065689124166965485,
0.0009569592657499015,
-0.037892553955316544,
-0.014333871193230152,
0.011562266387045383,
-0.09264323115348816,
3.167429405034149e-33,
-0.025955449789762497,
0.029972735792398453,
0.08212574571371078,
-0.0041818805038928986,
-0.004537216853350401,
-0.06255991011857986,
-0.009754249826073647,
0.02636885643005371,
-0.06837589293718338,
0.023817861452698708,
-0.06898610293865204,
-0.014923320151865482,
-0.02821233682334423,
0.11516274511814117,
-0.0021097606513649225,
0.011623202823102474,
0.004330783616751432,
0.08288582414388657,
-0.0033491728827357292,
-0.044540487229824066,
-0.025132514536380768,
0.016068389639258385,
-0.07519549131393433,
-0.03851938620209694,
-0.028875935822725296,
0.013374212197959423,
-0.02903646230697632,
0.01708461344242096,
-0.006661942694336176,
-0.010102590546011925,
-0.06808660179376602,
0.06656060367822647,
0.036145832389593124,
0.04270213842391968,
0.035074204206466675,
-0.021296560764312744,
0.021080540493130684,
-0.028213543817400932,
-0.04760630056262016,
-0.010975543409585953,
-0.0077367546036839485,
0.0286280307918787,
0.07034530490636826,
-0.05568481609225273,
-0.03572062775492668,
-0.05739753693342209,
-0.08251772820949554,
0.011649797670543194,
0.027570653706789017,
-0.04951247572898865,
0.059067659080028534,
-0.04644247516989708,
0.025448136031627655,
0.059384096413850784,
-0.09276257455348969,
0.02362886443734169,
-0.09903690963983536,
0.016823384910821915,
0.08851800858974457,
0.06638576090335846,
-0.006437350995838642,
-0.0009083315962925553,
-0.0325000025331974,
0.03319435939192772,
-0.01272535603493452,
0.053375594317913055,
-0.022874988615512848,
-0.043294165283441544,
-0.016931360587477684,
0.13079170882701874,
0.03143155574798584,
0.03951237350702286,
-0.048277050256729126,
-0.04822821542620659,
0.0639234110713005,
0.0013074890011921525,
0.047348398715257645,
-0.01968483068048954,
0.0037392261438071728,
0.055193059146404266,
0.0014694554265588522,
0.030619345605373383,
-0.010486170649528503,
-0.042884908616542816,
-0.08267565816640854,
-0.11400283128023148,
0.0644049346446991,
-0.1340140402317047,
-0.11319625377655029,
-0.02265072800219059,
-0.08338036388158798,
0.007935844361782074,
-0.052726034075021744,
-0.018557103350758553,
0.09520494937896729,
-3.389252535874007e-33,
-0.02564900927245617,
-0.0013308298075571656,
0.059230737388134,
0.008658637292683125,
0.022403659299016,
-0.036728307604789734,
-0.042280081659555435,
-0.003960581496357918,
-0.007803982589393854,
-0.009006657637655735,
0.043490130454301834,
-0.04515229910612106,
-0.024647004902362823,
-0.017032666131854057,
-0.09181924909353256,
0.04768205061554909,
0.017258139327168465,
0.06799145042896271,
0.04129697382450104,
0.023150930181145668,
0.010143931955099106,
0.024388689547777176,
-0.06139305979013443,
0.032335586845874786,
-0.02563123032450676,
-0.015534797683358192,
0.07426652312278748,
0.010324619710445404,
-0.027606097981333733,
-0.027823470532894135,
-0.12333499640226364,
0.007573781069368124,
-0.06377974152565002,
-0.018739625811576843,
-0.09757727384567261,
0.07391617447137833,
0.042021527886390686,
-0.01578742265701294,
0.06614227592945099,
0.026897644624114037,
-0.04596181586384773,
0.06039903685450554,
-0.09925617277622223,
0.010494238696992397,
0.01300579123198986,
-0.06900011748075485,
0.022775446996092796,
0.02639871835708618,
0.08747629076242447,
-0.007544484920799732,
-0.0007548690191470087,
0.03581945225596428,
-0.03507252410054207,
-0.008219641633331776,
-0.022637758404016495,
-0.08822150528430939,
-0.04151191562414169,
0.009843628853559494,
0.041660185903310776,
-0.04012003913521767,
-0.03775207698345184,
0.009109767153859138,
-0.028377914801239967,
0.04650958999991417,
0.009946975857019424,
-0.06985525786876678,
0.04447281360626221,
0.04953012615442276,
0.03814791142940521,
0.019303906708955765,
-0.01987280882894993,
0.041486408561468124,
0.1232416033744812,
0.04822961241006851,
-0.014014416374266148,
-0.0840529277920723,
0.03152768686413765,
-0.16460584104061127,
-0.028163127601146698,
-0.03115117736160755,
0.0711093619465828,
-0.056158021092414856,
-0.009576339274644852,
-0.05498679727315903,
0.0385495163500309,
0.10805216431617737,
0.07893536239862442,
0.06989100575447083,
0.03205490857362747,
0.017061371356248856,
-0.008616549894213676,
-0.0002285683003719896,
0.013562841340899467,
0.05301763117313385,
-0.051866672933101654,
-4.731367653221241e-8,
0.06725069880485535,
0.044726286083459854,
-0.020084362477064133,
0.042037151753902435,
-0.04157663509249687,
0.10207127779722214,
-0.07270306348800659,
0.043938588351011276,
-0.0774829089641571,
-0.02681301347911358,
-0.008487786166369915,
0.04399450495839119,
-0.08785104751586914,
0.023941509425640106,
-0.03906835988163948,
0.0016197174554690719,
-0.027951210737228394,
0.06165246665477753,
-0.004260600544512272,
0.0634588748216629,
-0.035577934235334396,
0.05715833604335785,
0.07649698108434677,
-0.03659942373633385,
0.11282121390104294,
-0.11014958471059799,
-0.003185178153216839,
0.01409568078815937,
0.07923244684934616,
0.032828789204359055,
-0.03159397095441818,
-0.0458933562040329,
-0.0126322777941823,
-0.05958804115653038,
0.030423473566770554,
0.092679463326931,
-0.05280616134405136,
-0.0034356059040874243,
-0.01057566236704588,
0.02021844871342182,
0.000008121835890051443,
0.03041948936879635,
0.0075575425289571285,
-0.010699116624891758,
0.05194789543747902,
-0.015149127691984177,
0.06596718728542328,
0.04085835814476013,
0.03570909425616264,
-0.06014157459139824,
0.05492491275072098,
-0.021726004779338837,
0.027064770460128784,
-0.06328054517507553,
0.08989955484867096,
-0.02178829349577427,
-0.1103217750787735,
-0.028440406545996666,
0.09371387958526611,
-0.09205102175474167,
0.032249756157398224,
-0.055171139538288116,
-0.04446740448474884,
0.014734146185219288
] | 0 | 0 | 0 | 0 |
|
GridMix: Exploring Spatial Modulation for Neural Fields in PDE Modeling | https://openreview.net/forum?id=Fur0DtynPX | [
"Honghui Wang",
"Shiji Song",
"Gao Huang"
] | Oral | Significant advancements have been achieved in PDE modeling using neural fields. Despite their effectiveness, existing methods rely on global modulation, limiting their ability to reconstruct local details. While spatial modulation with vanilla grid-based representations offers a promising alternative, it struggles with inadequate global information modeling and over-fitting to the training spatial domain. To address these challenges, we propose GridMix, a novel approach that models spatial modulation as a mixture of grid-based representations. GridMix effectively explores global structures while preserving locality for fine-grained modulation. Furthermore, we introduce spatial domain augmentation to enhance the robustness of the modulated neural fields against spatial domain variations.
With all these innovations,
our comprehensive approach culminates in MARBLE, a framework that significantly advancing the capabilities of neural fields in PDE modeling. The effectiveness of MARBLE is extensively
validated on diverse benchmarks encompassing dynamics modeling and geometric prediction. | Partial Differential Equations, Neural Fields | null | 2,066 | null | [
-0.025807183235883713,
-0.0940018743276596,
0.10497207194566727,
-0.0016753824893385172,
0.005661472678184509,
-0.030624505132436752,
-0.06659318506717682,
-0.06055043637752533,
0.010897466912865639,
-0.03867979347705841,
-0.06798253208398819,
-0.032068368047475815,
-0.04052305966615677,
0.06134567782282829,
-0.022569051012396812,
-0.017144784331321716,
0.038264594972133636,
0.07598007470369339,
-0.1104804128408432,
0.01841035671532154,
0.005643839482218027,
-0.08358272165060043,
-0.014444764703512192,
0.0051016672514379025,
0.1031699851155281,
0.06696272641420364,
0.06000930070877075,
0.01572047919034958,
0.020265717059373856,
-0.007119004614651203,
0.025651196017861366,
0.02598637156188488,
-0.10766007006168365,
-0.02081107348203659,
-0.056421712040901184,
0.05633486062288284,
-0.08815658092498779,
-0.006928471382707357,
-0.014547075144946575,
-0.05896424502134323,
-0.029432350769639015,
0.007075194269418716,
0.04347047209739685,
0.056503236293792725,
0.017059843987226486,
-0.03670376166701317,
0.052231062203645706,
-0.06045949459075928,
0.04541531577706337,
-0.016521936282515526,
-0.047697797417640686,
-0.029922686517238617,
-0.021695440635085106,
0.022055737674236298,
-0.015283399261534214,
0.09254541248083115,
0.06513216346502304,
0.017984727397561073,
0.06846325099468231,
-0.025051720440387726,
-0.07379282265901566,
0.024554437026381493,
-0.002680089557543397,
0.0338226780295372,
-0.03245621174573898,
0.02291668765246868,
-0.035613738000392914,
0.057428330183029175,
0.012423302978277206,
-0.05591882765293121,
0.034356169402599335,
0.015225773677229881,
-0.04850379005074501,
-0.07669387012720108,
0.06094074621796608,
0.06520296633243561,
-0.03603554517030716,
-0.006119034718722105,
0.05767344310879707,
-0.04431851580739021,
0.0575721338391304,
0.048582904040813446,
0.02745620161294937,
-0.04458821564912796,
0.06683499366044998,
0.03443131223320961,
-0.0029900032095611095,
0.053448326885700226,
0.006246350705623627,
0.01573500968515873,
0.03271714597940445,
-0.03367221727967262,
-0.01219131425023079,
0.02844906412065029,
-0.09815271198749542,
-0.04726067930459976,
0.0030527322087436914,
-0.08464676886796951,
-0.007775592617690563,
0.035994235426187515,
-0.02010810561478138,
0.0013812812976539135,
0.06032618507742882,
0.03639434650540352,
0.060588207095861435,
-0.08095552772283554,
0.0698789656162262,
0.08721888065338135,
-0.02471720613539219,
-0.003209042828530073,
0.0481145866215229,
-0.032209500670433044,
-0.005772160831838846,
0.007571876980364323,
-0.01261877454817295,
0.0032664237078279257,
-0.004775574896484613,
-0.04584713280200958,
-0.01222090981900692,
-0.03648561239242554,
-0.054368309676647186,
-0.09005597233772278,
-0.16124330461025238,
0.04230066016316414,
-0.0030393917113542557,
0.003928567282855511,
-0.08046571910381317,
3.5912795956826117e-35,
0.011962415650486946,
0.021877381950616837,
0.010963293723762035,
0.05106645077466965,
0.08488843590021133,
-0.020855318754911423,
0.00742232846096158,
-0.029416168108582497,
0.009806952439248562,
0.08487935364246368,
-0.14066477119922638,
0.04902656376361847,
-0.057544879615306854,
0.11328417807817459,
0.08338771015405655,
-0.04778117686510086,
0.020832188427448273,
0.01838972233235836,
-0.08528468757867813,
0.033788226544857025,
0.03576454147696495,
0.033478744328022,
-0.03686659038066864,
-0.05178148299455643,
0.022217391058802605,
0.021844932809472084,
-0.07100766152143478,
-0.0009417851106263697,
-0.004686752799898386,
0.014815517701208591,
-0.09554681181907654,
0.02858119085431099,
-0.08882065117359161,
0.025953950360417366,
0.0365750826895237,
-0.0038313022814691067,
0.003977706655859947,
-0.054297078400850296,
-0.0091117974370718,
-0.01582225412130356,
0.009339741431176662,
0.02098012901842594,
-0.0010239366674795747,
0.003706462914124131,
-0.04327631741762161,
-0.046210553497076035,
0.04544663429260254,
0.06308307498693466,
-0.06279054284095764,
-0.04882016405463219,
0.05578465014696121,
0.053502850234508514,
0.00990222953259945,
0.02600806951522827,
0.006626912858337164,
0.0013015124714002013,
0.03234552964568138,
-0.031184017658233643,
0.07136419415473938,
0.017749659717082977,
0.07123786956071854,
-0.011167341843247414,
0.043745096772909164,
-0.017833668738603592,
0.04110821336507797,
-0.055355459451675415,
-0.023232001811265945,
0.0411740280687809,
-0.0038359747268259525,
0.023895785212516785,
-0.005628604907542467,
0.04901547729969025,
-0.00830200593918562,
-0.023299239575862885,
0.04281209781765938,
-0.07742443680763245,
0.027988219633698463,
-0.040526244789361954,
0.039671048521995544,
0.00898790918290615,
-0.05132810026407242,
0.050137389451265335,
-0.02821260318160057,
0.017602626234292984,
-0.04337018355727196,
0.05212049186229706,
0.0978887677192688,
-0.008226747624576092,
-0.13947038352489471,
-0.12455850094556808,
-0.06652364879846573,
-0.09406760334968567,
0.06593852490186691,
-0.03548663854598999,
-0.053927816450595856,
-2.1866883868622054e-33,
-0.042540501803159714,
-0.02780575305223465,
-0.009552898816764355,
-0.02768597938120365,
-0.0584787055850029,
0.02957095392048359,
-0.014958322048187256,
0.022525206208229065,
-0.02674822136759758,
-0.018011678010225296,
-0.040575481951236725,
0.006049586925655603,
0.05590193346142769,
-0.012437206692993641,
0.05338016152381897,
-0.030724307522177696,
-0.05843859538435936,
-0.006023217923939228,
-0.007065473590046167,
-0.00713489530608058,
0.012139041908085346,
0.032754115760326385,
-0.0012053475948050618,
0.09869127720594406,
-0.06597854942083359,
-0.021317658945918083,
-0.04399694874882698,
0.03583889827132225,
-0.03372535482048988,
0.036871545016765594,
-0.07252020388841629,
0.0661761462688446,
0.02399662882089615,
-0.03349463269114494,
-0.06805535405874252,
0.11767789721488953,
0.09893184900283813,
-0.08559765666723251,
-0.020759806036949158,
-0.0305973868817091,
0.037602879106998444,
-0.042135559022426605,
-0.002438486088067293,
0.13049766421318054,
-0.02298550121486187,
0.015520837157964706,
-0.009160731919109821,
0.04640571400523186,
-0.030712667852640152,
0.10086824744939804,
-0.010762983933091164,
0.0345953032374382,
-0.06507356464862823,
-0.03050009161233902,
-0.02834394946694374,
-0.036504652351140976,
-0.05662764981389046,
-0.046549052000045776,
0.08779066801071167,
-0.08296762406826019,
-0.08161957561969757,
-0.07611845433712006,
0.024807820096611977,
0.032025810331106186,
0.0026131996419280767,
0.04547183960676193,
-0.059986647218465805,
-0.0002133549569407478,
0.09267253428697586,
0.0003134793078061193,
0.012481218203902245,
0.031522784382104874,
-0.012654904276132584,
0.03670356795191765,
0.06952463835477829,
-0.08534923940896988,
0.03925222158432007,
0.008704266510903835,
0.02113126963376999,
-0.015328719280660152,
0.07884064316749573,
0.011376483365893364,
0.042548540979623795,
0.0765225812792778,
0.06961765140295029,
0.06004511937499046,
0.006711689289659262,
0.030743297189474106,
-0.05226307362318039,
-0.01905638352036476,
0.018150031566619873,
0.050295837223529816,
-0.0034364538732916117,
0.04097854718565941,
-0.008094596676528454,
-3.774335866069123e-8,
-0.03803706541657448,
0.0820901095867157,
0.025737915188074112,
-0.02316983975470066,
-0.025325799360871315,
-0.03594471141695976,
0.0387980155646801,
0.005123730283230543,
-0.04945021867752075,
-0.025229590013623238,
0.08808103203773499,
-0.05492033064365387,
-0.03735683858394623,
-0.04394838213920593,
-0.017427213490009308,
0.109219029545784,
0.014351221732795238,
0.06275195628404617,
-0.03402259200811386,
0.0830804705619812,
-0.009123923256993294,
0.0579858161509037,
0.02594115398824215,
-0.0008966629393398762,
0.09112928807735443,
-0.02208068035542965,
-0.10000720620155334,
0.027615558356046677,
0.05465896055102348,
-0.0015500346198678017,
0.024282895028591156,
-0.04346787557005882,
0.02213636040687561,
0.010336093604564667,
-0.09430316835641861,
0.08599886298179626,
0.045819513499736786,
-0.017804190516471863,
-0.0660744458436966,
-0.0695016160607338,
-0.05650359019637108,
-0.019006092101335526,
-0.037802472710609436,
-0.021391723304986954,
-0.026755111292004585,
0.0820489227771759,
0.10725045949220657,
-0.025907600298523903,
0.03570932522416115,
0.0664830282330513,
-0.07746655493974686,
0.04656258970499039,
0.004677151329815388,
-0.04319149628281593,
0.01708466000854969,
0.025626298040151596,
-0.04413141682744026,
0.010588372126221657,
0.108921118080616,
-0.0068998560309410095,
-0.048057831823825836,
0.06481695920228958,
-0.14385733008384705,
-0.028177209198474884
] | 0 | 0 | 0 | 0 |
|
Data Selection via Optimal Control for Language Models | https://openreview.net/forum?id=dhAL5fy8wS | [
"Yuxian Gu",
"Li Dong",
"Hongning Wang",
"Yaru Hao",
"Qingxiu Dong",
"Furu Wei",
"Minlie Huang"
] | Oral | This work investigates the selection of high-quality pre-training data from massive corpora to enhance LMs' capabilities for downstream usage.
We formulate data selection as a generalized Optimal Control problem, which can be solved theoretically by Pontryagin's Maximum Principle (PMP), yielding a set of necessary conditions that characterize the relationship between optimal data selection and LM training dynamics.
Based on these theoretical results, we introduce **P**MP-based **D**ata **S**election (**PDS**), a framework that approximates optimal data selection by solving the PMP conditions.
In our experiments, we adopt PDS to select data from CommmonCrawl and show that the PDS-selected corpus accelerates the learning of LMs and constantly boosts their performance on a wide range of downstream tasks across various model sizes.
Moreover, the benefits of PDS extend to ~400B models trained on ~10T tokens, as evidenced by the extrapolation of the test loss curves according to the Scaling Laws.
PDS also improves data utilization when the pre-training data is limited, by reducing the data demand by 1.8 times, which helps mitigate the quick exhaustion of available web-crawled corpora. Our code, model, and data can be found at https://github.com/microsoft/LMOps/tree/main/data_selection. | Pre-training Language Models, Data Selection, Optimal Control | This paper introduces a framework to select high-quality pre-training data via optimal control. | 2,015 | 2410.07064 | [
-0.03354058414697647,
-0.11188038438558578,
0.07563654333353043,
0.046280503273010254,
0.061592936515808105,
0.058406006544828415,
-0.06856103986501694,
0.05168928951025009,
0.05737888440489769,
0.058210596442222595,
-0.02999817207455635,
-0.017912644892930984,
-0.029172969982028008,
-0.013898041099309921,
0.03015136532485485,
0.02792615070939064,
0.11062294244766235,
0.019767582416534424,
-0.10178900510072708,
-0.047708477824926376,
0.0034187287092208862,
0.038843657821416855,
-0.0024449857883155346,
0.04100437089800835,
0.051920730620622635,
-0.01872292160987854,
-0.002138780429959297,
-0.03527966886758804,
0.0325201079249382,
-0.03201378509402275,
0.0021424456499516964,
0.03031090274453163,
0.0486651174724102,
-0.01765376143157482,
-0.07096719741821289,
0.012652588076889515,
-0.06059926375746727,
-0.0312074925750494,
0.0036742747761309147,
0.009113088250160217,
0.023553693667054176,
-0.03803844749927521,
0.027346568182110786,
0.0051188161596655846,
0.02543427236378193,
0.03799173980951309,
0.015264423564076424,
0.01881321705877781,
0.015820641070604324,
-0.012442431412637234,
-0.06682264059782028,
-0.01363711804151535,
6.855545962025644e-7,
0.024270707741379738,
-0.03157626464962959,
-0.036272916942834854,
0.059261854737997055,
0.023450052365660667,
0.04337005317211151,
-0.02874225191771984,
-0.1145515888929367,
-0.08152029663324356,
-0.05800134688615799,
-0.024537887424230576,
-0.02975253388285637,
-0.001431410782970488,
0.005361870862543583,
0.04935004189610481,
-0.02237756922841072,
0.07531405985355377,
-0.023618850857019424,
0.055543042719364166,
-0.054369308054447174,
0.06419899314641953,
0.006990108173340559,
0.01751803793013096,
0.016432682052254677,
-0.023021195083856583,
0.08157863467931747,
-0.06078120321035385,
0.0041639842092990875,
0.06577842682600021,
0.0479217991232872,
-0.04385072737932205,
0.07372516393661499,
0.02281528152525425,
0.014523682184517384,
0.04661348834633827,
0.11245056241750717,
-0.00582774356007576,
0.016041487455368042,
-0.015670951455831528,
0.0007878946489654481,
0.03704163432121277,
-0.08585914969444275,
0.04090685024857521,
-0.012166817672550678,
-0.04892538860440254,
0.052644677460193634,
0.08607034385204315,
0.0705660730600357,
0.07430525124073029,
0.008037597872316837,
0.020064854994416237,
-0.04318368062376976,
-0.09194044023752213,
0.04258803650736809,
0.05117582902312279,
0.08460094779729843,
-0.052375711500644684,
0.09076932072639465,
0.04672478139400482,
0.015694154426455498,
-0.016892507672309875,
0.007381184957921505,
-0.016073264181613922,
0.0026867471169680357,
-0.08306180685758591,
-0.043790191411972046,
0.08130214363336563,
-0.05005559325218201,
-0.015973394736647606,
0.03975046053528786,
0.0052840434946119785,
-0.00652142520993948,
-0.019130589440464973,
-0.09620325267314911,
4.652605263022172e-34,
0.047669291496276855,
-0.0049803550355136395,
0.010877946391701698,
-0.01618688553571701,
-0.008727630600333214,
-0.004596530459821224,
0.0109602902084589,
-0.0013962221564725041,
-0.029985997825860977,
0.014620800502598286,
-0.048300229012966156,
-0.017734970897436142,
-0.06587003171443939,
0.04116341844201088,
0.03259969502687454,
-0.017475469037890434,
-0.011614173650741577,
0.08216655999422073,
-0.016417743638157845,
-0.019121097400784492,
0.05041954293847084,
-0.016480308026075363,
0.0014634410617873073,
-0.07115520536899567,
0.0388132706284523,
0.026304898783564568,
0.04859713464975357,
-0.008002636022865772,
-0.04617289826273918,
0.019928086549043655,
-0.12329959869384766,
-0.08752412348985672,
0.008518466725945473,
0.04703228920698166,
-0.006993339397013187,
0.030708860605955124,
-0.0765351951122284,
-0.07312452793121338,
-0.01473656203597784,
-0.007254318799823523,
-0.03292102739214897,
0.053186215460300446,
0.031846463680267334,
0.047032710164785385,
-0.1299174726009369,
-0.040845125913619995,
-0.019932186231017113,
-0.023341763764619827,
-0.077200748026371,
-0.08804509788751602,
0.07016407698392868,
-0.003621836192905903,
-0.002821830566972494,
-0.01048098411411047,
0.04156620055437088,
0.0437694750726223,
-0.035707730799913406,
0.0013994191540405154,
0.004914735443890095,
0.07428022474050522,
0.015373692847788334,
0.0017776153981685638,
0.018473736941814423,
0.10732569545507431,
0.015937790274620056,
0.03845197334885597,
0.011987810954451561,
0.05644809454679489,
0.10428425669670105,
-0.059121742844581604,
-0.005937255918979645,
-0.08101575821638107,
0.03378202021121979,
-0.05671193078160286,
0.13473671674728394,
-0.04803097993135452,
0.1497587114572525,
-0.07905326783657074,
-0.019486645236611366,
0.041366226971149445,
-0.03962016478180885,
-0.03976980969309807,
0.03738662600517273,
-0.11262938380241394,
0.027990300208330154,
-0.12386569380760193,
0.03547434136271477,
-0.0817466527223587,
-0.035900358110666275,
-0.06639554351568222,
-0.08350646495819092,
-0.002209014492109418,
-0.03727323189377785,
0.0275490190833807,
-0.010146810673177242,
-1.5227104946357324e-33,
-0.02199815958738327,
-0.031106989830732346,
0.004661868792027235,
0.07805758714675903,
-0.09538490325212479,
-0.03370820730924606,
0.03423914685845375,
0.004264303483068943,
-0.09035928547382355,
-0.0734962671995163,
-0.012961522676050663,
-0.11321613192558289,
0.06295933574438095,
-0.0597347728908062,
0.04184854403138161,
0.01883811131119728,
-0.003570514963939786,
-0.035263195633888245,
0.04967498779296875,
0.04346656799316406,
-0.02873004414141178,
0.0810137689113617,
-0.04516572132706642,
0.07231757789850235,
-0.03630540892481804,
-0.04340454563498497,
-0.12055965512990952,
0.05032649263739586,
0.010848823934793472,
-0.02456209994852543,
-0.09105223417282104,
-0.01648266613483429,
-0.07000306248664856,
0.015745753422379494,
-0.07943663001060486,
0.07474088668823242,
0.05716808885335922,
0.07279972732067108,
-0.020158853381872177,
0.18718376755714417,
0.09507187455892563,
-0.04031378775835037,
-0.06903667002916336,
0.007854660041630268,
-0.02834360860288143,
-0.026469402015209198,
-0.1217409297823906,
-0.018763529136776924,
0.06006842851638794,
-0.01770992949604988,
-0.06895656883716583,
0.00676043052226305,
0.020316869020462036,
0.04240623489022255,
-0.0652795359492302,
-0.09431985020637512,
0.00846182368695736,
-0.03263985365629196,
-0.015075311064720154,
0.01433289423584938,
-0.03990488499403,
-0.012845499441027641,
0.007713961880654097,
0.00797977950423956,
-0.011554056778550148,
-0.025985710322856903,
0.020442992448806763,
-0.0038714015390723944,
0.03836686909198761,
-0.057180024683475494,
-0.013248244300484657,
-0.04880458116531372,
0.03267166391015053,
0.031730301678180695,
-0.03242689371109009,
-0.025722483173012733,
-0.0024572708643972874,
-0.06048879027366638,
0.0036915047094225883,
-0.022777928039431572,
0.027818812057375908,
0.11104541271924973,
0.03895843029022217,
0.04514592885971069,
0.027807341888546944,
0.07727872580289841,
0.047809701412916183,
0.06860782951116562,
0.0019871345721185207,
-0.012335563078522682,
0.0047194561921060085,
-0.017865505069494247,
0.05917957425117493,
0.03597208857536316,
-0.0648702010512352,
-4.9777529653738384e-8,
-0.047056760638952255,
0.021794810891151428,
0.051002468913793564,
0.002656203228980303,
0.03531423956155777,
-0.039068616926670074,
-0.06470759958028793,
0.08275879174470901,
0.014085396192967892,
0.015912814065814018,
0.06250173598527908,
0.015466651879251003,
-0.07145777344703674,
-0.05116230994462967,
0.03166666254401207,
0.04708768054842949,
-0.01913996785879135,
-0.020555106922984123,
0.014788736589252949,
-0.0004158062220085412,
0.04278600960969925,
0.051692068576812744,
0.03080097772181034,
-0.025261573493480682,
0.08071308583021164,
-0.04916264861822128,
-0.013548275455832481,
0.11184671521186829,
0.04313226044178009,
0.015945741906762123,
-0.0366620235145092,
-0.002098448807373643,
-0.07714658975601196,
0.02385818026959896,
0.09028100967407227,
0.04938611760735512,
-0.027406657114624977,
-0.05834547430276871,
-0.032362792640924454,
0.023926623165607452,
0.013531667180359364,
0.07971851527690887,
-0.0814230665564537,
-0.024788815528154373,
-0.014220737852156162,
0.011340257711708546,
-0.02734151855111122,
-0.07041799277067184,
0.07335714250802994,
-0.04440727084875107,
0.06489518284797668,
0.0019199420930817723,
-0.01096514891833067,
0.05592086538672447,
0.08650202304124832,
0.0025126703549176455,
-0.02587343379855156,
0.02266005612909794,
-0.014547260478138924,
-0.021576279774308205,
0.01922498270869255,
0.03327396512031555,
-0.04041349142789841,
-0.024638159200549126
] | https://github.com/microsoft/lmops | 3,960 | 0 | 0 | 0 |
Simplifying, Stabilizing and Scaling Continuous-time Consistency Models | https://openreview.net/forum?id=LyJi5ugyJx | [
"Cheng Lu",
"Yang Song"
] | Oral | Consistency models (CMs) are a powerful class of diffusion-based generative models optimized for fast sampling. Most existing CMs are trained using discretized timesteps, which introduce additional hyperparameters and are prone to discretization errors. While continuous-time formulations can mitigate these issues, their success has been limited by training instability. To address this, we propose a simplified theoretical framework that unifies previous parameterizations of diffusion models and CMs, identifying the root causes of instability. Based on this analysis, we introduce key improvements in diffusion process parameterization, network architecture, and training objectives. These changes enable us to train continuous-time CMs at an unprecedented scale, reaching 1.5B parameters on ImageNet 512×512. Our proposed training algorithm, using only two sampling steps, achieves FID scores of 2.06 on CIFAR-10, 1.48 on ImageNet 64×64, and 1.88 on ImageNet 512×512, narrowing the gap in FID scores with the best existing diffusion models to within 10\%. | continuous-time consistency models, diffusion models, fast sampling | 2-step continuous-time consistency models reduce the gap to within 10\% in sample quality (FID) compared to best diffusion models | 1,982 | 2410.11081 | [
-0.046855464577674866,
-0.08203115314245224,
0.04462417587637901,
0.02031245082616806,
0.04179231822490692,
-0.06542231887578964,
-0.05780302733182907,
-0.01928197592496872,
-0.015129304490983486,
-0.08139815181493759,
-0.025306405499577522,
-0.0005929242470301688,
-0.00444372184574604,
0.032974839210510254,
-0.0735672190785408,
-0.05186014994978905,
0.025992698967456818,
0.04992661997675896,
-0.09795404970645905,
-0.03234608471393585,
-0.05126534029841423,
0.05886511504650116,
-0.012630291283130646,
-0.021552307531237602,
-0.0076390583999454975,
0.03449802100658417,
0.0285692997276783,
-0.0761670246720314,
0.03465273603796959,
-0.07831297069787979,
0.07277703285217285,
-0.03004324808716774,
-0.1101662665605545,
0.02694987878203392,
0.019494187086820602,
0.011677028611302376,
0.013383382931351662,
-0.030601168051362038,
0.06277572363615036,
-0.04642920568585396,
0.05735250189900398,
-0.01677274890244007,
0.015895497053861618,
-0.003627246944233775,
0.05600842833518982,
-0.044432591646909714,
0.028830764815211296,
0.013994653709232807,
-0.02196059748530388,
-0.0032834578305482864,
-0.037187639623880386,
-0.03163009509444237,
-0.028534740209579468,
0.07970523089170456,
0.028427572920918465,
0.03166956827044487,
0.054156944155693054,
0.09770095348358154,
0.0037696838844567537,
-0.01869797892868519,
-0.060550104826688766,
-0.058145392686128616,
-0.07109512388706207,
0.028535690158605576,
-0.017113910987973213,
0.01909206621348858,
0.06307821720838547,
0.024896498769521713,
0.02865016460418701,
-0.0005417261854745448,
0.016765154898166656,
0.09769720584154129,
-0.12831762433052063,
0.026609187945723534,
-0.026094790548086166,
0.0721558928489685,
0.04318702593445778,
-0.019871659576892853,
0.006825559306889772,
-0.050929341465234756,
0.04915222153067589,
0.02447439916431904,
0.06068435311317444,
-0.0837184265255928,
0.11869683861732483,
-0.03597996383905411,
0.01378246396780014,
0.0778447836637497,
0.018946180120110512,
-0.04422472417354584,
-0.021160777658224106,
0.14298465847969055,
-0.04263663291931152,
0.0031269739847630262,
0.029543645679950714,
0.03654094040393829,
0.024285685271024704,
-0.055001795291900635,
0.010945050977170467,
0.1283300817012787,
-0.020637203007936478,
0.06587491929531097,
0.010448819026350975,
0.002194379922002554,
0.08939437568187714,
-0.08468441665172577,
0.09902611374855042,
0.07798431813716888,
0.036590080708265305,
-0.022882161661982536,
0.022296838462352753,
0.016801204532384872,
-0.04201606288552284,
0.004419319797307253,
0.0530516542494297,
-0.08789660781621933,
-0.034980107098817825,
-0.028974277898669243,
0.042656347155570984,
0.051578033715486526,
-0.054667457938194275,
0.04479898884892464,
-0.09111515432596207,
-0.028815653175115585,
0.02085091359913349,
-0.05364410579204559,
-0.033670954406261444,
3.454095375659257e-33,
-0.04340023547410965,
-0.07226712256669998,
0.01606888882815838,
0.01505010761320591,
0.013010533526539803,
0.019897008314728737,
0.022927239537239075,
-0.027142131701111794,
-0.000916196673642844,
-0.06598310172557831,
-0.04954587668180466,
-0.0920364111661911,
-0.1525290310382843,
0.0445576012134552,
0.06177986040711403,
-0.03836318850517273,
-0.0036397557705640793,
0.10702831298112869,
0.03947838395833969,
0.008973052725195885,
0.09659004956483841,
-0.0787171795964241,
-0.05386850982904434,
-0.020812617614865303,
0.028205454349517822,
0.0034195822663605213,
0.05548917502164841,
0.07463745772838593,
0.05194859206676483,
0.021284019574522972,
-0.09350946545600891,
0.02355376072227955,
0.017838863655924797,
-0.021012669429183006,
0.06754513829946518,
0.0037216604687273502,
-0.0279558002948761,
-0.021744798868894577,
-0.022864481434226036,
-0.049934931099414825,
-0.05572020262479782,
0.07628723978996277,
0.012613832019269466,
-0.0024441941641271114,
-0.09635376185178757,
-0.04105555638670921,
-0.012280352413654327,
-0.05088907107710838,
-0.038145340979099274,
-0.031201092526316643,
0.01599835231900215,
0.017992420122027397,
0.06190739944577217,
-0.06214001029729843,
0.0003338706155773252,
-0.04606468975543976,
0.07264621555805206,
-0.02777227759361267,
-0.03784811496734619,
0.033473819494247437,
0.08774195611476898,
-0.017216244712471962,
-0.060557473450899124,
0.057170018553733826,
0.07534319162368774,
0.02667410857975483,
-0.026578200981020927,
0.07862930744886398,
0.017381949350237846,
0.036505211144685745,
-0.05575917661190033,
-0.006328476592898369,
-0.05372602865099907,
-0.11447550356388092,
0.08563075214624405,
-0.12589257955551147,
0.038823556154966354,
-0.051064785569906235,
-0.04173842817544937,
0.06508125364780426,
-0.07549570500850677,
-0.027852723374962807,
-0.06285130232572556,
-0.06772632896900177,
-0.01085264328867197,
-0.050971340388059616,
-0.016442928463220596,
0.01303036417812109,
-0.06044680252671242,
-0.030228391289711,
-0.015667570754885674,
0.04415528103709221,
0.041117604821920395,
0.04392753541469574,
-0.04062441736459732,
-3.301883183565171e-33,
-0.033543746918439865,
0.05383222922682762,
0.003989661578088999,
0.13158239424228668,
-0.025341862812638283,
-0.02793864719569683,
-0.030672140419483185,
0.06358495354652405,
0.019329924136400223,
-0.1105273887515068,
-0.018341153860092163,
-0.032198335975408554,
-0.02802409417927265,
-0.052773259580135345,
-0.018998926505446434,
-0.002085426589474082,
0.009520246647298336,
-0.013301337137818336,
0.02653304487466812,
0.03948863223195076,
0.015813084319233894,
0.00922191422432661,
-0.05082080513238907,
-0.006597609259188175,
-0.025817623361945152,
0.01656535640358925,
-0.05770976468920708,
0.05187875032424927,
0.04302556440234184,
-0.02447659522294998,
-0.021034905686974525,
-0.03786284849047661,
0.029971614480018616,
-0.038718730211257935,
-0.018135294318199158,
0.06025801599025726,
0.04158468544483185,
-0.012202968820929527,
0.035597287118434906,
0.11360418796539307,
0.07469005137681961,
-0.015095751732587814,
-0.10424493998289108,
0.06539642810821533,
0.010233358480036259,
0.05435938015580177,
0.019550789147615433,
-0.022913774475455284,
-0.06789103150367737,
0.008183115161955357,
-0.01413783896714449,
0.05152716487646103,
-0.07374095916748047,
0.05470913276076317,
0.04970238357782364,
-0.02906429022550583,
-0.06732876598834991,
-0.011819272302091122,
0.023433834314346313,
-0.023329075425863266,
-0.00892478134483099,
-0.03275712579488754,
-0.018970439210534096,
-0.030192388221621513,
-0.012991530820727348,
-0.014016937464475632,
-0.008252263069152832,
-0.013818983919918537,
-0.02789292111992836,
0.12627793848514557,
-0.03033825010061264,
0.030360491946339607,
-0.02616872265934944,
0.04465271905064583,
-0.03773518651723862,
-0.1001792848110199,
0.07448522746562958,
-0.049292609095573425,
0.05185273289680481,
-0.04054805636405945,
0.030780991539359093,
0.08181101083755493,
-0.030113087967038155,
0.01009363029152155,
0.025913340970873833,
0.03594724461436272,
0.10051006078720093,
-0.018941454589366913,
0.048379283398389816,
0.07232615351676941,
-0.014399225823581219,
-0.005634321365505457,
0.08102759718894958,
0.038062941282987595,
-0.11945745348930359,
-5.112327983169962e-8,
-0.10763812065124512,
0.07034869492053986,
-0.012321877293288708,
0.07461413741111755,
0.038611505180597305,
-0.005924364551901817,
0.04198590666055679,
0.03267665207386017,
0.09838598966598511,
-0.04104537516832352,
0.06660206615924835,
-0.02611869014799595,
-0.03278353810310364,
-0.08753915131092072,
0.010479500517249107,
0.040161021053791046,
-0.025249755010008812,
-0.03388751298189163,
0.004751156084239483,
-0.02944820374250412,
-0.05410376191139221,
0.007199032697826624,
0.015029460191726685,
-0.006549519952386618,
0.05012153461575508,
0.013621636666357517,
0.015509705059230328,
0.08277950435876846,
0.07899930328130722,
0.01712394692003727,
0.001833612099289894,
0.0017012374009937048,
0.03492400422692299,
-0.005244194529950619,
-0.0582147017121315,
0.07990310341119766,
0.0023482083342969418,
-0.009275839664041996,
0.0008891040924936533,
-0.06280957162380219,
-0.000154332781676203,
0.08177393674850464,
-0.023428354412317276,
-0.004823840223252773,
0.010717579163610935,
0.0020054602064192295,
0.029362119734287262,
0.004023927729576826,
-0.0403377003967762,
-0.021042656153440475,
0.021453052759170532,
-0.015287895686924458,
-0.01683337427675724,
0.06610995531082153,
0.07067682594060898,
-0.05679658427834511,
0.0674089789390564,
-0.0818544551730156,
0.0292473454028368,
0.0401344858109951,
-0.007936789654195309,
-0.021324345842003822,
-0.07978862524032593,
0.06824521720409393
] | 0 | 0 | 0 | 0 |
|
Dynamic Multimodal Evaluation with Flexible Complexity by Vision-Language Bootstrapping | https://openreview.net/forum?id=X1OfiRYCLn | [
"Yue Yang",
"Shuibo Zhang",
"Kaipeng Zhang",
"Yi Bin",
"Yu Wang",
"Ping Luo",
"Wenqi Shao"
] | Oral | Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities across multimodal tasks such as visual perception and reasoning, leading to good performance on various multimodal evaluation benchmarks. However, these benchmarks keep a static nature and overlap with the pre-training data, resulting in fixed complexity constraints and data contamination issues. This raises the concern regarding the validity of the evaluation. To address these two challenges, we introduce a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB). VLB provides a robust and comprehensive assessment for LVLMs with reduced data contamination and flexible complexity. To this end, VLB dynamically generates new visual question-answering samples through a multimodal bootstrapping module that modifies both images and language, while ensuring that newly generated samples remain consistent with the original ones by a judge module. By composing various bootstrapping strategies, VLB offers dynamic variants of existing benchmarks with diverse complexities, enabling the evaluation to co-evolve with the ever-evolving capabilities of LVLMs. Extensive experimental results across multiple benchmarks, including SEEDBench, MMBench, and MME, show that VLB significantly reduces data contamination and exposes performance limitations of LVLMs. | Dynamic Evaluation, Vision-Language Bootstrapping, data contamination, Flexible Complexity, Large Vision-Language Model | We develop the first dynamic multimodal evaluation protocol with flexible complexity via Vision-Language Bootstrapping. | 1,837 | 2410.08695 | [
-0.03247664496302605,
-0.06680597364902496,
-0.014137337915599346,
0.055269915610551834,
0.17144925892353058,
-0.05605047941207886,
-0.05000748485326767,
0.022639142349362373,
0.007429178338497877,
-0.05714920163154602,
-0.05867478996515274,
-0.14117690920829773,
0.02829761803150177,
0.08671276271343231,
-0.048247214406728745,
0.001373795559629798,
0.022830963134765625,
0.0558638721704483,
-0.10210878401994705,
-0.018631400540471077,
0.03744419291615486,
-0.04073462635278702,
0.06459775567054749,
-0.050212908536195755,
0.052090421319007874,
0.03117438778281212,
-0.029635367915034294,
-0.05337601527571678,
0.06374743580818176,
-0.036246318370103836,
0.03202618286013603,
-0.003577358089387417,
-0.03871294856071472,
0.08118465542793274,
-0.003591682529076934,
0.08592672646045685,
-0.023974265903234482,
0.015569382347166538,
-0.04586150497198105,
-0.028248416259884834,
-0.04704325273633003,
-0.023877564817667007,
0.05729242041707039,
-0.031013434752821922,
0.07488001137971878,
-0.00927654281258583,
-0.014325142838060856,
-0.04244820401072502,
-0.010841403156518936,
0.025141358375549316,
-0.08205465227365494,
-0.051798515021800995,
0.016613854095339775,
0.007290068082511425,
-0.058161888271570206,
0.021508987993001938,
-0.01392554584890604,
-0.006938017904758453,
0.010979567654430866,
0.018154289573431015,
-0.010266980156302452,
-0.017699571326375008,
-0.009050869382917881,
0.044740092009305954,
-0.01125202514231205,
-0.034308645874261856,
0.06319373846054077,
-0.07154856622219086,
0.010173751972615719,
-0.006999528035521507,
-0.08085538446903229,
0.07182879000902176,
0.02010502479970455,
0.0812205970287323,
-0.0023779990151524544,
0.04634307324886322,
0.030339166522026062,
-0.021773474290966988,
0.06735499948263168,
-0.10444159805774689,
0.059633225202560425,
-0.013323832303285599,
0.028338931500911713,
0.0038702767342329025,
0.14133122563362122,
0.013775036670267582,
-0.02027888037264347,
0.05664855241775513,
0.002117339288815856,
0.010493270121514797,
-0.10357485711574554,
-0.03152697533369064,
0.015303045511245728,
0.0019169208826497197,
0.059048496186733246,
0.012469032779335976,
0.0364849679172039,
-0.06890327483415604,
-0.05246758833527565,
0.10401888191699982,
0.02251063659787178,
0.06880982965230942,
-0.060068320482969284,
-0.043903984129428864,
0.04917564615607262,
-0.02552744746208191,
0.09409008920192719,
0.007046776823699474,
-0.0052056629210710526,
-0.0565267838537693,
0.07809509336948395,
0.057325124740600586,
0.04226964712142944,
-0.0036468524485826492,
0.013144858181476593,
-0.011018033139407635,
-0.01074040774255991,
0.029260756447911263,
0.052422791719436646,
-0.012299503199756145,
0.023001736029982567,
-0.005677841138094664,
0.018755009397864342,
-0.08371035009622574,
0.09455470740795135,
-0.04512768238782883,
0.00404568575322628,
1.2330392992343502e-33,
0.0745159164071083,
-0.010233460925519466,
0.06742969900369644,
0.029703011736273766,
0.06790731847286224,
-0.017490524798631668,
-0.0059922379441559315,
0.07342585176229477,
-0.07930967956781387,
-0.0009806104935705662,
0.011965884827077389,
0.003712220350280404,
-0.08095850795507431,
0.08628393709659576,
0.0407123863697052,
0.004009336233139038,
-0.06920355558395386,
0.015134157612919807,
-0.05877227708697319,
-0.010803534649312496,
0.04652462154626846,
-0.037015996873378754,
0.02045530080795288,
0.019290506839752197,
0.03607134893536568,
0.002010158495977521,
0.10295332968235016,
-0.09037686139345169,
-0.01840490661561489,
0.013803448528051376,
-0.1400582194328308,
-0.0411451980471611,
0.0021085820626467466,
0.011087988503277302,
0.061341941356658936,
-0.05008027330040932,
-0.02722708135843277,
-0.04419206455349922,
-0.003842699807137251,
-0.009002558887004852,
-0.027521513402462006,
-0.026647917926311493,
0.015935981646180153,
-0.016188958659768105,
-0.09961313754320145,
0.03326508030295372,
-0.00599683215841651,
0.06958300620317459,
-0.08367165923118591,
-0.07228740304708481,
0.005688325967639685,
0.004705914296209812,
0.0004646684101317078,
-0.05571933463215828,
-0.015379277057945728,
-0.046146973967552185,
0.023657672107219696,
0.037858378142118454,
0.05010189861059189,
-0.014603407122194767,
-0.05318743363022804,
0.08620866388082504,
-0.022148365154862404,
0.12104477733373642,
0.056263361126184464,
0.0078478017821908,
-0.0007949057617224753,
0.03399737924337387,
0.06493484228849411,
0.00782077107578516,
0.005297317635267973,
0.014884328469634056,
0.016759278252720833,
-0.06887190043926239,
0.035554200410842896,
0.026750938966870308,
0.041976023465394974,
-0.03977867588400841,
0.03990000858902931,
0.03733066841959953,
0.009816281497478485,
-0.04511471837759018,
0.038176730275154114,
-0.10175423324108124,
-0.005115610081702471,
-0.060143958777189255,
0.06896495074033737,
-0.12249723821878433,
-0.001490047317929566,
-0.04244646057486534,
-0.06902985274791718,
-0.008243505842983723,
0.013576040975749493,
-0.12855619192123413,
0.02728375978767872,
-6.855041784436542e-34,
0.048632536083459854,
0.03556813672184944,
-0.0530400313436985,
0.07386486977338791,
-0.042510733008384705,
-0.05473746731877327,
0.024415897205471992,
0.01590179279446602,
-0.008028285577893257,
-0.10011918842792511,
-0.022235525771975517,
-0.026959534734487534,
0.0023216954432427883,
0.019689561799168587,
0.02193831093609333,
0.022068241611123085,
-0.02582685463130474,
-0.002272701356559992,
0.0029133360367268324,
0.043692972511053085,
-0.020906943827867508,
0.10599350929260254,
-0.04363374784588814,
0.01333085261285305,
-0.10105035454034805,
0.06833597272634506,
-0.03081240877509117,
0.02704601176083088,
-0.08585254848003387,
-0.023190120235085487,
-0.0037369998171925545,
0.0035142062697559595,
-0.002054781187325716,
-0.06764393299818039,
-0.04708901420235634,
0.02184530720114708,
0.051479995250701904,
-0.09650985151529312,
-0.010305848903954029,
0.11453106254339218,
0.08630866557359695,
-0.050759635865688324,
-0.11922299861907959,
-0.054177843034267426,
-0.004256242886185646,
-0.056860025972127914,
-0.08770957589149475,
-0.05326785147190094,
0.012051091529428959,
0.05874893069267273,
-0.04793240875005722,
-0.008484628982841969,
-0.10905145853757858,
-0.005434283521026373,
-0.0019538968335837126,
-0.08610434085130692,
-0.10989021509885788,
0.008308622054755688,
0.05638622120022774,
0.003004526486620307,
-0.022046806290745735,
0.01013227365911007,
-0.0254181120544672,
-0.005631649401038885,
0.05696558207273483,
-0.004567923955619335,
-0.025857552886009216,
-0.07298322021961212,
0.026179539039731026,
0.003108373610302806,
-0.0217282697558403,
-0.03580399975180626,
0.01423622015863657,
0.050118088722229004,
0.05527046322822571,
0.03324585035443306,
0.06518173962831497,
0.016481535509228706,
0.0180953498929739,
-0.07333136349916458,
0.0669940859079361,
-0.06634321063756943,
0.081479512155056,
0.03588882461190224,
0.004190799780189991,
0.09273383766412735,
0.030708475038409233,
0.10389416664838791,
0.0497656986117363,
0.08332541584968567,
0.010274107567965984,
0.006374987307935953,
0.04784373939037323,
0.06239473447203636,
0.01925819180905819,
-3.728439423866803e-8,
-0.05307363346219063,
0.07653742283582687,
-0.02279975451529026,
0.0015736573841422796,
0.0037700955290347338,
-0.051995135843753815,
-0.0320507287979126,
0.07374262809753418,
0.03746730461716652,
-0.040513601154088974,
0.04659859463572502,
0.005243638996034861,
-0.009517727419734001,
-0.03749080002307892,
0.00014685986388940364,
0.09023109078407288,
-0.062067776918411255,
0.01555700320750475,
-0.029711343348026276,
-0.012032688595354557,
0.03738988935947418,
-0.0013458926696330309,
-0.008402280509471893,
0.042821817100048065,
-0.016185956075787544,
-0.016844840720295906,
-0.053925126791000366,
0.02924809418618679,
0.040498677641153336,
-0.05490914732217789,
0.023176807910203934,
0.05419458821415901,
0.034642864018678665,
-0.04891476407647133,
0.016008706763386726,
0.05123094841837883,
-0.03392375633120537,
-0.009331232868134975,
0.03667958080768585,
-0.011719728820025921,
0.023976940661668777,
0.09765037149190903,
-0.05642585828900337,
0.03947918117046356,
-0.010355610400438309,
-0.017486929893493652,
-0.05747110769152641,
-0.08030135929584503,
-0.030240517109632492,
-0.06069321185350418,
-0.042340099811553955,
-0.05760430544614792,
-0.011793467216193676,
0.0848366767168045,
0.06529365479946136,
-0.049808721989393234,
0.016990888863801956,
0.05002736300230026,
0.05360952392220497,
0.11097228527069092,
0.015057886950671673,
0.006050656549632549,
0.014338691718876362,
-0.024303974583745003
] | https://github.com/yangyue5114/DME | 4 | 0 | 0 | 0 |
Mind the Gap: Examining the Self-Improvement Capabilities of Large Language Models | https://openreview.net/forum?id=mtJSMcF3ek | [
"Yuda Song",
"Hanlin Zhang",
"Carson Eisenach",
"Sham M. Kakade",
"Dean Foster",
"Udaya Ghai"
] | Oral | Self-improvement is a mechanism in Large Language Model (LLM) pre-training, post-training and test-time inference. We explore a framework where the model verifies its own outputs, filters or reweights data based on this verification, and distills the filtered data. Despite several empirical successes, a fundamental understanding is still lacking. In this work, we initiate a comprehensive, modular and controlled study on LLM self-improvement. We provide a mathematical formulation for self-improvement, which is largely governed by a quantity which we formalize as the **generation-verification gap**. Through experiments with various model families and tasks, we discover a scaling phenomenon of self-improvement -- a variant of the generation-verification gap scales monotonically with the model pre-training flops. We also examine when self-improvement is possible, an iterative self-improvement procedure, and ways to improve its performance. Our findings not only advance understanding of LLM self-improvement with practical implications, but also open numerous avenues for future research into its capabilities and boundaries. | LLM, self-improvement, synthetic data, post-training, test-time optimization | We conduct a comprehensive examination on LLM self-improvement capability via the generation-verification gap. | 1,706 | 2412.02674 | [
-0.020280631259083748,
-0.027560031041502953,
0.020985351875424385,
0.05015300214290619,
0.031174257397651672,
0.00559353968128562,
-0.025326257571578026,
-0.0035131422337144613,
0.043742094188928604,
-0.05125589668750763,
-0.006267482880502939,
-0.03180236741900444,
-0.0020270473323762417,
0.025619782507419586,
0.02235632948577404,
-0.03727763891220093,
0.05192921310663223,
0.03321917727589607,
-0.10659083724021912,
-0.12628412246704102,
-0.007489646784961224,
0.018442483618855476,
0.05536920949816704,
0.049519263207912445,
0.053079839795827866,
-0.053570184856653214,
-0.07244948297739029,
-0.03344958275556564,
0.06837255507707596,
-0.008851280435919762,
0.07111376523971558,
0.015604308806359768,
0.05512417480349541,
0.046961598098278046,
-0.02216668426990509,
0.01882234588265419,
0.0035442316439002752,
0.046284571290016174,
0.022275280207395554,
-0.014427989721298218,
-0.011753460392355919,
-0.07834965735673904,
0.04323970153927803,
-0.011277777142822742,
0.05195578187704086,
0.011325765401124954,
0.008389153517782688,
0.007917017675936222,
-0.09758194535970688,
-0.002206922974437475,
-0.08656422048807144,
0.001449054223485291,
-0.030825935304164886,
-0.01690569519996643,
-0.008906476199626923,
0.04280688613653183,
0.07617029547691345,
0.09724345058202744,
-0.04624731466174126,
-0.04304683580994606,
-0.016341695562005043,
-0.09419775754213333,
-0.06820941716432571,
0.031126802787184715,
0.008655807003378868,
0.06155150383710861,
-0.002729997504502535,
0.03620191290974617,
-0.0262526273727417,
0.08174701035022736,
-0.03923781216144562,
0.05369087681174278,
0.0070557729341089725,
0.08636806160211563,
0.054161760956048965,
0.03136551380157471,
0.004683959297835827,
0.01727246679365635,
0.0820031464099884,
-0.03432638943195343,
-0.04269285127520561,
0.01630152389407158,
0.02760566771030426,
0.015957282856106758,
0.01079484447836876,
0.012675890699028969,
0.04446890950202942,
-0.0417785681784153,
0.023666756227612495,
-0.01977614127099514,
-0.014329049736261368,
-0.055518634617328644,
-0.07986199110746384,
0.01296464167535305,
-0.013631531037390232,
0.031212057918310165,
0.039750970900058746,
-0.07841704040765762,
0.025315528735518456,
0.07678080350160599,
0.06521771103143692,
0.09210144728422165,
0.008563063107430935,
-0.05042773112654686,
-0.00748833455145359,
-0.012482009828090668,
0.07831086963415146,
0.09174495935440063,
-0.013282572850584984,
-0.06251934170722961,
0.052038129419088364,
0.03216472640633583,
0.0009406448807567358,
-0.002270727651193738,
0.017626414075493813,
0.04310063272714615,
-0.0010447162203490734,
-0.01345641165971756,
0.04741519317030907,
0.04479309916496277,
-0.0015710035804659128,
0.020604535937309265,
0.0620640330016613,
-0.02637753076851368,
-0.03981846198439598,
-0.02239088900387287,
-0.047974973917007446,
-2.0030434758271318e-33,
0.07839377969503403,
0.10402023792266846,
0.023407669737935066,
0.024350518360733986,
-0.019319530576467514,
-0.007620592135936022,
0.014666528441011906,
-0.006741647608578205,
0.001327834790572524,
-0.0015727233840152621,
-0.0042478484101593494,
0.0036065001040697098,
-0.038919441401958466,
0.03882432356476784,
0.07726269215345383,
-0.0036565905902534723,
-0.04649537429213524,
-0.008415855467319489,
-0.018247507512569427,
-0.011222761124372482,
0.11761671304702759,
-0.023950528353452682,
0.03253370523452759,
-0.11847780644893646,
0.03564608097076416,
0.05402260646224022,
0.06541784107685089,
-0.003908121958374977,
-0.11409378796815872,
0.014641514979302883,
-0.08309166878461838,
0.015117025934159756,
-0.014847178012132645,
0.054466165602207184,
0.02000664919614792,
-0.04858826473355293,
0.026438724249601364,
-0.0610806904733181,
0.04363719001412392,
-0.03185826539993286,
-0.03591010719537735,
0.03088150918483734,
0.03558095544576645,
-0.006660956423729658,
-0.07336784899234772,
-0.10023956745862961,
-0.02955242432653904,
0.011808927170932293,
-0.05631394684314728,
-0.05053434520959854,
0.07521513849496841,
0.05504017695784569,
-0.06389377266168594,
-0.06316309422254562,
-0.03407502919435501,
0.031468454748392105,
0.009037389419972897,
0.07083968818187714,
0.06299791485071182,
0.1014142632484436,
0.06313721090555191,
0.01239702943712473,
-0.049684133380651474,
0.11317289620637894,
0.05216499790549278,
0.043998897075653076,
0.035706691443920135,
0.039739325642585754,
0.13272768259048462,
-0.047265589237213135,
-0.028418267145752907,
-0.08025800436735153,
-0.0868227630853653,
-0.054482538253068924,
0.1325545459985733,
-0.07661435008049011,
0.05419974401593208,
-0.09539661556482315,
-0.008889978751540184,
-0.01400022767484188,
0.07612413167953491,
0.009304982610046864,
0.02435017004609108,
-0.138896182179451,
0.01890398934483528,
-0.104592464864254,
0.00706071313470602,
-0.04560854285955429,
0.0223760437220335,
-0.036603815853595734,
-0.051789455115795135,
0.005376032553613186,
0.024953359737992287,
0.016397733241319656,
0.017977723851799965,
1.0773561845629734e-33,
-0.02458086796104908,
-0.01333161536604166,
-0.033094145357608795,
0.14013473689556122,
0.011720431968569756,
-0.09166943281888962,
0.08001063019037247,
0.09929315000772476,
-0.045652322471141815,
-0.059946414083242416,
-0.08477125316858292,
-0.1017875149846077,
0.08725988864898682,
0.03009755164384842,
-0.037518881261348724,
-0.04847201332449913,
-0.02509297803044319,
-0.01607455499470234,
0.05003391578793526,
0.03806362673640251,
-0.01712130196392536,
0.10736828297376633,
-0.022055530920624733,
0.05133845657110214,
-0.02998405322432518,
0.017375698313117027,
-0.08162660151720047,
0.04233616963028908,
-0.000012676438927883282,
-0.013109534978866577,
-0.015639036893844604,
0.013722349889576435,
-0.029371371492743492,
-0.003807919332757592,
0.0061528305523097515,
0.037556666880846024,
0.02355124056339264,
-0.033973630517721176,
-0.015437642112374306,
0.07298708707094193,
0.054062604904174805,
-0.06801982969045639,
-0.044389426708221436,
-0.024733688682317734,
-0.04017690569162369,
-0.05242234468460083,
-0.08766396343708038,
-0.013498236425220966,
0.022489581257104874,
-0.00044591998448595405,
-0.04348563775420189,
0.02481740526854992,
-0.04164789617061615,
0.04149693623185158,
-0.03889128193259239,
-0.08276284486055374,
0.013779081404209137,
-0.08510135859251022,
0.019143806770443916,
0.019867992028594017,
-0.11603150516748428,
-0.02984433062374592,
0.081771120429039,
-0.02601662650704384,
0.05644426494836807,
-0.06747794896364212,
-0.04931649938225746,
0.02271481417119503,
0.08147358149290085,
0.005879730451852083,
0.006433325819671154,
-0.061886996030807495,
-0.05297570675611496,
-0.0018529393710196018,
-0.05793846771121025,
-0.04366692155599594,
0.03724374622106552,
-0.05019020661711693,
-0.026578037068247795,
-0.024218127131462097,
-0.055293142795562744,
0.05423814430832863,
-0.011883484199643135,
0.04230426251888275,
0.01707886904478073,
0.02431020326912403,
0.038932789117097855,
0.1056893840432167,
-0.030422532930970192,
0.049971770495176315,
0.005598235875368118,
0.03282029181718826,
-0.06023253872990608,
0.04331343621015549,
-0.059478435665369034,
-4.583528223633948e-8,
-0.07254087924957275,
-0.02217523381114006,
0.008186310529708862,
0.15413208305835724,
0.047533098608255386,
-0.0161221195012331,
-0.026371851563453674,
0.09168503433465958,
-0.01727980375289917,
-0.05454057827591896,
0.04283996671438217,
0.03131534531712532,
-0.03557517006993294,
-0.0723385438323021,
0.009958273731172085,
0.02233305759727955,
-0.0222625732421875,
0.07351236045360565,
-0.033899545669555664,
-0.07791690528392792,
0.035343095660209656,
0.03479723259806633,
0.020700978115200996,
-0.05167068541049957,
0.017523057758808136,
-0.023674463853240013,
-0.038788214325904846,
0.03788404539227486,
0.0009111075196415186,
-0.07946381717920303,
0.005495324265211821,
-0.03251943364739418,
-0.01801646314561367,
-0.030053090304136276,
0.07287757843732834,
0.09598441421985626,
0.015162589028477669,
0.040561724454164505,
0.027258528396487236,
-0.0621764101088047,
0.007046844344586134,
0.053923819214105606,
-0.06830184906721115,
0.06468147784471512,
-0.001700992463156581,
-0.0671943798661232,
-0.07927010953426361,
-0.05735692009329796,
-0.044881779700517654,
-0.08774178475141525,
0.0563834048807621,
-0.017562376335263252,
-0.029225977137684822,
0.00019826881180051714,
0.025715259835124016,
0.04799257591366768,
-0.014971362426877022,
-0.014857352711260319,
-0.014738421887159348,
0.06911404430866241,
0.05876285955309868,
0.018562063574790955,
-0.0007556868949905038,
-0.022460438311100006
] | 0 | 0 | 0 | 0 |
|
SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers | https://openreview.net/forum?id=N8Oj1XhtYZ | [
"Enze Xie",
"Junsong Chen",
"Junyu Chen",
"Han Cai",
"Haotian Tang",
"Yujun Lin",
"Zhekai Zhang",
"Muyang Li",
"Ligeng Zhu",
"Yao Lu",
"Song Han"
] | Oral | We introduce Sana, a text-to-image framework that can efficiently generate images up to 4096$\times$4096 resolution. Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed, deployable on laptop GPU. Core designs include: (1) Deep compression autoencoder: unlike traditional AEs, which compress images only 8$\times$, we trained an AE that can compress images 32$\times$, effectively reducing the number of latent tokens. (2) Linear DiT: we replace all vanilla attention in DiT with linear attention, which is more efficient at high resolutions without sacrificing quality. (3) Decoder-only text encoder: we replaced T5 with modern decoder-only small LLM as the text encoder and designed complex human instruction with in-context learning to enhance the image-text alignment. (4) Efficient training and sampling: we propose Flow-DPM-Solver to reduce sampling steps, with efficient caption labeling and selection to accelerate convergence. As a result, Sana-0.6B is very competitive with modern giant diffusion model (e.g. Flux-12B), being 20 times smaller and 100+ times faster in measured throughput. Moreover, Sana-0.6B can be deployed on a 16GB laptop GPU, taking less than 1 second to generate a 1024$\times$1024 resolution image. Sana enables content creation at low cost. Code and model will be publicly released upon publication. | Efficient AI, Diffusion Models, Text to Image generation | Sana can synthesize high-resolution, high-quality images with strong text-image alignment at a remarkably fast speed. | 1,682 | null | [
-0.04485930874943733,
0.002568504773080349,
0.03892553970217705,
-0.026704560965299606,
0.04030961915850639,
-0.059771399945020676,
-0.041877564042806625,
0.02087998017668724,
0.06889510899782181,
0.009531729854643345,
-0.0028040423057973385,
-0.03864996135234833,
-0.02523234859108925,
0.0587199367582798,
0.009253161028027534,
0.016432706266641617,
0.012135474011301994,
0.09805229306221008,
-0.08297961205244064,
-0.05726730450987816,
-0.00020585588936228305,
-0.030771266669034958,
0.052180346101522446,
-0.03842981532216072,
0.04948534443974495,
0.04534054920077324,
0.011257339268922806,
-0.06106773763895035,
0.038066983222961426,
-0.06825332343578339,
0.058709628880023956,
0.04714307188987732,
0.03290795534849167,
0.04086839035153389,
-0.06287446618080139,
0.10640382766723633,
-0.014883973635733128,
-0.039209574460983276,
0.003285684622824192,
-0.06614742428064346,
0.04081403464078903,
0.05786675959825516,
0.007019898388534784,
0.044874269515275955,
0.11698410660028458,
-0.03284919261932373,
0.021585997194051743,
-0.03199795261025429,
0.05669422075152397,
-0.026460636407136917,
-0.02115972526371479,
0.0347462072968483,
-0.048411231487989426,
0.030228184536099434,
-0.06463297456502914,
0.009124867618083954,
0.039029598236083984,
-0.04533853381872177,
0.044726934283971786,
-0.006817929446697235,
-0.015576988458633423,
-0.01699835993349552,
0.04815712571144104,
-0.0016576393973082304,
-0.03506690263748169,
-0.03998807072639465,
0.12726891040802002,
-0.007756020408123732,
0.006208932492882013,
-0.011638058349490166,
-0.0354575477540493,
0.013078799471259117,
-0.01901811547577381,
0.051086872816085815,
-0.03306199237704277,
0.032639794051647186,
0.02950301207602024,
-0.0674511268734932,
0.07666537910699844,
0.015246974304318428,
0.044411078095436096,
0.004440319258719683,
0.08323590457439423,
0.024699149653315544,
0.03320322185754776,
-0.03941363841295242,
0.012701204046607018,
0.03477280214428902,
-0.006728806532919407,
0.0028690823819488287,
-0.0582675039768219,
-0.06362877786159515,
0.024167945608496666,
-0.01872817985713482,
0.01690801978111267,
0.005894670728594065,
-0.03357834368944168,
-0.07027950882911682,
-0.0731145590543747,
0.09062999486923218,
0.024692533537745476,
0.003139510517939925,
-0.01847764290869236,
-0.006811804138123989,
-0.011984585784375668,
-0.05005399510264397,
0.14021460711956024,
0.05831082537770271,
-0.017529549077153206,
-0.02183441072702408,
0.056741390377283096,
-0.0005075842491351068,
-0.0097309909760952,
0.002840822096914053,
0.07887347042560577,
-0.05878439545631409,
-0.04087565839290619,
-0.02502298355102539,
0.025565413758158684,
0.007664209697395563,
-0.06663497537374496,
0.02998310886323452,
-0.15104080736637115,
-0.06590048968791962,
-0.0010269097983837128,
-0.06391977518796921,
-0.026423588395118713,
3.567470336428128e-33,
0.028986027464270592,
0.06714396923780441,
0.06734160333871841,
0.046680793166160583,
0.0843309536576271,
0.014534170739352703,
0.03540262579917908,
0.0006559789180755615,
-0.03476753458380699,
-0.005304005462676287,
-0.031777288764715195,
-0.08164865523576736,
-0.06874604523181915,
0.15802116692066193,
0.014298957772552967,
-0.08630776405334473,
0.00411304272711277,
0.020730232819914818,
-0.015005908906459808,
0.021013258025050163,
0.002423489699140191,
0.01731826178729534,
0.027760613709688187,
-0.06734096258878708,
-0.00215367553755641,
-0.01778910495340824,
0.03326403722167015,
-0.06916575878858566,
0.035835232585668564,
-0.02500876598060131,
-0.19450505077838898,
-0.04940878972411156,
0.03727753460407257,
0.04798755422234535,
0.009375255554914474,
-0.05480312556028366,
0.021740073338150978,
-0.02401416003704071,
-0.011102976277470589,
-0.009519643150269985,
0.05766380950808525,
0.1211349293589592,
0.0020139494445174932,
-0.07933630049228668,
-0.089228056371212,
0.04984012246131897,
0.052945930510759354,
0.029112784191966057,
-0.004929667804390192,
-0.008956046774983406,
0.03510388359427452,
0.052631668746471405,
-0.09144571423530579,
-0.06101282313466072,
0.052706945687532425,
-0.06880086660385132,
0.07845168560743332,
0.017321182414889336,
0.05714355781674385,
0.10331737250089645,
0.017548901960253716,
-0.03554737567901611,
-0.06490614265203476,
0.1135881319642067,
0.05820711329579353,
0.009562501683831215,
0.013599843718111515,
0.019561313092708588,
-0.04678664356470108,
-0.007941706106066704,
-0.05612694472074509,
-0.01587323658168316,
0.027230706065893173,
-0.02939527854323387,
0.03620654344558716,
0.03977467492222786,
0.05144023150205612,
-0.04078160971403122,
-0.0019152001477777958,
0.015060675330460072,
-0.031049614772200584,
-0.003686846699565649,
-0.01512592751532793,
-0.007138032466173172,
-0.04648846015334129,
-0.01030761655420065,
-0.006168820429593325,
-0.09258588403463364,
0.005220682360231876,
-0.03954577073454857,
-0.010830950923264027,
-0.048448335379362106,
0.015608105808496475,
-0.005475434008985758,
-0.0360548309981823,
-7.713701377626355e-34,
0.06800621747970581,
0.02552603743970394,
-0.021965883672237396,
0.05969790369272232,
-0.0237739160656929,
0.01941000670194626,
0.0506737120449543,
0.012887243181467056,
0.07970292866230011,
-0.06355387717485428,
-0.05742703750729561,
-0.02592514641582966,
0.046126484870910645,
-0.08379141241312027,
0.05228988453745842,
-0.014678838662803173,
0.07670679688453674,
-0.01237664744257927,
-0.04243671894073486,
-0.053654301911592484,
0.014346711337566376,
0.12718786299228668,
0.010644515976309776,
0.0076234787702560425,
-0.002646389650180936,
0.04071690887212753,
-0.04490120708942413,
0.07271113991737366,
0.034049756824970245,
-0.043519146740436554,
-0.009659318253397942,
-0.027134599164128304,
0.061543889343738556,
-0.009033018723130226,
-0.06895019114017487,
0.05176723748445511,
0.0647045150399208,
-0.0010513818124309182,
-0.07716350257396698,
0.06583993136882782,
0.11602205038070679,
-0.061658233404159546,
-0.002180103212594986,
0.08407438546419144,
-0.03568625822663307,
0.007920859381556511,
-0.1040283590555191,
-0.11271791905164719,
0.0835302323102951,
0.11740994453430176,
-0.01735272817313671,
0.030206244438886642,
-0.053539078682661057,
-0.013049779459834099,
-0.03267709165811539,
-0.11157573014497757,
-0.08042627573013306,
0.012129266746342182,
0.037804778665304184,
-0.030159173533320427,
0.02642999403178692,
-0.003426267532631755,
0.024990135803818703,
-0.0979856327176094,
-0.023666050285100937,
-0.05350511893630028,
-0.045247238129377365,
0.0036976903211325407,
-0.00121203507296741,
0.03272199258208275,
0.02005714178085327,
-0.11111655086278915,
-0.01371191255748272,
0.06752592325210571,
-0.012666971422731876,
-0.04167911410331726,
0.01906893029808998,
-0.03592618182301521,
0.03723752871155739,
-0.06414532661437988,
0.005267025902867317,
0.0328507274389267,
-0.036369483917951584,
0.0748295858502388,
0.07318186014890671,
0.055268388241529465,
0.051050301641225815,
-0.02975277788937092,
0.04698435217142105,
0.11265652626752853,
-0.054530151188373566,
0.053925175219774246,
0.05333060398697853,
0.0647762268781662,
-0.029746105894446373,
-4.260817121348737e-8,
-0.052117571234703064,
-0.021734610199928284,
-0.07437803596258163,
0.024251263588666916,
-0.05719567835330963,
-0.10458513349294662,
0.03301700949668884,
0.05635425075888634,
0.04339556768536568,
-0.030589185655117035,
0.1436196267604828,
-0.03365960344672203,
-0.07083749026060104,
-0.03977694362401962,
-0.009706425480544567,
0.0735396295785904,
0.028290633112192154,
-0.021806960925459862,
0.03740999475121498,
-0.05580319091677666,
-0.04308909922838211,
0.05774950608611107,
0.008471163921058178,
-0.045166824012994766,
0.009949586354196072,
-0.02727246657013893,
-0.038832079619169235,
0.0849013552069664,
0.06894011050462723,
-0.022455265745520592,
0.014779926277697086,
0.01483636349439621,
0.01883408986032009,
-0.06957156211137772,
-0.008982730098068714,
-0.03040592558681965,
0.028793886303901672,
-0.0012529095401987433,
0.01842997781932354,
-0.0414486825466156,
0.032724808901548386,
0.05490686744451523,
-0.03291261941194534,
-0.02947860397398472,
-0.03948340192437172,
-0.0038921451196074486,
-0.02306927554309368,
-0.06540996581315994,
0.023224998265504837,
-0.03895504027605057,
0.04017585515975952,
0.00861220434308052,
0.009836395271122456,
0.059431180357933044,
0.031808316707611084,
-0.008063114248216152,
0.0447072871029377,
-0.0167938731610775,
0.07938370853662491,
0.0851183533668518,
0.02835994027554989,
-0.023556819185614586,
-0.000184681688551791,
-0.028208814561367035
] | 0 | 0 | 0 | 0 |
|
Can a MISL Fly? Analysis and Ingredients for Mutual Information Skill Learning | https://openreview.net/forum?id=xoIeVdFO7U | [
"Chongyi Zheng",
"Jens Tuyls",
"Joanne Peng",
"Benjamin Eysenbach"
] | Oral | Self-supervised learning has the potential of lifting several of the key challenges in reinforcement learning today, such as exploration, representation learning, and reward design. Recent work (METRA) has effectively argued that moving away from mutual information and instead optimizing a certain Wasserstein distance is important for good performance. In this paper, we argue that the benefits seen in that paper can largely be explained within the existing framework of mutual information skill learning (MISL).
Our analysis suggests a new MISL method (contrastive successor features) that retains the excellent performance of METRA with fewer moving parts, and highlights connections between skill learning, contrastive representation learning, and successor features. Finally, through careful ablation studies, we provide further insight into some of the key ingredients for both our method and METRA. | unsupervised learning, reinforcement learning, mutual information, successor feature | Through careful analysis of a prior method, we develop a new method called Contrastive Successor Features (CSF) that illustrates mutual information skill learning can be made highly effective. | 1,383 | 2412.08021 | [
-0.02917289175093174,
-0.04524433985352516,
0.05940844118595123,
0.03763046860694885,
-0.01472858339548111,
0.01373104378581047,
0.06519081443548203,
-0.05108792707324028,
0.03738586977124214,
-0.022839536890387535,
-0.048859041184186935,
-0.025704005733132362,
0.04237084463238716,
0.11435309052467346,
-0.048177726566791534,
0.06742693483829498,
0.06360121071338654,
-0.025781910866498947,
-0.06505054980516434,
-0.03578117489814758,
0.021509185433387756,
0.02078603208065033,
-0.006392257288098335,
0.03765000402927399,
-0.0025130605790764093,
0.029035592451691628,
-0.01685832068324089,
0.01410080585628748,
0.04504473879933357,
-0.09050970524549484,
0.08567359298467636,
-0.029784144833683968,
0.02122471109032631,
-0.0652158260345459,
-0.07600170373916626,
0.1312493234872818,
-0.037528492510318756,
0.06262855976819992,
-0.0024365794379264116,
-0.045753516256809235,
-0.019455716013908386,
0.061359599232673645,
0.04635613411664963,
0.060801953077316284,
0.018391340970993042,
0.08353091031312943,
-0.035234395414590836,
-0.02309439890086651,
0.009227324277162552,
0.011630363762378693,
-0.09064695239067078,
0.00416057463735342,
0.019393611699342728,
-0.012469420209527016,
0.06463862210512161,
0.10862289369106293,
-0.0010344113688915968,
0.04508813098073006,
-0.042894840240478516,
-0.02447618544101715,
0.02179022692143917,
-0.058110762387514114,
-0.006678675301373005,
-0.04529345780611038,
-0.01651424914598465,
-0.02912101149559021,
0.03130946680903435,
0.046379897743463516,
-0.001665921532548964,
-0.0067846872843801975,
-0.007020319811999798,
0.01142694242298603,
-0.05717609077692032,
-0.014712766744196415,
0.06372790038585663,
0.056191299110651016,
0.03388934209942818,
-0.0011680889874696732,
0.04082944989204407,
-0.021326139569282532,
-0.06967813521623611,
0.015205909498035908,
-0.03917492553591728,
-0.019596317782998085,
0.10122936218976974,
-0.049840036779642105,
-0.039320800453424454,
-0.03341778740286827,
0.06647028774023056,
-0.04047141969203949,
0.04560144618153572,
-0.0613672249019146,
-0.04771824926137924,
-0.009603604674339294,
-0.03720066323876381,
0.005803423933684826,
0.013657409697771072,
0.017744965851306915,
-0.07688836008310318,
0.07201561331748962,
-0.016997233033180237,
0.005647030659019947,
-0.08913819491863251,
0.0453399196267128,
-0.044214699417352676,
-0.09330884367227554,
0.06900724023580551,
-0.009727522730827332,
0.07470453530550003,
-0.07452079653739929,
0.004916472360491753,
-0.027690518647432327,
-0.025740385055541992,
0.054167572408914566,
-0.01986672170460224,
0.041397690773010254,
-0.0073886713944375515,
0.006924963556230068,
-0.000723390665370971,
0.03356559947133064,
-0.0002656120923347771,
-0.09106377512216568,
0.0510614775121212,
0.06586293131113052,
-0.013856923207640648,
-0.10625738650560379,
0.007004030980169773,
1.968780570565547e-33,
0.032130613923072815,
0.008117720484733582,
-0.014593156054615974,
0.06158928573131561,
-0.058343306183815,
-0.03189215064048767,
-0.00916129071265459,
-0.006112771574407816,
-0.01762254163622856,
-0.0004014903970528394,
-0.003531030146405101,
0.04606672003865242,
0.0026316193398088217,
0.06458252668380737,
0.03496812656521797,
-0.032716304063797,
-0.06478635966777802,
-0.03888719901442528,
0.04283422604203224,
-0.13538865745067596,
0.023566531017422676,
0.024458926171064377,
-0.006725299172103405,
-0.11641239374876022,
0.03459640219807625,
0.026191864162683487,
-0.005671867169439793,
-0.061307717114686966,
0.015873974189162254,
-0.015692858025431633,
-0.06237638741731644,
0.02509072981774807,
-0.04672963544726372,
0.03810381144285202,
0.017867468297481537,
0.08617082983255386,
-0.01827903836965561,
-0.0436343215405941,
0.004804091062396765,
0.009258590638637543,
0.0379299558699131,
0.0022482958156615496,
-0.004225036595016718,
-0.05415838584303856,
-0.04296964779496193,
-0.01680571399629116,
0.029979418963193893,
-0.0340753011405468,
-0.04382528364658356,
-0.04955677315592766,
-0.03604239597916603,
-0.034508414566516876,
0.02636474184691906,
-0.08946472406387329,
0.004758630879223347,
0.07987117022275925,
0.009525952860713005,
0.09527978301048279,
0.009597853757441044,
0.069028839468956,
0.021508488804101944,
0.03372525796294212,
-0.011506631970405579,
0.09597217291593552,
-0.005228972993791103,
0.09980595856904984,
-0.013628502376377583,
0.03744620829820633,
0.17725269496440887,
-0.04950222373008728,
-0.023945501074194908,
0.05196832865476608,
0.021758511662483215,
-0.08292854577302933,
0.04939791187644005,
-0.054276563227176666,
0.04929400980472565,
-0.044018663465976715,
0.05111932381987572,
-0.010411259718239307,
0.04722166806459427,
-0.012928089126944542,
-0.0376039557158947,
0.019424494355916977,
-0.05424773693084717,
-0.05060023069381714,
0.05154375731945038,
-0.08453863859176636,
-0.02519184909760952,
0.07577825337648392,
-0.037408456206321716,
-0.0015130611136555672,
0.025975601747632027,
0.04691581055521965,
0.04887999966740608,
-2.8899265965446927e-33,
-0.06075577810406685,
-0.036698322743177414,
0.04700084403157234,
0.005214223638176918,
0.013312825001776218,
0.012561438605189323,
0.06957463175058365,
-0.04292439669370651,
0.02394549734890461,
0.020092450082302094,
-0.012372314929962158,
0.03635450825095177,
-0.015193757601082325,
-0.017778070643544197,
-0.04573742672801018,
0.06438589841127396,
-0.040020450949668884,
0.012167626991868019,
0.07154904305934906,
-0.04187798872590065,
0.06051772087812424,
0.1592414528131485,
-0.09965576231479645,
0.05757373571395874,
0.00701427785679698,
0.03460090234875679,
0.039243947714567184,
0.04037385433912277,
-0.00875164195895195,
0.05766173452138901,
-0.006025792099535465,
-0.06920117139816284,
-0.005503604654222727,
-0.0006575350998900831,
-0.05545254051685333,
0.05740377679467201,
0.027180591598153114,
0.012744361534714699,
-0.0822257399559021,
0.07125702500343323,
-0.018452733755111694,
0.04950954020023346,
-0.0537913553416729,
-0.03873230889439583,
0.02349276840686798,
-0.05892370641231537,
-0.08814537525177002,
0.04245470464229584,
0.029293112456798553,
-0.0031704178545624018,
0.01874302513897419,
-0.0253666453063488,
-0.12741613388061523,
-0.05100336670875549,
-0.05904441326856613,
-0.050001055002212524,
0.06992679834365845,
0.0035482635721564293,
0.08227815479040146,
0.02414589561522007,
-0.07819254696369171,
-0.010379704646766186,
-0.04294486716389656,
0.04561099410057068,
0.011285492219030857,
-0.007549121510237455,
-0.053050704300403595,
0.0339667834341526,
0.041672106832265854,
-0.022258372977375984,
0.0026926007121801376,
-0.02183050848543644,
0.02826899290084839,
-0.057298578321933746,
0.03292039781808853,
-0.04472377896308899,
-0.0009463127935305238,
0.006571464706212282,
-0.04185917228460312,
-0.0007014386937953532,
-0.0319543331861496,
-0.016498364508152008,
0.04088740795850754,
0.07313335686922073,
0.07217002660036087,
0.11196573823690414,
-0.0014525725273415446,
-0.008052685298025608,
0.00891304761171341,
-0.11804243922233582,
-0.00041997781954705715,
0.03064500354230404,
0.06380746513605118,
-0.02336260862648487,
-0.07559540122747421,
-4.9242647293112896e-8,
-0.06446770578622818,
-0.016567520797252655,
0.10151265561580658,
-0.08043117076158524,
0.08210273087024689,
-0.02868516929447651,
0.004372741095721722,
0.015614456497132778,
-0.06035842373967171,
-0.020004170015454292,
0.06484585255384445,
0.022117391228675842,
0.021218642592430115,
-0.057887621223926544,
0.02351200021803379,
0.02403203956782818,
0.054151926189661026,
-0.04321246221661568,
-0.001598120667040348,
0.002345040440559387,
0.055660560727119446,
-0.05986196920275688,
-0.009418288245797157,
-0.12821698188781738,
-0.020556798204779625,
-0.10867053270339966,
-0.03238915279507637,
0.04310154542326927,
0.042711351066827774,
0.02497679553925991,
0.0005213969270698726,
0.0052248588763177395,
0.12930600345134735,
0.01486009731888771,
0.01546289399266243,
0.07259947806596756,
0.049410380423069,
-0.04119180515408516,
-0.11596079170703888,
0.11282900720834732,
-0.06901590526103973,
0.07467523217201233,
-0.005050843581557274,
0.03267644718289375,
-0.00004639440885512158,
0.0387091301381588,
-0.034297965466976166,
-0.12394469976425171,
0.03507746756076813,
-0.028037751093506813,
0.11046518385410309,
-0.005897064693272114,
-0.011600065045058727,
0.03166689723730087,
0.1288580447435379,
-0.030007196590304375,
-0.005278954282402992,
-0.03067690134048462,
-0.0041696964763104916,
0.03751841187477112,
0.04298972710967064,
0.031302567571401596,
-0.06563457101583481,
0.00991842895746231
] | https://github.com/Princeton-RL/contrastive-successor-features | 8 | 0 | 0 | 0 |
When Selection Meets Intervention: Additional Complexities in Causal Discovery | https://openreview.net/forum?id=xByvdb3DCm | [
"Haoyue Dai",
"Ignavier Ng",
"Jianle Sun",
"Zeyu Tang",
"Gongxu Luo",
"Xinshuai Dong",
"Peter Spirtes",
"Kun Zhang"
] | Oral | We address the common yet often-overlooked selection bias in interventional studies, where subjects are selectively enrolled into experiments. For instance, participants in a drug trial are usually patients of the relevant disease; A/B tests on mobile applications target existing users only, and gene perturbation studies typically focus on specific cell types, such as cancer cells. Ignoring this bias leads to incorrect causal discovery results. Even when recognized, the existing paradigm for interventional causal discovery still fails to address it. This is because subtle differences in _when_ and _where_ interventions happen can lead to significantly different statistical patterns. We capture this dynamic by introducing a graphical model that explicitly accounts for both the observed world (where interventions are applied) and the counterfactual world (where selection occurs while interventions have not been applied). We characterize the Markov property of the model, and propose a provably sound algorithm to identify causal relations as well as selection mechanisms up to the equivalence class, from data with soft interventions and unknown targets. Through synthetic and real-world experiments, we demonstrate that our algorithm effectively identifies true causal relations despite the presence of selection bias. | causal discovery, selection bias, experiments, interventions | null | 1,361 | 2503.07302 | [
-0.06619355827569962,
-0.012559035792946815,
0.021583598107099533,
0.03127691522240639,
0.10483676195144653,
0.011403881944715977,
-0.008386347442865372,
-0.01149010844528675,
0.054981790482997894,
0.023833123967051506,
-0.0752408429980278,
-0.0462322011590004,
0.05690644681453705,
0.044535256922245026,
-0.05159712955355644,
0.003566443920135498,
0.09980448335409164,
-0.016012998297810555,
-0.035774338990449905,
0.005422105547040701,
-0.03176020085811615,
0.04227082058787346,
0.11972058564424515,
0.0370749868452549,
-0.07081207633018494,
-0.03864428400993347,
0.0378725565969944,
0.02959069050848484,
-0.00505766412243247,
0.034995630383491516,
0.03754663094878197,
0.04390989989042282,
-0.10834050178527832,
-0.04572802409529686,
-0.02175552025437355,
0.0020285255741328,
0.028345536440610886,
0.10283946245908737,
-0.0247203316539526,
-0.006850663106888533,
0.042152103036642075,
-0.04396486282348633,
0.0139279430732131,
0.10586418956518173,
0.07839200645685196,
-0.07813321799039841,
-0.03058698959648609,
0.027534350752830505,
-0.03977462649345398,
0.006749942433089018,
-0.05850068852305412,
0.006211427506059408,
0.013698307797312737,
0.04145485907793045,
0.012790086679160595,
0.01722269505262375,
0.048168763518333435,
0.0038210724014788866,
0.010982741601765156,
0.013868392445147038,
0.006506761070340872,
-0.0816391259431839,
-0.07201378792524338,
-0.028014162555336952,
-0.04714522883296013,
0.07267088443040848,
-0.026886656880378723,
0.055250171571969986,
-0.041653599590063095,
0.017108110710978508,
-0.003978238441050053,
0.06356891244649887,
-0.0671696588397026,
0.05919985845685005,
0.06113765016198158,
0.0020076720975339413,
0.014355410821735859,
0.05566664785146713,
0.034989457577466965,
-0.0948733389377594,
0.026296013966202736,
-0.03559279069304466,
0.038482118397951126,
0.06440626084804535,
0.0744275227189064,
-0.02532709203660488,
0.05009032040834427,
0.02696978859603405,
0.057307105511426926,
0.0795891135931015,
0.011861946433782578,
0.0061421701684594154,
0.02452464960515499,
0.036766353994607925,
-0.024601662531495094,
0.02329477295279503,
0.018604518845677376,
-0.04238147661089897,
0.05508515611290932,
0.07384543865919113,
0.012251332402229309,
-0.031042542308568954,
-0.0017064184648916125,
0.008671393617987633,
-0.008544056676328182,
-0.044764768332242966,
-0.060594428330659866,
-0.014383588917553425,
0.04821378365159035,
-0.027625951915979385,
-0.01875297538936138,
-0.01689296029508114,
0.07079803943634033,
0.03368030861020088,
-0.01441377867013216,
-0.018013887107372284,
0.0317842960357666,
-0.0146211888641119,
-0.012706978246569633,
0.02993764914572239,
0.03292951360344887,
-0.014994186349213123,
0.0565047487616539,
-0.010502814315259457,
0.010204211808741093,
0.026240676641464233,
-0.13789622485637665,
1.1389256858409158e-34,
0.07679818570613861,
-0.058693766593933105,
0.09074515104293823,
0.040113665163517,
-0.034417521208524704,
-0.059556275606155396,
-0.05326106771826744,
-0.056976187974214554,
0.010133016854524612,
-0.025101305916905403,
-0.043771881610155106,
0.027590317651629448,
-0.014156908728182316,
0.07458892464637756,
0.059701401740312576,
0.05716497078537941,
-0.0758727416396141,
0.09711534529924393,
-0.04627751186490059,
-0.010613973252475262,
-0.051792699843645096,
-0.045684296637773514,
-0.050771743059158325,
-0.033766549080610275,
-0.0188447218388319,
0.03419036418199539,
-0.012278733775019646,
0.03642958402633667,
0.025103677064180374,
0.03242337703704834,
-0.06261669844388962,
-0.018274938687682152,
0.09039689600467682,
-0.0228432547301054,
-0.016808202490210533,
-0.006827245000749826,
-0.07617998123168945,
-0.07412052899599075,
-0.02846118062734604,
-0.009954343549907207,
-0.03216538205742836,
-0.03235641121864319,
0.040138836950063705,
0.014218590222299099,
-0.04033271223306656,
-0.13058660924434662,
-0.055588651448488235,
-0.02732025645673275,
-0.015278778038918972,
-0.02389535866677761,
0.055931806564331055,
0.03937904164195061,
0.03706891089677811,
-0.07965221256017685,
-0.07609400153160095,
0.02438882365822792,
-0.08515700697898865,
0.01582169160246849,
-0.06280794739723206,
0.02979879081249237,
0.05020713061094284,
-0.0063539776019752026,
-0.028269318863749504,
0.08925704658031464,
-0.04570203274488449,
0.05124066770076752,
-0.016787700355052948,
-0.11316380649805069,
-0.02825579047203064,
0.052626851946115494,
-0.030725499615073204,
0.053564831614494324,
-0.008846181444823742,
-0.06261160969734192,
0.03531414642930031,
-0.039534762501716614,
0.03183763101696968,
0.014762334525585175,
-0.0498700812458992,
0.025510678067803383,
-0.002584998030215502,
-0.12538132071495056,
-0.06076296791434288,
0.03502191603183746,
-0.07954110205173492,
-0.059236180037260056,
-0.033957187086343765,
-0.060482196509838104,
-0.07362349331378937,
-0.07614319771528244,
-0.04593118652701378,
-0.0010791616514325142,
-0.02988768368959427,
0.057090695947408676,
0.05243081599473953,
-2.6018660120749293e-33,
-0.029629748314619064,
-0.014965127222239971,
0.15976665914058685,
-0.07070240378379822,
0.05860593169927597,
-0.010816691443324089,
0.03593994677066803,
-0.14110709726810455,
0.0932360589504242,
-0.050314776599407196,
0.03564271330833435,
-0.019558649510145187,
0.0653347447514534,
0.020560242235660553,
-0.027936430647969246,
0.023013336583971977,
-0.0046409498900175095,
0.012450365349650383,
-0.03156191110610962,
-0.007809937000274658,
0.07324863970279694,
0.05220308154821396,
0.005200481973588467,
-0.019095325842499733,
-0.062424566596746445,
-0.043010514229536057,
0.0277642123401165,
0.06705145537853241,
-0.017390139400959015,
-0.04509246349334717,
-0.09350184351205826,
0.017748726531863213,
-0.04248599708080292,
0.008285426534712315,
-0.03647345304489136,
0.09628535807132721,
0.09027494490146637,
-0.0035734218545258045,
-0.05686432495713234,
-0.032656241208314896,
0.02850690484046936,
0.053089484572410583,
-0.12078408151865005,
0.040669284760951996,
0.02361852489411831,
-0.007841136306524277,
-0.026198428124189377,
0.0963558629155159,
0.06457770615816116,
-0.015909984707832336,
-0.05657868832349777,
0.07035879790782928,
0.047233760356903076,
0.04579615965485573,
-0.059254925698041916,
-0.08669636398553848,
0.03754076361656189,
-0.08830452710390091,
-0.013487201184034348,
0.05911053717136383,
-0.14380888640880585,
-0.0029375937301665545,
-0.03788284584879875,
0.03814108669757843,
-0.01663747802376747,
-0.002282668836414814,
0.06545795500278473,
-0.02617298811674118,
0.044060006737709045,
-0.03583727777004242,
0.05776965618133545,
0.07507079839706421,
-0.024545641615986824,
-0.06544727832078934,
0.03231910988688469,
-0.001017928821966052,
-0.03300086408853531,
-0.02103482186794281,
-0.09094377607107162,
-0.020374231040477753,
0.04958060756325722,
-0.016003767028450966,
0.06367503851652145,
-0.07594144344329834,
-0.050125524401664734,
0.056412070989608765,
0.04203413426876068,
0.04559636488556862,
0.02135133370757103,
-0.008374221622943878,
-0.10279466956853867,
-0.03357037901878357,
-0.04729815199971199,
0.006532661151140928,
-0.11054234206676483,
-4.467914038741583e-8,
0.05555654317140579,
0.02945968508720398,
0.03699491545557976,
0.020702214911580086,
0.05107468366622925,
0.011730227619409561,
-0.13823269307613373,
0.08951816707849503,
-0.023568525910377502,
-0.018181731924414635,
0.013199041597545147,
0.08290668576955795,
0.014018782414495945,
-0.012653063051402569,
-0.010613879188895226,
0.026757434010505676,
0.0044077676720917225,
0.012139161117374897,
-0.01600112020969391,
0.009391143918037415,
0.028576398268342018,
0.004360673949122429,
0.038056060671806335,
0.04420130327343941,
0.1060701310634613,
-0.04390844702720642,
-0.03854617103934288,
0.016412027180194855,
0.07436645030975342,
0.014124740846455097,
0.042304959148168564,
-0.06132717803120613,
-0.009328383952379227,
-0.01835499331355095,
-0.021937064826488495,
0.014507077634334564,
0.030994189903140068,
-0.0931033343076706,
-0.03476761281490326,
-0.04971105232834816,
0.04018428921699524,
0.09436935186386108,
-0.01853764057159424,
0.030617324635386467,
0.0331960953772068,
-0.04536987096071243,
-0.060735587030649185,
-0.06664171069860458,
0.04026300087571144,
-0.023933714255690575,
0.0025058931205421686,
-0.05472966656088829,
0.023426441475749016,
-0.039421748369932175,
0.05532359704375267,
0.05304483696818352,
-0.012793980538845062,
-0.027849579229950905,
0.042320068925619125,
-0.023555468767881393,
0.08924484997987747,
0.05824335664510727,
0.0001207866589538753,
-0.010058913379907608
] | https://github.com/MarkDana/CDIS | 0 | 0 | 0 | 0 |
LVSM: A Large View Synthesis Model with Minimal 3D Inductive Bias | https://openreview.net/forum?id=QQBPWtvtcn | [
"Haian Jin",
"Hanwen Jiang",
"Hao Tan",
"Kai Zhang",
"Sai Bi",
"Tianyuan Zhang",
"Fujun Luan",
"Noah Snavely",
"Zexiang Xu"
] | Oral | We propose the Large View Synthesis Model (LVSM), a novel transformer-based approach for scalable and generalizable novel view synthesis from sparse-view inputs. We introduce two architectures: (1) an encoder-decoder LVSM, which encodes input image tokens into a fixed number of 1D latent tokens, functioning as a fully learned scene representation, and decodes novel-view images from them; and (2) a decoder-only LVSM, which directly maps input images to novel-view outputs, completely eliminating intermediate scene representations. Both models bypass the 3D inductive biases used in previous methods---from 3D representations (e.g., NeRF, 3DGS) to network designs (e.g., epipolar projections, plane sweeps)---addressing novel view synthesis with a fully data-driven approach. While the encoder-decoder model offers faster inference due to its independent latent representation, the decoder-only LVSM achieves superior quality, scalability, and zero-shot generalization, outperforming previous state-of-the-art methods by 1.5 to 3.5 dB PSNR. Comprehensive evaluations across multiple datasets demonstrate that both LVSM variants achieve state-of-the-art novel view synthesis quality, delivering superior performance even with reduced computational resources (1-2 GPUs). Please see our anonymous website for more details: https://haian-jin.github.io/projects/LVSM/ | novel view synthesis, transformer, large model | We put forward a purely transformer-based large view synthesis model, which achieves impressive novel view synthesis results on both object-level and scene-level with minimal 3D inductive bias. | 1,355 | 2410.17242 | [
-0.018347762525081635,
-0.11615794152021408,
-0.008552836254239082,
-0.025914231315255165,
0.08558452129364014,
0.009698232635855675,
-0.10853935033082962,
-0.04824516549706459,
0.07078743726015091,
0.0070868125185370445,
-0.057551391422748566,
-0.02681593969464302,
0.05048982799053192,
0.05421148240566254,
-0.021533239632844925,
0.023740755394101143,
0.031139371916651726,
0.11292144656181335,
-0.07359768450260162,
-0.004894231911748648,
-0.033417265862226486,
-0.0034212812315672636,
-0.024955272674560547,
-0.0393131859600544,
0.0745813176035881,
0.03227180987596512,
0.10229799151420593,
-0.049495358020067215,
-0.02169611118733883,
-0.05239275470376015,
0.049263808876276016,
0.02452922612428665,
-0.03546615689992905,
-0.025560515001416206,
-0.013234531506896019,
0.01695236936211586,
-0.03974158316850662,
-0.014719750732183456,
-0.03967949375510216,
-0.03052252158522606,
0.0335858017206192,
0.04856020584702492,
-0.009186108596622944,
-0.0069797709584236145,
0.038413457572460175,
-0.012608111836016178,
0.02412625215947628,
-0.10256990790367126,
0.009471498429775238,
-0.029006358236074448,
0.0027233788277953863,
0.011459145694971085,
-0.06512563675642014,
0.000705094775184989,
-0.06117865815758705,
0.001987476134672761,
-0.028078977018594742,
-0.03883298859000206,
-0.0078196469694376,
0.023994140326976776,
0.014977768994867802,
0.006733593065291643,
0.007973993197083473,
0.002982276026159525,
-0.05241931602358818,
-0.009515495970845222,
0.06856187433004379,
-0.009610513225197792,
-0.04469146206974983,
-0.04582713544368744,
-0.028360208496451378,
0.004590448923408985,
-0.006014110986143351,
-0.03691135346889496,
0.05016817897558212,
-0.007185210939496756,
0.11628483235836029,
-0.06299276649951935,
0.07606381922960281,
-0.061429306864738464,
0.048416025936603546,
0.031302209943532944,
-0.03202793374657631,
0.04309980943799019,
0.04293934628367424,
0.03344346955418587,
-0.052920468151569366,
0.11834871768951416,
0.013883539475500584,
0.01585954613983631,
-0.042726825922727585,
-0.05717926844954491,
-0.049129169434309006,
-0.021415425464510918,
0.005217184312641621,
0.00014756772725377232,
0.0434439517557621,
-0.06275789439678192,
-0.015013453550636768,
0.10025931894779205,
-0.030747726559638977,
0.016054771840572357,
0.05480039119720459,
-0.029788238927721977,
0.04776066169142723,
-0.03149881586432457,
0.10722242295742035,
0.07122743129730225,
-0.010104791261255741,
-0.0520523339509964,
0.0035456812474876642,
-0.0002948169130831957,
0.020029880106449127,
0.03581516072154045,
0.05017716810107231,
-0.007399999536573887,
0.055380430072546005,
-0.05746816471219063,
0.03471606224775314,
0.032651592046022415,
-0.03992629051208496,
-0.01825425587594509,
-0.04457607865333557,
0.06409330666065216,
-0.027732718735933304,
-0.01653730869293213,
-0.02953827753663063,
7.441924328651358e-34,
-0.022799083963036537,
0.09130726754665375,
0.047844987362623215,
0.034712571650743484,
0.08298689872026443,
0.028743978589773178,
0.06180433928966522,
0.0310552679002285,
0.004441339988261461,
0.005652149673551321,
-0.046774715185165405,
-0.05306046083569527,
0.03174087405204773,
0.1545971781015396,
-0.01290064211934805,
0.011415943503379822,
-0.05592132732272148,
0.04713550955057144,
-0.029237624257802963,
-0.006687608081847429,
-0.04490507394075394,
-0.05066539719700813,
-0.015800097957253456,
0.008438299410045147,
-0.09296105057001114,
-0.006596481427550316,
0.09061222523450851,
-0.019708050414919853,
-0.07240522652864456,
-0.024613523855805397,
-0.11072326451539993,
0.0032567065209150314,
0.07909098267555237,
-0.03292103111743927,
0.0007068037521094084,
-0.043073222041130066,
0.008725841529667377,
-0.053672946989536285,
0.010429666377604008,
-0.04461994767189026,
-0.018146291375160217,
0.1277511566877365,
-0.035090796649456024,
-0.027513917535543442,
-0.1096935048699379,
0.05217012017965317,
0.039227765053510666,
0.09536051750183105,
-0.0890258401632309,
-0.010106965899467468,
0.14520145952701569,
0.017596742138266563,
-0.044257864356040955,
-0.04715484753251076,
-0.0163793433457613,
-0.02879602275788784,
0.025283601135015488,
0.0008773223962634802,
0.07236822694540024,
-0.012386515736579895,
-0.039706405252218246,
0.0596606470644474,
-0.020945286378264427,
0.10381364077329636,
-0.002140994882211089,
0.04690546169877052,
0.058938656002283096,
0.06818056106567383,
0.011709525249898434,
0.012456889264285564,
-0.052213214337825775,
0.0037354074884206057,
-0.010236935690045357,
-0.05664213374257088,
0.07535133510828018,
-0.022539833560585976,
0.02061474323272705,
-0.050407059490680695,
0.027343343943357468,
0.1210898756980896,
-0.02834629826247692,
0.03179660066962242,
0.004477058537304401,
-0.01343773677945137,
-0.005753646604716778,
-0.014135399833321571,
0.011690150946378708,
-0.07564938813447952,
0.03996552526950836,
-0.03431303799152374,
-0.05765349790453911,
-0.09110818803310394,
-0.03381646052002907,
-0.05230163410305977,
0.02251814305782318,
3.1868496737356868e-34,
0.03636929765343666,
0.03827141225337982,
-0.06275081634521484,
-0.04301691800355911,
-0.031617920845746994,
-0.09739413112401962,
-0.029669327661395073,
0.014186609536409378,
0.057676028460264206,
-0.045632269233465195,
-0.03272375836968422,
-0.010381536558270454,
0.05134957656264305,
-0.016225669533014297,
0.059558890759944916,
-0.025220459327101707,
0.010379673913121223,
-0.05970005691051483,
-0.048451587557792664,
-0.016359983012080193,
0.07334621250629425,
0.0775449126958847,
-0.04761151969432831,
0.03064700961112976,
-0.048357024788856506,
0.10762707144021988,
-0.04216613247990608,
0.0711035504937172,
0.04358736798167229,
0.013567295856773853,
0.02156135067343712,
-0.05783679336309433,
0.07573104649782181,
-0.024713264778256416,
-0.07234576344490051,
0.09503012150526047,
0.04488974064588547,
0.015877598896622658,
-0.05418973043560982,
0.014983625151216984,
-0.007782686967402697,
-0.04823913425207138,
-0.01448895875364542,
0.08710431307554245,
-0.031058887019753456,
-0.011951835826039314,
0.03416832908987999,
-0.01693529076874256,
0.028223449364304543,
0.11971110105514526,
-0.10297928005456924,
0.09758659452199936,
-0.06631428003311157,
0.03964008018374443,
0.028209621086716652,
-0.0963120386004448,
-0.06622692197561264,
-0.037705592811107635,
0.13828228414058685,
-0.041843339800834656,
0.001341061433777213,
-0.08909448236227036,
-0.09735437482595444,
-0.0775132104754448,
-0.03240002691745758,
-0.041177619248628616,
-0.0884317010641098,
-0.013042144477367401,
0.07673045247793198,
0.07657656818628311,
0.008396168239414692,
-0.014632758684456348,
-0.0020067235454916954,
-0.00893043726682663,
-0.0016096762847155333,
-0.06665819138288498,
0.009881449863314629,
-0.0023953146301209927,
0.052150875329971313,
-0.04122022166848183,
0.01886684074997902,
-0.039368610829114914,
0.06573747098445892,
0.05891769379377365,
0.0872083380818367,
0.06360326707363129,
0.0017344111111015081,
0.0032696679700165987,
0.04437457025051117,
0.013477656058967113,
-0.02502121962606907,
0.011652362532913685,
-0.001135185593739152,
0.08340969681739807,
-0.0271427184343338,
-3.903564049778652e-8,
-0.06524382531642914,
0.0563068725168705,
0.03114904649555683,
-0.014529751613736153,
-0.07357137650251389,
-0.05634373053908348,
0.004754567984491587,
0.07373877614736557,
0.028688639402389526,
-0.04409634694457054,
0.0788525640964508,
-0.05094347149133682,
-0.05054376646876335,
-0.020592784509062767,
0.05389358475804329,
0.08562854677438736,
0.041101839393377304,
0.01868097297847271,
-0.008092326112091541,
0.03917185962200165,
-0.0426756888628006,
0.04947904497385025,
-0.015153788961470127,
-0.05174225941300392,
0.05537198856472969,
-0.02924300916492939,
-0.054136764258146286,
0.05384647473692894,
0.09050234407186508,
-0.04921438917517662,
-0.01544918678700924,
0.06341417133808136,
0.041571784764528275,
-0.044305622577667236,
0.0022958379704505205,
0.00017730030231177807,
0.04392112419009209,
-0.007898244075477123,
0.008141105994582176,
-0.058260414749383926,
-0.046683043241500854,
0.06263013184070587,
-0.03474133461713791,
0.02656402811408043,
-0.10097385197877884,
-0.002473460743203759,
0.03439079597592354,
-0.06134523078799248,
-0.01888071373105049,
-0.05656088888645172,
-0.040403421968221664,
-0.04274639114737511,
0.020915422588586807,
0.0710221454501152,
0.03049643523991108,
0.05560002103447914,
0.06770501285791397,
-0.061616454273462296,
0.06022407487034798,
0.033760733902454376,
-0.028739027678966522,
0.03432730585336685,
-0.0364968366920948,
0.036520738154649734
] | 0 | 0 | 0 | 0 |
|
Flow Matching with General Discrete Paths: A Kinetic-Optimal Perspective | https://openreview.net/forum?id=tcvMzR2NrP | [
"Neta Shaul",
"Itai Gat",
"Marton Havasi",
"Daniel Severo",
"Anuroop Sriram",
"Peter Holderrieth",
"Brian Karrer",
"Yaron Lipman",
"Ricky T. Q. Chen"
] | Oral | The design space of discrete-space diffusion or flow generative models are significantly less well-understood than their continuous-space counterparts, with many works focusing only on a simple masked construction.
In this work, we aim to take a holistic approach to the construction of discrete generative models based on continuous-time Markov chains, and for the first time, allow the use of arbitrary discrete probability paths, or colloquially, corruption processes.
Through the lens of optimizing the symmetric kinetic energy, we propose velocity formulas that can be applied to any given probability path, completely decoupling the probability and velocity, and giving the user the freedom to specify any desirable probability path based on expert knowledge specific to the data domain.
Furthermore, we find that a special construction of mixture probability paths optimizes the symmetric kinetic energy for the discrete case.
We empirically validate the usefulness of this new design space across multiple modalities: text generation, inorganic material generation, and image generation. We find that we can outperform the mask construction even in text with kinetic-optimal mixture paths, while we can make use of domain-specific constructions of the probability path over the visual domain. | flow matching, discrete generative modeling | Through the lens of kinetic optimality, we expand the design space of Discrete Flow Matching, allowing the use of any probability path and simultaneously justifying existing mixture paths. | 1,351 | 2412.03487 | [
-0.08232829719781876,
-0.029599810019135475,
0.042892441153526306,
0.02114635892212391,
0.016266822814941406,
-0.07999720424413681,
-0.014288770034909248,
-0.04638759419322014,
0.09634754061698914,
-0.014438855461776257,
-0.03805340826511383,
-0.07930932939052582,
0.03109164349734783,
0.03571615740656853,
-0.002683594124391675,
0.027239464223384857,
0.04612656310200691,
0.09852169454097748,
-0.04940218850970268,
-0.05869192257523537,
0.06399904936552048,
-0.0025763388257473707,
0.02802848629653454,
-0.05576396733522415,
0.06492055207490921,
0.0756990984082222,
0.04850089177489281,
-0.03471778333187103,
0.04461868479847908,
-0.02539851889014244,
0.024917297065258026,
-0.02254638820886612,
-0.021922871470451355,
-0.038583092391490936,
0.024031512439250946,
0.07691031694412231,
-0.07150231301784515,
0.09821633994579315,
0.012515036389231682,
0.01693420670926571,
-0.04390957951545715,
0.03248590603470802,
0.029664108529686928,
0.03612547367811203,
0.04781630262732506,
0.00726156635209918,
-0.0023772839922457933,
-0.043807391077280045,
-0.03646576777100563,
-0.01738177053630352,
-0.04485337436199188,
-0.020993296056985855,
-0.0825178399682045,
-0.010587792843580246,
0.013089547865092754,
0.04837895929813385,
0.0923817902803421,
-0.032889656722545624,
-0.007427689153701067,
-0.0586908794939518,
-0.011262859217822552,
-0.09982327371835709,
-0.0516553670167923,
-0.035180047154426575,
0.011653031222522259,
-0.03298409655690193,
0.026873359456658363,
0.08132368326187134,
0.03290271759033203,
0.06951745599508286,
-0.06392937153577805,
0.04254486411809921,
-0.06332332640886307,
0.014480261132121086,
-0.06179303675889969,
0.008315972983837128,
0.013371158391237259,
-0.05843738093972206,
-0.04005710035562515,
-0.09247662872076035,
0.0666668564081192,
-0.009426716715097427,
0.08498948812484741,
-0.024167222902178764,
0.06150805950164795,
0.07221360504627228,
0.018617723137140274,
0.10081518441438675,
0.08858750760555267,
-0.005797321442514658,
-0.13671845197677612,
0.0788840726017952,
0.03922656178474426,
-0.06245184689760208,
0.06991838663816452,
0.05093584209680557,
0.01322996336966753,
-0.07128307223320007,
0.11123662441968918,
0.057282835245132446,
0.026893045753240585,
0.002413532231003046,
0.009184696711599827,
-0.026112303137779236,
0.08935580402612686,
-0.018889611586928368,
-0.014284535311162472,
0.10152251273393631,
0.020477809011936188,
0.03190458565950394,
0.06266383826732635,
0.05057547241449356,
0.05899382382631302,
-0.013464212417602539,
-0.05448940023779869,
-0.028438467532396317,
-0.031367141753435135,
-0.04266306385397911,
0.03882185369729996,
0.10503391176462173,
-0.025021761655807495,
0.007794778794050217,
-0.04426886886358261,
0.004927933681756258,
-0.006099147256463766,
-0.08521457761526108,
0.032729797065258026,
3.566615853054126e-34,
-0.006543833762407303,
0.0027785717975348234,
0.044128987938165665,
0.08954042941331863,
0.09139949828386307,
0.03408975526690483,
0.008634107187390327,
-0.06119502708315849,
0.00867307186126709,
-0.029315244406461716,
-0.04321286827325821,
-0.07699772715568542,
-0.11448866128921509,
0.11448454856872559,
0.0038569194730371237,
-0.006789239589124918,
-0.009026139043271542,
0.04040806367993355,
-0.007161248940974474,
-0.02757725864648819,
0.0177396759390831,
-0.0337865948677063,
-0.06304991245269775,
-0.010327399708330631,
-0.013288374058902264,
0.035497020930051804,
0.027014855295419693,
-0.03597173094749451,
0.007952706888318062,
-0.023824729025363922,
-0.09795340150594711,
0.009825180284678936,
-0.006822959054261446,
-0.01482611708343029,
0.04173494502902031,
0.04324186593294144,
0.0037769870832562447,
-0.06967630237340927,
-0.00020395526371430606,
-0.008055844344198704,
-0.06459144502878189,
0.05518433824181557,
0.009163801558315754,
-0.016267646104097366,
-0.1645611673593521,
0.03220656141638756,
0.046843525022268295,
0.06979721784591675,
-0.0764286071062088,
-0.04775224253535271,
0.09095992892980576,
0.008409779518842697,
0.02798672392964363,
-0.03663622587919235,
-0.023476125672459602,
-0.055547792464494705,
0.025027990341186523,
-0.013253183104097843,
0.05545869842171669,
0.104545459151268,
0.008410695008933544,
0.14878849685192108,
-0.006087790243327618,
0.06805362552404404,
0.04627114161849022,
0.05066080763936043,
-0.031947776675224304,
0.008852392435073853,
0.08122475445270538,
-0.024146586656570435,
-0.024323493242263794,
0.029690729454159737,
-0.06862077116966248,
-0.10622280091047287,
0.04207390546798706,
-0.013171727769076824,
-0.001360535272397101,
-0.02966109849512577,
-0.0391157791018486,
0.03142566978931427,
-0.11101816594600677,
0.0359254851937294,
-0.07054729759693146,
-0.10087188333272934,
0.010177827440202236,
-0.009940806776285172,
-0.015352778136730194,
-0.06131304055452347,
-0.07179135084152222,
-0.06411776691675186,
-0.012272316962480545,
-0.05603937804698944,
0.0007688686600886285,
0.050021570175886154,
0.04613828659057617,
-1.4341277199161611e-33,
-0.04933498054742813,
0.06609869003295898,
0.06309010088443756,
0.0899689644575119,
0.026898538693785667,
0.0002476135559845716,
0.01245533861219883,
-0.05194784700870514,
0.06554069370031357,
-0.0705621987581253,
-0.05909295380115509,
0.021600309759378433,
-0.06415481865406036,
-0.07211335003376007,
-0.012495295144617558,
-0.08137860894203186,
0.021229131147265434,
0.0008758082403801382,
-0.04247565567493439,
0.029464244842529297,
0.007408532779663801,
0.05242513492703438,
-0.13453461229801178,
-0.019537657499313354,
0.006224998738616705,
0.08696898072957993,
-0.012887656688690186,
0.006640250328928232,
0.013215182349085808,
0.03749970346689224,
-0.0018339219968765974,
0.00834878720343113,
-0.04321102052927017,
-0.013199275359511375,
-0.1046486422419548,
0.030373044312000275,
0.03883406147360802,
0.0957924947142601,
-0.05358884856104851,
0.03682537004351616,
-0.004918495658785105,
-0.03580758348107338,
-0.05808274447917938,
0.05363783240318298,
-0.044835109263658524,
-0.03636905923485756,
-0.004878748673945665,
-0.040438462048769,
0.017398478463292122,
-0.008824325166642666,
0.016956916078925133,
0.03581327944993973,
-0.08094412088394165,
0.020280275493860245,
-0.025089994072914124,
0.018922150135040283,
-0.10915755480527878,
-0.032221242785453796,
0.1120985895395279,
-0.0386708602309227,
-0.00970251951366663,
-0.058042217046022415,
-0.0092029282823205,
0.03500596433877945,
-0.016151173040270805,
-0.06071718409657478,
-0.0474386028945446,
-0.0168859101831913,
-0.052961740642786026,
0.01703997701406479,
0.014477638527750969,
0.037629153579473495,
0.03055570274591446,
0.02882186509668827,
0.07858099043369293,
-0.054724618792533875,
0.11545680463314056,
0.04414735361933708,
0.002244931412860751,
0.04114927724003792,
-0.01407523825764656,
0.06711135804653168,
0.003868646454066038,
0.039676155894994736,
0.017866602167487144,
-0.011404168792068958,
-0.0436963215470314,
-0.06012545898556709,
0.05151282995939255,
-0.005563089624047279,
-0.0271857101470232,
-0.02911151386797428,
0.03956913575530052,
0.06976369768381119,
-0.11749386042356491,
-4.9514014222040714e-8,
-0.029794966802001,
-0.01293261256068945,
-0.017823051661252975,
-0.02013700269162655,
0.0190866868942976,
0.08613218367099762,
0.0041940114460885525,
0.031732816249132156,
0.03449314832687378,
-0.05725942179560661,
0.13003699481487274,
0.007023940794169903,
-0.027255045250058174,
-0.015898559242486954,
-0.05968576669692993,
0.05290349945425987,
-0.017591889947652817,
-0.008954163640737534,
-0.013122514821588993,
-0.03939901292324066,
-0.02990056574344635,
0.02170604281127453,
-0.07862521708011627,
0.0766473114490509,
0.03888994827866554,
-0.005541592836380005,
-0.006324305664747953,
-0.04785076156258583,
0.07143110781908035,
0.026094546541571617,
-0.03447719290852547,
-0.010439426638185978,
0.04243296757340431,
0.09130745381116867,
0.007106690667569637,
-0.005597297102212906,
-0.00822945311665535,
-0.018583988770842552,
-0.03900869935750961,
-0.042692191898822784,
0.01376164797693491,
0.030955534428358078,
0.0052058277651667595,
0.015465336851775646,
0.055794235318899155,
0.0166423711925745,
0.004698894917964935,
-0.06495574116706848,
-0.038686614483594894,
0.08880043774843216,
-0.02699529007077217,
-0.056534428149461746,
-0.013543887995183468,
0.026405787095427513,
0.07678980380296707,
0.005549674853682518,
0.02954978682100773,
-0.03671390563249588,
0.03330681845545769,
0.011491527780890465,
0.02064697816967964,
0.002116176765412092,
0.02642357535660267,
-0.03844229876995087
] | 0 | 0 | 0 | 0 |
|
Cut Your Losses in Large-Vocabulary Language Models | https://openreview.net/forum?id=E4Fk3YuG56 | [
"Erik Wijmans",
"Brody Huval",
"Alexander Hertzberg",
"Vladlen Koltun",
"Philipp Kraehenbuehl"
] | Oral | As language models grow ever larger, so do their vocabularies.
This has shifted the memory footprint of LLMs during training disproportionately to one single layer: the cross-entropy in the loss computation.
Cross-entropy builds up a logit matrix with entries for each pair of input tokens and vocabulary items and, for small models, consumes an order of magnitude more memory than the rest of the LLM combined.
We propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss without materializing the logits for all tokens into global memory.
Rather, CCE only computes the logit for the correct token and evaluates the log-sum-exp over all logits on the fly.
We implement a custom kernel that performs the matrix multiplications and the log-sum-exp reduction over the vocabulary in flash memory, making global memory consumption for the cross-entropy computation negligible. This has a dramatic effect. Taking the Gemma 2 (2B) model as an example, CCE reduces the memory footprint of the loss computation from 24 GB to 1 MB, and the total training-time memory consumption of the classifier head from 28 GB to 1 GB.
To improve the throughput of CCE, we leverage the inherent sparsity of softmax and propose to skip elements of the gradient computation that have a negligible (i.e. below numerical precision) contribution to the gradient.
Experiments demonstrate that the dramatic reduction in memory consumption is accomplished without sacrificing training speed or convergence. | large language model, large vocabulary, efficient | We propose Cut Cross-Entropy (CCE), a method that computes the cross-entropy loss with negligible memory consumption. | 1,344 | 2411.09009 | [
-0.012221810407936573,
-0.06043341010808945,
-0.010034522973001003,
0.03961368277668953,
0.054927024990320206,
0.001089678262360394,
-0.011740797199308872,
0.01757964678108692,
0.050272565335035324,
-0.033203475177288055,
0.009031226858496666,
-0.0652845948934555,
0.029008880257606506,
0.008466391824185848,
0.024728072807192802,
-0.013696655631065369,
0.011622393503785133,
0.1672549545764923,
-0.06983451545238495,
-0.03934253379702568,
0.0817032903432846,
-0.012212459929287434,
0.004874196834862232,
0.04476867616176605,
0.025430351495742798,
-0.028453992679715157,
0.004307394847273827,
-0.02593567781150341,
0.01521874126046896,
-0.02060595154762268,
0.033672548830509186,
0.056876249611377716,
-0.03391318395733833,
0.07532574236392975,
-0.0450102835893631,
0.027029378339648247,
-0.05933831259608269,
-0.006431092042475939,
-0.005328573752194643,
-0.05199427157640457,
-0.04013669118285179,
0.03919505700469017,
0.04143466800451279,
0.05246548727154732,
0.05327342450618744,
0.015992939472198486,
0.023345910012722015,
-0.031828273087739944,
-0.01142758596688509,
0.013048119843006134,
0.0037305732257664204,
0.020130813121795654,
-0.03733766824007034,
0.08102492243051529,
-0.06586611270904541,
-0.008918827399611473,
-0.01344771683216095,
0.001419086242094636,
-0.06133665516972542,
0.0019264434231445193,
-0.09457001090049744,
-0.09196021407842636,
-0.02134542539715767,
-0.00829367432743311,
-0.034411948174238205,
-0.07802419364452362,
0.020693335682153702,
0.056868623942136765,
0.021044574677944183,
0.0825420394539833,
-0.09691153466701508,
0.0407518669962883,
-0.03163798898458481,
0.0476670041680336,
-0.04719768464565277,
0.06008477881550789,
0.060785893350839615,
0.010497861541807652,
0.05582381784915924,
-0.07518642395734787,
0.04250006377696991,
0.048670392483472824,
0.08468945324420929,
-0.028595881536602974,
0.12852345407009125,
-0.04542987793684006,
0.07974733412265778,
-0.0041048103012144566,
0.030997259542346,
-0.032909493893384933,
0.019957102835178375,
-0.034590937197208405,
0.10010961443185806,
0.008495236746966839,
0.0507015734910965,
-0.027662811800837517,
0.019631505012512207,
-0.009418905712664127,
0.002661811886355281,
0.07113861292600632,
0.01828772947192192,
0.10268080234527588,
0.0017085166182368994,
-0.008686820045113564,
0.02241862565279007,
-0.03929327428340912,
0.006220535840839148,
-0.020990479737520218,
0.08168314397335052,
-0.048808854073286057,
0.06424310058355331,
0.13408958911895752,
-0.060462187975645065,
0.022649969905614853,
0.04406598210334778,
0.054693326354026794,
-0.014574549160897732,
-0.0676058754324913,
0.06826488673686981,
0.13808032870292664,
-0.09764780104160309,
0.06818553060293198,
-0.002431793138384819,
-0.037264134734869,
-0.04175996780395508,
-0.016348252072930336,
-0.03239818289875984,
2.1695182714875304e-33,
0.010200947523117065,
0.07375548779964447,
-0.056558478623628616,
-0.010466812178492546,
0.056986741721630096,
0.027241071686148643,
0.042611122131347656,
-0.007376720197498798,
-0.03987564891576767,
-0.03888315334916115,
-0.05411230027675629,
0.05580059438943863,
-0.026499897241592407,
0.07665292918682098,
0.09810806810855865,
0.013314560055732727,
-0.026030607521533966,
0.016945429146289825,
0.007792887277901173,
-0.05842011049389839,
0.03379714488983154,
0.018885385245084763,
0.03216417506337166,
-0.04630541801452637,
0.006837871391326189,
0.005105176474899054,
0.057463113218545914,
-0.10872510075569153,
-0.01668792963027954,
0.0601910762488842,
-0.07772994786500931,
-0.04912453517317772,
-0.0019484669901430607,
0.012656139209866524,
0.08089777827262878,
-0.039721306413412094,
-0.03546472638845444,
-0.030943090096116066,
0.05395174399018288,
-0.1418575495481491,
-0.0563025064766407,
0.01134531944990158,
0.01542099379003048,
-0.10071410238742828,
-0.12399612367153168,
-0.09732875972986221,
0.023606987670063972,
0.005607821512967348,
-0.011236762627959251,
-0.050125911831855774,
0.11034733057022095,
-0.044700201600790024,
-0.10341740399599075,
0.011963999830186367,
0.023339273408055305,
-0.02504045143723488,
0.024115033447742462,
0.014623520895838737,
0.07494490593671799,
0.09540290385484695,
-0.07292576134204865,
0.046876683831214905,
0.06639368087053299,
0.07383748143911362,
0.04631872847676277,
0.042808618396520615,
-0.008885174989700317,
0.022072384133934975,
0.015835024416446686,
-0.05957603082060814,
-0.03886914998292923,
-0.015688957646489143,
0.022264664992690086,
-0.07139242440462112,
0.016397593542933464,
-0.008690527640283108,
0.09566528350114822,
-0.10703417658805847,
-0.047925982624292374,
0.05830040201544762,
0.011255864053964615,
-0.018626989796757698,
0.03805726021528244,
-0.08513801544904709,
-0.004923896864056587,
-0.039840217679739,
0.005083179567009211,
-0.08393208682537079,
-0.058718834072351456,
0.03373552858829498,
-0.015026910230517387,
-0.0619322843849659,
0.057309918105602264,
-0.04079080745577812,
-0.05310508981347084,
-2.522346390635694e-33,
-0.057310380041599274,
0.0011095308000221848,
-0.003707227297127247,
0.042347609996795654,
-0.07516700774431229,
-0.11517365276813507,
0.018243467435240746,
0.021393870934844017,
-0.05351242423057556,
-0.05564514547586441,
-0.008828788995742798,
-0.02550465054810047,
-0.009949823841452599,
0.04321404919028282,
0.07458172738552094,
-0.004736957140266895,
-0.026530446484684944,
0.04460575059056282,
0.07656095176935196,
0.03854422643780708,
0.010365581139922142,
0.0949220359325409,
-0.06096865236759186,
0.036413565278053284,
-0.016339082270860672,
0.015231378376483917,
-0.0219790730625391,
0.059731464833021164,
0.0477689690887928,
-0.03687172755599022,
-0.043691348284482956,
-0.015866100788116455,
-0.029897356405854225,
-0.025653038173913956,
-0.046881821006536484,
-0.02133364789187908,
0.034012068063020706,
0.005153834354132414,
-0.02545197866857052,
0.07047522068023682,
0.03224547579884529,
-0.019628090783953667,
-0.050648562610149384,
-0.029179638251662254,
-0.0011772997677326202,
-0.01788298785686493,
-0.1658944934606552,
0.006550735794007778,
0.07232637703418732,
0.03004138357937336,
0.029123086482286453,
-0.029258204624056816,
-0.05890386924147606,
0.055532556027173996,
-0.04821013659238815,
-0.09143542498350143,
-0.03202536329627037,
-0.05291016399860382,
-0.058715399354696274,
-0.034039177000522614,
-0.026278609409928322,
-0.04855724424123764,
0.025634104385972023,
-0.05965977534651756,
0.05310255289077759,
-0.03344426676630974,
-0.09281592071056366,
0.0057591767981648445,
-0.030472245067358017,
0.010048167780041695,
0.06833875924348831,
0.06785444915294647,
-0.04275534674525261,
0.011000231839716434,
-0.013021477498114109,
0.02426554635167122,
0.0353366881608963,
-0.09769175201654434,
-0.008528330363333225,
-0.026605861261487007,
-0.06186208501458168,
0.07703500241041183,
0.04622487351298332,
0.01642761006951332,
0.0331006683409214,
0.010267527773976326,
0.0003146085946355015,
0.0355607308447361,
0.004352285992354155,
-0.01697617582976818,
-0.010106202214956284,
0.014416404999792576,
0.09246283769607544,
0.07382597029209137,
-0.05609771981835365,
-4.4716500724462094e-8,
-0.03735177963972092,
-0.001043241354636848,
0.035059042274951935,
0.04254836216568947,
0.06352567672729492,
-0.08784948289394379,
-0.04092491790652275,
0.12051676958799362,
0.00493134930729866,
-0.028016168624162674,
0.08710943907499313,
0.02181694470345974,
-0.023457063362002373,
-0.03829458728432655,
-0.040274038910865784,
0.07153728604316711,
-0.016730807721614838,
-0.009803642518818378,
0.007539261598140001,
-0.05635154992341995,
0.028162546455860138,
0.015105842612683773,
0.019000202417373657,
0.06616954505443573,
-0.0012493389658629894,
-0.048893142491579056,
0.05459264665842056,
0.09247338026762009,
0.05324225872755051,
-0.05501455441117287,
-0.028321946039795876,
0.0074081020429730415,
-0.043747782707214355,
0.008831433020532131,
0.0012144324136897922,
0.056713301688432693,
-0.001053619314916432,
-0.04407829791307449,
-0.010404447093605995,
0.04052784666419029,
0.05361445993185043,
-0.01702086813747883,
0.018698034808039665,
0.041924264281988144,
0.032842036336660385,
-0.017145909368991852,
-0.03872109204530716,
-0.07833091169595718,
0.05386741831898689,
-0.03963002935051918,
-0.013784799724817276,
0.04329651966691017,
-0.036021579056978226,
0.08956305682659149,
0.055866360664367676,
-0.006719318684190512,
-0.05750183388590813,
-0.014275182969868183,
0.05182382091879845,
0.024805206805467606,
-0.03752091899514198,
-0.07558044046163559,
-0.06382406502962112,
0.019148562103509903
] | https://github.com/apple/ml-cross-entropy | 430 | 0 | 0 | 0 |
AFlow: Automating Agentic Workflow Generation | https://openreview.net/forum?id=z5uVAKwmjf | [
"Jiayi Zhang",
"Jinyu Xiang",
"Zhaoyang Yu",
"Fengwei Teng",
"Xiong-Hui Chen",
"Jiaqi Chen",
"Mingchen Zhuge",
"Xin Cheng",
"Sirui Hong",
"Jinlin Wang",
"Bingnan Zheng",
"Bang Liu",
"Yuyu Luo",
"Chenglin Wu"
] | Oral | Large language models (LLMs) have demonstrated remarkable potential in solving complex tasks across diverse domains, typically by employing agentic workflows that follow detailed instructions and operational sequences. However, constructing these workflows requires significant human effort, limiting scalability and generalizability. Recent research has sought to automate the generation and optimization of these workflows, but existing methods still rely on initial manual setup and fall short of achieving fully automated and effective workflow generation. To address this challenge, we reformulate workflow optimization as a search problem over code-represented workflows, where LLM-invoking nodes are connected by edges. We introduce AFLOW, an automated framework that efficiently explores this space using Monte Carlo Tree Search, iteratively refining workflows through code modification, tree-structured experience, and execution feedback. Empirical evaluations across six benchmark datasets demonstrate AFLOW's efficacy, yielding a 5.7% average improvement over state-of-the-art baselines. Furthermore, AFLOW enables smaller models to outperform GPT-4o on specific tasks at 4.55% of its inference cost in dollars. The code is available at https://github.com/geekan/MetaGPT. | LLM Agent; Prompt Optimization; Workflow Generation | We introduce the field of Agentic Workflow Optimization and propose an effective search algorithm called AFLOW, enabling it to surpass manually constructed workflows on six reasoning datasets. | 1,308 | 2410.10762 | [
-0.07306510955095291,
-0.011645564809441566,
-0.026818418875336647,
0.06152576580643654,
0.023103367537260056,
-0.08543524146080017,
-0.060491107404232025,
-0.04503011330962181,
-0.0064043872989714146,
-0.0720835030078888,
-0.11643650382757187,
-0.103748619556427,
0.0120268315076828,
0.004756688140332699,
0.019341789186000824,
-0.001962347887456417,
0.0518806055188179,
0.049514587968587875,
-0.03889702633023262,
-0.10291492938995361,
-0.001996278762817383,
0.04685923829674721,
-0.013654613867402077,
-0.03433879092335701,
-0.0019776313565671444,
0.01025662012398243,
-0.00880077388137579,
-0.0898209884762764,
0.000316678051603958,
0.03559977933764458,
0.02925603650510311,
0.06304479390382767,
0.02344919554889202,
0.09619700908660889,
-0.014772710390388966,
0.06488527357578278,
-0.11104551702737808,
0.017049547284841537,
0.01442398689687252,
-0.040921855717897415,
-0.037638336420059204,
0.024634109809994698,
0.025692414492368698,
-0.02529490552842617,
0.02340359427034855,
-0.05976111814379692,
-0.02194378152489662,
-0.029477128759026527,
-0.025092586874961853,
-0.012917534448206425,
-0.13296714425086975,
-0.0781872570514679,
-0.03098643384873867,
0.047729283571243286,
-0.013555316254496574,
0.010435987263917923,
0.1152653694152832,
-0.01720891147851944,
0.014838849194347858,
-0.04830218106508255,
-0.013395184651017189,
-0.10021047294139862,
-0.023993251845240593,
-0.021038511767983437,
-0.05336211621761322,
-0.029416481032967567,
-0.007026514504104853,
0.012082609347999096,
-0.010874174535274506,
-0.02411969192326069,
-0.04528331756591797,
0.05459184572100639,
-0.06565507501363754,
0.030343540012836456,
0.01831570267677307,
0.07973318547010422,
-0.02545376680791378,
-0.041032303124666214,
0.06330230832099915,
-0.04903063923120499,
-0.03570230305194855,
-0.03013504110276699,
0.012287660501897335,
0.09994576126337051,
0.053339261561632156,
0.04344319552183151,
0.04412372410297394,
0.06739875674247742,
0.06518706679344177,
0.08755338191986084,
-0.044209424406290054,
-0.03744002804160118,
0.03535548225045204,
0.0036177777219563723,
0.014556084759533405,
0.04406433179974556,
0.0238313227891922,
-0.13434000313282013,
-0.0594194158911705,
0.059569891542196274,
-0.012527008540928364,
0.10682215541601181,
0.061838407069444656,
-0.11680921912193298,
-0.07395279407501221,
0.0458957701921463,
0.038882918655872345,
0.0444033145904541,
0.004733526613563299,
-0.06250385195016861,
-0.03916709125041962,
0.08072603493928909,
0.05206373333930969,
-0.034279465675354004,
0.03321298956871033,
0.024751072749495506,
0.0223429873585701,
-0.027262693271040916,
-0.03679879382252693,
0.1488381326198578,
0.0511351153254509,
0.07403728365898132,
-0.037470169365406036,
0.015023953281342983,
0.05490976199507713,
-0.02052886411547661,
-0.04575800150632858,
2.7713232573085548e-33,
0.10297286510467529,
0.060489900410175323,
0.02733571082353592,
0.07696452736854553,
0.12866173684597015,
-0.022702544927597046,
0.07942670583724976,
0.03511057794094086,
0.003990571480244398,
-0.014500346966087818,
-0.008536305278539658,
0.03836634010076523,
-0.096017025411129,
0.01198827289044857,
0.007900549098849297,
-0.03603097051382065,
-0.09907583147287369,
0.06677411496639252,
-0.05211622640490532,
0.02288462221622467,
0.04343205690383911,
-0.022160865366458893,
0.016045475378632545,
-0.006524425931274891,
0.05727013573050499,
0.05630706250667572,
-0.007941045798361301,
-0.05428903549909592,
0.05184313654899597,
0.012543978169560432,
-0.08773980289697647,
-0.0271851047873497,
0.02523029036819935,
0.08310402184724808,
-0.040120210498571396,
0.013685924001038074,
-0.04074672982096672,
-0.10042769461870193,
0.01387586910277605,
0.023286031559109688,
-0.03360305353999138,
0.024545922875404358,
0.0639636293053627,
-0.002826235955581069,
-0.04465768113732338,
-0.09307366609573364,
-0.08754748106002808,
0.0006847664481028914,
0.04324980452656746,
0.027139484882354736,
0.11430609971284866,
0.0239343773573637,
0.052540943026542664,
0.02283649891614914,
-0.01780529133975506,
0.03467624634504318,
-0.008795981295406818,
0.051446203142404556,
0.04245050624012947,
0.0894479900598526,
0.04979710653424263,
-0.017082370817661285,
-0.01730252616107464,
0.06879431754350662,
0.06955979019403458,
0.009019528515636921,
-0.0064505706541240215,
0.08102913200855255,
0.07540580630302429,
0.027626633644104004,
-0.010029313154518604,
0.057900719344615936,
0.02059052512049675,
0.013287756592035294,
0.02406870573759079,
-0.012713149189949036,
0.07066633552312851,
-0.10359065979719162,
-0.019433295354247093,
-0.06615623086690903,
-0.042866457253694534,
0.05440353602170944,
0.04344595596194267,
-0.0009894693503156304,
0.051056087017059326,
-0.09139787405729294,
-0.014023451134562492,
-0.026909250766038895,
-0.045162882655858994,
0.004206482321023941,
-0.055187687277793884,
-0.04754622280597687,
-0.010311487130820751,
0.0050888084806501865,
0.027488868683576584,
-3.4301770043559e-33,
-0.03220396116375923,
-0.06574909389019012,
-0.04040537774562836,
0.08791572600603104,
-0.05307284742593765,
-0.1124534159898758,
-0.03697450831532478,
-0.04985300451517105,
0.024286560714244843,
-0.011231377720832825,
-0.040203385055065155,
-0.013788473792374134,
0.026150552555918694,
-0.037624895572662354,
0.030872346833348274,
-0.0962463766336441,
0.01832002028822899,
0.014410505071282387,
0.03727767616510391,
0.11065606772899628,
-0.030519092455506325,
0.11463817209005356,
-0.07915128767490387,
-0.013951019383966923,
-0.02806808240711689,
-0.01740160584449768,
0.006830642465502024,
-0.01759357564151287,
0.033866997808218,
-0.010906040668487549,
0.010739512741565704,
0.06981878727674484,
-0.04674576222896576,
-0.005348305683583021,
-0.011247603222727776,
0.0380525104701519,
0.03770190104842186,
0.03944861516356468,
-0.0143143180757761,
0.06126158684492111,
0.0763038620352745,
-0.06611999869346619,
-0.0424358993768692,
0.020572513341903687,
-0.014935364946722984,
0.0690242126584053,
-0.08121035248041153,
0.07125555723905563,
0.010927792638540268,
-0.03667787089943886,
-0.05516030266880989,
-0.0022323643788695335,
-0.057677969336509705,
-0.010209680534899235,
-0.021492239087820053,
-0.056675609201192856,
0.013917751610279083,
-0.013317890465259552,
-0.03206290304660797,
-0.023594960570335388,
-0.09908789396286011,
0.013810877688229084,
0.07590749859809875,
-0.07667743414640427,
-0.00306494883261621,
-0.05990438908338547,
-0.08231192082166672,
-0.03658042848110199,
-0.04317598417401314,
0.0015283963875845075,
0.03692707046866417,
-0.05132870748639107,
-0.05660784989595413,
0.052863430231809616,
0.026251373812556267,
-0.0024223432410508394,
0.030660610646009445,
-0.087355837225914,
0.03250948339700699,
-0.0035316641442477703,
-0.009006365202367306,
0.11532516777515411,
0.016109809279441833,
-0.013355080038309097,
0.001805144944228232,
0.017468007281422615,
-0.021562539041042328,
0.09478262811899185,
0.021896935999393463,
0.022675661370158195,
-0.0724979117512703,
0.0032351776026189327,
0.004231736063957214,
-0.04309139773249626,
-0.10096819698810577,
-4.172006384806082e-8,
-0.05232272297143936,
0.0824764221906662,
-0.011295368894934654,
0.02977723255753517,
-0.018835406750440598,
-0.02409856766462326,
0.000919716723728925,
0.1265566051006317,
0.014348365366458893,
-0.041306886821985245,
0.07368574291467667,
-0.046374231576919556,
0.01568126305937767,
-0.05017489567399025,
-0.00774496840313077,
0.04023561254143715,
-0.045101407915353775,
-0.031036730855703354,
-0.018008412793278694,
-0.027455337345600128,
-0.01953883282840252,
0.046978648751974106,
-0.001009143888950348,
0.004889786243438721,
0.00028707896126434207,
-0.07085077464580536,
0.02542436495423317,
0.04862959310412407,
0.05978836491703987,
0.01782088167965412,
0.021668752655386925,
0.0034588295966386795,
-0.001192001043818891,
0.00040526839438825846,
-0.003186812624335289,
-0.029302215203642845,
0.003184361383318901,
-0.03993877395987511,
0.036932945251464844,
-0.04824904724955559,
0.10809390246868134,
0.017672564834356308,
-0.013158928602933884,
-0.03095897100865841,
-0.005337063688784838,
0.00038624712033197284,
-0.0322803296148777,
-0.038885243237018585,
0.08988072723150253,
-0.07039553672075272,
-0.013584460131824017,
-0.018190676346421242,
-0.04702996462583542,
-0.003868376137688756,
0.03403428569436073,
0.07348217070102692,
-0.027387987822294235,
-0.061110809445381165,
0.10656553506851196,
0.026503300294280052,
-0.03649396076798439,
0.030250834301114082,
-0.0058257984928786755,
0.013952291570603848
] | https://github.com/geekan/metagpt | 55,029 | 0 | 0 | 0 |
Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models | https://openreview.net/forum?id=uAFHCZRmXk | [
"Simon Schrodi",
"David T. Hoffmann",
"Max Argus",
"Volker Fischer",
"Thomas Brox"
] | Oral | Contrastive vision-language models (VLMs), like CLIP, have gained popularity for their versatile applicability to various downstream tasks. Despite their successes in some tasks, like zero-shot object recognition, they perform surprisingly poor on other tasks, like attribute recognition. Previous work has attributed these challenges to the modality gap, a separation of image and text in the shared representation space, and to a bias towards objects over other factors, such as attributes. In this analysis paper, we investigate both phenomena thoroughly. We evaluated off-the-shelf VLMs and while the gap's influence on performance is typically overshadowed by other factors, we find indications that closing the gap indeed leads to improvements. Moreover, we find that, contrary to intuition, only few embedding dimensions drive the gap and that the embedding spaces are differently organized. To allow for a clean study of object bias, we introduce a definition and a corresponding measure of it. Equipped with this tool, we find that object bias does not lead to worse performance on other concepts, such as attributes per se. However, why do both phenomena, modality gap and object bias, emerge in the first place? To answer this fundamental question and uncover some of the inner workings of contrastive VLMs, we conducted experiments that allowed us to control the amount of shared information between the modalities. These experiments revealed that the driving factor behind both the modality gap and the object bias, is an information imbalance between images and captions, and unveiled an intriguing connection between the modality gap and entropy of the logits. | CLIP, modality gap, object bias, contrastive loss, data-centric, vision language models, VLM | We find that an information imbalance between images and texts leads to the modality gap and object bias of contrastive VLMs. We study both phenomena in depth, eliminate common misconceptions, and improve the understanding of both of them. | 1,079 | 2404.07983 | [
0.0036473653744906187,
-0.08835086226463318,
-0.008940078318119049,
0.03907368704676628,
0.13992200791835785,
-0.011739449575543404,
-0.023980403319001198,
0.0017079781973734498,
0.07646632194519043,
-0.08083972334861755,
-0.038621943444013596,
-0.052276208996772766,
0.05463770031929016,
0.04422226548194885,
0.006944007705897093,
-0.007915576919913292,
0.14051946997642517,
0.03398583084344864,
-0.14601446688175201,
0.00860938336700201,
0.01512966863811016,
-0.000031117917387746274,
0.05381682515144348,
-0.05445105955004692,
0.05276105925440788,
0.036097608506679535,
-0.04706291854381561,
-0.05252304673194885,
0.016906317323446274,
-0.059068553149700165,
0.005501884035766125,
0.015148916281759739,
0.013588489033281803,
0.033291835337877274,
-0.017031695693731308,
0.07238616794347763,
-0.002837160136550665,
0.06411829590797424,
-0.07733156532049179,
-0.04663512483239174,
-0.055178381502628326,
0.02132742665708065,
-0.001131347264163196,
0.04176979139447212,
0.02714698389172554,
0.00031342136207967997,
0.04978660121560097,
-0.0043451664969325066,
-0.00037459374289028347,
-0.04228430613875389,
-0.08380688726902008,
-0.026062771677970886,
-0.033632081001996994,
0.050496555864810944,
-0.0915340781211853,
0.011433916166424751,
0.018563857302069664,
0.06974923610687256,
-0.013688718900084496,
-0.04548526182770729,
-0.011308900080621243,
-0.02096279338002205,
0.043212007731199265,
0.04009276628494263,
0.009394095279276371,
-0.044335171580314636,
0.03736579418182373,
-0.021715693175792694,
-0.0760488510131836,
-0.004896538332104683,
0.019488627091050148,
0.10951690375804901,
-0.023528583347797394,
0.0008776507456786931,
0.022076217457652092,
0.0155947832390666,
0.06701745837926865,
-0.05507507175207138,
0.058034446090459824,
-0.1344228982925415,
0.05542311444878578,
-0.01012421678751707,
0.0046075028367340565,
0.020577646791934967,
0.11877568811178207,
0.0613323338329792,
0.0020611269865185022,
0.0032703098841011524,
-0.08657220751047134,
-0.031576141715049744,
-0.04110698401927948,
-0.06756457686424255,
0.04733092337846756,
-0.007646620739251375,
0.09591470658779144,
-0.025796689093112946,
0.004284742288291454,
-0.000560196116566658,
-0.009174751117825508,
0.08863718807697296,
-0.030913567170500755,
0.04798733443021774,
0.03712006285786629,
-0.005229164846241474,
0.047630395740270615,
0.0023172374349087477,
0.03423120453953743,
-0.0409901961684227,
-0.035998228937387466,
-0.07470010221004486,
0.019716471433639526,
0.0420708991587162,
-0.0245370976626873,
-0.003228507237508893,
0.03461204469203949,
-0.07242424786090851,
0.019452493637800217,
-0.039860498160123825,
0.05333796143531799,
-0.020339153707027435,
-0.016643473878502846,
0.010008860379457474,
-0.015032760798931122,
-0.024232422932982445,
0.04507682844996452,
0.00874491035938263,
-0.006686693523079157,
3.6597791182950063e-35,
0.03263731300830841,
-0.015275602228939533,
-0.004559477791190147,
0.07147958129644394,
0.08164376765489578,
-0.024989858269691467,
-0.023705199360847473,
0.07164094597101212,
-0.0167536623775959,
0.04154389724135399,
-0.015281197614967823,
0.053048327565193176,
-0.020371217280626297,
0.10351556539535522,
0.08614625781774521,
0.019357984885573387,
-0.006323274690657854,
0.005480694584548473,
-0.05568237230181694,
0.03243667259812355,
-0.01799554005265236,
-0.041604552417993546,
-0.024383598938584328,
0.03938087075948715,
-0.05901148170232773,
-0.022915750741958618,
0.07025083154439926,
-0.10351859033107758,
-0.06022808328270912,
-0.00043443936738185585,
-0.16896510124206543,
0.053400058299303055,
0.05255410075187683,
0.013158117420971394,
0.012851836159825325,
-0.03481235355138779,
-0.0065828533843159676,
-0.050227195024490356,
0.0714997947216034,
-0.043896764516830444,
-0.000038169782783370465,
0.06642866134643555,
-0.03700888156890869,
-0.0888127014040947,
-0.08520503342151642,
-0.03323899209499359,
0.03993835672736168,
0.05223962664604187,
-0.06419609487056732,
0.04018157348036766,
0.046915456652641296,
0.046677377074956894,
0.04381144791841507,
-0.04789818078279495,
-0.01770855113863945,
-0.03632428124547005,
0.011866016313433647,
0.020331602543592453,
0.014551632106304169,
-0.017026379704475403,
-0.024287540465593338,
0.05086476728320122,
0.006768071558326483,
0.06659740209579468,
0.011307800188660622,
0.06703288853168488,
-0.02350587770342827,
0.02343188226222992,
0.007784165907651186,
0.017691390588879585,
-0.06867874413728714,
0.05057956278324127,
-0.10310016572475433,
-0.046796102076768875,
0.04669073596596718,
-0.0034943970385938883,
0.022217171266674995,
-0.13229753077030182,
0.05254027619957924,
0.07982787489891052,
-0.02926274575293064,
-0.022305671125650406,
0.0343472845852375,
-0.09862971305847168,
-0.11885745823383331,
-0.0411311611533165,
0.060897450894117355,
-0.12758974730968475,
0.015043884515762329,
-0.036865465342998505,
-0.017935916781425476,
-0.06071964651346207,
-0.042665332555770874,
-0.007297789677977562,
0.0725540891289711,
-1.8782998316468783e-33,
-0.0176229365170002,
0.022462429478764534,
-0.037439975887537,
0.03310847654938698,
-0.014868443831801414,
-0.04608646035194397,
0.09892316162586212,
0.009548082016408443,
-0.02260693535208702,
-0.03552533686161041,
0.005448743235319853,
-0.020393574610352516,
0.01280074380338192,
-0.0070143453776836395,
0.026662353426218033,
-0.012632163241505623,
0.02796294540166855,
0.04377714917063713,
0.05795682594180107,
0.016306832432746887,
0.07450677454471588,
0.08591408282518387,
-0.07412024587392807,
0.026864834129810333,
-0.061336029320955276,
0.06499989330768585,
-0.07419440895318985,
-0.07383738458156586,
-0.058794982731342316,
-0.08883614838123322,
-0.015507500618696213,
0.02420576475560665,
-0.03311914950609207,
-0.05139933153986931,
-0.06440700590610504,
0.03612174838781357,
-0.03479190170764923,
-0.04055039584636688,
-0.011090898886322975,
0.07132215797901154,
0.07187727838754654,
-0.02207421138882637,
-0.10533107817173004,
-0.04097285866737366,
-0.029710879549384117,
-0.049100663512945175,
-0.06368239969015121,
-0.037936732172966,
0.05362154170870781,
-0.013912892900407314,
-0.024208685383200645,
0.017858978360891342,
-0.04454057291150093,
0.042936425656080246,
-0.04517178609967232,
-0.05039789527654648,
-0.020231278613209724,
-0.01722203567624092,
0.08767259865999222,
0.06713183224201202,
0.0059346966445446014,
-0.002647785935550928,
-0.05501651391386986,
-0.040532249957323074,
0.008621115237474442,
-0.02215823531150818,
0.0017961191479116678,
-0.027845796197652817,
0.06591271609067917,
-0.02238345704972744,
0.0569051094353199,
-0.052498709410429,
0.02268310636281967,
0.023539669811725616,
-0.053758975118398666,
0.0014590552309527993,
0.05380877107381821,
0.003464752109721303,
-0.02207784354686737,
-0.029283717274665833,
0.003679197747260332,
-0.08511970192193985,
0.03210027515888214,
0.02409079484641552,
0.028282925486564636,
0.1271262764930725,
-0.041413210332393646,
-0.0382668562233448,
0.035947177559137344,
-0.00965226162225008,
-0.07338431477546692,
-0.01260791253298521,
-0.016559239476919174,
0.11594489216804504,
-0.0018197501776739955,
-4.973048461920371e-8,
-0.04809807986021042,
0.05760572850704193,
-0.026223966851830482,
-0.0015483418246731162,
-0.05524243786931038,
-0.08222194015979767,
-0.09367061406373978,
0.0685657188296318,
-0.01432149950414896,
0.01861981675028801,
0.040999650955200195,
0.011129084974527359,
-0.06945619732141495,
-0.05724947527050972,
0.0066899508237838745,
0.0872756689786911,
-0.00931194331496954,
0.03130956366658211,
0.02541845291852951,
0.027402957901358604,
0.0742570161819458,
0.012919149361550808,
-0.01814190112054348,
-0.06384647637605667,
0.018244605511426926,
-0.09381197392940521,
-0.03017445281147957,
0.052522290498018265,
0.1174248605966568,
-0.07363811880350113,
0.01892358437180519,
0.10999132692813873,
-0.00979361031204462,
-0.07406619936227798,
0.01872706227004528,
0.06526345759630203,
0.02720639295876026,
-0.012225048616528511,
-0.04703185334801674,
0.027177566662430763,
0.10922719538211823,
0.053558968007564545,
-0.005575529299676418,
0.013958673924207687,
0.05153372883796692,
0.04327429085969925,
0.014959241263568401,
-0.055535636842250824,
-0.03924480453133583,
-0.02802749164402485,
0.027844678610563278,
0.003491021227091551,
-0.03840890899300575,
0.1180979385972023,
0.0027557318098843098,
-0.046936649829149246,
0.07186505943536758,
-0.013806438073515892,
0.019104044884443283,
0.005811379291117191,
0.028329815715551376,
0.03766850754618645,
-0.06548615545034409,
-0.028841307386755943
] | https://github.com/lmb-freiburg/two-effects-one-trigger | 10 | 0 | 0 | 0 |
FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference | https://openreview.net/forum?id=OfjIlbelrT | [
"Xunhao Lai",
"Jianqiao Lu",
"Yao Luo",
"Yiyuan Ma",
"Xun Zhou"
] | Oral | Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold.
FlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference. | Large Language Models (LLMs), LLM inference, Long-context LLMs, Sparse Attention Mechanism | FlexPrefill is a novel sparse attention mechanism for large language models that dynamically adapts attention patterns and computational budgets in real-time to optimize performance for each input and attention head. | 1,022 | 2502.20766 | [
0.0013292825315147638,
-0.03994413837790489,
-0.04920182749629021,
0.08447371423244476,
0.05321108549833298,
0.08391931653022766,
-0.010985404253005981,
0.027987999841570854,
0.03878720849752426,
-0.05880086123943329,
-0.043614018708467484,
-0.01350377220660448,
0.000026059908122988418,
-0.019913198426365852,
-0.005212624091655016,
-0.036935243755578995,
0.05381270870566368,
0.033755429089069366,
-0.033849019557237625,
-0.03625757247209549,
0.058403708040714264,
-0.005386121571063995,
0.04623052850365639,
-0.06043165549635887,
-0.014773409813642502,
-0.01339646615087986,
-0.04960862174630165,
-0.03599220886826515,
0.054236311465501785,
0.04186767339706421,
0.007613961584866047,
0.06904753297567368,
0.09081143140792847,
0.08203022181987762,
-0.07061140984296799,
0.04414912313222885,
-0.06547772139310837,
0.024658847600221634,
-0.026435131207108498,
-0.03525843471288681,
-0.04806293919682503,
-0.02280190773308277,
-0.04905620962381363,
0.08839666098356247,
0.044438187032938004,
-0.026743963360786438,
0.019617080688476562,
-0.018175393342971802,
0.023010512813925743,
-0.010248196311295033,
-0.055662602186203,
-0.012173444963991642,
0.043780986219644547,
0.03533092513680458,
-0.030578620731830597,
0.020669613033533096,
-0.0778576210141182,
0.03659379482269287,
-0.013070137239992619,
-0.005820437800139189,
-0.08088001608848572,
-0.16605274379253387,
0.030567102134227753,
-0.0058167967945337296,
-0.01992936059832573,
0.02697073668241501,
0.042323194444179535,
0.006245007738471031,
-0.014116497710347176,
0.00015641764912288636,
-0.020719463005661964,
0.08486001193523407,
-0.014447544701397419,
0.03263275325298309,
0.04605591669678688,
0.06180984154343605,
-0.027541689574718475,
-0.037283044308423996,
0.09889781475067139,
-0.005339037626981735,
-0.03675660490989685,
-0.045045509934425354,
0.019678110256791115,
-0.05118073523044586,
0.15164470672607422,
0.03500567004084587,
0.056742798537015915,
-0.02350931242108345,
0.001142283552326262,
0.027674632146954536,
-0.035842519253492355,
-0.10380823165178299,
0.0258480217307806,
-0.03942650184035301,
-0.026585835963487625,
0.08886132389307022,
0.003489473834633827,
-0.03714464604854584,
-0.04667781665921211,
0.03129971772432327,
0.04913363605737686,
0.04011866822838783,
0.00936484057456255,
-0.008537942543625832,
-0.029440242797136307,
-0.06575077772140503,
0.0033308416604995728,
0.07345429062843323,
0.006918185856193304,
-0.05801911652088165,
-0.002394020324572921,
0.06253091245889664,
0.08346447348594666,
0.015797549858689308,
-0.0056665982119739056,
-0.04132929816842079,
0.01996413618326187,
-0.0632706955075264,
0.04869193583726883,
0.09746343642473221,
0.05936987325549126,
0.058512263000011444,
-0.05597531422972679,
-0.03479720279574394,
0.05490527302026749,
-0.0013455416774377227,
-0.0618119016289711,
2.83434169312713e-33,
0.06413900852203369,
0.08247972279787064,
0.04996241629123688,
-0.0036110938526690006,
-0.0627780631184578,
-0.016397546976804733,
0.0400061272084713,
0.07411111891269684,
-0.0701272115111351,
0.04995289817452431,
0.02114323154091835,
0.031361427158117294,
-0.012783178128302097,
0.0872926190495491,
-0.02222876064479351,
-0.01404591929167509,
-0.0071921199560165405,
0.08414395153522491,
-0.01024193037301302,
-0.03865078091621399,
-0.034410715103149414,
0.006621190346777439,
0.05334821343421936,
-0.07254798710346222,
0.045685164630413055,
-0.023565569892525673,
0.05457165837287903,
-0.0732334777712822,
0.008788791485130787,
0.03659648820757866,
-0.13004766404628754,
-0.007908991537988186,
-0.06843341886997223,
-0.018334152176976204,
0.017788754776120186,
-0.06020282208919525,
0.007236349396407604,
-0.03493526950478554,
0.026061173528432846,
0.013009399175643921,
-0.0320809967815876,
0.014596768654882908,
0.0815521702170372,
-0.030978703871369362,
-0.12594561278820038,
-0.07132109254598618,
-0.019892927259206772,
0.0567299909889698,
0.024422388523817062,
0.05621287226676941,
0.049431998282670975,
0.0019602717366069555,
-0.05633008852601051,
-0.00502318749204278,
0.011437605135142803,
-0.01578575372695923,
0.04430496692657471,
0.04949250444769859,
-0.009655284695327282,
0.09796543419361115,
-0.008266828022897243,
-0.01328252162784338,
-0.0031761438585817814,
0.08395808190107346,
0.06299428641796112,
0.02928287722170353,
-0.03035086952149868,
0.1127636730670929,
0.053503554314374924,
0.01615912653505802,
0.023718679323792458,
-0.032846055924892426,
0.016817647963762283,
-0.06476107984781265,
0.032727859914302826,
0.039561934769153595,
0.08664639294147491,
-0.10937675833702087,
0.016592975705862045,
-0.0024667128454893827,
-0.029923075810074806,
0.013249161653220654,
0.047086384147405624,
-0.029902637004852295,
-0.07484190911054611,
-0.01763291284441948,
0.07626764476299286,
-0.044003039598464966,
-0.03328093886375427,
-0.11710268259048462,
0.04048645868897438,
-0.003362085437402129,
-0.05111565813422203,
-0.08493366092443466,
-0.023775575682520866,
-1.2431356936819598e-33,
0.05326998978853226,
-0.028861500322818756,
-0.02645270712673664,
0.002224432537332177,
-0.009292887523770332,
-0.004759134724736214,
-0.008583585731685162,
0.017171943560242653,
-0.04310931637883186,
-0.08453014492988586,
0.0047550443559885025,
-0.05861305072903633,
0.10704075545072556,
-0.039602071046829224,
0.02925664186477661,
0.053414177149534225,
-0.04175660014152527,
0.03021232597529888,
-0.002256934531033039,
0.08198946714401245,
-0.016798613592982292,
0.012420585379004478,
-0.06303740292787552,
-0.005207202862948179,
-0.06087631732225418,
-0.0033669706899672747,
-0.059339288622140884,
0.057939764112234116,
-0.02133028209209442,
-0.05082884803414345,
-0.10814499109983444,
-0.009080257266759872,
-0.04420975223183632,
-0.07732336223125458,
-0.07317893952131271,
0.011182715184986591,
0.0006854331004433334,
0.006284528411924839,
-0.04817057400941849,
0.08285683393478394,
0.09016592055559158,
-0.020487679168581963,
-0.012745561078190804,
-0.006537342444062233,
-0.06642472743988037,
-0.043627578765153885,
-0.07414650917053223,
-0.0013070899294689298,
0.007692619692534208,
0.05725745111703873,
-0.04708195850253105,
0.017980871722102165,
-0.033085897564888,
0.011523690074682236,
-0.06359055638313293,
-0.049250103533267975,
-0.10307901352643967,
-0.0028847106732428074,
0.019782062619924545,
-0.06453923881053925,
-0.009590054862201214,
0.05168754607439041,
0.004833925981074572,
-0.07790980488061905,
0.1109461560845375,
-0.04598401114344597,
0.04207638278603554,
-0.12715350091457367,
0.011823076754808426,
-0.07136692851781845,
0.07060116529464722,
0.01844032295048237,
0.02906961739063263,
-0.002347515197470784,
0.048613909631967545,
0.06643541157245636,
0.027804836630821228,
-0.09508218616247177,
-0.04137991741299629,
0.01160628441721201,
-0.002702662255614996,
0.08449672162532806,
-0.05228295177221298,
0.013593236915767193,
-0.025037232786417007,
0.15577800571918488,
0.06265069544315338,
0.09139542281627655,
0.0347701720893383,
0.014910993166267872,
-0.057547930628061295,
0.03222107142210007,
0.04584761708974838,
0.05501869320869446,
-0.05045066773891449,
-4.097323014207177e-8,
-0.08292830735445023,
0.059697363525629044,
-0.056997813284397125,
0.03576577827334404,
0.06396158039569855,
-0.08720265328884125,
-0.009220171719789505,
0.09441995620727539,
-0.02743532881140709,
-0.05613337457180023,
0.018648521974682808,
0.04708891361951828,
0.005323304794728756,
-0.05846264958381653,
0.006381465587764978,
0.07227535545825958,
0.01828225702047348,
0.005466277711093426,
-0.052550058811903,
-0.0687679871916771,
-0.0025114223826676607,
0.0495823509991169,
-0.005614462308585644,
0.015676574781537056,
0.053053099662065506,
-0.036924753338098526,
-0.025934064760804176,
0.08385962247848511,
0.12367269396781921,
-0.0344235859811306,
-0.030065003782510757,
-0.017343202605843544,
-0.04476824402809143,
-0.07371405512094498,
-0.021734466776251793,
0.04541783779859543,
0.048840850591659546,
-0.015256263315677643,
-0.022140396758913994,
-0.014455383643507957,
0.04786289855837822,
-0.03737638518214226,
-0.0029793379362672567,
0.044534772634506226,
-0.04479910433292389,
0.021262990310788155,
-0.02019476518034935,
-0.051170170307159424,
0.07385274767875671,
-0.09425157308578491,
0.025635432451963425,
-0.048668984323740005,
-0.0031521772034466267,
0.020136114209890366,
-0.0050003123469650745,
0.07768256962299347,
0.037475600838661194,
0.03236259147524834,
0.03346201777458191,
0.018042150884866714,
0.040938444435596466,
0.028694577515125275,
-0.033807575702667236,
0.041078198701143265
] | https://github.com/bytedance/FlexPrefill | 97 | 0 | 0 | 0 |
REGENT: A Retrieval-Augmented Generalist Agent That Can Act In-Context in New Environments | https://openreview.net/forum?id=NxyfSW6mLK | [
"Kaustubh Sridhar",
"Souradeep Dutta",
"Dinesh Jayaraman",
"Insup Lee"
] | Oral | Building generalist agents that can rapidly adapt to new environments is a key challenge for deploying AI in the digital and real worlds. Is scaling current agent architectures the most effective way to build generalist agents? We propose a novel approach to pre-train relatively small policies on relatively small datasets and adapt them to unseen environments via in-context learning, without any finetuning. Our key idea is that retrieval offers a powerful bias for fast adaptation. Indeed, we demonstrate that even a simple retrieval-based 1-nearest neighbor agent offers a surprisingly strong baseline for today's state-of-the-art generalist agents. From this starting point, we construct a semi-parametric agent, REGENT, that trains a transformer-based policy on sequences of queries and retrieved neighbors. REGENT can generalize to unseen robotics and game-playing environments via retrieval augmentation and in-context learning, achieving this with up to 3x fewer parameters and up to an order-of-magnitude fewer pre-training datapoints, significantly outperforming today's state-of-the-art generalist agents. | Generalist Agent, Retrieval, In-Context Learning, VLA, Imitation Learning, Reinforcement Learning | We propose a retrieval-augmented generalist agent that can adapt to new environments via in-context learning | 961 | 2412.04759 | [
-0.06906356662511826,
-0.08283304423093796,
-0.014713240787386894,
0.0028555728495121002,
0.06883330643177032,
0.025359241291880608,
0.021162664517760277,
-0.014648092910647392,
-0.06794021278619766,
0.001651493483223021,
-0.026993807405233383,
-0.03523176908493042,
0.05305951461195946,
0.042933814227581024,
0.06641917675733566,
-0.01652056910097599,
0.15209804475307465,
0.06346862018108368,
-0.09580868482589722,
-0.013407654128968716,
-0.029474224895238876,
0.048223987221717834,
0.04006041958928108,
-0.05912246182560921,
-0.021303536370396614,
-0.03236572444438934,
0.05316784977912903,
-0.0948827862739563,
0.01647595874965191,
-0.007606062572449446,
0.004690511152148247,
-0.015132344327867031,
-0.053704634308815,
0.042348653078079224,
0.019176432862877846,
0.08597708493471146,
-0.08406071364879608,
0.06779743731021881,
-0.02717711217701435,
-0.05526124686002731,
0.02454105205833912,
0.012031849473714828,
-0.010351174511015415,
0.03954359143972397,
0.0294069591909647,
0.03194577246904373,
-0.008340708911418915,
-0.01711456850171089,
0.0702391192317009,
-0.024700893089175224,
-0.07119192183017731,
-0.01680731400847435,
-0.0063565075397491455,
0.02694670855998993,
-0.05524159222841263,
0.034942615777254105,
0.035409510135650635,
0.008661003783345222,
0.004017608240246773,
-0.030917400494217873,
0.01688786782324314,
-0.106163389980793,
0.008240529336035252,
-0.021145496517419815,
-0.025797955691814423,
-0.06836406886577606,
0.0406370684504509,
0.006381419952958822,
0.022695718333125114,
-0.05053368955850601,
0.0015604201471433043,
0.04058659076690674,
0.023088062182068825,
0.032471638172864914,
0.04130108654499054,
0.008287555538117886,
-0.03637387976050377,
-0.05254180729389191,
0.03521407023072243,
-0.022705895826220512,
-0.0038200076669454575,
-0.05640852078795433,
0.010047602467238903,
0.1359800398349762,
0.09307979047298431,
-0.0020238650031387806,
0.01872042380273342,
-0.0047666155733168125,
0.04225672408938408,
0.016492726281285286,
-0.017302948981523514,
-0.025930991396307945,
0.06891661137342453,
0.001741735264658928,
0.04938727989792824,
0.05272107198834419,
0.010337747633457184,
-0.1529301553964615,
-0.05681515857577324,
0.13972027599811554,
-0.03196300193667412,
0.06224185228347778,
-0.02786494418978691,
0.02752581611275673,
-0.028071677312254906,
0.04125217720866203,
0.029120327904820442,
0.036359649151563644,
0.023456256836652756,
-0.05527551472187042,
0.007843171246349812,
0.014030086807906628,
0.050288327038288116,
-0.04271473363041878,
0.025316502898931503,
-0.017068246379494667,
0.04449455440044403,
0.024395927786827087,
-0.025610176846385002,
0.039569687098264694,
0.06615213304758072,
0.016574855893850327,
-0.02186567336320877,
0.03247827664017677,
0.07892675697803497,
0.06785506010055542,
-0.02304164133965969,
-1.834332485835286e-33,
0.11293977499008179,
0.06263907998800278,
0.1005578264594078,
0.021517908200621605,
-0.013367479667067528,
-0.017314275726675987,
0.05556124448776245,
0.05254663899540901,
-0.04335402324795723,
0.009555757977068424,
0.019539540633559227,
0.09805773943662643,
-0.07672568410634995,
0.09332892298698425,
-0.00865806732326746,
0.007711364887654781,
-0.06470496952533722,
0.04491341859102249,
0.03273322060704231,
-0.006108982488512993,
0.002159731462597847,
-0.021666323766112328,
0.026122791692614555,
-0.028334828093647957,
0.03498182073235512,
-0.027614176273345947,
0.005913008935749531,
-0.04525570571422577,
0.07032015174627304,
-0.027866320684552193,
-0.1128130778670311,
0.010383500717580318,
-0.0011213257675990462,
0.05471831560134888,
0.0020085894502699375,
-0.013146228156983852,
-0.05531057342886925,
-0.06328041851520538,
0.02006971649825573,
0.013966196216642857,
-0.009586342610418797,
0.03565960004925728,
0.06363997608423233,
-0.05171250179409981,
-0.10509191453456879,
-0.12137454748153687,
0.026474783197045326,
0.027327869087457657,
-0.04294168949127197,
-0.025562671944499016,
0.05279738828539848,
0.023188339546322823,
-0.07186296582221985,
-0.04796585813164711,
0.007808747235685587,
-0.037025872617959976,
0.022018026560544968,
0.031247666105628014,
0.012835172936320305,
-0.02148498222231865,
0.08298854529857635,
0.0032976295333355665,
0.04016928747296333,
0.14489130675792694,
-0.0023869844153523445,
0.011298006400465965,
0.03913050517439842,
0.0815417617559433,
0.08406375348567963,
-0.02352435141801834,
-0.03862186521291733,
0.024141734465956688,
0.01744845323264599,
-0.03167833015322685,
0.05112303048372269,
0.023082954809069633,
0.03058166429400444,
-0.12299981713294983,
0.030720772221684456,
-0.04137260839343071,
0.008546989411115646,
-0.040335893630981445,
-0.07145483791828156,
0.014456872828304768,
-0.06498395651578903,
-0.07611270248889923,
0.02124909684062004,
-0.07746636122465134,
0.015779266133904457,
-0.060428161174058914,
-0.08897125720977783,
-0.017277389764785767,
-0.1056651696562767,
0.010042538866400719,
-0.015188376419246197,
5.983817258662391e-34,
0.010474630631506443,
-0.1108233779668808,
-0.05243382975459099,
0.01167573593556881,
-0.0017061014659702778,
-0.05516444146633148,
-0.055400650948286057,
-0.03451669588685036,
-0.043389879167079926,
-0.04016939178109169,
-0.09130319207906723,
0.02065187133848667,
0.08809508383274078,
-0.02946392633020878,
-0.030905326828360558,
0.0030544763430953026,
-0.007930074818432331,
-0.0016898729372769594,
-0.034021906554698944,
-0.022190911695361137,
0.08156121522188187,
0.08020149171352386,
-0.08189785480499268,
0.0033630940597504377,
-0.005700866691768169,
-0.00618008291348815,
-0.020342862233519554,
0.06736353784799576,
0.0032641368452459574,
0.061491552740335464,
-0.03975199535489082,
-0.025182189419865608,
0.02511373534798622,
-0.013530263677239418,
-0.04455270990729332,
0.12442448735237122,
0.04381261020898819,
-0.009602607227861881,
-0.12605628371238708,
0.11351149529218674,
0.017836514860391617,
-0.10963888466358185,
-0.06367097795009613,
0.01917414367198944,
-0.01847034879028797,
0.019651047885417938,
-0.035438403487205505,
0.031488217413425446,
-0.004772912710905075,
0.015958644449710846,
-0.01523626409471035,
0.06960885226726532,
-0.07910313457250595,
-0.08953993022441864,
-0.02176591381430626,
-0.05674087628722191,
-0.03195582702755928,
0.018474997952580452,
0.054926548153162,
0.0018104100599884987,
-0.05900437757372856,
-0.03562070056796074,
-0.07508228719234467,
0.04952908679842949,
-0.037530068308115005,
-0.028066933155059814,
-0.024069149047136307,
0.042656004428863525,
-0.014094612561166286,
-0.020349768921732903,
-0.005940412636846304,
-0.050286032259464264,
-0.0035556009970605373,
-0.05907992646098137,
-0.017261063680052757,
0.04760025069117546,
0.03677654638886452,
-0.005343619268387556,
0.032829999923706055,
-0.0016940805362537503,
-0.00031379767460748553,
0.02087075635790825,
0.037993308156728745,
0.062108661979436874,
0.05386493727564812,
0.09583363682031631,
0.038551151752471924,
0.0515260212123394,
0.023385299369692802,
0.03212101384997368,
-0.0542297288775444,
0.030227134004235268,
-0.03406963124871254,
0.020719332620501518,
-0.0941556990146637,
-3.910176005206267e-8,
-0.084989532828331,
0.10078296810388565,
0.04219639301300049,
0.03458573669195175,
-0.026285400614142418,
-0.027096794918179512,
-0.02542489394545555,
0.09807290136814117,
-0.07544557750225067,
-0.01142524741590023,
0.041644468903541565,
0.0000746817167964764,
0.06893790513277054,
-0.03712768852710724,
0.058187294751405716,
0.029906053096055984,
0.004097724333405495,
-0.04754725471138954,
-0.026521828025579453,
-0.0038197359535843134,
0.021001897752285004,
0.042975060641765594,
-0.01255618967115879,
-0.06427794694900513,
-0.013609049841761589,
-0.06212973594665527,
-0.03893202915787697,
0.03733314946293831,
0.06910324841737747,
0.07388775050640106,
0.008432270027697086,
-0.003490232862532139,
0.07303844392299652,
-0.04948592185974121,
-0.00030811349279247224,
0.037552934139966965,
-0.03196737542748451,
-0.024918297305703163,
-0.006090614944696426,
-0.09471205621957779,
-0.033709973096847534,
0.06530268490314484,
-0.02773279696702957,
-0.007884647697210312,
-0.023112526163458824,
0.031598541885614395,
-0.0025992575101554394,
-0.10563084483146667,
0.04030181095004082,
-0.03457818552851677,
-0.028283433988690376,
-0.05478690564632416,
0.00794167723506689,
0.045942798256874084,
0.13922765851020813,
0.03786221519112587,
0.0013026864035055041,
-0.022523250430822372,
0.0603986531496048,
0.07105287164449692,
-0.07533928006887436,
0.02572454698383808,
-0.1244472861289978,
0.03599132224917412
] | 0 | 0 | 0 | 0 |
|
MMIE: Massive Multimodal Interleaved Comprehension Benchmark for Large Vision-Language Models | https://openreview.net/forum?id=HnhNRrLPwm | [
"Peng Xia",
"Siwei Han",
"Shi Qiu",
"Yiyang Zhou",
"Zhaoyang Wang",
"Wenhao Zheng",
"Zhaorun Chen",
"Chenhang Cui",
"Mingyu Ding",
"Linjie Li",
"Lijuan Wang",
"Huaxiu Yao"
] | Oral | Interleaved multimodal comprehension and generation, enabling models to produce and interpret both images and text in arbitrary sequences, have become a pivotal area in multimodal learning. Despite significant advancements, the evaluation of this capability remains insufficient. Existing benchmarks suffer from limitations in data scale, scope, and evaluation depth, while current evaluation metrics are often costly or biased, lacking in reliability for practical applications. To address these challenges, we introduce MMIE, a large-scale knowledge-intensive benchmark for evaluating interleaved multimodal comprehension and generation in Large Vision-Language Models (LVLMs). MMIE comprises 20K meticulously curated multimodal queries, spanning 3 categories, 12 fields, and 102 subfields, including mathematics, coding, physics, literature, health, and arts. It supports both interleaved inputs and outputs, offering a mix of multiple-choice and open-ended question formats to evaluate diverse competencies. Moreover, we propose a reliable automated evaluation metric, leveraging a scoring model fine-tuned with human-annotated data and systematic evaluation criteria, aimed at reducing bias and improving evaluation accuracy. Extensive experiments demonstrate the effectiveness of our benchmark and metrics in providing a comprehensive evaluation of interleaved LVLMs. Specifically, we evaluate eight LVLMs, revealing that even the best models show significant room for improvement, with most achieving only moderate results. We believe MMIE will drive further advancements in the development of interleaved LVLMs. | large vision-language model, interleaved text-and-image evaluation | null | 944 | 2410.10139 | [
0.059920910745859146,
-0.1139662116765976,
0.0025939862243831158,
0.02466413378715515,
0.07945513725280762,
0.02202664688229561,
-0.05320969596505165,
0.027842003852128983,
0.04698674753308296,
-0.012205911800265312,
-0.04634024575352669,
-0.16053535044193268,
0.006686648819595575,
0.07121086865663528,
-0.020790478214621544,
0.031237924471497536,
-0.022170720621943474,
0.05064159259200096,
-0.11186574399471283,
-0.020770682021975517,
0.07096340507268906,
0.05744220316410065,
0.08830451965332031,
-0.08042358607053757,
0.04769260063767433,
0.0042114234529435635,
-0.02162887342274189,
-0.072786346077919,
-0.0025670102331787348,
-0.0055851717479527,
0.002635406795889139,
-0.026251638308167458,
-0.023838357999920845,
0.06978322565555573,
0.018559472635388374,
0.06016945466399193,
-0.07811427116394043,
0.02416408248245716,
0.0071282642893493176,
-0.0327509343624115,
-0.01408949214965105,
-0.019453927874565125,
0.022752907127141953,
-0.01695065014064312,
0.09308280795812607,
-0.06782055646181107,
-0.05701475962996483,
-0.05631881579756737,
0.008395574986934662,
0.038150325417518616,
-0.07505341619253159,
-0.021698597818613052,
-0.03225415199995041,
0.054152123630046844,
-0.002199179260060191,
0.05502442643046379,
-0.045853711664676666,
-0.013407574966549873,
-0.010674020275473595,
-0.04327458143234253,
-0.06120181828737259,
-0.1143554076552391,
0.008349482901394367,
0.02380872145295143,
-0.008746202103793621,
-0.054726168513298035,
0.018650704994797707,
-0.015940362587571144,
-0.02048555761575699,
0.004166445694863796,
-0.05715488642454147,
0.05163285881280899,
0.014586839824914932,
0.052275463938713074,
0.06043893098831177,
0.10760597884654999,
0.04698646441102028,
-0.07608481496572495,
0.049713779240846634,
-0.049253612756729126,
0.029134945943951607,
-0.008843670599162579,
0.07978807389736176,
-0.031818270683288574,
0.13968369364738464,
0.029981980100274086,
0.03150247782468796,
0.06917259097099304,
0.0017021623207256198,
0.020459050312638283,
-0.05920332670211792,
-0.040099598467350006,
0.01977699249982834,
0.013672013767063618,
0.1014254093170166,
0.04502248764038086,
0.0011715063592419028,
-0.07906990498304367,
-0.06267052888870239,
0.0428449846804142,
0.06924810260534286,
0.07533044368028641,
-0.015025167725980282,
-0.054933641105890274,
-0.008202563039958477,
-0.01067473366856575,
0.10213402658700943,
0.013914414681494236,
-0.0041557070799171925,
-0.08952569961547852,
0.042383912950754166,
0.021234046667814255,
0.03046589344739914,
0.0262495968490839,
0.026937050744891167,
0.03150905296206474,
0.0213884599506855,
0.046906910836696625,
0.07225613296031952,
0.04227251932024956,
-0.033878546208143234,
0.0013379724696278572,
0.007497426122426987,
-0.025321481749415398,
0.01457150187343359,
-0.026766130700707436,
-0.013035055249929428,
8.795629133188074e-34,
0.010223064571619034,
-0.005769344978034496,
0.07872296124696732,
0.03831152245402336,
0.06763927638530731,
-0.044201646000146866,
-0.045357123017311096,
0.10313770174980164,
-0.07780174165964127,
-0.007515438832342625,
-0.041547033935785294,
0.019315719604492188,
-0.09705359488725662,
0.12603913247585297,
-0.0016787665663287044,
-0.018327007070183754,
-0.11464496701955795,
0.043064601719379425,
-0.04858667030930519,
0.06597668677568436,
0.03528079763054848,
-0.05653250962495804,
-0.01983480155467987,
-0.003434382379055023,
0.021684784442186356,
0.03132891654968262,
0.058911968022584915,
-0.07663161307573318,
-0.05748426169157028,
0.0020062190014868975,
-0.1625470221042633,
-0.02651647850871086,
-0.009982386603951454,
-0.03934061899781227,
0.03403487801551819,
-0.04706628993153572,
0.020143955945968628,
-0.009381134063005447,
-0.025427892804145813,
0.00347495567984879,
-0.030307896435260773,
-0.0240316279232502,
0.013659789226949215,
0.020612318068742752,
-0.08804880827665329,
-0.017158985137939453,
-0.010727408342063427,
0.062277402728796005,
-0.059491947293281555,
-0.043588943779468536,
0.03286541625857353,
0.02092924900352955,
-0.05264601111412048,
-0.08509030193090439,
0.03146732598543167,
-0.012671300210058689,
0.040135011076927185,
0.04434267058968544,
0.029098613187670708,
0.008344599977135658,
0.020832328125834465,
0.07394620776176453,
0.00530312629416585,
0.11326234042644501,
0.032349709421396255,
0.016133859753608704,
-0.05830669030547142,
0.05029982700943947,
0.0686643198132515,
0.012264077551662922,
-0.04891592636704445,
0.011546044610440731,
-0.00216613058000803,
-0.09662047773599625,
0.013410476967692375,
0.02522827498614788,
0.032402437180280685,
-0.05732949450612068,
0.03412166237831116,
0.07025374472141266,
0.03677351772785187,
0.010828386060893536,
-0.0018152380362153053,
-0.12911631166934967,
-0.033278048038482666,
-0.050783611834049225,
0.053297076374292374,
-0.12915541231632233,
0.003361067734658718,
-0.07549019157886505,
-0.006341876462101936,
0.013143178075551987,
-0.016305530443787575,
-0.06157943233847618,
0.004073664546012878,
-1.3750698161432065e-35,
-0.004110952373594046,
0.012359851971268654,
-0.021189136430621147,
0.08622454851865768,
0.00518434215337038,
-0.05333330109715462,
0.00853839609771967,
0.03954987972974777,
0.04849155619740486,
-0.09530706703662872,
0.006022103130817413,
-0.014905814081430435,
-0.02443879283964634,
-0.021745240315794945,
0.055425938218832016,
-0.015123961493372917,
-0.018553799018263817,
-0.009121047332882881,
0.011383689008653164,
0.05478927493095398,
0.01445184089243412,
0.12378240376710892,
-0.006392082665115595,
0.012324532493948936,
-0.12549404799938202,
0.1167958453297615,
-0.05671176314353943,
-0.05313171446323395,
-0.07405868917703629,
0.0069145746529102325,
0.022992154583334923,
-0.049338534474372864,
0.032243285328149796,
-0.02631961926817894,
-0.026192914694547653,
0.043748099356889725,
0.10475751012563705,
-0.1298247128725052,
-0.06399178504943848,
0.11379662901163101,
0.09529060125350952,
-0.008294293656945229,
-0.0740576907992363,
-0.0041267154738307,
-0.06081051006913185,
-0.027593405917286873,
-0.048968154937028885,
-0.05661459639668465,
0.025911064818501472,
0.09659827500581741,
-0.040456030517816544,
0.006322561297565699,
-0.10658570379018784,
-0.019910113885998726,
-0.029052412137389183,
-0.08328749984502792,
-0.06829546391963959,
-0.032409876585006714,
0.03620562702417374,
-0.0053603872656822205,
-0.030703900381922722,
-0.004932637792080641,
-0.029979117214679718,
-0.025675714015960693,
0.046711649745702744,
-0.013988841325044632,
-0.02897936850786209,
-0.03166988119482994,
-0.00624329736456275,
0.01569119095802307,
-0.049885671585798264,
-0.05047135800123215,
0.03336464613676071,
0.0031045195646584034,
0.03433237597346306,
0.013093234971165657,
0.054566338658332825,
-0.006933937780559063,
0.02597212791442871,
-0.043262701481580734,
0.030739927664399147,
-0.04977161064743996,
0.0021522713359445333,
0.1250234693288803,
0.033421628177165985,
0.06676862388849258,
0.0613473579287529,
0.08470515161752701,
0.0581279918551445,
0.05127371475100517,
0.0003478490689303726,
0.0011972790816798806,
0.02376343496143818,
0.05901927500963211,
-0.05052866414189339,
-3.7175198031036416e-8,
-0.04084639251232147,
0.0946493074297905,
-0.04009019210934639,
0.03779396787285805,
-0.010605348274111748,
-0.0581345297396183,
-0.10423076897859573,
0.03441478684544563,
0.012685923837125301,
0.01047392375767231,
0.022503774613142014,
-0.005488985683768988,
-0.032267842441797256,
-0.03693088889122009,
0.03484763205051422,
0.0565706267952919,
0.001612008665688336,
0.07464557141065598,
0.013420664705336094,
-0.04114465042948723,
0.06135556474328041,
0.052881017327308655,
0.019869722425937653,
0.00880231149494648,
-0.037138860672712326,
0.024919331073760986,
-0.06234090402722359,
0.04887673631310463,
0.043090060353279114,
-0.03752759471535683,
0.020408201962709427,
0.009879741817712784,
0.009189101867377758,
-0.03902769461274147,
-0.02051592245697975,
0.03788790851831436,
-0.004204135853797197,
-0.013558004051446915,
0.007273419760167599,
0.023557966575026512,
0.009399307891726494,
0.030392024666070938,
0.016739407554268837,
0.03573468700051308,
-0.0032364653889089823,
-0.04399777576327324,
-0.04965971037745476,
-0.020444152876734734,
0.004406224936246872,
-0.008036844432353973,
-0.05865676328539848,
-0.04507957398891449,
0.017590204253792763,
0.0880822017788887,
0.06488590687513351,
-0.05702674761414528,
0.011058537289500237,
0.059210844337940216,
0.015795893967151642,
0.05660032108426094,
0.022403638809919357,
0.016707872971892357,
-0.05927126109600067,
0.004614026285707951
] | https://github.com/Lillianwei-h/MMIE | 33 | 0 | 0 | 0 |
Do as We Do, Not as You Think: the Conformity of Large Language Models | https://openreview.net/forum?id=st77ShxP1K | [
"Zhiyuan Weng",
"Guikun Chen",
"Wenguan Wang"
] | Oral | Recent advancements in large language models (LLMs) revolutionize the field of intelligent agents, enabling collaborative multi-agent systems capable of tackling complex problems across various domains. However, the potential of conformity within these systems, analogous to phenomena like conformity bias and group-think in human group dynamics, remains largely unexplored, raising concerns about their collective problem-solving capabilities and possible ethical implications. This paper presents a comprehensive study on conformity in LLM-driven multi-agent systems, focusing on three aspects: the existence of conformity, the factors influencing conformity, and potential mitigation strategies. In particular, we introduce BenchForm, a new conformity-oriented benchmark, featuring reasoning-intensive tasks and five distinct interaction protocols designed to probe LLMs’ behavior in collaborative scenarios. Several representative LLMs are evaluated on BenchForm, using metrics such as conformity rate and independence rate to quantify conformity’s impact. Our analysis delves into factors influencing conformity, including interaction time and majority size, and examines how the subject agent rationalize its conforming behavior. Furthermore, we explore two strategies to mitigate conformity effects, i.e., developing enhanced persona and implementing a reflection mechanism. Several interesting findings regarding LLMs’ conformity are derived from empirical results and case studies. We hope that these insights can pave the way for more robust and ethically-aligned collaborative AI systems. Our benchmark and code are available at BenchForm. | Large Language Models, Conformity, Multi-agent System | null | 934 | 2501.13381 | [
-0.005781632848083973,
-0.05690199136734009,
-0.05209159478545189,
-0.02528863586485386,
0.027912562713027,
-0.053811945021152496,
-0.04889007285237312,
-0.021795259788632393,
0.08951134979724884,
-0.007957988418638706,
-0.06116344407200813,
-0.041439857333898544,
0.035939108580350876,
0.06501995772123337,
0.06317169964313507,
-0.02481999062001705,
0.06344690918922424,
0.0045339185744524,
-0.14320601522922516,
-0.05597696825861931,
0.00485689751803875,
-0.027761954814195633,
-0.011341196484863758,
-0.019689850509166718,
-0.07469804584980011,
-0.015695534646511078,
-0.0517263300716877,
-0.11873798072338104,
-0.00557672418653965,
0.005074457731097937,
-0.05103006958961487,
0.076143279671669,
0.02045002393424511,
0.027257559821009636,
-0.07575850933790207,
0.06530063599348068,
0.011796760372817516,
-0.004376192111521959,
0.010217701084911823,
-0.02292519435286522,
-0.05793223902583122,
0.010364099405705929,
0.00277214078232646,
-0.03799205273389816,
0.020448530092835426,
-0.03903376683592796,
-0.03583591803908348,
0.029138099402189255,
-0.07606422901153564,
0.005288618616759777,
-0.1000240370631218,
-0.07173775136470795,
0.04985525831580162,
0.04076695069670677,
0.02124546840786934,
-0.034860774874687195,
0.03647010028362274,
0.07145067304372787,
-0.001101061818189919,
-0.11536680161952972,
-0.07100511342287064,
-0.08218028396368027,
0.027163363993167877,
0.0560714416205883,
0.017799900844693184,
0.06443551927804947,
0.007031710352748632,
0.0442153662443161,
-0.11985325068235397,
0.030507026240229607,
-0.034362103790044785,
-0.00503574451431632,
-0.03232327103614807,
-0.015878036618232727,
0.04496189206838608,
0.00043468730291351676,
-0.03119323030114174,
-0.042601920664310455,
0.0734235867857933,
-0.09525449573993683,
-0.013272418640553951,
0.026963235810399055,
0.028316985815763474,
0.07240621000528336,
0.07156066596508026,
-0.019997375085949898,
0.00266016717068851,
0.04789772257208824,
0.09821825474500656,
0.058520521968603134,
-0.03225652873516083,
-0.0014649132499471307,
0.03126551955938339,
0.040331676602363586,
0.05588432401418686,
0.07823523879051208,
-0.046231918036937714,
0.016675205901265144,
0.015561380423605442,
0.07523713260889053,
0.004752508830279112,
0.11412728577852249,
-0.0010459671029821038,
-0.031854622066020966,
-0.021128347143530846,
-0.022237742319703102,
0.06801814585924149,
-0.01017615757882595,
0.01689421758055687,
-0.04254668951034546,
-0.017539121210575104,
0.01937786117196083,
0.0710967630147934,
0.018079617992043495,
0.031964756548404694,
-0.09499318152666092,
0.12187688052654266,
0.020309584215283394,
-0.003909836523234844,
0.007647944148629904,
0.030357951298356056,
-0.018709233030676842,
0.08465787023305893,
0.028116386383771896,
0.10059021413326263,
0.0772600844502449,
-0.029739342629909515,
3.565592805626926e-34,
0.020095380023121834,
0.015153499320149422,
0.008893119171261787,
0.079891636967659,
0.09546267241239548,
-0.034141283482313156,
-0.010950853116810322,
0.051429007202386856,
-0.08387359976768494,
0.009105263277888298,
0.022919999435544014,
0.058308109641075134,
0.04962143674492836,
-0.019061677157878876,
0.030509505420923233,
0.01115609984844923,
-0.0035085726995021105,
0.0636296421289444,
-0.10324418544769287,
-0.08840535581111908,
0.058216534554958344,
-0.007482842542231083,
0.03023499622941017,
-0.012448610737919807,
0.015211801044642925,
-0.01698893867433071,
-0.0032920376397669315,
-0.0263921320438385,
0.07426170259714127,
-0.006559740751981735,
-0.08589417487382889,
-0.011429467238485813,
-0.033679962158203125,
0.07516913115978241,
-0.07352355867624283,
0.03464095667004585,
-0.0468730553984642,
-0.05013227462768555,
0.07281803339719772,
0.01610402949154377,
-0.01804199628531933,
0.03232065215706825,
0.01511058397591114,
-0.045224953442811966,
-0.030068285763263702,
-0.03248806670308113,
0.03791197016835213,
0.03665275126695633,
-0.04217154160141945,
-0.06369663774967194,
0.022245004773139954,
0.067179836332798,
0.005705210845917463,
-0.027826393023133278,
0.06278268992900848,
0.006147532258182764,
-0.010760258883237839,
0.0008139855344779789,
-0.0382244735956192,
0.019971853122115135,
-0.01441627461463213,
0.0447227880358696,
-0.015628937631845474,
0.05495043098926544,
0.07940168678760529,
-0.05528085678815842,
-0.10286913067102432,
0.04497218877077103,
0.08861307799816132,
-0.0000314763055939693,
0.03387710824608803,
0.040137987583875656,
-0.07740256935358047,
0.05453234165906906,
-0.04068077355623245,
-0.035195037722587585,
0.04176419600844383,
-0.033598870038986206,
0.08800166100263596,
-0.0023821115028113127,
-0.05044517293572426,
-0.0298151858150959,
-0.036419786512851715,
-0.02229880727827549,
-0.07572975754737854,
-0.06506744027137756,
-0.03889966756105423,
-0.032326873391866684,
-0.04227014631032944,
0.005157109349966049,
-0.07025732845067978,
-0.019806379452347755,
0.013505632989108562,
0.0015776511281728745,
-0.01967969909310341,
-3.012796245590346e-33,
-0.08125465363264084,
-0.03557811677455902,
-0.025968801230192184,
0.03918027505278587,
0.02347084879875183,
-0.05579199641942978,
-0.05061223357915878,
-0.04205415025353432,
0.018883056938648224,
-0.050153519958257675,
-0.021511906757950783,
-0.04022641479969025,
0.08358969539403915,
-0.0069175297394394875,
0.05082867667078972,
-0.08469223976135254,
0.036019422113895416,
0.02603038027882576,
0.07430052012205124,
0.004108932800590992,
0.0482577383518219,
0.01876484975218773,
-0.010932470671832561,
0.06255179643630981,
-0.006547511555254459,
0.013008627109229565,
-0.01758408546447754,
-0.08062020689249039,
0.07622960209846497,
0.05024123191833496,
-0.03232698142528534,
-0.05189676210284233,
-0.056791409850120544,
-0.013446792028844357,
-0.024701636284589767,
0.07946459203958511,
-0.06769809871912003,
0.03601142391562462,
-0.019177304580807686,
0.021690644323825836,
0.0392785519361496,
-0.06623208522796631,
-0.11247590184211731,
-0.027032386511564255,
-0.017486797645688057,
-0.002560908906161785,
-0.061163995414972305,
-0.019807279109954834,
-0.005514710675925016,
-0.01796334981918335,
-0.006801714189350605,
-0.03285208344459534,
0.05256999656558037,
-0.041807226836681366,
-0.09510131925344467,
0.024242684245109558,
0.038277141749858856,
-0.05994521826505661,
0.0413682647049427,
0.0020277334842830896,
-0.013705642893910408,
0.019976194947957993,
0.048789896070957184,
-0.014989431947469711,
-0.051888130605220795,
0.03598325327038765,
-0.016829252243041992,
0.08964995294809341,
0.0340450219810009,
0.019414765760302544,
0.06273813545703888,
-0.056704360991716385,
-0.05846614018082619,
-0.0017064310377463698,
0.10204661637544632,
0.028354834765195847,
-0.04277490824460983,
-0.09302831441164017,
-0.02161092311143875,
-0.047809045761823654,
0.053918782621622086,
0.08927919715642929,
0.03344322368502617,
0.03123602271080017,
-0.04063630849123001,
-0.03657655417919159,
0.011731572449207306,
0.14926202595233917,
0.02445550635457039,
0.05262579396367073,
0.01894940622150898,
-0.02271227166056633,
-0.002081295009702444,
0.05570749565958977,
-0.12373149394989014,
-4.712534362738552e-8,
-0.02468843199312687,
0.0064852251671254635,
0.05498697608709335,
0.041032277047634125,
-0.006336862687021494,
-0.035622403025627136,
-0.059031084179878235,
-0.051493722945451736,
-0.0028349095955491066,
0.0942041426897049,
0.028405211865901947,
-0.030996356159448624,
-0.04209534451365471,
-0.00874530803412199,
0.07930357754230499,
0.06485186517238617,
-0.0729416087269783,
-0.049252405762672424,
-0.0657491683959961,
-0.050946664065122604,
0.003683842485770583,
-0.009102076292037964,
-0.05678178742527962,
-0.02066137082874775,
0.0225672647356987,
-0.034044791013002396,
-0.11502676457166672,
0.06221088021993637,
-0.014750050380825996,
0.00005186941780266352,
-0.033652786165475845,
0.05367802456021309,
-0.03639320284128189,
0.050785429775714874,
0.003942734561860561,
-0.008180481381714344,
-0.09999293088912964,
-0.044671542942523956,
0.0681527853012085,
0.03171670436859131,
0.04362531751394272,
0.07068701833486557,
0.0071884142234921455,
0.05899382010102272,
0.07103197276592255,
-0.03269920125603676,
-0.10989893972873688,
-0.03702973574399948,
0.003063908778131008,
-0.03377813845872879,
-0.059581175446510315,
-0.035574063658714294,
-0.030514884740114212,
0.1272592842578888,
0.06178692728281021,
0.056751564145088196,
0.019184337928891182,
0.012220007367432117,
0.10407385230064392,
-0.019995441660284996,
0.04815715178847313,
0.09457394480705261,
-0.030173301696777344,
-0.059004608541727066
] | https://github.com/zhiyuan-weng/benchform | 14 | 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.