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
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
AdvPaint: Protecting Images from Inpainting Manipulation via Adversarial Attention Disruption | https://openreview.net/forum?id=m73tETvFkX | [
"Joonsung Jeon",
"Woo Jae Kim",
"Suhyeon Ha",
"Sooel Son",
"Sung-eui Yoon"
] | Poster | The outstanding capability of diffusion models in generating high-quality images poses significant threats when misused by adversaries. In particular, we assume malicious adversaries exploiting diffusion models for inpainting tasks, such as replacing a specific region with a celebrity. While existing methods for protecting images from manipulation in diffusion-based generative models have primarily focused on image-to-image and text-to-image tasks, the challenge of preventing unauthorized inpainting has been rarely addressed, often resulting in suboptimal protection performance. To mitigate inpainting abuses, we propose ADVPAINT, a novel defensive framework that generates adversarial perturbations that effectively disrupt the adversary’s inpainting tasks. ADVPAINT targets the self- and cross-attention blocks in a target diffusion inpainting model to distract semantic understanding and prompt interactions during image generation. ADVPAINT also employs a two-stage perturbation strategy, dividing the perturbation region based on an enlarged bounding box around the object, enhancing robustness across diverse masks of varying shapes and sizes. Our experimental results demonstrate that ADVPAINT’s perturbations are highly effective in disrupting the adversary’s inpainting tasks, outperforming existing methods; ADVPAINT attains over a 100-point increase in FID and substantial decreases in precision. | Adversarial Example, Adversarial Attack, Inpainting, Image Protection | null | 13,119 | 2503.10081 | [
-0.023467592895030975,
0.0013463417999446392,
0.0048040347173810005,
0.08721697330474854,
0.1321849375963211,
-0.04221610724925995,
-0.008522252552211285,
-0.06480365246534348,
0.06504695862531662,
-0.046693768352270126,
-0.007185041904449463,
-0.041715022176504135,
0.056858841329813004,
-0.004835710860788822,
-0.0239825788885355,
-0.007748928386718035,
0.062187500298023224,
0.08624810725450516,
-0.10640321671962738,
-0.01336468756198883,
0.047591302543878555,
-0.034353334456682205,
0.0688486397266388,
-0.02259465865790844,
-0.03181268647313118,
-0.031172990798950195,
-0.03930596634745598,
-0.07759403437376022,
0.003252932569012046,
-0.04829353839159012,
0.020972900092601776,
-0.055419664829969406,
-0.046890739351511,
0.0817289873957634,
-0.12373658269643784,
0.047835178673267365,
-0.08247698098421097,
0.07054276019334793,
0.0510353222489357,
-0.09884314984083176,
0.014202453196048737,
-0.040442243218421936,
-0.04506561532616615,
0.014404002577066422,
0.0414409413933754,
-0.052457962185144424,
0.031678300350904465,
-0.020012477412819862,
0.030323248356580734,
-0.02930869348347187,
-0.009677533991634846,
-0.05098208785057068,
-0.06548311561346054,
0.042615484446287155,
-0.06888758391141891,
-0.0009897589916363358,
0.05286305397748947,
0.051349371671676636,
0.033037152141332626,
0.022741029039025307,
-0.04043775796890259,
-0.03180763125419617,
0.01570562645792961,
-0.015445660799741745,
-0.01151951588690281,
-0.048405081033706665,
0.04383491352200508,
0.030989887192845345,
-0.04758407548069954,
0.048313207924366,
0.08043602108955383,
0.046091046184301376,
-0.03919617459177971,
0.004546677228063345,
0.10738503187894821,
0.06957152485847473,
-0.026720115914940834,
-0.02830064669251442,
0.04203800484538078,
-0.06363857537508011,
0.13160406053066254,
-0.015275704674422741,
0.0760316327214241,
0.016793955117464066,
0.09106892347335815,
0.03657986596226692,
-0.032731495797634125,
0.031078411266207695,
0.09418096393346786,
-0.00855084415525198,
-0.11728211492300034,
-0.07732739299535751,
0.0017924564890563488,
0.057657334953546524,
0.08655262738466263,
-0.016611795872449875,
-0.0635828971862793,
-0.08094453811645508,
-0.045937467366456985,
0.04495539143681526,
0.014457987621426582,
-0.040334511548280716,
-0.06451061367988586,
-0.00842504296451807,
0.10395321995019913,
-0.03285866603255272,
0.08815141767263412,
-0.027689876034855843,
-0.01711389608681202,
0.024136727675795555,
0.049703214317560196,
0.04798838868737221,
0.03786827623844147,
-0.026576165109872818,
0.05147145688533783,
0.0036249745171517134,
-0.01977705769240856,
0.027767116203904152,
0.0232979878783226,
-0.015158703550696373,
0.045072443783283234,
0.036045096814632416,
0.006661153864115477,
0.020984722301363945,
-0.0028364108875393867,
-0.08697747439146042,
-0.012318826280534267,
7.099769147195684e-34,
-0.0041898987255990505,
-0.015433317981660366,
0.026297446340322495,
0.02908032387495041,
0.07468477636575699,
-0.028252316638827324,
-0.026825351640582085,
0.0035048373974859715,
-0.04684493690729141,
-0.04769430309534073,
-0.005922823678702116,
-0.0598330944776535,
-0.05300605669617653,
0.12036528438329697,
0.07608486711978912,
0.02289300225675106,
-0.00023228304053191096,
0.03081607259809971,
-0.005292551591992378,
0.030604174360632896,
-0.03068852238357067,
-0.050279952585697174,
-0.04255910962820053,
0.06054706498980522,
0.004511163104325533,
-0.019576027989387512,
0.02282196469604969,
-0.08452121913433075,
0.04285706207156181,
0.04901505261659622,
-0.15684236586093903,
0.02803991176187992,
0.019306549802422523,
0.07559408992528915,
-0.03936128318309784,
0.014943077228963375,
-0.045552220195531845,
-0.05317512899637222,
0.03625066205859184,
0.019876232370734215,
0.011159620247781277,
0.07416976243257523,
0.0014648025389760733,
0.020732440054416656,
-0.054180748760700226,
0.001436230493709445,
0.0584811195731163,
0.1341010481119156,
-0.10235888510942459,
0.0037955823354423046,
0.019968055188655853,
0.032806701958179474,
0.017322441563010216,
-0.08873382210731506,
-0.11875691264867783,
-0.020342634990811348,
0.0807313472032547,
-0.04822421818971634,
0.01198606938123703,
-0.0002572437224444002,
0.034904420375823975,
0.05277971178293228,
-0.009405150078237057,
0.10319690406322479,
-0.026826372370123863,
-0.06760097295045853,
-0.02073088474571705,
0.02107842080295086,
-0.019784005358815193,
0.020233256742358208,
-0.08704511821269989,
0.02924029342830181,
-0.024458998814225197,
0.0010209963656961918,
-0.03983260691165924,
0.007285523694008589,
0.05489688366651535,
0.01642119325697422,
0.0841745063662529,
0.007323306053876877,
-0.02315942756831646,
-0.057521600276231766,
0.020342553034424782,
-0.06692907214164734,
-0.10970032960176468,
-0.03024390898644924,
0.025527942925691605,
-0.06428855657577515,
-0.015691975131630898,
-0.03276930749416351,
0.0027137198485434055,
0.01486615464091301,
-0.04667235165834427,
0.03829485550522804,
-0.02492060884833336,
-9.511404173793073e-34,
-0.014120480976998806,
0.06535866111516953,
-0.03614843264222145,
0.019460586830973625,
-0.06846745312213898,
-0.001217738608829677,
0.02635466493666172,
0.03224317356944084,
0.05612625926733017,
-0.0728994756937027,
-0.10058829933404922,
0.005866933614015579,
0.019617823883891106,
-0.06098116934299469,
-0.024572227150201797,
-0.08327198773622513,
0.016382351517677307,
0.09653250873088837,
-0.12083961814641953,
-0.021514102816581726,
0.03370758518576622,
0.08053179085254669,
0.007199003826826811,
0.025115113705396652,
-0.045120030641555786,
0.012478379532694817,
-0.04701578617095947,
0.023413825780153275,
0.010562704876065254,
-0.0039655608125030994,
-0.052893925458192825,
0.053124330937862396,
0.03306528553366661,
-0.030745994299650192,
-0.08092913031578064,
0.07210397720336914,
0.05329369008541107,
-0.00921225268393755,
-0.02469645068049431,
0.05084184557199478,
0.05789632722735405,
-0.03320473060011864,
-0.10132399201393127,
0.029372617602348328,
-0.09268726408481598,
-0.056002527475357056,
-0.028905736282467842,
-0.012628634460270405,
0.02452237345278263,
0.05268301069736481,
-0.010822094976902008,
-0.01774490810930729,
-0.007853548973798752,
0.012486554682254791,
-0.07449852675199509,
-0.03661402314901352,
-0.07459429651498795,
0.015480952337384224,
0.11834599077701569,
-0.02646775357425213,
-0.010707170702517033,
-0.007339905947446823,
-0.03768696263432503,
-0.06891312450170517,
-0.015691682696342468,
0.0012411599745973945,
-0.037687238305807114,
-0.003441344713792205,
0.06719290465116501,
0.0023217687848955393,
0.07192222028970718,
-0.02160748839378357,
0.03729410097002983,
0.03897248953580856,
0.02286934293806553,
-0.0352543368935585,
0.011621818877756596,
-0.05081803724169731,
-0.009172087535262108,
-0.03693101182579994,
0.03830787539482117,
-0.05567650869488716,
0.021444717422127724,
0.03562268614768982,
0.01481269858777523,
0.036802273243665695,
0.02502432093024254,
0.018478846177458763,
-0.023590004071593285,
0.05283178761601448,
-0.006026013288646936,
0.0012750898022204638,
-0.018920211121439934,
0.04900650680065155,
0.027325255796313286,
-3.520797520195629e-8,
-0.00437067961320281,
0.004802688956260681,
-0.0174112506210804,
0.04263576120138168,
0.004251479171216488,
-0.0333985798060894,
-0.06006814166903496,
0.04948753863573074,
0.025830458849668503,
-0.02308199740946293,
0.027838733047246933,
-0.012586424127221107,
0.0012067839270457625,
-0.12324810028076172,
-0.0564558170735836,
0.05923406407237053,
-0.008027606643736362,
0.06404993683099747,
-0.03993098810315132,
-0.0271308533847332,
-0.08629301935434341,
0.003986028954386711,
-0.03368997573852539,
-0.11057747900485992,
0.0377289280295372,
-0.049899354577064514,
-0.05495133250951767,
-0.002813669154420495,
0.039363425225019455,
0.001025254256092012,
0.005883470643311739,
-0.03844930976629257,
-0.0002473342465236783,
0.0658566802740097,
-0.04497363045811653,
0.14681865274906158,
0.002238450339064002,
-0.030995288863778114,
-0.007193483412265778,
-0.03692765161395073,
0.048695314675569534,
-0.012505275197327137,
0.0699966698884964,
-0.010831783525645733,
-0.03286924213171005,
0.07108275592327118,
0.050748616456985474,
-0.07430113852024078,
0.11102592945098877,
0.017360122874379158,
0.020233307033777237,
-0.009432964026927948,
0.019544700160622597,
0.11879372596740723,
-0.020489446818828583,
-0.03986858204007149,
0.064704030752182,
0.03627151995897293,
0.07131746411323547,
0.08783300966024399,
0.06893271207809448,
0.021990099921822548,
0.03666871041059494,
-0.01695728860795498
] | https://github.com/joonsungjeon/advpaint | 7 | 0 | 0 | 0 |
Mastering Task Arithmetic: τJp as a Key Indicator for Weight Disentanglement | https://openreview.net/forum?id=1VwWi6zbxs | [
"Kotaro Yoshida",
"Yuji Naraki",
"Takafumi Horie",
"Ryosuke Yamaki",
"Ryotaro Shimizu",
"Yuki Saito",
"Julian McAuley",
"Hiroki Naganuma"
] | Poster | Model-editing techniques using task arithmetic have rapidly gained attention.
Through task arithmetic, simply through arithmetic operations on the weights of pre-trained and fine-tuned models create desired models, such as multi-task models, models in which specific tasks are unsolvable, or domain-transferred models.
However, task arithmetic faces challenges, such as poor reproducibility and the high cost associated with adjusting coefficients in the arithmetic operations on model parameters, which have limited its practical success. In this paper, we present three key contributions in the context of task addition and task negation within task arithmetic. First, we propose a new metric called $\tau$Jp which is based on the product of the task vector ($\tau$) and the Jacobian of the pre-trained model with respect to its weights. We show that $\tau$Jp has a causal relationship with the interference that occurs from arithmetic operations. Second, we show that introducing regularization to minimize $\tau$Jp significantly mitigates interference between task inference, which leads to the elimination of coefficient tuning and improved accuracy on each task.
Third, in the context of incremental learning, we demonstrate that our $\tau$Jp regularization achieves more robust performance in environments where access to future tasks is unavailable, thus validating the scalability of the approach.
Finally, we demonstrate that the $\tau$Jp regularizer further reinforces the performance of task arithmetic by leveraging publicly available fine-tuned models, offering practical benefits for real-world applications.
Our code is available at https://github.com/katoro8989/tau-Jp_Task_Arithmetic | task arithmetic, model editing, task vector | Our proposed $\tau$Jp regularizer improve the performance of task arithmetic and lead to its practical applications. | 13,117 | null | [
-0.061212535947561264,
-0.02127782441675663,
0.03991280123591423,
0.01798759214580059,
0.008786534890532494,
0.004098673816770315,
0.01992678828537464,
-0.027314087375998497,
0.0717761367559433,
0.013357494957745075,
-0.027723463252186775,
-0.021861683577299118,
0.04629994183778763,
0.06354522705078125,
-0.05384443700313568,
-0.0516754649579525,
0.10154593735933304,
0.07971972227096558,
-0.1340421438217163,
-0.09513898938894272,
0.013253824785351753,
-0.00635790778324008,
0.021004322916269302,
0.023490386083722115,
0.04020866006612778,
0.0017938000382855535,
-0.04063981771469116,
-0.06633930653333664,
0.01425580121576786,
0.0010949013521894813,
-0.07959058880805969,
-0.04470574110746384,
0.002141093136742711,
-0.02775595709681511,
-0.10131200402975082,
0.08450431376695633,
-0.02134438045322895,
0.0734829530119896,
-0.004329105839133263,
-0.03758527338504791,
-0.0035834971349686384,
-0.02395557053387165,
0.0023381831124424934,
0.04969720542430878,
0.025753580033779144,
0.0009447272168472409,
-0.034431155771017075,
-0.037827104330062866,
-0.012471463531255722,
0.006087479647248983,
-0.014325705356895924,
0.015637004747986794,
-0.056010693311691284,
0.03662026673555374,
0.02707839012145996,
0.018213504925370216,
0.07777464389801025,
0.008034178987145424,
-0.022334719076752663,
-0.08830231428146362,
-0.07654627412557602,
-0.05021938681602478,
-0.07545216381549835,
-0.07150590419769287,
-0.005970777943730354,
-0.015244183130562305,
-0.030612105503678322,
0.029447903856635094,
-0.024730287492275238,
0.0972428023815155,
-0.011551338247954845,
0.10589667409658432,
-0.019899418577551842,
0.010408738628029823,
-0.03346806392073631,
-0.0004611785407178104,
0.009365290403366089,
0.018484007567167282,
0.045569226145744324,
-0.019663242623209953,
0.010263586416840553,
-0.028844282031059265,
-0.006604312919080257,
-0.04116840288043022,
0.19615864753723145,
-0.012390420772135258,
0.007959874346852303,
0.013048539869487286,
0.14240913093090057,
-0.03458936884999275,
-0.043791722506284714,
-0.00005080656410427764,
-0.004196722526103258,
-0.0005583715974353254,
-0.012565018609166145,
-0.005472068674862385,
0.017107419669628143,
0.05967530608177185,
-0.06260775029659271,
0.10822898149490356,
0.04568146914243698,
0.03272716701030731,
-0.0002147593186236918,
0.002483759308233857,
0.041985295712947845,
0.010155286639928818,
0.005611160304397345,
0.04036097228527069,
0.0024124684277921915,
-0.14245910942554474,
0.01750006154179573,
-0.028056534007191658,
0.01120062917470932,
0.03491370752453804,
-0.011777698062360287,
-0.03591563180088997,
0.07093226909637451,
-0.0030796027276664972,
-0.03242775797843933,
0.033129919320344925,
0.035241805016994476,
0.00857151485979557,
0.036282435059547424,
0.03384951874613762,
-0.029962508007884026,
-0.021405186504125595,
-0.0832420065999031,
1.896799725005921e-33,
0.045874740928411484,
0.07513327151536942,
0.03645126149058342,
0.011648123152554035,
0.02612404339015484,
0.02179625630378723,
-0.05781107768416405,
0.030627459287643433,
0.0014147459296509624,
0.0034215187188237906,
0.002258959459140897,
0.00485250586643815,
-0.0906396359205246,
0.07901439070701599,
0.11442255228757858,
0.023171046748757362,
0.017553966492414474,
0.016915367916226387,
0.01975599303841591,
-0.025367004796862602,
0.03037779964506626,
-0.07525703310966492,
-0.0742432177066803,
-0.019444134086370468,
0.003572911024093628,
-0.0022183842957019806,
0.03340110182762146,
-0.09440864622592926,
-0.07981014251708984,
0.01292706560343504,
-0.06741973012685776,
-0.045146871358156204,
0.04526740685105324,
0.06026527285575867,
-0.020251678302884102,
-0.05152856931090355,
0.047903142869472504,
-0.06465902179479599,
0.07278299331665039,
-0.05979807302355766,
0.023578938096761703,
-0.0035663468297570944,
0.014400172978639603,
-0.03983280435204506,
-0.06605375558137894,
-0.07663600146770477,
0.059387579560279846,
0.026669751852750778,
-0.019100641831755638,
-0.055743828415870667,
0.060071974992752075,
0.05434505268931389,
0.0032110849861055613,
-0.09956254810094833,
-0.07715708762407303,
0.0177016444504261,
0.058477915823459625,
0.01781492494046688,
-0.004390336107462645,
0.07447529584169388,
0.002425504382699728,
0.08372455090284348,
-0.03788316622376442,
0.10409803688526154,
-0.02146431989967823,
0.056164465844631195,
-0.010517019778490067,
0.09716706722974777,
0.07770747691392899,
-0.013482071459293365,
-0.07435378432273865,
-0.0005504392902366817,
0.003117564832791686,
0.031891096383333206,
0.1341831535100937,
-0.013467026874423027,
0.05739831551909447,
-0.16263654828071594,
0.00928069744259119,
0.009537403471767902,
0.019597411155700684,
0.018911568447947502,
-0.009445075877010822,
-0.041052937507629395,
-0.02001594938337803,
-0.08909603208303452,
0.022462816908955574,
-0.009490549564361572,
-0.02853189781308174,
-0.03939129412174225,
0.038036152720451355,
0.05398224666714668,
-0.06645244359970093,
0.025600222870707512,
0.008055627346038818,
-1.195231635656563e-33,
-0.05673179402947426,
-0.03037738800048828,
-0.0611429288983345,
0.04163553565740585,
-0.025586044415831566,
-0.0007552167517133057,
0.043117377907037735,
-0.07180546224117279,
0.0007823145133443177,
-0.11023027449846268,
-0.024674581363797188,
-0.05252733454108238,
-0.06400406360626221,
0.04209696501493454,
-0.000005686915756086819,
0.008915923535823822,
-0.052481863647699356,
0.05471361428499222,
-0.03003154695034027,
-0.0504857562482357,
0.024795038625597954,
0.0969289168715477,
-0.06054643914103508,
0.024407515302300453,
-0.020791804417967796,
0.01773817464709282,
-0.052070967853069305,
0.08172297477722168,
0.008504432626068592,
0.026348795741796494,
-0.04364888370037079,
-0.023250125348567963,
0.022726142778992653,
-0.03152679651975632,
-0.047426488250494,
0.007365991827100515,
0.06641550362110138,
0.011384997516870499,
-0.04241885617375374,
0.027475276961922646,
0.048180919140577316,
-0.04935341328382492,
-0.0006944366032257676,
0.05579303577542305,
0.045842744410037994,
-0.002621991792693734,
-0.09782756119966507,
-0.03289290517568588,
-0.037970490753650665,
-0.04678422585129738,
-0.022110946476459503,
0.004764084238559008,
-0.06202420964837074,
0.051172636449337006,
-0.05798589438199997,
-0.05109986662864685,
-0.004441423807293177,
-0.06375736743211746,
0.10051308572292328,
-0.0037014116533100605,
-0.03767700493335724,
0.019377319142222404,
0.03325720876455307,
-0.002371580805629492,
0.04193338379263878,
0.024630224332213402,
0.024607622995972633,
0.011729702353477478,
0.02696821838617325,
0.03126688674092293,
-0.01327673252671957,
0.025080177932977676,
0.042090289294719696,
-0.025338998064398766,
-0.022054195404052734,
-0.05584197863936424,
0.05910393223166466,
0.007596142590045929,
-0.010987501591444016,
-0.032494332641363144,
-0.025729231536388397,
-0.04214050993323326,
0.041967153549194336,
-0.007520031649619341,
-0.07875943183898926,
0.048249751329422,
0.06550994515419006,
0.10552632808685303,
0.003969782032072544,
0.005152233876287937,
-0.053263768553733826,
0.06245027855038643,
0.018602708354592323,
0.03667711839079857,
-0.0644037202000618,
-4.794278041231337e-8,
-0.0361490398645401,
0.027456054463982582,
-0.011281349696218967,
-0.021308349445462227,
0.019776344299316406,
-0.009874610230326653,
-0.05199478194117546,
0.04799366369843483,
0.04058094695210457,
-0.01423933170735836,
0.0462634451687336,
-0.012436043471097946,
-0.020113306120038033,
-0.027073021978139877,
-0.10082566738128662,
0.08404524624347687,
0.059602465480566025,
0.028597252443432808,
0.0023394301533699036,
-0.05407578498125076,
0.06770698726177216,
0.11240344494581223,
-0.02043243497610092,
-0.02025236189365387,
0.05324225872755051,
-0.12506389617919922,
-0.019190117716789246,
0.15508653223514557,
0.12085709720849991,
0.024983109906315804,
0.010602371767163277,
0.015456425957381725,
0.007571752183139324,
0.05393281951546669,
-0.013291660696268082,
0.021906962618231773,
0.1015474796295166,
0.007233085110783577,
-0.01208692230284214,
-0.05303617939352989,
0.05259888991713524,
0.05841388553380966,
-0.056672997772693634,
0.0469866581261158,
0.09004051238298416,
-0.045262329280376434,
-0.08244037628173828,
-0.13241726160049438,
0.023970404639840126,
-0.04609764739871025,
0.07860597223043442,
0.05641484260559082,
0.002187176840379834,
0.0392439030110836,
-0.010758117772638798,
0.03006020560860634,
0.033146172761917114,
-0.07163123041391373,
-0.03791769593954086,
0.04759860411286354,
0.0488019622862339,
0.0041382466442883015,
-0.04960693418979645,
0.028216177597641945
] | 0 | 0 | 0 | 0 |
|
Generating CAD Code with Vision-Language Models for 3D Designs | https://openreview.net/forum?id=BLWaTeucYX | [
"Kamel Alrashedy",
"Pradyumna Tambwekar",
"Zulfiqar Haider Zaidi",
"Megan Langwasser",
"Wei Xu",
"Matthew Gombolay"
] | Poster | Generative AI has transformed the fields of Design and Manufacturing by providing
efficient and automated methods for generating and modifying 3D objects. One
approach involves using Large Language Models (LLMs) to generate Computer-
Aided Design (CAD) scripting code, which can then be executed to render a 3D
object; however, the resulting 3D object may not meet the specified requirements.
Testing the correctness of CAD generated code is challenging due to the complexity
and structure of 3D objects (e.g., shapes, surfaces, and dimensions) that are not
feasible in code. In this paper, we introduce CADCodeVerify, a novel approach to
iteratively verify and improve 3D objects generated from CAD code. Our approach
works by producing ameliorative feedback by prompting a Vision-Language Model
(VLM) to generate and answer a set of validation questions to verify the generated
object and prompt the VLM to correct deviations. To evaluate CADCodeVerify, we
introduce, CADPrompt, the first benchmark for CAD code generation, consisting of
200 natural language prompts paired with expert-annotated scripting code for 3D
objects to benchmark progress. Our findings show that CADCodeVerify improves
VLM performance by providing visual feedback, enhancing the structure of the 3D
objects, and increasing the success rate of the compiled program. When applied to
GPT-4, CADCodeVerify achieved a 7.30% reduction in Point Cloud distance and a
5.0% improvement in success rate compared to prior work. | Code Generation, Self-refinement | we develop a Self-Verification approach, wherein LLMs generate 3D objects through scripting code, and then verify whether the generated 3D objects meet the requirements. | 13,116 | 2410.05340 | [
-0.05616603046655655,
-0.01811128854751587,
0.0033673434518277645,
0.03683460131287575,
0.013162609189748764,
-0.08955145627260208,
-0.037913840264081955,
-0.013676327653229237,
-0.04188466817140579,
-0.0011998021509498358,
-0.09334636479616165,
-0.1453821361064911,
0.054627254605293274,
0.03687871992588043,
-0.02500302717089653,
-0.011310435831546783,
0.06785464286804199,
0.06524757295846939,
-0.061981528997421265,
0.05188388377428055,
0.04682469740509987,
-0.012567216530442238,
-0.0050315107218921185,
-0.007196739781647921,
0.033921368420124054,
0.013182936236262321,
0.07630757987499237,
-0.07470883429050446,
0.02404019609093666,
-0.007444585207849741,
0.025077370926737785,
-0.007395969238132238,
-0.057824913412332535,
0.04065573215484619,
0.07996129989624023,
0.022292234003543854,
-0.03222700208425522,
0.0317772775888443,
-0.0023168136831372976,
-0.08264889568090439,
-0.0843953937292099,
0.026005703955888748,
0.03303124010562897,
-0.011643809266388416,
0.1180543527007103,
0.01728149689733982,
-0.0487842783331871,
-0.05307189002633095,
-0.032918717712163925,
-0.01720518432557583,
-0.040326304733753204,
-0.08896614611148834,
-0.035131119191646576,
-0.08136893808841705,
-0.018214482814073563,
0.0770624652504921,
0.0826987773180008,
-0.04706038162112236,
0.023839574307203293,
-0.0006831082282587886,
-0.0011141131399199367,
-0.03172069042921066,
0.013188274577260017,
-0.02554154023528099,
-0.025148892775177956,
-0.03781872242689133,
-0.019355295225977898,
0.00021639675833284855,
-0.010081560350954533,
0.019014446064829826,
-0.011565730907022953,
0.06419682502746582,
0.037793055176734924,
0.047817815095186234,
0.06965292990207672,
0.028056887909770012,
0.008575170300900936,
-0.032078973948955536,
0.01728256605565548,
-0.13551433384418488,
-0.026184216141700745,
0.06292534619569778,
-0.0378887876868248,
0.05770738050341606,
0.03884992003440857,
0.07224881649017334,
0.01939459890127182,
0.15113060176372528,
0.004615685436874628,
0.03699616715312004,
-0.03190179541707039,
-0.011203810572624207,
-0.07051004469394684,
-0.0035405224189162254,
-0.038454025983810425,
-0.018900061026215553,
0.05862351506948471,
-0.11462442576885223,
-0.0030448727775365114,
0.028525816276669502,
-0.0325113907456398,
0.021742962300777435,
0.08097105473279953,
-0.06254519522190094,
-0.0026642177253961563,
0.04858345538377762,
0.08018838614225388,
-0.008235686458647251,
0.0010916647734120488,
-0.007535888813436031,
0.03329324722290039,
0.04004933685064316,
-0.018797652795910835,
-0.025119757279753685,
-0.025039341300725937,
0.019681237637996674,
-0.05254069343209267,
-0.0004931074217893183,
0.05311386287212372,
0.024724820628762245,
0.02809913270175457,
0.042728323489427567,
0.012487051077187061,
0.020899953320622444,
0.020337559282779694,
-0.04842915013432503,
0.024842603132128716,
-1.214024116906365e-33,
-0.010755924507975578,
0.03373612463474274,
0.019548775628209114,
0.09883274137973785,
0.13405710458755493,
-0.028032425791025162,
0.05373256281018257,
0.08037113398313522,
-0.016554592177271843,
-0.029159871861338615,
-0.06217312440276146,
-0.09508949518203735,
-0.06494207680225372,
0.136872798204422,
0.05252540484070778,
0.020498130470514297,
-0.01595531776547432,
-0.0033964712638407946,
-0.12846529483795166,
0.04423562437295914,
0.02650318667292595,
-0.09973014146089554,
-0.009065553545951843,
0.03179842233657837,
0.0038302841130644083,
0.05138060823082924,
0.039640508592128754,
0.04445475712418556,
-0.05081845074892044,
0.0024957233108580112,
-0.07225731760263443,
-0.030272917822003365,
0.0306372232735157,
0.03256366029381752,
-0.011076651513576508,
-0.007423369213938713,
0.009242512285709381,
-0.055147185921669006,
-0.016028018668293953,
0.03883010894060135,
0.004079448990523815,
0.040011681616306305,
-0.0025642041582614183,
-0.009668060578405857,
-0.05671852082014084,
0.020183250308036804,
0.027319297194480896,
0.04732077941298485,
-0.06835278123617172,
0.02679104171693325,
0.05217992141842842,
0.07839997112751007,
0.056255362927913666,
-0.028706775978207588,
-0.06631538271903992,
-0.09416284412145615,
-0.01664666086435318,
-0.02686278149485588,
0.09117624163627625,
0.02199827879667282,
-0.026797480881214142,
0.0733301118016243,
-0.062217455357313156,
0.06856972724199295,
-0.06388585269451141,
0.03381671756505966,
0.05603451281785965,
0.07266777008771896,
0.03153419867157936,
0.013824574649333954,
-0.06469245254993439,
-0.006091601215302944,
-0.049456916749477386,
-0.021471228450536728,
0.02349667251110077,
-0.02545928955078125,
0.011141850613057613,
-0.028482096269726753,
0.01034774724394083,
-0.02376404218375683,
-0.04198925197124481,
0.03793851286172867,
0.010084664449095726,
-0.06025857478380203,
0.015559828840196133,
-0.10624594241380692,
0.06436601281166077,
-0.03288327530026436,
-0.03471300005912781,
-0.013529415242373943,
-0.06519698351621628,
-0.11112135648727417,
0.03248395770788193,
-0.05447063967585564,
0.02431790716946125,
5.065618704411873e-36,
0.028233887627720833,
-0.012464604340493679,
-0.026638953015208244,
0.051269616931676865,
-0.06401742994785309,
-0.12663334608078003,
-0.08911088109016418,
-0.009595105424523354,
-0.0266816895455122,
-0.07133840024471283,
-0.013698743656277657,
0.01259777694940567,
0.028757838532328606,
0.06021759659051895,
0.028574280440807343,
0.0065484801307320595,
-0.08902619034051895,
-0.07000282406806946,
-0.004517646040767431,
0.010147078894078732,
0.03879838436841965,
0.06847086548805237,
-0.08475096523761749,
-0.050662387162446976,
-0.13339035212993622,
0.059993550181388855,
-0.07948064059019089,
0.001612532651051879,
0.009642373770475388,
0.04741273075342178,
0.03389247506856918,
0.010159909725189209,
-0.007055594585835934,
0.04416089132428169,
0.00772269070148468,
0.004623747430741787,
0.08096307516098022,
-0.03109547682106495,
0.010513256303966045,
0.05469155311584473,
0.040408920496702194,
-0.0407431535422802,
-0.04084962606430054,
0.0876014307141304,
-0.056215789169073105,
0.037004705518484116,
0.00425518536940217,
-0.017091713845729828,
-0.005255965515971184,
0.04323434457182884,
-0.045624859631061554,
0.04256965592503548,
0.0008198348805308342,
-0.04134756699204445,
0.024319887161254883,
-0.09585950523614883,
-0.03803993761539459,
-0.04276758432388306,
0.07503967732191086,
-0.011006283573806286,
-0.05848987028002739,
-0.06584925949573517,
0.04098977893590927,
-0.015164656564593315,
0.0186616163700819,
-0.0019523946102708578,
-0.04884272813796997,
-0.05545230582356453,
0.04715961590409279,
-0.003982103895395994,
0.06339817494153976,
-0.0031615409534424543,
-0.067375048995018,
0.059499092400074005,
-0.01936386525630951,
-0.061675991863012314,
0.0473700575530529,
-0.007805258966982365,
0.04405420273542404,
0.004684735089540482,
0.08632563799619675,
-0.05040151998400688,
0.08515843003988266,
0.09932910650968552,
0.013867123983800411,
0.037984780967235565,
-0.0390634648501873,
0.03357711806893349,
-0.02218318171799183,
0.08033396303653717,
-0.05252150446176529,
0.09843539446592331,
0.037142254412174225,
0.035108532756567,
-0.05349517613649368,
-3.255819791547765e-8,
-0.07308454066514969,
0.012876149266958237,
0.019546186551451683,
-0.03130492568016052,
-0.04441118240356445,
0.023922789841890335,
-0.02455572970211506,
0.050668735057115555,
-0.013609998859465122,
-0.07676069438457489,
0.07781938463449478,
-0.0636155903339386,
-0.005846961867064238,
0.013338150456547737,
0.05069119483232498,
0.053049780428409576,
-0.03757980465888977,
0.02116699516773224,
-0.02347680926322937,
-0.025511126965284348,
-0.039741020649671555,
0.0800536721944809,
0.03869704157114029,
-0.02408590167760849,
0.015345311723649502,
-0.0545770600438118,
-0.00885059405118227,
0.05870673805475235,
0.03285451978445053,
-0.03498058766126633,
0.040136173367500305,
0.022275356575846672,
0.06704524159431458,
0.05915812775492668,
0.026768038049340248,
0.01881282962858677,
0.004628578666597605,
0.012963316403329372,
0.057105403393507004,
-0.07791997492313385,
0.0690373033285141,
0.11631683260202408,
-0.041013557463884354,
-0.01916072517633438,
0.039762746542692184,
-0.052882712334394455,
0.04854727163910866,
-0.10931509733200073,
-0.045141786336898804,
-0.04196963086724281,
-0.11769187450408936,
-0.019623752683401108,
-0.05620996654033661,
0.06592392176389694,
-0.04788213223218918,
0.07490964978933334,
-0.07236094027757645,
-0.019602518528699875,
0.07937662303447723,
0.04764264449477196,
-0.019517386332154274,
0.046184323728084564,
0.02187662571668625,
-0.02665082924067974
] | 0 | 0 | 0 | 0 |
|
A Formal Framework for Understanding Length Generalization in Transformers | https://openreview.net/forum?id=U49N5V51rU | [
"Xinting Huang",
"Andy Yang",
"Satwik Bhattamishra",
"Yash Sarrof",
"Andreas Krebs",
"Hattie Zhou",
"Preetum Nakkiran",
"Michael Hahn"
] | Poster | A major challenge for transformers is generalizing to sequences longer than those observed during training. While previous works have empirically shown that transformers can either succeed or fail at length generalization depending on the task, theoretical understanding of this phenomenon remains limited. In this work, we introduce a rigorous theoretical framework to analyze length generalization in causal transformers with learnable absolute positional encodings. In particular, we characterize those functions that are identifiable in the limit from sufficiently long inputs with absolute positional encodings under an idealized inference scheme using a norm-based regularizer. This enables us to prove the possibility of length generalization for a rich family of problems. We experimentally validate the theory as a predictor of success and failure of length generalization across a range of algorithmic and formal language tasks. Our theory not only explains a broad set of empirical observations but also opens the way to provably predicting length generalization capabilities in transformers. | transformers, theory, length generalization, expressivity, analysis, algorithmic reasoning, systematic generalization, formal languages | We introduce a theoretical framework for understanding which problems transformers length-generalize on. | 13,115 | 2410.02140 | [
-0.08294325321912766,
0.01914122886955738,
0.06131984293460846,
0.0006118035526014864,
0.0718516856431961,
0.06837427616119385,
-0.01689865253865719,
-0.030260367318987846,
-0.00845863576978445,
-0.04082471504807472,
-0.03327402472496033,
0.023943215608596802,
0.04856076091527939,
0.04556605964899063,
-0.019216664135456085,
-0.01714913919568062,
-0.00907929427921772,
0.058375172317028046,
-0.12327896803617477,
-0.05630787834525108,
0.03767913579940796,
0.044344715774059296,
-0.04387260228395462,
0.024694079533219337,
0.03432003781199455,
-0.01724901795387268,
-0.03236943110823631,
-0.046197354793548584,
0.05199109762907028,
0.0009594623115845025,
-0.002471241168677807,
0.013275058940052986,
-0.035752370953559875,
0.01014899741858244,
0.0008836094639264047,
0.06860116124153137,
-0.06302371621131897,
0.0010095317848026752,
0.020123042166233063,
-0.03030388057231903,
0.04101347178220749,
0.010955561883747578,
0.03893337771296501,
0.054159633815288544,
0.014166424050927162,
0.0574050173163414,
0.062292203307151794,
-0.058829985558986664,
-0.0735926628112793,
-0.05481168255209923,
-0.07178764790296555,
-0.009050367400050163,
-0.02205309458076954,
0.10757649689912796,
0.008035310544073582,
-0.05839446559548378,
0.020594626665115356,
0.01636837050318718,
-0.04377538710832596,
0.001963687362149358,
-0.1009676456451416,
-0.1128050833940506,
-0.08200307935476303,
-0.05548323690891266,
-0.02793380804359913,
-0.01894756592810154,
0.057317622005939484,
0.08392509818077087,
-0.07728767395019531,
0.1052592322230339,
-0.09493209421634674,
0.05107094347476959,
-0.10652383416891098,
0.15569138526916504,
0.1281910389661789,
0.026721514761447906,
0.0564066581428051,
-0.004370678216218948,
-0.03773314133286476,
-0.09417849779129028,
-0.02058017998933792,
-0.0535193607211113,
0.040988292545080185,
0.04905315861105919,
0.09040157496929169,
-0.016074003651738167,
-0.013439170084893703,
0.015085268765687943,
0.026134520769119263,
0.05434383079409599,
-0.008086705580353737,
-0.12326843291521072,
0.036758195608854294,
0.03915923461318016,
0.06461425125598907,
-0.004546819720417261,
-0.0024401014670729637,
-0.05299103260040283,
0.02902057021856308,
0.10333140939474106,
-0.007688680663704872,
0.022769272327423096,
0.022971607744693756,
-0.013304540887475014,
-0.018704606220126152,
-0.03854775056242943,
-0.05484229326248169,
0.021457750350236893,
0.0026526956353336573,
-0.12333830446004868,
0.028198478743433952,
-0.03218012675642967,
0.04390588775277138,
-0.015623998828232288,
-0.003827325999736786,
-0.07565868645906448,
0.016185788437724113,
-0.03247882053256035,
-0.03554404899477959,
-0.014247484505176544,
0.042240701615810394,
0.0251605287194252,
-0.023581285029649734,
0.0702337697148323,
0.028684107586741447,
-0.000824698363430798,
0.01694243773818016,
2.019899881146931e-33,
0.011245228350162506,
0.05417004972696304,
0.04966282472014427,
0.06666696071624756,
-0.022076232358813286,
0.048655103892087936,
-0.015210689045488834,
0.04978698119521141,
0.05597827583551407,
0.03390328213572502,
0.0077247098088264465,
0.060311220586299896,
-0.05123867094516754,
0.039981111884117126,
0.05021163076162338,
0.03220618888735771,
0.028819041326642036,
0.07675351202487946,
-0.062489066272974014,
-0.05068625509738922,
0.08422114700078964,
-0.01796558126807213,
-0.035312481224536896,
-0.07450038939714432,
-0.04926242679357529,
-0.05617227405309677,
0.01644420623779297,
-0.06316682696342468,
-0.05086119472980499,
-0.03873826563358307,
-0.12757869064807892,
0.024543913081288338,
0.04414944350719452,
0.0614304393529892,
0.03281594067811966,
-0.07408278435468674,
0.004948022309690714,
-0.06579223275184631,
-0.03509240224957466,
0.015184884890913963,
-0.03206295892596245,
-0.010444301180541515,
-0.003918772097676992,
0.033545613288879395,
-0.008463677950203419,
-0.08688332140445709,
-0.005723865237087011,
0.02846066839993,
-0.05272969603538513,
-0.021671054884791374,
0.02114233374595642,
0.01573442667722702,
-0.028962289914488792,
-0.04419371485710144,
0.09183045476675034,
0.07743728905916214,
-0.015904288738965988,
0.017840703949332237,
-0.0032264527399092913,
0.00041781392064876854,
-0.03473682329058647,
0.05391514673829079,
0.022483371198177338,
0.029583346098661423,
-0.01007321197539568,
0.024693608283996582,
-0.0717933252453804,
-0.007049876730889082,
0.09477312862873077,
-0.008371811360120773,
-0.013023455627262592,
-0.024024195969104767,
-0.06251081079244614,
-0.02124345302581787,
0.06374899297952652,
-0.050483573228120804,
0.03288457170128822,
-0.1052265539765358,
-0.006323933135718107,
0.006609650794416666,
-0.04188837856054306,
-0.024388713762164116,
0.06539726257324219,
-0.027068806812167168,
-0.005307353567332029,
-0.06334661692380905,
0.023780979216098785,
-0.14128385484218597,
-0.05808083713054657,
-0.08065870404243469,
-0.04477328434586525,
-0.0009007106418721378,
-0.07163308560848236,
0.03079058788716793,
0.014171777293086052,
-4.3438156621951226e-33,
-0.05503101274371147,
0.025021426379680634,
-0.02171182632446289,
0.04890724644064903,
-0.00005202738248044625,
-0.07992463558912277,
-0.0348382368683815,
0.09687923640012741,
-0.017388273030519485,
-0.02889241836965084,
-0.01156520377844572,
-0.034890592098236084,
0.024774543941020966,
-0.012774969451129436,
0.013786344788968563,
0.00959736667573452,
-0.004807882476598024,
-0.04199380427598953,
0.02696908265352249,
0.039207153022289276,
0.07014421373605728,
0.08494608849287033,
-0.06352411210536957,
0.0028740321286022663,
-0.011925380676984787,
0.0013380099553614855,
-0.046153243631124496,
0.1223231852054596,
-0.017382845282554626,
0.030620209872722626,
-0.03660092130303383,
-0.051473263651132584,
-0.06118172034621239,
-0.020131727680563927,
-0.03546864166855812,
0.00003103741619270295,
0.030767278745770454,
0.1071186289191246,
-0.0412086546421051,
0.094367615878582,
0.039995815604925156,
-0.047201696783304214,
-0.02956010028719902,
0.04237313196063042,
-0.016454439610242844,
-0.08909376710653305,
-0.010487240739166737,
-0.016352294012904167,
0.05765005573630333,
-0.030655862763524055,
-0.016681749373674393,
0.050130460411310196,
-0.03461776673793793,
-0.008919867686927319,
-0.028116853907704353,
-0.046436380594968796,
0.0048093367367982864,
-0.041537463665008545,
0.09152989089488983,
-0.0077018216252326965,
-0.07206276059150696,
-0.02474643476307392,
0.06873118132352829,
-0.04872828349471092,
0.02186029590666294,
-0.024810930714011192,
-0.013185207732021809,
-0.03549799695611,
0.08031933754682541,
0.028809141367673874,
0.03265611454844475,
0.025329187512397766,
-0.08957025408744812,
-0.010148673318326473,
-0.07240090519189835,
0.02254425175487995,
0.016901450231671333,
-0.055037230253219604,
-0.05403025448322296,
0.00713673559948802,
0.028695669025182724,
0.008071855641901493,
0.07395042479038239,
0.012013785541057587,
-0.06839793920516968,
0.06671164184808731,
0.07203216105699539,
0.11797335743904114,
0.07767178118228912,
-0.018947124481201172,
-0.002577983308583498,
0.0616823174059391,
0.005316814873367548,
0.04463544860482216,
-0.08367270231246948,
-4.3454519982333295e-8,
-0.06420043110847473,
0.03580985590815544,
-0.047305136919021606,
0.010923881083726883,
0.09659925848245621,
-0.044713184237480164,
-0.012194225564599037,
0.07250930368900299,
-0.039126910269260406,
-0.0005843667313456535,
0.06662995368242264,
0.04014945030212402,
-0.004949474241584539,
-0.021962860599160194,
-0.016945011913776398,
0.07567699998617172,
-0.024395987391471863,
-0.03136633336544037,
-0.05468841642141342,
0.0642390325665474,
0.005177691113203764,
0.05309353768825531,
-0.051457345485687256,
-0.03012625314295292,
-0.014340223744511604,
-0.08786509186029434,
-0.03059406764805317,
0.09500686079263687,
0.06211632490158081,
0.03319495543837547,
0.015220989473164082,
0.07294738292694092,
0.005978414323180914,
-0.05527183786034584,
-0.051757100969552994,
0.06866295635700226,
0.04624566435813904,
-0.05820849537849426,
-0.018167803063988686,
-0.07919885963201523,
0.011949364095926285,
0.036938559263944626,
-0.07828385382890701,
0.05752559006214142,
0.04927804321050644,
-0.030537979677319527,
0.015448391437530518,
-0.09237830340862274,
-0.03537329286336899,
0.0002652264083735645,
0.08135630190372467,
0.0008515375084243715,
-0.01591898873448372,
0.024143945425748825,
0.09120651334524155,
0.030072201043367386,
0.014273115433752537,
-0.06176357716321945,
-0.07885223627090454,
0.03281411901116371,
-0.024698907509446144,
0.09312997758388519,
-0.00211949716322124,
0.02582373097538948
] | https://github.com/lacoco-lab/length_generalization | 0 | 0 | 0 | 0 |
PALMBENCH: A COMPREHENSIVE BENCHMARK OF COMPRESSED LARGE LANGUAGE MODELS ON MOBILE PLATFORMS | https://openreview.net/forum?id=xzSUdw6s76 | [
"Yilong Li",
"Jingyu Liu",
"Hao Zhang",
"M Badri Narayanan",
"Utkarsh Sharma",
"Shuai Zhang",
"Yijing Zeng",
"Jayaram Raghuram",
"Suman Banerjee"
] | Poster | Deploying large language models (LLMs) locally on mobile devices is advantageous in scenarios where transmitting data to remote cloud servers is either undesirable due to privacy concerns or impractical due to network connection. Recent advancements have facilitated the local deployment of LLMs. However, local deployment also presents challenges, particularly in balancing quality (generative performance), latency, and throughput within the hardware constraints of mobile devices. In this paper, we introduce our lightweight, all-in-one automated benchmarking framework that allows users to evaluate LLMs on mobile devices. We provide a comprehensive benchmark of various popular LLMs with different quantization configurations (both weights and activations) across multiple mobile platforms with varying hardware capabilities. Unlike traditional benchmarks that assess full-scale models on high-end GPU clusters, we focus on evaluating resource efficiency (memory and power consumption) and harmful output for compressed models on mobile devices. Our key observations include: i) differences in energy efficiency and throughput across mobile platforms; ii) the impact of quantization on memory usage, GPU execution time, and power consumption; and iii) accuracy and performance degradation of quantized models compared to their non-quantized counterparts; and iv) the frequency of hallucinations and toxic content generated by compressed LLMs on
mobile devices. | Mobile Platforms, Large Language Models, Quantization, Benchmark | Benchmarking compressed Large Language Models on mobile devices | 13,109 | 2410.05315 | [
-0.0376126654446125,
-0.009056419134140015,
0.02825232595205307,
0.003840901656076312,
0.05602719634771347,
-0.07016567140817642,
-0.07159971445798874,
-0.0656726211309433,
0.05360092222690582,
-0.013918845914304256,
-0.004996291361749172,
-0.06333421170711517,
0.016260460019111633,
-0.021473905071616173,
0.032706305384635925,
-0.009195619262754917,
0.08870011568069458,
0.05418666824698448,
-0.09142773598432541,
-0.03531866893172264,
0.014790721237659454,
-0.016701428219676018,
-0.002220453228801489,
-0.0008201130549423397,
0.15522433817386627,
-0.018371345475316048,
-0.012556607834994793,
-0.0141910370439291,
0.049474090337753296,
0.011887134984135628,
-0.015686875209212303,
0.021079692989587784,
-0.026180412620306015,
0.0987679660320282,
-0.0011347390245646238,
-0.012691321782767773,
0.013578204438090324,
-0.0722295343875885,
-0.06189831346273422,
-0.09448307752609253,
-0.005316041875630617,
-0.044350381940603256,
0.026842357590794563,
0.06538228690624237,
-0.020251931622624397,
0.006860804744064808,
0.014101330190896988,
0.028328528627753258,
-0.03675597906112671,
-0.03195589780807495,
0.047146014869213104,
0.0051062386482954025,
-0.05889724940061569,
0.016750765964388847,
-0.1459416002035141,
0.03753631189465523,
0.06598200649023056,
0.026689015328884125,
0.05575471743941307,
0.0865461453795433,
-0.031433455646038055,
-0.06059848144650459,
0.005382164381444454,
0.09087024629116058,
-0.030876079574227333,
-0.026248903945088387,
0.13229718804359436,
0.004810438957065344,
0.030069706961512566,
0.005981111433357,
-0.009060988202691078,
-0.010805891826748848,
0.006050179246813059,
0.07155518233776093,
-0.07477401942014694,
0.033037636429071426,
0.021326441317796707,
-0.02589188702404499,
0.02402237243950367,
-0.007516474928706884,
0.00198503490537405,
-0.05106116831302643,
0.02535928785800934,
0.021684790030121803,
0.10681826621294022,
-0.07689878344535828,
-0.01845560595393181,
0.0720735713839531,
0.011984902434051037,
-0.01734720543026924,
-0.03630514442920685,
0.08185377717018127,
0.011414802633225918,
0.050638679414987564,
0.01945279724895954,
0.009929533116519451,
0.028216494247317314,
-0.07189055532217026,
-0.07589846104383469,
0.09654687345027924,
0.009851260110735893,
0.08644681423902512,
0.051584742963314056,
-0.022624893113970757,
-0.013279254548251629,
0.004396346863359213,
0.01460326835513115,
0.03976292908191681,
-0.01704569160938263,
0.03204231709241867,
0.04513024538755417,
0.05185828357934952,
0.0022901741322129965,
-0.0036270921118557453,
0.04092118889093399,
0.0028369559440761805,
-0.006802938878536224,
-0.03913315013051033,
0.029492536559700966,
0.043739497661590576,
-0.08148852735757828,
-0.04236076399683952,
-0.03810913488268852,
-0.050611406564712524,
-0.004741040989756584,
0.0089083481580019,
-0.0943913385272026,
1.7802569906325762e-33,
0.0490092895925045,
0.022916901856660843,
0.007380439899861813,
-0.03183858096599579,
0.04759764298796654,
-0.04205724969506264,
0.07459522783756256,
0.06316118687391281,
-0.04489447548985481,
0.008136003278195858,
-0.012579706497490406,
0.009398306719958782,
-0.043868597596883774,
0.057465966790914536,
0.10806629061698914,
-0.016642030328512192,
-0.00008933447679737583,
0.051872558891773224,
0.0020481403917074203,
0.00029978776001371443,
0.04089982435107231,
-0.04095739871263504,
-0.002364296466112137,
-0.04898807778954506,
0.026733485981822014,
0.053925517946481705,
0.03786815330386162,
-0.061824701726436615,
0.01909574493765831,
0.02815161645412445,
-0.0682576522231102,
-0.09793973714113235,
0.006627324037253857,
-0.02695039100944996,
0.021677562966942787,
-0.03299243003129959,
-0.009738829918205738,
-0.05766468495130539,
0.02247917465865612,
-0.029012629762291908,
-0.050902508199214935,
-0.003983825910836458,
0.0022930342238396406,
-0.12243952602148056,
-0.14962826669216156,
-0.03954465687274933,
-0.046594467014074326,
-0.04151696339249611,
-0.0387716218829155,
-0.03431200236082077,
0.03556772693991661,
-0.006896648555994034,
-0.06251002103090286,
0.05438215658068657,
-0.007383966352790594,
-0.06505589932203293,
0.029959743842482567,
0.025908786803483963,
0.03617607429623604,
0.12650246918201447,
-0.056440919637680054,
0.016838179901242256,
-0.07183720916509628,
0.06109530106186867,
0.013816247694194317,
-0.05703018233180046,
-0.001288555795326829,
0.04072566330432892,
-0.04033900052309036,
0.051491040736436844,
-0.025930313393473625,
-0.024343501776456833,
0.07855722308158875,
-0.01334387343376875,
0.09289631992578506,
0.03161029890179634,
0.025366192683577538,
-0.037048500031232834,
-0.07973155379295349,
-0.02364080585539341,
0.00822095200419426,
0.058196768164634705,
-0.00389662547968328,
-0.01176449190825224,
-0.04606517776846886,
-0.08230029791593552,
-0.03137746825814247,
-0.07837401330471039,
-0.05298228934407234,
0.07844461500644684,
-0.08356469869613647,
-0.05100414901971817,
0.0405733585357666,
-0.051421474665403366,
-0.09102798998355865,
-2.942053893531907e-33,
-0.08280939608812332,
-0.01023675873875618,
0.01080227643251419,
0.13369128108024597,
0.029877007007598877,
-0.11070306599140167,
0.005838671699166298,
0.0344068743288517,
-0.04975294694304466,
-0.10825855284929276,
-0.08212681859731674,
-0.005745369009673595,
0.057071004062891006,
0.0023305462673306465,
0.03240039944648743,
0.014257700182497501,
-0.04893812909722328,
-0.050814978778362274,
-0.014909463003277779,
0.021235879510641098,
-0.043893564492464066,
0.07817164063453674,
0.02606368064880371,
-0.013754122890532017,
-0.010365422815084457,
-0.005181562155485153,
-0.08676651865243912,
-0.005921975243836641,
0.04154711961746216,
-0.03793247789144516,
-0.003852462163195014,
0.01773674227297306,
-0.022262075915932655,
-0.06259681284427643,
-0.0023222332820296288,
0.025539780035614967,
0.07023826241493225,
0.09660206735134125,
0.005679871421307325,
0.058711763471364975,
0.14742864668369293,
-0.07307504117488861,
-0.08298107236623764,
0.002134449779987335,
0.0531587116420269,
0.019397644326090813,
-0.1416858285665512,
-0.020613638684153557,
0.004615772981196642,
-0.0629538968205452,
0.0457206554710865,
0.05912800505757332,
0.007483722176402807,
0.03646467626094818,
0.027298301458358765,
-0.08424205332994461,
-0.07101462781429291,
-0.04620758071541786,
-0.018004247918725014,
-0.029608672484755516,
0.009190280921757221,
-0.08095669001340866,
-0.06266237795352936,
-0.06630486994981766,
-0.01289457269012928,
-0.033297449350357056,
0.006059726700186729,
0.014439179562032223,
-0.02664717100560665,
0.03307301551103592,
-0.03261403366923332,
-0.03520771861076355,
-0.00565384142100811,
0.06931417435407639,
-0.05011003091931343,
0.021317491307854652,
0.0414416566491127,
-0.08238313347101212,
0.014528707601130009,
0.015315746888518333,
0.054363567382097244,
0.13914571702480316,
-0.013203968293964863,
-0.06855085492134094,
0.07535239309072495,
0.02335408329963684,
0.04501218721270561,
0.03663632273674011,
-0.008142312057316303,
0.0031182158272713423,
-0.07021859288215637,
0.07544330507516861,
-0.04318932071328163,
0.036708299070596695,
-0.06794391572475433,
-3.8503877419771015e-8,
0.025384075939655304,
0.04997570067644119,
0.011045917868614197,
-0.00621513556689024,
-0.05145315080881119,
0.00014314395957626402,
0.020436367020010948,
0.039607834070920944,
0.08943650871515274,
0.0026119640097022057,
0.09752720594406128,
-0.07592230290174484,
-0.0373002290725708,
-0.0035465622786432505,
0.007098439149558544,
0.0879548192024231,
-0.004876554943621159,
0.0038958978839218616,
-0.008288098499178886,
-0.045735765248537064,
-0.028800511732697487,
0.12159435451030731,
-0.00585763156414032,
0.01432455237954855,
0.027407333254814148,
-0.016964025795459747,
0.02671550214290619,
0.06526919454336166,
0.09726684540510178,
0.027595767751336098,
-0.06069093197584152,
0.010216097347438335,
-0.0008945309091359377,
-0.028927529230713844,
0.028924033045768738,
0.03233518451452255,
-0.03005259670317173,
-0.017653688788414,
0.1056598350405693,
-0.0007247450994327664,
0.019892720505595207,
-0.0000127679650177015,
-0.05488305911421776,
0.004262034315615892,
0.020688733085989952,
-0.0710851177573204,
-0.044984593987464905,
0.003918944392353296,
-0.013247271068394184,
0.05868026986718178,
-0.014227615669369698,
-0.010830575600266457,
-0.09324418008327484,
0.01690979115664959,
0.0408039316534996,
0.005143801681697369,
-0.06023668497800827,
-0.05449903756380081,
0.08825308829545975,
0.025385284796357155,
0.059017427265644073,
-0.03168546408414841,
-0.03474872559309006,
-0.006634063553065062
] | 0 | 0 | 0 | 0 |
|
On Stochastic Contextual Bandits with Knapsacks in Small Budget Regime | https://openreview.net/forum?id=FCMpUOZkxi | [
"Hengquan Guo",
"Xin Liu"
] | Poster | This paper studies stochastic contextual bandits with knapsack constraints (CBwK), where a learner observes a context, takes an action, receives a reward, and incurs a vector of costs at every round. The learner aims to maximize the cumulative rewards across $T$ rounds under the knapsack constraints with an initial budget of $B$. We study CBwK in the small budget regime where the budget $B = \Omega(\sqrt{T})$
and propose an Adaptive and Universal Primal--Dual algorithm (AUPD) that achieves strong regret performance:
i) AUPD achieves $\tilde{O}((1 + \frac{\nu^*}{\delta b})\sqrt{T})$ regret under the strict feasibility assumption without any prior information, matching the best-known bounds;
ii) AUPD achieves $\tilde{O}(\sqrt{T}+ \frac{\nu^*}{\sqrt{b}}T^{\frac{3}{4}})$ regret without strict feasibility assumption,
which, to the best of our knowledge, is the first result in the literature. Here, the parameter $\nu^*$ represents the optimal average reward; $b=B/T$ is the average budget and $\delta b$ is the feasibility/safety margin.
We establish these strong results through the adaptive budget-aware design, which effectively balances reward maximization and budget consumption. We provide a new perspective on analyzing budget consumption using the Lyapunov drift method, along with a refined analysis of its cumulative variance. Our theory is further supported by experiments conducted on a large-scale dataset. | Contextual bandits with knapsacks, small budget | null | 13,108 | null | [
-0.08777470141649246,
0.022903284057974815,
-0.07501205056905746,
-0.0038354655262082815,
0.01364306453615427,
-0.014331337064504623,
0.02405409701168537,
0.02317538857460022,
0.0469307079911232,
0.04637366160750389,
-0.05586175620555878,
0.028904035687446594,
0.06781665235757828,
-0.04634334519505501,
0.02544378861784935,
0.02168806456029415,
0.09248458594083786,
0.00438415352255106,
-0.00008743281796341762,
0.027232255786657333,
-0.05343933403491974,
-0.03466275706887245,
0.04146169498562813,
-0.06518244743347168,
0.058262310922145844,
-0.07300212234258652,
-0.0007445942610502243,
0.009503157809376717,
0.035218823701143265,
0.0025966239627450705,
0.045353829860687256,
0.00430174358189106,
0.07072510570287704,
-0.06114145368337631,
0.017860544845461845,
0.013808023184537888,
-0.12082656472921371,
0.021294791251420975,
-0.02208804152905941,
0.009160034358501434,
-0.020676523447036743,
0.02665896899998188,
-0.05718114972114563,
0.02553398907184601,
0.0193308275192976,
-0.028807662427425385,
-0.023962855339050293,
0.019936826080083847,
0.026621107012033463,
-0.025967374444007874,
0.005611035972833633,
-0.004691805224865675,
-0.03087368980050087,
-0.03192659839987755,
-0.05654340237379074,
0.011342973448336124,
-0.026193656027317047,
0.00019497644098009914,
0.009646765887737274,
0.02041744627058506,
-0.007180294021964073,
-0.1316545307636261,
-0.07132773846387863,
-0.03348318487405777,
0.03218993544578552,
0.018155040219426155,
0.019690103828907013,
-0.013903587125241756,
0.029239268973469734,
0.052796173840761185,
0.031130008399486542,
-0.001085919444449246,
-0.07842651009559631,
0.0484301820397377,
0.047660961747169495,
-0.028393544256687164,
0.052841588854789734,
-0.03625421226024628,
-0.036864157766103745,
0.026894010603427887,
-0.06438353657722473,
0.027728641405701637,
-0.0019406687933951616,
-0.08930868655443192,
0.08566303551197052,
-0.07736038416624069,
0.008839312009513378,
0.005111544858664274,
0.15597829222679138,
-0.08194484561681747,
0.038491200655698776,
-0.02650557830929756,
-0.09042972326278687,
0.008576089516282082,
-0.006940808612853289,
-0.022388078272342682,
-0.0017698785522952676,
0.022071629762649536,
-0.03322751820087433,
0.08625862747430801,
0.07736123353242874,
0.007235317956656218,
-0.007788216695189476,
-0.05080704018473625,
0.046735286712646484,
-0.05623626708984375,
0.04602998495101929,
0.0380757674574852,
0.08108378201723099,
-0.060623738914728165,
0.027538277208805084,
0.07257506251335144,
0.08371013402938843,
0.031717631965875626,
-0.022403541952371597,
-0.022725025191903114,
0.007759916130453348,
-0.03266172111034393,
-0.0158824622631073,
0.06722069531679153,
-0.03181207552552223,
0.003018107730895281,
0.056251160800457,
0.0161772221326828,
-0.029979193583130836,
-0.035506729036569595,
-0.012349169701337814,
1.9417164661767945e-33,
0.032446395605802536,
0.041138503700494766,
0.029604962095618248,
-0.14503251016139984,
-0.0005436812061816454,
0.023226220160722733,
0.06965257972478867,
-0.06268772482872009,
-0.07401945441961288,
0.062273163348436356,
0.08014962077140808,
0.01721826381981373,
-0.08275102078914642,
0.010558311827480793,
0.16038653254508972,
-0.062445372343063354,
-0.013221876695752144,
0.03527921438217163,
0.11091616749763489,
-0.05731892213225365,
0.054217878729104996,
-0.013672740198671818,
-0.03160826116800308,
-0.043427418917417526,
-0.03238777816295624,
-0.06023497134447098,
0.04712533950805664,
-0.04946315288543701,
0.034544315189123154,
0.054215095937252045,
-0.0920664370059967,
0.02022751234471798,
-0.07862484455108643,
0.05029655247926712,
0.00007610380271216854,
-0.027524227276444435,
-0.10047631710767746,
0.02683320827782154,
-0.0905173122882843,
-0.02520528994500637,
-0.04008219763636589,
0.05408337339758873,
0.029812194406986237,
0.025574788451194763,
-0.09186096489429474,
-0.1273912489414215,
0.043577343225479126,
0.04195807874202728,
-0.050951018929481506,
-0.06129647418856621,
-0.04054182767868042,
-0.03190264105796814,
-0.005791733507066965,
0.010860830545425415,
0.012716970406472683,
-0.003598226699978113,
0.023870917037129402,
0.00760230515152216,
-0.001372125931084156,
0.05300195887684822,
0.10190055519342422,
-0.10253928601741791,
-0.022858895361423492,
0.01714974269270897,
-0.04077720269560814,
0.06634626537561417,
0.014989473856985569,
0.008642124012112617,
0.054515570402145386,
-0.04346807673573494,
-0.0111034344881773,
-0.009949886240065098,
0.07856259495019913,
-0.08969046175479889,
0.07816155254840851,
-0.031905174255371094,
0.1295471340417862,
0.06578321754932404,
0.001231942675076425,
-0.0859222486615181,
0.027186769992113113,
0.03072603978216648,
0.016084514558315277,
0.009402473457157612,
-0.05929668992757797,
0.023618336766958237,
0.024936765432357788,
-0.07967672497034073,
-0.041021376848220825,
-0.055475782603025436,
-0.050177134573459625,
-0.026893727481365204,
-0.07285764068365097,
0.010215367190539837,
-0.013501610606908798,
-1.58091169950334e-33,
0.08439559489488602,
-0.05511477589607239,
0.0360959954559803,
0.008871560916304588,
0.07063856720924377,
0.0025688274763524532,
-0.06709624081850052,
-0.07279569655656815,
-0.04786195233464241,
-0.027685120701789856,
-0.04610714316368103,
0.009526737034320831,
0.016147835180163383,
0.02274635247886181,
0.043395716696977615,
-0.09162691980600357,
0.08141697198152542,
0.08414110541343689,
0.0024588159285485744,
-0.024732155725359917,
-0.004468064289540052,
0.04262958839535713,
-0.061750058084726334,
0.011576218530535698,
0.025559483096003532,
0.013038390316069126,
-0.04331756383180618,
0.12928485870361328,
-0.08283628523349762,
0.04414333775639534,
-0.025299038738012314,
0.0355362631380558,
-0.05210328474640846,
0.016515448689460754,
-0.039801664650440216,
0.014118960127234459,
0.05458196625113487,
0.060242123901844025,
-0.047060202807188034,
0.12347108870744705,
0.06572239845991135,
0.010221044532954693,
0.028352659195661545,
-0.0015253617893904448,
0.036335501819849014,
-0.02379162423312664,
-0.021945960819721222,
-0.05985402315855026,
0.07247425615787506,
-0.04645106941461563,
0.05678436532616615,
0.030817318707704544,
-0.020435858517885208,
0.07617443799972534,
-0.04326728358864784,
-0.043722379952669144,
-0.011594973504543304,
-0.05620574951171875,
0.10804808139801025,
-0.023090941831469536,
-0.07466932386159897,
0.03769315034151077,
-0.03716466948390007,
0.04558295011520386,
0.009797482751309872,
-0.005507992580533028,
-0.019735140725970268,
0.016781078651547432,
0.00284752924926579,
0.00783661101013422,
-0.05412930250167847,
0.0940772071480751,
0.007677053101360798,
-0.023610683158040047,
0.018304603174328804,
0.0862513855099678,
0.05280798673629761,
0.031244555488228798,
-0.01546472404152155,
-0.03461542725563049,
-0.027076568454504013,
0.03748149052262306,
0.07609807699918747,
-0.013971437700092793,
-0.017797349020838737,
0.052915964275598526,
0.052075568586587906,
0.019122295081615448,
-0.0160097386687994,
-0.044966861605644226,
-0.003075482090935111,
-0.03418199345469475,
0.038474731147289276,
0.0495617538690567,
-0.01622677966952324,
-4.61171900667523e-8,
-0.04496803879737854,
0.09292297810316086,
-0.044218700379133224,
0.014814911410212517,
0.12141360342502594,
0.017585979774594307,
-0.06665647774934769,
-0.006376070901751518,
-0.0641581267118454,
-0.012985815294086933,
0.019153164699673653,
0.02448098361492157,
-0.011199573054909706,
-0.025773797184228897,
-0.09353243559598923,
0.031573306769132614,
-0.018532302230596542,
-0.06782504916191101,
0.012141214683651924,
-0.009035566821694374,
0.033382825553417206,
0.04845757037401199,
0.06409351527690887,
-0.04459771141409874,
-0.007524203974753618,
0.0005519898841157556,
-0.03239027038216591,
0.08089987933635712,
0.11630196869373322,
0.09690111875534058,
-0.0539456270635128,
0.013006803579628468,
-0.027892310172319412,
-0.01849380135536194,
-0.028303764760494232,
-0.025583038106560707,
-0.0206120815128088,
-0.025193605571985245,
0.0020333107095211744,
0.05082964152097702,
0.060528893023729324,
-0.03403691202402115,
-0.04994351789355278,
0.0065606739372015,
0.005059235729277134,
0.06303036212921143,
-0.043862372636795044,
-0.0542290173470974,
0.002312047639861703,
-0.051351871341466904,
0.04198606684803963,
0.006033124402165413,
0.000944384140893817,
0.060515303164720535,
0.1290983408689499,
-0.025980694219470024,
-0.022276736795902252,
-0.10774493217468262,
-0.031433943659067154,
0.0678461417555809,
0.004151279106736183,
-0.046432893723249435,
-0.05825887992978096,
0.04144059494137764
] | 0 | 0 | 0 | 0 |
|
From Few to Many: Self-Improving Many-Shot Reasoners Through Iterative Optimization and Generation | https://openreview.net/forum?id=JBXO05r4AV | [
"Xingchen Wan",
"Han Zhou",
"Ruoxi Sun",
"Sercan O Arik"
] | Poster | Recent advances in long-context large language models (LLMs) have led to the emerging paradigm of many-shot in-context learning (ICL), where it is observed that scaling many more demonstrating examples beyond the conventional few-shot setup in the context can lead to performance benefits. However, despite its promise, it is unclear what aspects dominate the benefits and whether simply scaling to more examples is the most effective way of improving many-shot ICL. In this work, we first provide an analysis on the factors driving many-shot ICL, and we find that 1) many-shot performance can still be attributed to often a few disproportionately influential examples and 2) identifying such influential examples ("optimize") and using them as demonstrations to regenerate new examples ("generate") can lead to further improvements. Inspired by the findings, we propose BRIDGE, an algorithm that alternates between the optimize step with Bayesian optimization to discover the influential sets of examples and the generate step to reuse this set to expand the reasoning paths of the examples back to the many-shot regime automatically. On Gemini, Claude, and Mistral LLMs of different sizes, we show BRIDGE led to significant improvements across a diverse set of tasks including symbolic reasoning, numerical reasoning and code generation. | many-shot, in-context learning, large language models | null | 13,102 | 2502.00330 | [
0.008937553502619267,
-0.026242880150675774,
0.03969605267047882,
0.01857423596084118,
0.14321790635585785,
0.0039670090191066265,
0.046474505215883255,
0.03570891171693802,
0.06536421924829483,
-0.015344375744462013,
-0.07226066291332245,
-0.020743902772665024,
0.06753453612327576,
0.05451773852109909,
0.03990384191274643,
0.0011251965770497918,
0.09822206944227219,
0.039782606065273285,
-0.13545078039169312,
-0.10636522620916367,
-0.06504750996828079,
-0.0514492504298687,
0.09357402473688126,
-0.03044815920293331,
-0.010659567080438137,
-0.030551908537745476,
0.01820610836148262,
0.03311461955308914,
0.09034930169582367,
0.010632311925292015,
0.012968828901648521,
0.05976724624633789,
-0.03917926922440529,
-0.0028111040592193604,
0.008684086613357067,
0.0760083794593811,
-0.04477545991539955,
0.12163244187831879,
-0.008930916897952557,
-0.021334534510970116,
-0.006673777010291815,
0.030859988182783127,
0.06799966841936111,
-0.0033392617478966713,
0.056044019758701324,
0.010647439397871494,
-0.03625723719596863,
-0.011317654512822628,
0.062166500836610794,
-0.006257889326661825,
-0.13664351403713226,
-0.019439226016402245,
-0.07409418374300003,
0.017798051238059998,
0.0030262207146734,
-0.04561048746109009,
0.052957721054553986,
0.0317835807800293,
-0.00856938585639,
-0.07929687201976776,
-0.041741881519556046,
-0.12190970778465271,
-0.07073085010051727,
0.015783095732331276,
-0.015878768637776375,
0.010806945152580738,
0.00520627386868,
0.024792542681097984,
-0.05883018672466278,
0.08115913718938828,
0.003941988572478294,
0.08797279745340347,
-0.007986988872289658,
0.07359182089567184,
-0.02715182863175869,
0.038328178226947784,
0.010456902906298637,
-0.03184022009372711,
0.015370381996035576,
-0.045419733971357346,
0.032339707016944885,
-0.010697155259549618,
-0.005039113573729992,
-0.029744036495685577,
0.09690971672534943,
0.01671171560883522,
0.010995497927069664,
-0.03314130753278732,
0.04531737416982651,
0.08632391691207886,
-0.027469931170344353,
-0.07686526328325272,
0.02977728098630905,
-0.019456347450613976,
0.05357450619339943,
0.03353404998779297,
-0.022634830325841904,
-0.07291097193956375,
-0.02222086302936077,
0.09938615560531616,
-0.00874311849474907,
0.05881527438759804,
0.031662169843912125,
-0.09942587465047836,
0.01927088014781475,
-0.005925739649683237,
-0.031716082245111465,
0.03888922184705734,
0.06493564695119858,
-0.11552207171916962,
0.0044647143222391605,
0.05946897715330124,
0.010678675957024097,
-0.021906722337007523,
0.0379086434841156,
-0.039202574640512466,
0.016440842300653458,
-0.06554507464170456,
0.0006941434112377465,
0.06616542488336563,
0.011784251779317856,
0.03187353163957596,
-0.04632752388715744,
-0.013529731892049313,
-0.04190820828080177,
-0.03962985426187515,
-0.03547193855047226,
6.1451207975341334e-34,
0.13102322816848755,
0.03564075008034706,
0.08120003342628479,
0.09329462796449661,
0.034895218908786774,
0.020731132477521896,
-0.02750631608068943,
0.03688345104455948,
-0.0713910236954689,
-0.02386968582868576,
0.05225207284092903,
-0.016514893621206284,
-0.07018444687128067,
0.01533552072942257,
0.03032223880290985,
-0.040582068264484406,
-0.06971492618322372,
-0.015658000484108925,
-0.04662475734949112,
0.023185521364212036,
-0.041797734797000885,
-0.04990578442811966,
0.012025849893689156,
-0.024740394204854965,
-0.01855127140879631,
0.034186094999313354,
0.08508529514074326,
-0.014753496274352074,
-0.06850262731313705,
0.0004778460424859077,
-0.09156189858913422,
-0.030017497017979622,
-0.0517011433839798,
0.08197811990976334,
0.04118955880403519,
-0.09571122378110886,
-0.025930654257535934,
-0.05526331067085266,
0.06387689709663391,
-0.012302846647799015,
-0.031117284670472145,
0.010788824409246445,
0.02484922483563423,
0.026633448898792267,
-0.06699935346841812,
-0.05541679635643959,
-0.09827369451522827,
0.027432609349489212,
-0.07051835209131241,
-0.006315791048109531,
0.1180204376578331,
0.03262091800570488,
-0.01879057101905346,
-0.008949430659413338,
0.0003909537917934358,
-0.020422643050551414,
-0.05218110606074333,
-0.014071309007704258,
0.04319002106785774,
0.03237200900912285,
0.06649304181337357,
0.0449327789247036,
-0.05492100492119789,
0.1623954474925995,
0.07927335798740387,
0.11547575145959854,
0.010376926511526108,
0.06786195188760757,
0.03224991261959076,
-0.034714341163635254,
0.012816522270441055,
0.0044607436284422874,
-0.034953415393829346,
-0.09660674631595612,
0.08727115392684937,
-0.0015150067629292607,
0.014734068885445595,
-0.054671380668878555,
0.004809215199202299,
0.043153066188097,
-0.04524160176515579,
-0.01441260613501072,
0.003822159254923463,
0.014454696327447891,
-0.022221393883228302,
-0.0417487695813179,
0.009957836009562016,
-0.11127226799726486,
0.017077403143048286,
-0.047501467168331146,
-0.05638658627867699,
-0.01158610638231039,
-0.030716564506292343,
-0.01771223545074463,
-0.060213603079319,
-7.2407885208535295e-34,
-0.010723601095378399,
-0.051427438855171204,
-0.048490989953279495,
0.010497095994651318,
-0.008558603934943676,
-0.06522034108638763,
0.009348021820187569,
-0.037182554602622986,
-0.007360251620411873,
-0.04468313977122307,
-0.022110577672719955,
-0.02984604798257351,
0.028179943561553955,
-0.06151394546031952,
-0.011414986103773117,
-0.022410141304135323,
0.0343143567442894,
0.02002597041428089,
-0.03868240490555763,
0.040632229298353195,
0.07032331824302673,
0.03803149238228798,
-0.028316454961895943,
0.0297382939606905,
-0.023283448070287704,
-0.004608852323144674,
-0.03617589920759201,
0.030930008739233017,
0.003541664220392704,
-0.01410421822220087,
-0.03822503611445427,
0.035034675151109695,
-0.007351280190050602,
0.006835281848907471,
-0.08673346787691116,
0.09866435825824738,
0.01965794712305069,
-0.013770410791039467,
-0.03408794477581978,
0.10321415960788727,
0.10374249517917633,
-0.0429350771009922,
-0.02377006597816944,
-0.05660025775432587,
-0.047491591423749924,
-0.0020971649792045355,
-0.08450357615947723,
-0.01916985586285591,
0.00007967653800733387,
0.06376659870147705,
-0.10650522261857986,
0.02031814306974411,
-0.05952032282948494,
0.06983953714370728,
-0.046059779822826385,
-0.07930715382099152,
0.004305867478251457,
-0.06299564242362976,
0.0789594054222107,
-0.009006479755043983,
-0.04332355782389641,
-0.020021634176373482,
0.040967825800180435,
-0.03955839201807976,
0.06351564079523087,
-0.01412347424775362,
-0.03993380442261696,
-0.021126706153154373,
-0.07590610533952713,
-0.007604303304105997,
0.01401627529412508,
-0.02837497554719448,
-0.05306399613618851,
-0.0016028019599616528,
-0.046593520790338516,
0.05194677412509918,
0.0406191349029541,
-0.05448264628648758,
-0.015216697007417679,
-0.08272603154182434,
0.015238421969115734,
-0.07880948483943939,
-0.0029888548888266087,
0.0945688858628273,
0.03178868442773819,
0.08234928548336029,
0.04945008084177971,
0.08912444859743118,
0.05966392531991005,
0.06297867745161057,
-0.025678128004074097,
0.012265525758266449,
0.04362904652953148,
0.046233806759119034,
-0.029300790280103683,
-4.655491991911731e-8,
-0.15471850335597992,
0.057708825916051865,
-0.04843923822045326,
0.061866164207458496,
0.05378055199980736,
-0.01503459271043539,
-0.04741041362285614,
0.10240507870912552,
0.007792897522449493,
-0.05435938015580177,
0.049256667494773865,
0.0409332774579525,
-0.006981374695897102,
-0.03985046595335007,
-0.029154127463698387,
0.04974044859409332,
-0.018407046794891357,
0.0069810920394957066,
0.005618984345346689,
0.013421997427940369,
0.04401161149144173,
-0.002033902332186699,
-0.006857807748019695,
-0.0019380294252187014,
-0.011415848508477211,
-0.09555799514055252,
0.0075485422275960445,
0.03658972680568695,
0.06407095491886139,
-0.08050914853811264,
-0.04683595523238182,
0.00207249796949327,
0.014801939018070698,
-0.044464778155088425,
0.02758420817553997,
0.06787963956594467,
-0.025930296629667282,
-0.0030346254352480173,
0.023762201890349388,
-0.057620543986558914,
0.0037529224064201117,
0.02906448021531105,
-0.02709917165338993,
0.0415041446685791,
0.02109360881149769,
0.06873612105846405,
0.021170103922486305,
-0.053828414529561996,
-0.03325222432613373,
0.005237814504653215,
-0.03074166364967823,
-0.015370666049420834,
0.021189885213971138,
-0.00664783688262105,
0.047893863171339035,
0.07529906928539276,
0.02401716262102127,
-0.01839873008430004,
0.07034003734588623,
0.007071378640830517,
0.023949820548295975,
0.09641376882791519,
-0.024688705801963806,
0.0742141380906105
] | 0 | 0 | 0 | 0 |
|
Learning Neural Networks with Distribution Shift: Efficiently Certifiable Guarantees | https://openreview.net/forum?id=ed7zI29lRF | [
"Gautam Chandrasekaran",
"Adam Klivans",
"Lin Lin Lee",
"Konstantinos Stavropoulos"
] | Poster | We give the first provably efficient algorithms for learning neural networks with respect to distribution shift. We work in the Testable Learning with Distribution Shift framework (TDS learning) of Klivans et al. (2024), where the learner receives labeled examples from a training distribution and unlabeled examples from a test distribution and must either output a hypothesis with low test error or reject if distribution shift is detected. No assumptions are made on the test distribution.
All prior work in TDS learning focuses on classification, while here we must handle the setting of nonconvex regression. Our results apply to real-valued networks with arbitrary Lipschitz activations and work whenever the training distribution has strictly sub-exponential tails. For training distributions that are bounded and hypercontractive, we give a fully polynomial-time algorithm for TDS learning one hidden-layer networks with sigmoid activations. We achieve this by importing classical kernel methods into the TDS framework using data-dependent feature maps and a type of kernel matrix that couples samples from both train and test distributions. | pac learning, distribution shift, distribution testing, testable learning, neural networks, kernel methods | null | 13,095 | 2502.16021 | [
-0.1547764092683792,
-0.12105982005596161,
0.012738184072077274,
0.03265174850821495,
0.09812597930431366,
0.04735149070620537,
0.022973351180553436,
-0.08426476269960403,
0.01302163116633892,
-0.004947534762322903,
0.05128633603453636,
-0.021062549203634262,
-0.0234825499355793,
0.0389726422727108,
-0.046231064945459366,
-0.058502230793237686,
0.07549997419118881,
0.087417371571064,
-0.022321386262774467,
0.06488120555877686,
-0.0069211870431900024,
0.027276581153273582,
-0.022391902282834053,
0.011946440674364567,
0.09872203320264816,
-0.007956596091389656,
-0.023150663822889328,
-0.032406773418188095,
-0.002698286436498165,
-0.006743398029357195,
0.03846023231744766,
-0.054544541984796524,
-0.06467851251363754,
0.01251845434308052,
-0.048985064029693604,
0.04447666555643082,
-0.017896736040711403,
0.033783528953790665,
0.04491666704416275,
0.016621887683868408,
0.024128355085849762,
-0.06816858798265457,
0.0011460317764431238,
-0.009164671413600445,
0.04922305792570114,
-0.001815551775507629,
0.02448463812470436,
-0.0016498995246365666,
0.0029370111878961325,
0.030005965381860733,
0.002826924668624997,
0.036872364580631256,
0.0022986414842307568,
0.062406864017248154,
-0.050820011645555496,
-0.012408253736793995,
0.08070674538612366,
0.032176755368709564,
-0.009670124389231205,
0.08639998733997345,
-0.013993469066917896,
-0.11948657035827637,
0.03929535672068596,
0.009706423617899418,
-0.09061987698078156,
0.030495746061205864,
-0.026216629892587662,
0.03217734768986702,
0.007131172344088554,
0.037088893353939056,
-0.03937499225139618,
0.12217842042446136,
-0.11015236377716064,
0.04540371522307396,
0.044810473918914795,
-0.02074919454753399,
0.14569611847400665,
0.00019325190805830061,
0.03498067334294319,
-0.09096359461545944,
0.023467592895030975,
0.03489917516708374,
0.018986649811267853,
-0.029113488271832466,
0.06829924881458282,
0.05452476069331169,
-0.004017493687570095,
0.09774472564458847,
-0.047546032816171646,
-0.021458499133586884,
-0.0345214381814003,
0.022084807977080345,
-0.06393598765134811,
-0.018200255930423737,
0.014333941042423248,
-0.05824240297079086,
0.057917628437280655,
-0.06151742860674858,
-0.013774177059531212,
0.1178639754652977,
-0.07535179704427719,
0.025385184213519096,
0.008236541412770748,
-0.04651004821062088,
0.0312662310898304,
-0.04566853493452072,
0.019770793616771698,
-0.040087006986141205,
0.01183219626545906,
-0.07267146557569504,
0.05598156899213791,
0.026887105777859688,
-0.01532311737537384,
-0.03939340263605118,
-0.018877260386943817,
-0.020177742466330528,
-0.029671072959899902,
-0.0012963144108653069,
-0.03344076871871948,
0.024423852562904358,
-0.07610961049795151,
0.023846806958317757,
-0.05233951658010483,
0.04101765900850296,
0.005257151089608669,
0.027307748794555664,
-0.04398316890001297,
3.19892898021634e-33,
-0.01726934313774109,
0.011666825972497463,
0.03714384138584137,
-0.06217839941382408,
0.03357426077127457,
-0.009311258792877197,
0.041439563035964966,
0.05425703525543213,
0.04019675776362419,
0.037914324551820755,
-0.04696173965930939,
0.0008614638936705887,
-0.06483978778123856,
0.057834334671497345,
0.03421272709965706,
0.023040521889925003,
-0.02439475804567337,
-0.020003506913781166,
0.08652232587337494,
-0.035156719386577606,
0.05244022607803345,
-0.13342219591140747,
-0.031261567026376724,
-0.032052841037511826,
-0.062897227704525,
-0.019544046372175217,
0.005485488101840019,
0.006564153824001551,
0.00778547115623951,
0.017443183809518814,
-0.08903700113296509,
0.03362644836306572,
-0.017477864399552345,
-0.02019774727523327,
0.03723274543881416,
0.003236771561205387,
0.0036855859216302633,
-0.019777506589889526,
-0.010647939518094063,
-0.02882850170135498,
-0.04156222939491272,
0.04349729046225548,
-0.031672604382038116,
-0.018702663481235504,
-0.05085684359073639,
-0.13411617279052734,
0.02588515356183052,
0.022928457707166672,
0.06761132180690765,
-0.13671381771564484,
0.03077106364071369,
-0.0003787208115682006,
-0.04577081277966499,
-0.14140668511390686,
0.006644926965236664,
0.050699930638074875,
0.057782307267189026,
0.05460570752620697,
0.05005893483757973,
0.07236326485872269,
-0.01822073757648468,
-0.046871304512023926,
-0.01337704248726368,
0.0644085705280304,
-0.0007130706799216568,
0.059304188936948776,
-0.0689355805516243,
-0.0260346457362175,
0.0340396985411644,
-0.028355801478028297,
0.012115269899368286,
-0.004492139909416437,
-0.01201038807630539,
-0.05216570198535919,
0.06003320589661598,
0.07155943661928177,
-0.021747957915067673,
-0.044908396899700165,
0.0025922460481524467,
0.01965346746146679,
0.07781199365854263,
0.0031175287440419197,
-0.0064543322660028934,
-0.01978698931634426,
0.023351265117526054,
-0.045806653797626495,
-0.01487649418413639,
0.011091354303061962,
0.0013721606228500605,
-0.05346338450908661,
-0.0866016298532486,
-0.019870521500706673,
-0.02794298157095909,
-0.026051703840494156,
0.030238913372159004,
-2.818988634879437e-33,
-0.060726579278707504,
0.12234841287136078,
-0.08025848865509033,
0.05347435176372528,
0.015707477927207947,
0.029315264895558357,
-0.049295056611299515,
0.07779420912265778,
-0.03706350922584534,
0.009355809539556503,
-0.021432435140013695,
-0.02946040779352188,
-0.05294429138302803,
-0.048624128103256226,
-0.03948821499943733,
-0.014040112495422363,
-0.0181459691375494,
0.031290967017412186,
-0.016632160171866417,
-0.035211414098739624,
0.042148690670728683,
0.1056152731180191,
-0.11301309615373611,
0.09916956722736359,
-0.06866832822561264,
-0.01755300536751747,
-0.0997019037604332,
0.1987372189760208,
-0.054665226489305496,
0.05425828695297241,
-0.08359701931476593,
-0.03623811528086662,
-0.05072895064949989,
-0.005503425374627113,
0.04744948074221611,
0.045748770236968994,
-0.036893922835588455,
0.06076626479625702,
-0.03414875268936157,
0.12405301630496979,
0.04091494530439377,
-0.02392236888408661,
-0.07012635469436646,
0.032029420137405396,
0.09151773899793625,
0.0014271748950704932,
0.038384199142456055,
-0.04581182450056076,
0.010822043754160404,
0.058084290474653244,
-0.007829670794308186,
0.06819767504930496,
-0.06405556946992874,
0.044297490268945694,
-0.034441325813531876,
0.006914027035236359,
-0.10609273612499237,
0.008339458145201206,
0.06677237898111343,
-0.00743835186585784,
-0.016383543610572815,
0.0037393574602901936,
0.0003989874676335603,
0.05975615233182907,
0.015707392245531082,
-0.06192322075366974,
-0.03611389547586441,
0.020602397620677948,
0.02964034117758274,
0.04315251484513283,
0.031139522790908813,
-0.03684067353606224,
0.01497625932097435,
-0.04661331698298454,
-0.012293608859181404,
-0.03009450063109398,
-0.016698328778147697,
0.006994200870394707,
-0.004946767818182707,
0.010211387649178505,
0.07643719017505646,
-0.018284065648913383,
0.03880377113819122,
0.04471656307578087,
0.14531271159648895,
0.1360926628112793,
0.055087532848119736,
-0.0328950509428978,
0.0930558294057846,
-0.05773793160915375,
0.00840628333389759,
0.04315773770213127,
-0.04266143962740898,
-0.028998831287026405,
0.026147928088903427,
-4.1824858243444396e-8,
-0.025717463344335556,
0.035096291452646255,
0.02481677383184433,
-0.020107820630073547,
0.07941628247499466,
-0.022196749225258827,
0.06125802546739578,
0.01687312126159668,
-0.04600954055786133,
-0.0482122041285038,
0.0020738677121698856,
0.03494776785373688,
0.0021344847045838833,
-0.04065664857625961,
-0.01841847039759159,
0.03148486837744713,
-0.005556218326091766,
0.021453335881233215,
0.02043035440146923,
-0.01428359467536211,
0.01555660367012024,
0.03810720518231392,
0.07846582680940628,
0.0035374683793634176,
-0.0004524110700003803,
-0.09362727403640747,
-0.0058725858107209206,
-0.0042311325669288635,
0.08425943553447723,
0.05778086557984352,
-0.06029148027300835,
-0.01971990056335926,
0.0011684703640639782,
-0.027802661061286926,
0.08126698434352875,
0.10758774727582932,
0.013656669296324253,
0.05195313319563866,
0.008933281525969505,
-0.005747951101511717,
-0.05986829474568367,
0.0029461842495948076,
-0.07272736728191376,
-0.03432530164718628,
-0.05982794612646103,
0.012223792262375355,
-0.004212460946291685,
-0.05284891650080681,
-0.004312729928642511,
-0.054774899035692215,
0.0748252123594284,
-0.015317915007472038,
-0.00565599137917161,
0.025039449334144592,
0.0017568019684404135,
0.009092661552131176,
-0.02176891639828682,
-0.09334275126457214,
-0.03782172128558159,
0.04341566562652588,
-0.002613064367324114,
0.04660548269748688,
0.009198002517223358,
-0.017190564423799515
] | 0 | 0 | 0 | 0 |
|
Forgetting Transformer: Softmax Attention with a Forget Gate | https://openreview.net/forum?id=q2Lnyegkr8 | [
"Zhixuan Lin",
"Evgenii Nikishin",
"Xu He",
"Aaron Courville"
] | Poster | An essential component of modern recurrent sequence models is the forget gate. While Transformers do not have an explicit recurrent form, we show that a forget gate can be naturally incorporated into Transformers by down-weighting the unnormalized attention scores in a data-dependent way. We name this attention mechanism the Forgetting Attention and the resulting model the Forgetting Transformer (FoX). We show that FoX outperforms the Transformer on long-context language modeling, length extrapolation, and short-context downstream tasks, while performing on par with the Transformer on long-context downstream tasks. Moreover, it is compatible with the FlashAttention algorithm and does not require any positional embeddings. Several analyses, including the needle-in-the-haystack test, show that FoX also retains the Transformer's superior long-context capabilities over recurrent sequence models such as Mamba-2, HGRN2, and DeltaNet. We also introduce a ``Pro'' block design that incorporates some common architectural components in recurrent sequence models and find it significantly improves the performance of both FoX and the Transformer.
Our code is available at [`https://github.com/zhixuan-lin/forgetting-transformer`](https://github.com/zhixuan-lin/forgetting-transformer). | sequence model, long-context sequence modeling, Transformer, softmax attention, linear attention, RNN, language model | We propose the Forgetting Transformer, a Transformer variant with a forget gate. | 13,091 | 2503.02130 | [
-0.07464299350976944,
-0.07576732337474823,
-0.008946006186306477,
0.04641010984778404,
0.024073578417301178,
0.053181275725364685,
-0.02011563815176487,
-0.0025287254247814417,
0.04931715875864029,
-0.12362530827522278,
-0.02586298994719982,
0.022749464958906174,
0.01530213188380003,
0.020902901887893677,
-0.030567705631256104,
0.003428129479289055,
-0.03477080538868904,
0.0905361920595169,
-0.039455387741327286,
-0.1151791512966156,
0.06812728196382523,
-0.007020269986242056,
0.08446130901575089,
-0.026684142649173737,
0.011154720559716225,
-0.04899510368704796,
-0.05601595342159271,
-0.030989525839686394,
0.07293123751878738,
-0.07647735625505447,
0.0660526379942894,
0.055826857686042786,
-0.021577486768364906,
0.0570494681596756,
-0.07604419440031052,
0.0829475149512291,
-0.08413784950971603,
-0.0005315361777320504,
-0.03303370624780655,
-0.020764566957950592,
0.045109525322914124,
-0.029754970222711563,
0.03923732414841652,
0.01493604201823473,
0.0016105591785162687,
-0.06432940810918808,
0.04102033004164696,
-0.0833323746919632,
0.026033060625195503,
0.0004900554777123034,
-0.03179957717657089,
0.013663497753441334,
-0.008346505463123322,
0.13247306644916534,
-0.021017253398895264,
0.05958624929189682,
0.024841420352458954,
0.02195275016129017,
-0.028631284832954407,
-0.023883387446403503,
-0.07313008606433868,
-0.06809848546981812,
-0.014595040120184422,
-0.06063765659928322,
-0.12174338847398758,
-0.04037914052605629,
0.05762898176908493,
0.019002048298716545,
-0.08950735628604889,
0.02836093306541443,
-0.04156750440597534,
0.01876319944858551,
-0.02281646803021431,
0.041588228195905685,
0.06526149064302444,
0.020201904699206352,
0.08165677636861801,
-0.08018171042203903,
0.05164916813373566,
-0.0001710480428300798,
0.04826761782169342,
-0.07773511111736298,
0.031207595020532608,
-0.029197415336966515,
0.12273337692022324,
0.004621817730367184,
-0.002463715383782983,
0.014689326286315918,
-0.01977609470486641,
0.030612831935286522,
-0.05004968121647835,
-0.12090026587247849,
0.06625504791736603,
0.0009697217028588057,
0.0625983253121376,
0.07348082214593887,
0.03960203751921654,
-0.0010540832299739122,
-0.09490137547254562,
0.06747759878635406,
-0.005736867897212505,
-0.01753224991261959,
0.004885379690676928,
-0.00867787841707468,
-0.03007359243929386,
-0.06529785692691803,
-0.017520520836114883,
0.04977045953273773,
0.025502828881144524,
-0.05216363072395325,
0.032413702458143234,
0.042445335537195206,
0.015275951474905014,
0.052920833230018616,
0.061854347586631775,
-0.11900172382593155,
-0.011134929955005646,
-0.048219405114650726,
-0.011633834801614285,
0.054279766976833344,
0.01653265208005905,
0.0872010812163353,
-0.032416265457868576,
0.06468332558870316,
0.02633790858089924,
-0.03805261477828026,
-0.030540646985173225,
7.90582490900872e-34,
-0.03143681213259697,
0.04699055105447769,
-0.005616310052573681,
-0.020619191229343414,
0.03794899955391884,
-0.0027840451803058386,
0.036806412041187286,
0.07726512849330902,
-0.032198719680309296,
0.01745583489537239,
-0.013889406807720661,
0.03419870138168335,
-0.06292299926280975,
0.07403519749641418,
-0.02467644214630127,
-0.05425189062952995,
0.018347950652241707,
0.014534199610352516,
-0.05424167215824127,
0.015568112954497337,
0.01631636917591095,
0.05826571211218834,
0.02037130482494831,
-0.04920808598399162,
0.019760407507419586,
-0.013060687109827995,
0.07608753442764282,
-0.0642755851149559,
-0.053675659000873566,
0.021335355937480927,
-0.15851706266403198,
0.02020508050918579,
-0.018398799002170563,
0.006637399550527334,
0.04442000389099121,
-0.08126885443925858,
0.010549938306212425,
-0.042859580367803574,
0.016953257843852043,
-0.03207194060087204,
-0.0010591360041871667,
0.05851177126169205,
-0.06848608702421188,
0.022603942081332207,
-0.11325329542160034,
-0.12695401906967163,
0.0856422632932663,
0.02262396551668644,
0.02832101099193096,
0.010366641916334629,
-0.001061365008354187,
0.006175185088068247,
-0.06451903283596039,
-0.030688509345054626,
0.011365463025867939,
-0.009917325340211391,
0.023754358291625977,
0.05638740211725235,
0.026182157918810844,
0.04376053810119629,
0.003636743873357773,
0.051492419093847275,
0.016278820112347603,
0.08018580824136734,
0.05008021369576454,
0.011966357007622719,
-0.01477537490427494,
0.05625580623745918,
-0.0033803670667111874,
0.021517805755138397,
-0.007558640092611313,
-0.01611451618373394,
-0.0013686259044334292,
0.005547305103391409,
0.07010587304830551,
0.003670408157631755,
0.07819584757089615,
-0.10459044575691223,
-0.0190627072006464,
0.00999824795871973,
0.01835574582219124,
-0.01467374712228775,
0.01699032261967659,
0.02474319376051426,
-0.038522668182849884,
-0.03602663800120354,
0.07367514073848724,
-0.09207265824079514,
-0.05102423205971718,
-0.02462718077003956,
-0.017951086163520813,
-0.03287634998559952,
-0.024059772491455078,
-0.013682675547897816,
-0.01975966803729534,
-1.0556702423357215e-33,
0.003745183115825057,
0.03466039523482323,
-0.04743485525250435,
0.021374408155679703,
0.004168755374848843,
-0.03990612179040909,
-0.0039105662144720554,
0.041449014097452164,
-0.04105732589960098,
-0.036020901054143906,
-0.012439124286174774,
-0.02092728391289711,
-0.014073381200432777,
-0.05719439312815666,
0.06426708400249481,
-0.013608425855636597,
-0.04671598970890045,
0.016377752646803856,
-0.10106296092271805,
0.012949023395776749,
0.01826629973948002,
0.12061342597007751,
-0.0679565966129303,
0.06597422808408737,
-0.08816253393888474,
0.010532708838582039,
-0.04522531479597092,
0.04650858789682388,
-0.005464904010295868,
-0.060856081545352936,
-0.05057103931903839,
-0.01981661655008793,
0.009623561054468155,
-0.03546677157282829,
-0.07277191430330276,
0.005709985736757517,
0.04754989966750145,
0.004524266812950373,
-0.04605482891201973,
0.042999424040317535,
0.1396407037973404,
-0.00976668018847704,
-0.04486053064465523,
0.049906354397535324,
-0.040191683918237686,
0.0026945089921355247,
-0.07709717750549316,
-0.010312710888683796,
0.02259618043899536,
0.06378421932458878,
-0.0637064278125763,
0.026418432593345642,
-0.09889092296361923,
0.002449487801641226,
-0.06162340193986893,
-0.08644114434719086,
-0.002442394383251667,
-0.06172412261366844,
0.04522262513637543,
-0.048315808176994324,
-0.016217725351452827,
0.006192771717905998,
0.026141485199332237,
-0.10118522495031357,
0.05430495738983154,
-0.0033816611394286156,
0.05439317226409912,
-0.03251595050096512,
0.042033664882183075,
0.030217407271265984,
0.09463269263505936,
0.04987344518303871,
-0.007108221296221018,
0.008310855366289616,
0.011258553713560104,
0.016199953854084015,
-0.03866669535636902,
-0.05288466438651085,
-0.019014455378055573,
-0.0652366578578949,
-0.046607017517089844,
-0.00707900058478117,
-0.05841260030865669,
0.03688518702983856,
0.041678160429000854,
0.08017057180404663,
0.11348073184490204,
0.06910025328397751,
0.05875053629279137,
-0.01383802480995655,
-0.03023986890912056,
0.0691811814904213,
-0.040024761110544205,
0.03990568965673447,
-0.02873767726123333,
-4.179375068247282e-8,
-0.09446059912443161,
0.06654170155525208,
-0.06368295103311539,
0.00526684895157814,
0.017480812966823578,
-0.08676562458276749,
-0.0038565495051443577,
0.07834085077047348,
-0.021728621795773506,
-0.022242333739995956,
0.11797592043876648,
-0.017917431890964508,
-0.0009760356624610722,
-0.050930071622133255,
-0.011030339635908604,
0.09240604937076569,
0.050032828003168106,
0.054127730429172516,
0.0010388134978711605,
0.016190117225050926,
0.01663646101951599,
0.06539566814899445,
-0.0025509430561214685,
0.014247052371501923,
0.004449265543371439,
-0.031312454491853714,
-0.031038856133818626,
0.11998878419399261,
0.04332425445318222,
-0.002581721404567361,
0.008283519186079502,
0.031542811542749405,
-0.005418218206614256,
-0.030625227838754654,
-0.08229807764291763,
0.04985039681196213,
0.10518932342529297,
-0.022574715316295624,
-0.019075635820627213,
0.04624832794070244,
0.07169953733682632,
0.03613051027059555,
-0.026779163628816605,
0.050052326172590256,
-0.07921239733695984,
0.03094678558409214,
-0.027209771797060966,
-0.11532514542341232,
0.06857273727655411,
-0.05168214812874794,
0.08341559767723083,
-0.009747620671987534,
-0.031191347166895866,
0.06989891827106476,
0.05172014236450195,
0.04095311462879181,
0.0419248566031456,
-0.03774905949831009,
-0.0015087976353242993,
0.038471467792987823,
0.03628886118531227,
0.05701087415218353,
-0.0279901884496212,
0.023556966334581375
] | https://github.com/zhixuan-lin/forgetting-transformer | 96 | 7 | 0 | 0 |
MaxInfoRL: Boosting exploration in reinforcement learning through information gain maximization | https://openreview.net/forum?id=R4q3cY3kQf | [
"Bhavya Sukhija",
"Stelian Coros",
"Andreas Krause",
"Pieter Abbeel",
"Carmelo Sferrazza"
] | Poster | Reinforcement learning (RL) algorithms aim to balance exploiting the current best strategy with exploring new options that could lead to higher rewards. Most common RL algorithms use undirected exploration, i.e., select random sequences of actions.
Exploration can also be directed using intrinsic rewards, such as curiosity or model epistemic uncertainty. However, effectively balancing task and intrinsic rewards is challenging and often task-dependent. In this work, we introduce a framework, MaxInfoRL, for balancing intrinsic and extrinsic exploration. MaxInfoRL steers exploration towards informative transitions, by maximizing intrinsic rewards such as the information gain about the underlying task.
When combined with Boltzmann exploration, this approach naturally trades off maximization of the value function with that of the entropy over states, rewards, and actions. We show that our approach achieves sublinear regret in the simplified setting of multi-armed bandits. We then apply this general formulation to a variety of off-policy model-free RL methods for continuous state-action spaces, yielding novel algorithms that achieve superior performance across hard exploration problems and complex scenarios such as visual control tasks. | Reinforcement learning, Exploration in off-policy methods, Continuous control | We propose a systematic way of combining directed exploration bonuses with extrinsic task rewards in RL and evaluate it across several continuous and visual control tasks. | 13,089 | 2412.12098 | [
-0.10698331892490387,
-0.06930194050073624,
0.01982194371521473,
0.020574260503053665,
0.045154206454753876,
-0.012441016733646393,
0.036224015057086945,
0.01728595420718193,
0.06046930328011513,
-0.003855431452393532,
-0.09646812081336975,
-0.06437429785728455,
0.04324698448181152,
0.04840761795639992,
-0.00858457013964653,
0.027113867923617363,
0.046791333705186844,
0.05733519792556763,
-0.03252224624156952,
0.004414254799485207,
0.02825275994837284,
-0.011947429738938808,
0.055151473730802536,
-0.008551200851798058,
-0.03569655492901802,
-0.025681685656309128,
0.0036942074075341225,
-0.03512870892882347,
0.03417954593896866,
-0.021733293309807777,
0.0823582112789154,
-0.04265035688877106,
0.06773789972066879,
-0.10332342237234116,
-0.06799936294555664,
0.05901254713535309,
-0.15813030302524567,
-0.022694453597068787,
-0.02775258757174015,
-0.03617727756500244,
-0.009598292410373688,
0.01790783554315567,
-0.025135666131973267,
0.031633857637643814,
0.07229007035493851,
0.037071727216243744,
-0.010281513445079327,
-0.06944847851991653,
0.08223605901002884,
-0.006159677170217037,
0.003993752878159285,
-0.04593880847096443,
0.00005230415627011098,
-0.040132589638233185,
-0.0035836463794112206,
0.06631103903055191,
0.023622935637831688,
-0.02050960250198841,
0.00140327715780586,
-0.07131531089544296,
0.06962451338768005,
-0.13529448211193085,
-0.02368060126900673,
-0.04182517156004906,
-0.04095584154129028,
-0.04144037887454033,
0.031626466661691666,
0.04822590574622154,
0.021686788648366928,
0.08044841885566711,
0.042716704308986664,
0.014329900033771992,
-0.036234088242053986,
0.03222312778234482,
0.07874095439910889,
0.007745279464870691,
0.01210930198431015,
-0.08336394280195236,
-0.008956016041338444,
0.06271805614233017,
-0.03746475279331207,
0.0012758984230458736,
-0.007615957409143448,
0.0007792066317051649,
0.0703984797000885,
-0.03630286455154419,
0.02159779518842697,
0.04482465237379074,
0.1385449916124344,
-0.018074173480272293,
-0.019508643075823784,
-0.008581419475376606,
-0.01052076742053032,
0.02393658272922039,
0.012184297665953636,
0.03750014305114746,
-0.04611637443304062,
-0.07495924085378647,
-0.044531967490911484,
0.05090884119272232,
0.032867107540369034,
0.017916586250066757,
-0.009981164708733559,
0.009890488348901272,
-0.003232371062040329,
-0.09076276421546936,
-0.011758091859519482,
0.06112195923924446,
0.05447682738304138,
-0.08795595169067383,
0.027532052248716354,
-0.008655558340251446,
0.08376506716012955,
0.033895257860422134,
-0.045898739248514175,
0.0701235830783844,
-0.0003849585773423314,
0.0011794576421380043,
-0.018616845831274986,
0.08748606592416763,
0.017520705237984657,
-0.025099098682403564,
0.019647229462862015,
0.03939318656921387,
-0.035472672432661057,
-0.03605087101459503,
0.004862088710069656,
1.1752885474753974e-33,
0.03712129965424538,
-0.04347499459981918,
0.059419065713882446,
-0.005197018850594759,
0.05033259466290474,
0.0694466382265091,
0.07364688813686371,
0.002782616298645735,
-0.03592069074511528,
0.026350008323788643,
0.03556622192263603,
0.07358383387327194,
-0.05288441851735115,
0.05500670149922371,
0.06339685618877411,
-0.05517837032675743,
-0.09654990583658218,
0.015063831582665443,
0.03118409402668476,
-0.07683917880058289,
0.0829434022307396,
-0.05710594728589058,
-0.0009303271654061973,
-0.04366305470466614,
0.03343592956662178,
-0.01178732980042696,
0.005920426454395056,
-0.020391544327139854,
-0.04012269526720047,
-0.02942318096756935,
-0.06351693719625473,
-0.016474878415465355,
-0.04914138466119766,
0.020538531243801117,
-0.001790461828932166,
0.059550702571868896,
-0.01655387133359909,
0.013583051040768623,
-0.0526031069457531,
0.004744414705783129,
-0.01687673106789589,
0.024253496900200844,
-0.026478055864572525,
0.01923445612192154,
-0.10065878182649612,
-0.05805474519729614,
-0.004256643354892731,
-0.010818560607731342,
-0.12240706384181976,
-0.020840099081397057,
0.009984447620809078,
0.004607659764587879,
0.004085169639438391,
-0.03166656941175461,
-0.014951196499168873,
0.03306637331843376,
0.02753288298845291,
0.053663428872823715,
-0.036839667707681656,
-0.01004733331501484,
0.02298956923186779,
-0.047529466450214386,
0.04640870913863182,
0.07137055695056915,
-0.029235294088721275,
0.12321735918521881,
-0.03489546850323677,
0.03425302356481552,
0.040506377816200256,
-0.011284451000392437,
-0.006490948610007763,
0.04395356774330139,
0.06373004615306854,
-0.122235007584095,
0.09816846251487732,
-0.009137232787907124,
0.07547371089458466,
0.0064919572323560715,
0.060475435107946396,
0.009120429866015911,
-0.014650997705757618,
-0.026752980425953865,
-0.02820226550102234,
0.00940045528113842,
-0.013455191627144814,
-0.00912715494632721,
-0.01349367294460535,
-0.10757915675640106,
-0.03535252809524536,
0.012726128101348877,
-0.109872005879879,
-0.024293886497616768,
0.016750238835811615,
0.007792109157890081,
-0.003509324509650469,
-1.782207025557995e-33,
0.058385275304317474,
-0.0993499904870987,
0.02289467677474022,
-0.01360491756349802,
0.05295248329639435,
0.011746250092983246,
-0.0335506834089756,
-0.08444706350564957,
-0.02407032623887062,
-0.07820656895637512,
-0.04207555204629898,
0.09891003370285034,
0.052630990743637085,
0.03128821775317192,
0.004449476022273302,
-0.039013203233480453,
0.02730896696448326,
0.03819337114691734,
-0.016324017196893692,
0.022399285808205605,
-0.04231325536966324,
0.09213437885046005,
-0.14522872865200043,
0.0015747552970424294,
0.03473612666130066,
0.03507067635655403,
0.021649302914738655,
0.09145618975162506,
0.0019919155165553093,
0.013748953118920326,
-0.05254611372947693,
-0.027910949662327766,
-0.019078481942415237,
0.022211799398064613,
-0.03521056845784187,
0.09607778489589691,
0.13397416472434998,
0.025676105171442032,
-0.08687646687030792,
0.13949252665042877,
0.006194133777171373,
-0.04614664241671562,
-0.00896704662591219,
0.00113506184425205,
0.03480501100420952,
-0.018129222095012665,
-0.05620959773659706,
0.0024582650512456894,
-0.017162343487143517,
0.020743651315569878,
0.0020628338679671288,
0.04815291985869408,
-0.1052158921957016,
0.02669883519411087,
-0.059306852519512177,
-0.0769079253077507,
-0.08007568120956421,
0.01107701100409031,
0.06711658090353012,
0.05738396197557449,
-0.05224081128835678,
0.022734692320227623,
-0.05905139446258545,
0.05123334750533104,
-0.07359054684638977,
0.010056630708277225,
-0.041089966893196106,
0.016763905063271523,
-0.0009716041968204081,
-0.024081692099571228,
-0.041906993836164474,
0.011893875896930695,
0.05629664659500122,
-0.04132556542754173,
0.05825144052505493,
0.007096709217876196,
0.04333832114934921,
0.017682060599327087,
0.032233141362667084,
-0.08844999969005585,
-0.0031591278966516256,
0.015873804688453674,
0.0744721069931984,
0.032060299068689346,
0.022965366020798683,
0.102205291390419,
-0.02341608889400959,
0.08085903525352478,
0.011448971927165985,
-0.10092689841985703,
-0.04273007810115814,
-0.06179997697472572,
0.04355674982070923,
0.014009784907102585,
-0.036376580595970154,
-3.88915708526838e-8,
-0.06158997118473053,
0.03802997246384621,
0.07549060136079788,
-0.014849730767309666,
0.09909623861312866,
0.06760281324386597,
-0.009991818107664585,
-0.014576070010662079,
-0.043124664574861526,
-0.03623301908373833,
0.10064727813005447,
-0.00335411773994565,
0.003691315185278654,
-0.044213224202394485,
-0.044073786586523056,
0.01033096108585596,
0.012821387499570847,
-0.0634346604347229,
0.025335662066936493,
-0.00897484552115202,
-0.007332145236432552,
0.007405349984765053,
-0.0033089062198996544,
-0.09374936670064926,
0.0011739021865651011,
-0.055175699293613434,
-0.00926696602255106,
0.08133193850517273,
0.1341429203748703,
-0.011169125325977802,
0.013569200411438942,
-0.01113628875464201,
0.0601227767765522,
0.06983069330453873,
0.057328976690769196,
0.00029309812816791236,
0.010142920538783073,
-0.015257389284670353,
-0.04164878651499748,
-0.00378667120821774,
0.01647924818098545,
0.07188713550567627,
0.018409488722682,
0.000978817348368466,
-0.07245231419801712,
0.06548566371202469,
-0.0941721573472023,
-0.03489834815263748,
0.11057557165622711,
-0.03596654534339905,
0.00004082379746250808,
-0.04240361973643303,
-0.017986871302127838,
0.07830711454153061,
0.06442534923553467,
0.035116568207740784,
-0.02649327740073204,
-0.03488249331712723,
-0.0030098012648522854,
0.08243302255868912,
0.021784421056509018,
-0.06603366136550903,
-0.11904571950435638,
0.0016516994219273329
] | 0 | 0 | 0 | 0 |
|
LancBiO: Dynamic Lanczos-aided Bilevel Optimization via Krylov Subspace | https://openreview.net/forum?id=wLmJIs1uqG | [
"Yan Yang",
"Bin Gao",
"Ya-xiang Yuan"
] | Poster | Bilevel optimization, with broad applications in machine learning, has an intricate hierarchical structure. Gradient-based methods have emerged as a common approach to large-scale bilevel problems. However, the computation of the hyper-gradient, which involves a Hessian inverse vector product, confines the efficiency and is regarded as a bottleneck. To circumvent the inverse, we construct a sequence of low-dimensional approximate Krylov subspaces with the aid of the Lanczos process. As a result, the constructed subspace is able to dynamically and incrementally approximate the Hessian inverse vector product with less effort and thus leads to a favorable estimate of the hyper-gradient. Moreover, we propose a provable subspace-based framework for bilevel problems where one central step is to solve a small-size tridiagonal linear system. To the best of our knowledge, this is the first time that subspace techniques are incorporated into bilevel optimization. This successful trial not only enjoys $\mathcal{O}(\epsilon^{-1})$ convergence rate but also demonstrates efficiency in a synthetic problem and two deep learning tasks. | Bilevel Optimization, Lanczos Process, Krylov Subspace | null | 13,086 | 2404.03331 | [
-0.04265693947672844,
-0.06424026936292648,
-0.013932941481471062,
-0.015837758779525757,
-0.013303945772349834,
-0.002820865949615836,
0.0031819308642297983,
-0.027176205068826675,
-0.05846802890300751,
0.019527191296219826,
0.026691781356930733,
0.06464681774377823,
0.015670891851186752,
0.008456367999315262,
-0.07544568181037903,
0.09796611219644547,
0.03335462510585785,
0.025707025080919266,
-0.017137866467237473,
-0.02688581682741642,
-0.03451545536518097,
-0.04358445480465889,
0.02008834481239319,
0.013799089938402176,
0.03466176986694336,
0.020238740369677544,
0.08383788913488388,
0.05861304700374603,
-0.051295205950737,
-0.0059785726480185986,
0.08622913807630539,
-0.02091739885509014,
0.03908460959792137,
-0.12038729339838028,
-0.07656171917915344,
0.05910685658454895,
0.015228943899273872,
0.029760761186480522,
-0.02292581833899021,
0.04327216371893883,
-0.005411603022366762,
0.10200657695531845,
-0.020300325006246567,
0.015554036945104599,
0.04850715026259422,
-0.04925898462533951,
0.051142171025276184,
-0.08041546493768692,
0.0912766233086586,
-0.07845549285411835,
-0.02213253639638424,
-0.032852329313755035,
-0.036210790276527405,
0.06541897356510162,
-0.03208775818347931,
-0.03343769535422325,
0.025370702147483826,
-0.029248353093862534,
0.03070264309644699,
-0.019646571949124336,
0.0013871120754629374,
-0.09384586662054062,
0.03160431981086731,
-0.03153045102953911,
-0.02500239573419094,
-0.014744208194315434,
-0.02328338660299778,
0.020812420174479485,
-0.0724433958530426,
0.053885798901319504,
-0.07086136192083359,
0.05778000131249428,
-0.08116137981414795,
0.036647044122219086,
0.008227179758250713,
-0.024316053837537766,
0.10374719649553299,
0.0011698695598170161,
0.10799822956323624,
-0.030258895829319954,
0.01630425825715065,
0.04005613550543785,
-0.04157140105962753,
-0.07850202172994614,
0.0814371332526207,
-0.00045971080544404685,
-0.022550396621227264,
-0.03338950499892235,
0.05198581516742706,
-0.07319597899913788,
0.06068175658583641,
0.011280984617769718,
-0.013143203221261501,
-0.04484687000513077,
-0.03403245285153389,
-0.11232849955558777,
0.10051746666431427,
0.022729039192199707,
0.05018750950694084,
0.09441576898097992,
0.037795912474393845,
-0.02176540717482567,
-0.01588851772248745,
-0.024319114163517952,
-0.00001774985867086798,
-0.011960667558014393,
0.027727682143449783,
0.08725948631763458,
0.023232534527778625,
-0.10664009302854538,
0.004233907908201218,
0.00901077687740326,
0.0005395606276579201,
-0.005268387030810118,
-0.05914979428052902,
0.03743577003479004,
0.07658001780509949,
-0.059001486748456955,
0.08960238844156265,
0.08627985417842865,
-0.06350753456354141,
-0.025705702602863312,
0.023622306063771248,
0.09241462498903275,
0.026745937764644623,
0.006058789324015379,
-0.08127496391534805,
9.237491949532402e-34,
-0.009522251784801483,
0.0455799363553524,
-0.023203756660223007,
0.00020604496239684522,
-0.009607200510799885,
0.0069716800935566425,
0.11140039563179016,
-0.0025254404172301292,
-0.019857263192534447,
-0.02211662009358406,
-0.038037221878767014,
0.007742463611066341,
-0.0771210640668869,
0.021074969321489334,
0.06110723316669464,
0.026405757293105125,
-0.01589154824614525,
0.042509641498327255,
-0.06577487289905548,
-0.10542789846658707,
0.044744912534952164,
-0.07113170623779297,
-0.04704437032341957,
-0.027702730149030685,
-0.004297097213566303,
-0.03970762714743614,
0.1184253990650177,
-0.03874480724334717,
-0.10334429144859314,
0.04624165967106819,
-0.05263179913163185,
0.020747574046254158,
-0.00985780730843544,
0.11155629903078079,
0.013813150115311146,
0.05708306282758713,
-0.001901037059724331,
-0.002347403671592474,
0.051945168524980545,
-0.03715088590979576,
0.007172788493335247,
-0.0004118379729334265,
-0.0005359924980439246,
-0.03002767823636532,
-0.11905138194561005,
-0.030677303671836853,
-0.008620389737188816,
0.07031238079071045,
-0.00003260477024014108,
-0.076871357858181,
-0.034272681921720505,
0.09228338301181793,
0.03275477513670921,
0.0740571990609169,
-0.026039255782961845,
-0.01721857488155365,
0.06881208717823029,
-0.06144896149635315,
0.07171213626861572,
0.004526528064161539,
0.019838349893689156,
0.03903931751847267,
-0.04009149968624115,
0.024481521919369698,
-0.057852696627378464,
-0.007432977668941021,
0.057056259363889694,
-0.039876729249954224,
0.006675658747553825,
-0.017874177545309067,
-0.03220546990633011,
0.017088089138269424,
0.038127969950437546,
-0.035735275596380234,
0.030070865526795387,
-0.02754810079932213,
0.0513480082154274,
-0.0636969804763794,
-0.08490970730781555,
-0.011155874468386173,
-0.037741519510746,
0.09375273436307907,
-0.0024385128635913134,
-0.04990643262863159,
-0.004585930146276951,
-0.0780840516090393,
0.06662251800298691,
-0.06977659463882446,
-0.09980066120624542,
-0.0672781690955162,
-0.13631977140903473,
0.044048745185136795,
-0.040927544236183167,
0.0032595188822597265,
0.046486757695674896,
-4.153663930750767e-34,
-0.006613856181502342,
0.008108802139759064,
-0.059398334473371506,
-0.05397068336606026,
0.004423877689987421,
-0.021076541393995285,
0.01652628183364868,
-0.04235106334090233,
-0.04076562449336052,
-0.0435074046254158,
0.006628234405070543,
-0.04854397475719452,
0.004735031630843878,
0.004168929066509008,
0.06836070865392685,
0.058172788470983505,
-0.030721688643097878,
0.027395905926823616,
0.004236636217683554,
-0.02001865766942501,
-0.04543827474117279,
0.053679194301366806,
0.013359950855374336,
0.058542944490909576,
-0.08683229237794876,
0.02243988774716854,
0.011102491989731789,
0.05034837871789932,
-0.020039493218064308,
0.04836694896221161,
-0.10655274987220764,
0.009455095045268536,
-0.07488136738538742,
0.08237127959728241,
-0.0773174986243248,
0.09877142310142517,
-0.03334830328822136,
-0.037216633558273315,
0.01127232052385807,
0.06601503491401672,
0.07580631971359253,
0.014190913178026676,
-0.04609993100166321,
0.014508596621453762,
-0.011302202939987183,
-0.016329236328601837,
0.020825233310461044,
-0.04388823360204697,
0.020271120592951775,
0.044965267181396484,
0.016630850732326508,
-0.026668574661016464,
-0.03256494551897049,
0.1085464358329773,
-0.044356729835271835,
0.044166870415210724,
-0.0629354864358902,
0.06140081211924553,
0.0948236733675003,
-0.07749796658754349,
-0.07891080528497696,
-0.03523772954940796,
-0.03969745337963104,
0.04956561326980591,
-0.0022702529095113277,
-0.01801377721130848,
-0.0029553056228905916,
-0.011761005036532879,
0.0444052591919899,
0.014615540392696857,
-0.08067021518945694,
0.0672493651509285,
0.07639181613922119,
0.05059240758419037,
-0.06687508523464203,
0.0102577218785882,
0.0038790625985711813,
-0.007916083559393883,
-0.03411252051591873,
-0.0360737107694149,
0.08936721086502075,
-0.004574617370963097,
0.07918890565633774,
0.050054945051670074,
0.05100177600979805,
0.051656611263751984,
-0.021815355867147446,
0.06534462422132492,
0.00558141153305769,
-0.10462646931409836,
0.08092906326055527,
-0.04455528035759926,
-0.03168535605072975,
-0.03432778641581535,
-0.008964866399765015,
-3.868258957595572e-8,
-0.06852460652589798,
0.0016518530901521444,
-0.015931637957692146,
0.006285015493631363,
0.04252955690026283,
-0.018386783078312874,
0.036102812737226486,
-0.009005208499729633,
0.04150915518403053,
-0.06922750920057297,
0.06335921585559845,
0.015568897128105164,
0.0004446480015758425,
-0.05588344857096672,
-0.11975175887346268,
0.11369212716817856,
-0.05773823708295822,
0.019089221954345703,
0.015097162686288357,
-0.019860299304127693,
-0.0017076822696253657,
0.0919191837310791,
0.009611976332962513,
0.03371868655085564,
0.044678885489702225,
-0.016732536256313324,
-0.05042377859354019,
0.022434357553720474,
0.10894181579351425,
-0.014542336575686932,
-0.0351647213101387,
0.017831606790423393,
-0.02244853973388672,
0.05602336302399635,
0.0360092893242836,
0.1252973973751068,
0.004031449556350708,
0.03093208745121956,
-0.080854631960392,
-0.025888267904520035,
-0.07437900453805923,
0.06432183086872101,
-0.030668802559375763,
-0.0008770679123699665,
0.06999274343252182,
0.05949152261018753,
-0.014179859310388565,
-0.02667064405977726,
0.07072746753692627,
0.0074679432436823845,
0.03658147528767586,
0.026080289855599403,
0.06716207414865494,
0.043291978538036346,
0.0339207798242569,
0.008338079787790775,
0.0023589669726788998,
-0.0666763111948967,
-0.010032868944108486,
0.004014409147202969,
0.008428790606558323,
-0.02340678870677948,
0.023658370599150658,
0.01383232232183218
] | https://github.com/ucas-yanyang/lancbio | 2 | 0 | 0 | 0 |
Toward Understanding In-context vs. In-weight Learning | https://openreview.net/forum?id=aKJr5NnN8U | [
"Bryan Chan",
"Xinyi Chen",
"András György",
"Dale Schuurmans"
] | Poster | It has recently been demonstrated empirically that in-context learning emerges in transformers when certain distributional properties are present in the training data, but this ability can also diminish upon further training. We provide a new theoretical understanding of these phenomena by identifying simplified distributional properties that give rise to the emergence and eventual disappearance of in-context learning. We do so by first analyzing a simplified model that uses a gating mechanism to choose between an in-weight and an in-context predictor. Through a combination of a generalization error and regret analysis we identify conditions where in-context and in-weight learning emerge. These theoretical findings are then corroborated experimentally by comparing the behaviour of a full transformer on the simplified distributions to that of the stylized model, demonstrating aligned results. We then extend the study to a full large language model, showing how fine-tuning on various collections of natural language prompts can elicit similar in-context and in-weight learning behaviour. | In-context learning, generalization error, transformers | We propose a theoretical explanation on when a model will choose in-context learning vs in-weight learning and support the explanation with experiments. | 13,085 | 2410.23042 | [
-0.08275657892227173,
-0.038399334996938705,
0.06132369488477707,
0.08708355575799942,
0.1025480180978775,
-0.04830717667937279,
0.08313922584056854,
0.014452684670686722,
0.10587304830551147,
-0.03173883259296417,
0.004405007231980562,
-0.03722570091485977,
0.06693657487630844,
0.02765015698969364,
0.07375828921794891,
-0.05524470657110214,
0.07562369108200073,
0.05375494435429573,
-0.13601745665073395,
-0.030628878623247147,
0.004328976385295391,
0.040178727358579636,
0.020187757909297943,
0.021893233060836792,
0.04939613863825798,
0.0006339284591376781,
-0.053525280207395554,
-0.01882101409137249,
0.0725744292140007,
0.03101823478937149,
0.019553570076823235,
0.06907808035612106,
-0.043249551206827164,
0.025914857164025307,
-0.0868564173579216,
0.06428119540214539,
-0.041521623730659485,
0.04199992120265961,
0.053725168108940125,
-0.018443481996655464,
-0.044712040573358536,
-0.013269559480249882,
0.01071550976485014,
0.056102387607097626,
0.06706295907497406,
-0.03958596661686897,
-0.006204196251928806,
-0.013938415795564651,
-0.06362757086753845,
-0.012058201245963573,
-0.08026013523340225,
-0.01879114843904972,
-0.0006408055196516216,
0.05948929488658905,
-0.04258811101317406,
0.07473360002040863,
0.059132568538188934,
0.0638248547911644,
-0.010349396616220474,
-0.02735004387795925,
-0.10491893440485,
-0.14207744598388672,
-0.0807817205786705,
0.04114999994635582,
-0.040408965200185776,
-0.01970125548541546,
0.0037490774411708117,
0.04926856979727745,
-0.06528284400701523,
0.00777755631133914,
0.05631031468510628,
0.034814901649951935,
-0.032775457948446274,
0.1235702633857727,
0.0952812060713768,
0.047480080276727676,
0.00577638391405344,
-0.0625094547867775,
0.0007017058087512851,
0.007808713708072901,
0.047899819910526276,
0.021869326010346413,
0.06942853331565857,
-0.03576623648405075,
0.11784432828426361,
0.0014406067784875631,
0.05456945300102234,
0.023411426693201065,
-0.017685217782855034,
0.04385204613208771,
-0.016817893832921982,
-0.06589949131011963,
-0.013272225856781006,
0.05804084241390228,
-0.023139698430895805,
0.0702049508690834,
-0.02735810913145542,
-0.07853096723556519,
-0.017620470374822617,
0.1202661469578743,
0.01174229010939598,
0.013376077637076378,
0.040451530367136,
0.019751789048314095,
-0.051707036793231964,
-0.0648595467209816,
-0.03728100284934044,
0.03278391435742378,
0.01378319039940834,
-0.07129031419754028,
-0.000983195728622377,
0.012698724865913391,
0.023754622787237167,
0.01736919768154621,
-0.03359515964984894,
-0.015429730527102947,
0.06379114836454391,
-0.042741913348436356,
-0.025624843314290047,
0.10816095024347305,
0.0060533881187438965,
0.04806412383913994,
-0.06579238921403885,
-0.021433150395751,
-0.05531640723347664,
-0.05030721053481102,
-0.06035137176513672,
9.23752592866598e-34,
0.0442974790930748,
-0.02502969652414322,
0.0269978865981102,
0.061190295964479446,
0.0065461136400699615,
0.027786580845713615,
-0.004805654287338257,
0.006028315518051386,
0.009261573664844036,
-0.0029968738090246916,
0.08848965913057327,
0.10638662427663803,
-0.022533610463142395,
0.08253984898328781,
0.009186826646327972,
0.01003708504140377,
-0.0536063015460968,
0.050575561821460724,
0.006464000791311264,
-0.02927209995687008,
0.010337700136005878,
-0.009949252009391785,
0.032142091542482376,
-0.08270259946584702,
-0.04215323179960251,
0.009055281057953835,
0.037763938307762146,
-0.05110170692205429,
-0.06649017333984375,
0.00943618081510067,
-0.10408395528793335,
0.010057251900434494,
-0.052622802555561066,
-0.008123801089823246,
-0.020783012732863426,
-0.08846498280763626,
0.04694316163659096,
-0.05673639848828316,
-0.002851742785423994,
-0.05213077366352081,
-0.011234421283006668,
0.03185799717903137,
0.04671184718608856,
0.01217381190508604,
-0.09432189166545868,
-0.001827500993385911,
0.043019089847803116,
-0.016558540984988213,
-0.03497723489999771,
-0.03562608361244202,
0.024232778698205948,
-0.003924163058400154,
-0.04212287813425064,
-0.002411277499049902,
0.06322576105594635,
0.10947474092245102,
-0.004794435575604439,
0.02868390642106533,
0.007393488194793463,
-0.034705184400081635,
0.006823049392551184,
0.07522552460432053,
0.01888153702020645,
0.020569933578372,
0.02084130235016346,
-0.0055464887991547585,
-0.01567394845187664,
0.003619341878220439,
0.010973285883665085,
-0.0269829872995615,
-0.01548257376998663,
0.007944701239466667,
-0.10500598698854446,
-0.027663305401802063,
0.07584258913993835,
-0.018266141414642334,
0.063804492354393,
-0.06741078943014145,
0.01042503397911787,
0.042876578867435455,
0.03505180776119232,
-0.035167451947927475,
0.07009222358465195,
-0.02708553709089756,
-0.0777311623096466,
-0.027549171820282936,
0.04012460261583328,
-0.10976522415876389,
-0.0321025587618351,
-0.04797070845961571,
-0.07671576738357544,
-0.01745644025504589,
-0.06933348625898361,
-0.03687155619263649,
0.0026389905251562595,
-2.5730155236149117e-33,
-0.051978789269924164,
-0.023596959188580513,
-0.05013950169086456,
0.02397693321108818,
0.00175092660356313,
-0.018206322565674782,
0.023861976340413094,
0.012015504762530327,
-0.019666459411382675,
-0.08565258234739304,
0.012666167691349983,
-0.0030126520432531834,
-0.010377180762588978,
0.03193072974681854,
0.048350948840379715,
-0.00525441812351346,
0.011768745258450508,
0.06080174446105957,
0.022668562829494476,
-0.015333004295825958,
0.09950742870569229,
0.039979465305805206,
-0.1302732229232788,
0.05882655829191208,
-0.009708165191113949,
-0.06415637582540512,
-0.0907682403922081,
0.10483680665493011,
-0.017160747200250626,
0.021832454949617386,
-0.0614347830414772,
0.010141403414309025,
0.017759498208761215,
0.012434451840817928,
-0.09790458530187607,
0.05754604935646057,
0.035736892372369766,
-0.021930702030658722,
-0.07062524557113647,
0.09412576258182526,
0.013989732600748539,
-0.01873059943318367,
-0.013831636868417263,
-0.009715113788843155,
-0.04202011600136757,
-0.08314671367406845,
-0.03209080547094345,
-0.07053378224372864,
0.05832388252019882,
0.0027158050797879696,
-0.0103380735963583,
0.04330938681960106,
-0.08064106851816177,
0.026685692369937897,
-0.08674709498882294,
-0.06733652204275131,
0.019840311259031296,
-0.07571857422590256,
0.0486302450299263,
-0.014888671226799488,
-0.056527279317379,
-0.014272232539951801,
0.018004216253757477,
-0.08428313583135605,
-0.03840995207428932,
-0.07588419318199158,
-0.0055461362935602665,
-0.026942627504467964,
0.022607382386922836,
-0.020648838952183723,
0.006728613283485174,
-0.021764103323221207,
-0.015160657465457916,
-0.03229255601763725,
-0.03368914872407913,
0.018159840255975723,
0.06821300834417343,
-0.07135344296693802,
-0.0404980294406414,
-0.06266198307275772,
0.02563914842903614,
0.004018990322947502,
0.022532928735017776,
0.01985090784728527,
0.0736389011144638,
0.11822415888309479,
-0.01605307124555111,
0.10286679118871689,
0.06651245802640915,
-0.002432470442727208,
-0.011260299012064934,
0.026854949072003365,
0.024626445025205612,
0.052848927676677704,
-0.1040794625878334,
-4.69193679464297e-8,
-0.11666068434715271,
0.038462311029434204,
-0.011742853559553623,
0.07581450790166855,
0.07765243947505951,
0.039603471755981445,
0.011866502463817596,
0.08327661454677582,
-0.05148245394229889,
-0.02989869937300682,
-0.001684745540842414,
0.03811706602573395,
-0.01787109486758709,
-0.03945045545697212,
0.01350671611726284,
0.13101094961166382,
-0.01722511649131775,
0.011374950408935547,
-0.009655673988163471,
0.018121257424354553,
0.10296618193387985,
0.055227018892765045,
-0.01759529858827591,
-0.044651180505752563,
0.006676970981061459,
-0.06149350106716156,
-0.06872323900461197,
0.07897239178419113,
0.041850388050079346,
0.036349568516016006,
-0.014431878924369812,
0.012755298987030983,
-0.057127419859170914,
-0.0730329379439354,
-0.015112830325961113,
0.08838152885437012,
0.060127392411231995,
-0.08827953785657883,
0.003864564001560211,
-0.027258340269327164,
0.0186319500207901,
0.01173276174813509,
0.008994203992187977,
0.045073412358760834,
-0.000996473478153348,
0.03916444256901741,
-0.03450077399611473,
-0.07029169797897339,
-0.006193090230226517,
-0.03338351100683212,
0.056068286299705505,
-0.01479492150247097,
-0.056404706090688705,
0.048270754516124725,
0.06777192652225494,
-0.016297928988933563,
-0.028860555961728096,
-0.03245437517762184,
-0.03440586104989052,
0.0730757862329483,
-0.01397782750427723,
0.030787494033575058,
-0.023457814007997513,
0.047213658690452576
] | 0 | 0 | 0 | 0 |
|
RouteLLM: Learning to Route LLMs from Preference Data | https://openreview.net/forum?id=8sSqNntaMr | [
"Isaac Ong",
"Amjad Almahairi",
"Vincent Wu",
"Wei-Lin Chiang",
"Tianhao Wu",
"Joseph E. Gonzalez",
"M Waleed Kadous",
"Ion Stoica"
] | Poster | Large language models (LLMs) excel at a wide range of tasks, but choosing the right model often involves balancing performance and cost. Powerful models offer better results but are expensive, while smaller models are more cost-effective but less capable. To address this trade-off, we introduce a training framework for learning efficient router models that dynamically select between a stronger and weaker LLM during inference. Our framework leverages human preference data and employs data augmentation techniques to enhance performance. Evaluations on public benchmarks show that our approach can reduce costs by over 2 times without sacrificing response quality. Moreover, our routers exhibit strong generalization capabilities, maintaining performance even when routing between LLMs not included in training. This highlights the potential of our framework to deliver cost-effective, high-performance LLM solutions. | Large language models, query routing | null | 13,083 | null | [
-0.03185151889920235,
-0.08971942216157913,
0.038572605699300766,
0.06731142103672028,
0.015620602294802666,
-0.006044314242899418,
-0.02852974459528923,
-0.011875767260789871,
-0.037359971553087234,
-0.025124656036496162,
-0.07677415013313293,
-0.0033032726496458054,
-0.02594241127371788,
0.0029732994735240936,
0.06721293926239014,
0.07228690385818481,
0.14757193624973297,
0.04816370829939842,
-0.05016294866800308,
-0.053665641695261,
-0.015565005131065845,
0.012889501638710499,
0.031334396451711655,
-0.019302494823932648,
0.03974207863211632,
-0.023196114227175713,
0.03647435083985329,
0.013665303587913513,
0.023986762389540672,
-0.031510129570961,
0.058948345482349396,
0.007921396754682064,
-0.07604996114969254,
0.027699584141373634,
-0.060548216104507446,
-0.030241668224334717,
-0.09476017206907272,
-0.02586045302450657,
-0.013505905866622925,
-0.00540558248758316,
-0.027369454503059387,
-0.025110701099038124,
-0.017217082902789116,
0.0037642924580723047,
0.03169865906238556,
0.012450603768229485,
0.026367707177996635,
0.010737217031419277,
0.014927350915968418,
0.0597921647131443,
-0.018800891935825348,
-0.02498672716319561,
-0.002791851991787553,
0.008723915554583073,
-0.04212430864572525,
0.02539454773068428,
0.0053682466968894005,
0.025411337614059448,
-0.05155806615948677,
-0.03583656996488571,
-0.07343222200870514,
-0.13679534196853638,
-0.005405230447649956,
0.021536851301789284,
-0.0900067463517189,
-0.06651493906974792,
0.04209795221686363,
0.05859041586518288,
-0.03360262140631676,
0.047482337802648544,
0.0202003326267004,
0.07897599041461945,
-0.033497296273708344,
0.08642436563968658,
0.03155634179711342,
0.03146716207265854,
-0.00623224675655365,
-0.07838951796293259,
0.05651146546006203,
0.012340737506747246,
-0.04520324990153313,
-0.014811521396040916,
-0.006037544459104538,
0.05299714207649231,
0.15994080901145935,
-0.05332932621240616,
0.02233988046646118,
0.03291654214262962,
-0.007028346415609121,
-0.002007524948567152,
-0.0552348755300045,
-0.02000478468835354,
0.0854298323392868,
-0.005848463624715805,
-0.07817655056715012,
0.05199126526713371,
-0.03258340433239937,
-0.12367690354585648,
-0.06055586040019989,
0.08828528225421906,
-0.0328480526804924,
0.0912710651755333,
-0.00004681048812926747,
-0.03390317037701607,
-0.0114840567111969,
0.004693625494837761,
0.0426802895963192,
0.0934491828083992,
0.04917975887656212,
-0.011212448589503765,
0.006837996654212475,
0.044188786298036575,
-0.005200461950153112,
-0.017019787803292274,
0.0027122653555125,
-0.07671091705560684,
0.013446670956909657,
-0.06803018599748611,
-0.0007750002550892532,
0.0352821871638298,
-0.06745569407939911,
0.01845012605190277,
0.03578278049826622,
0.016657164320349693,
-0.0019244012655690312,
-0.045390717685222626,
-0.0066154212690889835,
3.4985084909692e-33,
0.04330950230360031,
0.06834235787391663,
0.002426178427413106,
-0.02976793237030506,
0.05255032703280449,
0.01280918251723051,
0.059780001640319824,
0.045482005923986435,
-0.05192701891064644,
-0.015405507758259773,
-0.06480585038661957,
-0.011131239123642445,
-0.013751329854130745,
0.0742853432893753,
0.05462392419576645,
-0.014006122015416622,
-0.05390460044145584,
0.06856264919042587,
-0.017203155905008316,
-0.03938799723982811,
0.08472593873739243,
-0.013161291368305683,
0.039764124900102615,
-0.045488256961107254,
0.08770174533128738,
-0.025645356625318527,
0.09638546407222748,
0.01914123445749283,
-0.00293348147533834,
-0.039271097630262375,
-0.009380445815622807,
-0.03530912473797798,
0.025261290371418,
0.03095255233347416,
0.04541321098804474,
-0.0031897490844130516,
-0.03966932371258736,
-0.013484809547662735,
0.015786122530698776,
0.003216022625565529,
-0.024982620030641556,
0.054791148751974106,
0.016111532226204872,
-0.023819919675588608,
-0.13749390840530396,
-0.06873724609613419,
-0.022198760882019997,
0.048234760761260986,
-0.0450478233397007,
-0.05309043824672699,
0.052499812096357346,
-0.0319114625453949,
-0.05278176814317703,
0.041252367198467255,
-0.04004204645752907,
-0.06178879737854004,
0.011956513859331608,
0.09289153665304184,
0.03127298131585121,
0.018804505467414856,
-0.010655011050403118,
-0.03470971807837486,
0.017021488398313522,
0.07362887263298035,
0.043621473014354706,
-0.014237400144338608,
0.03224647417664528,
0.06575222313404083,
0.050391197204589844,
-0.05339619517326355,
-0.00881507620215416,
0.008449040353298187,
0.03304426744580269,
0.009275151416659355,
0.03823965787887573,
0.00921198632568121,
0.007379730697721243,
-0.06646282225847244,
0.06763159483671188,
-0.007777777500450611,
0.04942291975021362,
0.029383856803178787,
0.010315113700926304,
-0.016152141615748405,
-0.008972176350653172,
-0.06447946280241013,
0.06577087193727493,
-0.12852157652378082,
-0.02297297865152359,
0.08021776378154755,
-0.07129375636577606,
0.015125511214137077,
-0.05659749358892441,
0.053244903683662415,
0.015455658547580242,
-2.9622889267553588e-33,
-0.051042698323726654,
0.02511896938085556,
0.019602464511990547,
0.13581259548664093,
-0.053793374449014664,
-0.07887859642505646,
0.0342562235891819,
0.024321656674146652,
0.007134702987968922,
-0.050308942794799805,
-0.07464933395385742,
-0.030933400616049767,
0.10511382669210434,
-0.034590814262628555,
0.06739768385887146,
-0.0751010924577713,
-0.03785489872097969,
-0.036057084798812866,
0.02857000194489956,
-0.014127164147794247,
-0.08138567209243774,
0.11095964163541794,
-0.09217623621225357,
-0.07010577619075775,
0.005393010564148426,
0.020576318725943565,
-0.033376727253198624,
0.05372476205229759,
-0.025966571643948555,
0.04331710934638977,
-0.06040949374437332,
-0.027878593653440475,
0.01289727259427309,
-0.10695755481719971,
-0.007690925616770983,
0.09798120707273483,
0.029866240918636322,
0.08512908220291138,
-0.006843633018434048,
0.12439731508493423,
0.06846124678850174,
-0.060833510011434555,
-0.07251480966806412,
0.04855296015739441,
-0.0019902775529772043,
-0.030296675860881805,
-0.0837511196732521,
-0.027218444272875786,
0.003184176981449127,
-0.047462549060583115,
-0.0058879125863313675,
0.04274957627058029,
-0.06630165874958038,
0.04006015509366989,
0.020869603380560875,
-0.09138481318950653,
-0.00045217733713798225,
-0.09817080199718475,
0.0618533156812191,
-0.029470626264810562,
-0.0717524066567421,
-0.013684296980500221,
0.050737954676151276,
0.026454495266079903,
0.022399019449949265,
-0.0628073588013649,
-0.000249307369813323,
-0.05214054509997368,
-0.04035263881087303,
-0.03446968272328377,
0.005854946561157703,
-0.04117627069354057,
0.0094450106844306,
0.05625966191291809,
0.048153072595596313,
-0.03703106939792633,
0.025641150772571564,
0.0030700773932039738,
-0.02745283767580986,
-0.005401546135544777,
0.01001310907304287,
-0.010253389365971088,
-0.026284435763955116,
-0.008719312027096748,
0.06754732131958008,
0.034771766513586044,
0.0725264921784401,
0.06725560873746872,
-0.017625877633690834,
0.019306449219584465,
0.0008374212775379419,
0.06280270218849182,
-0.02694578282535076,
0.011137718334794044,
-0.08571913093328476,
-3.982679075420492e-8,
-0.11520364880561829,
0.07146982103586197,
0.04793308675289154,
0.10860349982976913,
0.05147287994623184,
-0.0453622043132782,
0.013764838688075542,
0.15228389203548431,
0.006079552229493856,
-0.009757699444890022,
0.07097649574279785,
0.04860370606184006,
-0.03532687574625015,
-0.0417507030069828,
0.03916032984852791,
0.0707298144698143,
-0.005475040525197983,
-0.011014792136847973,
-0.023401422426104546,
0.00013226915325503796,
-0.01967426761984825,
0.031067779287695885,
0.03496832773089409,
-0.03824688866734505,
0.09549791365861893,
-0.0027918298728764057,
-0.02964489348232746,
0.09794685244560242,
0.036329686641693115,
0.04184548184275627,
-0.05926261097192764,
0.007559084799140692,
0.0029880707152187824,
-0.03672381862998009,
0.07822123169898987,
0.089444100856781,
-0.02102936990559101,
-0.03824693709611893,
0.007783709559589624,
-0.044794950634241104,
-0.0002657763834577054,
0.05458511784672737,
-0.0655248612165451,
0.06264947354793549,
0.0443371944129467,
-0.05559233948588371,
-0.04676526039838791,
-0.06611750274896622,
0.012636237777769566,
-0.08679553121328354,
0.030545253306627274,
-0.06097391992807388,
0.025059262290596962,
-0.009931115433573723,
0.03322732821106911,
-0.00462069408968091,
-0.009387386962771416,
-0.07138940691947937,
0.015907637774944305,
0.06567622721195221,
0.034813474863767624,
-0.08142103999853134,
-0.11541566997766495,
0.017515329644083977
] | 0 | 0 | 0 | 0 |
|
Score-based Self-supervised MRI Denoising | https://openreview.net/forum?id=uNd289HjLi | [
"Jiachen Tu",
"Yaokun Shi",
"Fan Lam"
] | Poster | Magnetic resonance imaging (MRI) is a powerful noninvasive diagnostic imaging tool that provides unparalleled soft tissue contrast and anatomical detail. Noise contamination, especially in accelerated and/or low-field acquisitions, can significantly degrade image quality and diagnostic accuracy. Supervised learning based denoising approaches have achieved impressive performance but require high signal-to-noise ratio (SNR) labels, which are often unavailable. Self-supervised learning holds promise to address the label scarcity issue, but existing self-supervised denoising methods tend to oversmooth fine spatial features and often yield inferior performance than supervised methods. We introduce Corruption2Self (C2S), a novel score-based self-supervised framework for MRI denoising. At the core of C2S is a generalized denoising score matching (GDSM) loss, which extends denoising score matching to work directly with noisy observations by modeling the conditional expectation of higher-SNR images given further corrupted observations. This allows the model to effectively learn denoising across multiple noise levels directly from noisy data. Additionally, we incorporate a reparameterization of noise levels to stabilize training and enhance convergence, and introduce a detail refinement extension to balance noise reduction with the preservation of fine spatial features. Moreover, C2S can be extended to multi-contrast denoising by leveraging complementary information across different MRI contrasts. We demonstrate that our method achieves state-of-the-art performance among self-supervised methods and competitive results compared to supervised counterparts across varying noise conditions and MRI contrasts on the M4Raw and fastMRI dataset. The project website is available at: https://jiachentu.github.io/Corruption2Self-Self-Supervised-Denoising/. | Generalized Denoising Score Matching; Self-supervised Learning; Self-supervised Denoising; Score-based denoising; Medical Image Denoising; | Corruption2Self introduces a score-based self-supervised MRI denoising framework that leverages Generalized Denoising Score Matching (GDSM) to achieve high-quality denoising without clean reference images. | 13,077 | null | [
-0.00966398324817419,
-0.09584823250770569,
0.04017084836959839,
0.00003185620880685747,
0.029930882155895233,
-0.028539041057229042,
0.006773619446903467,
-0.10454156249761581,
-0.02699575386941433,
-0.07779935747385025,
-0.0934891626238823,
-0.0027130465023219585,
0.03772234171628952,
0.05542152374982834,
-0.09094352275133133,
-0.01948496140539646,
0.07914824038743973,
0.02454800345003605,
-0.011811490170657635,
0.0037980168126523495,
-0.05660908296704292,
0.019456522539258003,
-0.018003402277827263,
0.04859078302979469,
0.08794449269771576,
-0.0041579678654670715,
0.04385064169764519,
-0.027263091877102852,
-0.023189086467027664,
-0.04101793095469475,
0.10927315801382065,
-0.029921546578407288,
-0.01888342760503292,
-0.005966354627162218,
-0.03003416769206524,
-0.01410047709941864,
-0.032243773341178894,
0.09206323325634003,
-0.002412430476397276,
-0.0023316561710089445,
0.06054197996854782,
0.02088688313961029,
0.01801009476184845,
-0.03129719942808151,
0.0027665498200803995,
0.0364689975976944,
0.012588794343173504,
-0.09531187266111374,
0.03548416495323181,
0.020305940881371498,
-0.021250639110803604,
0.020139653235673904,
-0.13354748487472534,
0.04904645308852196,
-0.001198886428028345,
0.04342033714056015,
0.08261945098638535,
-0.01064261794090271,
-0.07486966997385025,
0.046888962388038635,
0.07810234278440475,
-0.02306789718568325,
0.04557853937149048,
0.0336298905313015,
0.04784167557954788,
0.025193708017468452,
-0.014924619346857071,
0.004582226276397705,
-0.06286520510911942,
0.06546656042337418,
-0.010561101138591766,
0.014144212007522583,
0.008893081918358803,
0.06967369467020035,
-0.061126332730054855,
0.02958439476788044,
-0.0473148450255394,
0.001569707877933979,
0.1263679563999176,
-0.039423637092113495,
-0.02456882782280445,
0.056746356189250946,
0.03156999498605728,
-0.018350688740611076,
0.10791563987731934,
0.010277376510202885,
0.049336232244968414,
0.0002278158935951069,
-0.06449730694293976,
0.06279430538415909,
0.011440971866250038,
-0.03266604617238045,
-0.14312569797039032,
0.07013845443725586,
-0.015533069148659706,
-0.0000865592883201316,
0.0788869708776474,
-0.00744366180151701,
0.06157328933477402,
0.020229483023285866,
-0.06006604805588722,
0.05298471450805664,
-0.05832371488213539,
0.01342055480927229,
0.03763315826654434,
-0.041834887117147446,
0.0691371038556099,
0.007099892012774944,
0.06252967566251755,
-0.012897158972918987,
0.027071340009570122,
0.014202425256371498,
-0.040619224309921265,
-0.009474921971559525,
0.07480206340551376,
-0.01887074112892151,
0.07165102660655975,
0.03010418824851513,
0.0006642999360337853,
0.007886751554906368,
-0.0008891416946426034,
-0.0735035166144371,
-0.04665538668632507,
-0.04763210564851761,
0.0309905968606472,
0.037840232253074646,
-0.07401137053966522,
2.862981510958961e-33,
0.06238356605172157,
-0.003609647275879979,
0.008067963644862175,
-0.007381589617580175,
-0.02190571092069149,
-0.056521303951740265,
-0.15387380123138428,
-0.02534199133515358,
0.0695643499493599,
0.08497065305709839,
0.029950430616736412,
0.007314698770642281,
-0.005295114126056433,
0.013070285320281982,
0.013335592113435268,
0.058783967047929764,
-0.018363378942012787,
-0.00635354220867157,
-0.0033257484901696444,
-0.09276078641414642,
0.028305884450674057,
0.021014442667365074,
-0.025904517620801926,
-0.03652394935488701,
0.03649125620722771,
0.00008745901141082868,
0.01559364702552557,
-0.004152292385697365,
0.061720434576272964,
0.006941957399249077,
-0.07743626087903976,
-0.0049687535502016544,
0.029567714780569077,
0.027291905134916306,
0.044840019196271896,
-0.06194494292140007,
0.024686163291335106,
-0.01267753541469574,
-0.0026639667339622974,
0.026372596621513367,
0.02688262239098549,
0.004707939922809601,
-0.027189821004867554,
-0.0510077103972435,
0.016319410875439644,
-0.02343250811100006,
0.016117949038743973,
-0.008327517658472061,
-0.03934560343623161,
-0.05518633499741554,
0.021464768797159195,
-0.03603953868150711,
-0.03336624056100845,
-0.03175611421465874,
-0.09055094420909882,
0.002320242812857032,
0.03515404462814331,
0.0033663264475762844,
0.016373278573155403,
-0.003725222544744611,
0.027186784893274307,
-0.046298377215862274,
-0.0005658745067194104,
0.041643176227808,
0.06390616297721863,
0.01995469629764557,
0.021328236907720566,
0.031892918050289154,
0.012665677815675735,
-0.06096230819821358,
0.00859227403998375,
0.06128336861729622,
-0.02773173898458481,
-0.06090610474348068,
0.015303462743759155,
-0.014219601638615131,
0.07148734480142593,
0.03956318274140358,
-0.0035876017063856125,
-0.04022057726979256,
-0.023419685661792755,
0.007271114271134138,
-0.065880686044693,
-0.045435432344675064,
-0.030791383236646652,
0.05158407241106033,
-0.011018470861017704,
-0.03525139018893242,
-0.04346412420272827,
-0.06335672736167908,
-0.04846545681357384,
0.032153233885765076,
0.0019011228578165174,
-0.037966612726449966,
-0.07418954372406006,
-2.3947242546638644e-33,
0.026533914729952812,
0.05617092177271843,
-0.014860927127301693,
0.002037202939391136,
0.027487123385071754,
-0.03141092136502266,
0.025834722444415092,
0.12602518498897552,
0.04995979741215706,
-0.03418644517660141,
-0.04124697297811508,
-0.05442670360207558,
0.007524607703089714,
0.009005711413919926,
-0.022865014150738716,
-0.0667787492275238,
-0.09629083424806595,
-0.001009574392810464,
-0.11357970535755157,
0.010818985290825367,
0.03437339887022972,
0.11478742212057114,
0.05143164470791817,
0.08481542021036148,
-0.06051527336239815,
0.09472194314002991,
-0.0785265564918518,
0.004977077711373568,
0.03108830377459526,
0.027097295969724655,
0.019038576632738113,
0.03724272921681404,
-0.014026358723640442,
-0.10149729251861572,
-0.0308239683508873,
0.05133910849690437,
0.020060228183865547,
-0.13425132632255554,
-0.0434148833155632,
0.07659868896007538,
0.04769701883196831,
0.06846437603235245,
-0.07948324829339981,
0.00005105474701849744,
0.033232495188713074,
-0.11761453747749329,
-0.02587473951280117,
0.026528000831604004,
0.08378523588180542,
0.0901968777179718,
-0.06388538330793381,
-0.005184983368963003,
-0.1336008459329605,
0.07504743337631226,
-0.07088639587163925,
-0.07786936312913895,
-0.034779392182826996,
0.0025250879116356373,
0.04050787165760994,
0.06569518893957138,
-0.058473799377679825,
0.015676075592637062,
-0.06268971413373947,
-0.000010216785994998645,
-0.055228106677532196,
0.032884396612644196,
0.019145987927913666,
0.039793461561203,
0.00892186351120472,
0.10771824419498444,
-0.03835582733154297,
-0.02215684950351715,
-0.017138291150331497,
-0.007539453916251659,
0.041083890944719315,
-0.03532550111413002,
0.008505135774612427,
0.014941873960196972,
-0.010792573913931847,
0.014478120021522045,
0.11693160235881805,
-0.059038061648607254,
0.022780606523156166,
0.11000030487775803,
0.08688528090715408,
-0.01578482612967491,
-0.04215488210320473,
-0.061183832585811615,
-0.004914323799312115,
-0.052110861986875534,
0.01585749350488186,
0.0037618728820234537,
-0.015218282118439674,
-0.0452214814722538,
-0.10297118872404099,
-4.174529877332134e-8,
-0.11398457735776901,
-0.0040860227309167385,
-0.01236980315297842,
0.03629312291741371,
-0.00013966600818093866,
-0.10191838443279266,
-0.024692397564649582,
0.07725393772125244,
0.04421025142073631,
-0.025008076801896095,
0.08878142386674881,
-0.08063709735870361,
-0.14875690639019012,
-0.06532365083694458,
-0.0037179619539529085,
0.029566336423158646,
0.007566771470010281,
0.07307486236095428,
0.00865117460489273,
0.0015137996524572372,
-0.030703263357281685,
-0.03306269645690918,
0.028514979407191277,
-0.07124453783035278,
0.03953643515706062,
0.031045377254486084,
0.027760079130530357,
-0.016179176047444344,
0.0340186171233654,
0.002678612479940057,
0.06564682722091675,
0.011868591420352459,
0.14059767127037048,
-0.0025973853189498186,
-0.03731014207005501,
0.07982167601585388,
0.0991196259856224,
0.019536249339580536,
-0.06322955340147018,
0.030329611152410507,
-0.04264337196946144,
0.06193627417087555,
-0.045214924961328506,
-0.015636373311281204,
-0.0515461266040802,
-0.006749827414751053,
0.046940285712480545,
-0.010129294358193874,
0.023007124662399292,
-0.04866616800427437,
0.036959871649742126,
-0.04168080538511276,
0.05054156482219696,
-0.026105761528015137,
0.08948303759098053,
-0.0011706487275660038,
0.026796597987413406,
0.06716535985469818,
0.043006572872400284,
0.05174078047275543,
-0.015456300228834152,
0.022873224690556526,
0.01070671807974577,
0.01861831359565258
] | 0 | 0 | 0 | 0 |
|
ReMoE: Fully Differentiable Mixture-of-Experts with ReLU Routing | https://openreview.net/forum?id=4D0f16Vwc3 | [
"Ziteng Wang",
"Jun Zhu",
"Jianfei Chen"
] | Poster | Sparsely activated Mixture-of-Experts (MoE) models are widely adopted to scale up model capacity without increasing the computation budget. However, vanilla TopK routers are trained in a discontinuous, non-differentiable way, limiting their performance and scalability.
To address this issue, we propose ReMoE, a fully differentiable MoE architecture that offers a simple yet effective drop-in replacement for the conventional TopK+Softmax routing, utilizing ReLU as the router instead. We further propose methods to regulate the router's sparsity while balancing the load among experts. ReMoE’s continuous nature enables efficient dynamic allocation of computation across tokens and layers, while also exhibiting domain specialization. Our experiments demonstrate that ReMoE consistently outperforms vanilla TopK-routed MoE across various model sizes, expert counts, and levels of granularity. Furthermore, ReMoE exhibits superior scalability with respect to the number of experts, surpassing traditional MoE architectures. The implementation based on Megatron-LM is available at https://github.com/thu-ml/ReMoE. | Mixture-of-Experts, Differentiable Routing, Sparsity | We propose ReMoE, a fully differentiable MoE with ReLU routing. ReMoE consistently outperforms vanilla TopK-routed MoE and exhibits superior scalability w.r.t. the number of experts. | 13,061 | 2412.14711 | [
-0.03395066782832146,
-0.08890299499034882,
0.030174532905220985,
0.0831470638513565,
-0.0032089450396597385,
-0.05927000939846039,
-0.03850226476788521,
0.04556252434849739,
-0.035429276525974274,
-0.05452846363186836,
-0.12218073010444641,
-0.003228861838579178,
0.0170042235404253,
-0.008334018290042877,
-0.052023082971572876,
0.0852324366569519,
0.1250453144311905,
0.12353736907243729,
-0.10462456196546555,
-0.07803241163492203,
-0.018053926527500153,
-0.05170723795890808,
0.002987861167639494,
-0.014607585035264492,
-0.01905428245663643,
-0.018638646230101585,
-0.0025641380343586206,
-0.017789917066693306,
0.02914137952029705,
-0.06759913265705109,
0.036362577229738235,
0.0025122188962996006,
-0.0402759350836277,
0.0026173607911914587,
-0.04843292012810707,
0.09350381046533585,
-0.08450943231582642,
-0.003585053840652108,
-0.005065509118139744,
-0.07515118271112442,
0.002463945420458913,
0.003175236051902175,
-0.0368192195892334,
0.046239182353019714,
0.02831215225160122,
0.004447076469659805,
0.03513270616531372,
0.003212419105693698,
0.042348768562078476,
0.03355136886239052,
0.025058036670088768,
-0.06050475686788559,
-0.00560214975848794,
0.03208141028881073,
-0.012728661298751831,
0.05814795196056366,
0.009128300473093987,
0.029321007430553436,
0.0015411147614941,
0.052982427179813385,
-0.11230941861867905,
-0.07173236459493637,
0.029768936336040497,
-0.026358501985669136,
0.0010929522104561329,
-0.038165584206581116,
0.012556493282318115,
0.03599761798977852,
0.007120752707123756,
0.033116914331912994,
0.049892544746398926,
0.053339891135692596,
-0.04033513367176056,
0.01838037185370922,
-0.00022108317352831364,
-0.04262217879295349,
0.04463464394211769,
0.03868770971894264,
0.04676389694213867,
0.017346154898405075,
0.02772100456058979,
-0.00511018605902791,
-0.046743035316467285,
-0.04834996536374092,
0.14112697541713715,
-0.08294487744569778,
-0.058824729174375534,
0.0007979016518220305,
0.06886222213506699,
-0.030549997463822365,
-0.052491072565317154,
0.009653993882238865,
0.08472003042697906,
-0.024193905293941498,
0.015551571734249592,
0.05416949838399887,
0.002831661608070135,
-0.07782608270645142,
-0.04493475332856178,
0.12065214663743973,
-0.08519812673330307,
-0.013801222667098045,
0.05373760312795639,
-0.025934157893061638,
0.011423921212553978,
-0.0027561846654862165,
0.06258799135684967,
0.09689391404390335,
0.01846858114004135,
-0.031177325174212456,
-0.003159491578117013,
0.03661980852484703,
0.10275305062532425,
-0.019950732588768005,
-0.0229340773075819,
0.01515401341021061,
-0.008080760948359966,
-0.023863159120082855,
-0.020212212577462196,
0.060954850167036057,
-0.08648370206356049,
-0.0030697719193995,
0.044166769832372665,
-0.04082762822508812,
0.06388118118047714,
0.0178015548735857,
-0.044719845056533813,
3.461589886829718e-33,
0.06092274188995361,
0.09150621294975281,
-0.0534416064620018,
0.0044271680526435375,
0.01948186196386814,
-0.023603787645697594,
0.07681871205568314,
0.006414369679987431,
-0.03738417476415634,
-0.03691471368074417,
-0.04510042816400528,
0.03451971337199211,
-0.0573735274374485,
0.010074585676193237,
0.09709619730710983,
-0.022537676617503166,
-0.0027164395432919264,
0.01755492202937603,
-0.014776048250496387,
-0.05780681595206261,
-0.00515262084081769,
-0.011679382063448429,
-0.049156948924064636,
-0.01875687763094902,
-0.011481981724500656,
0.021893683820962906,
0.08195880800485611,
-0.0391029417514801,
-0.022515008226037025,
0.017946694046258926,
-0.03327906131744385,
-0.013924657367169857,
0.015369236469268799,
0.06697124987840652,
0.05432150140404701,
0.004904516506940126,
-0.01734347455203533,
-0.02718224935233593,
0.014607281424105167,
-0.0008112366776913404,
-0.014527082443237305,
0.05948857590556145,
-0.06422283500432968,
-0.017376035451889038,
-0.15059055387973785,
-0.009188487194478512,
0.0021113588009029627,
0.04930111765861511,
-0.05817781761288643,
-0.009075741283595562,
0.0521203987300396,
0.008646367117762566,
0.028523042798042297,
-0.0166348684579134,
-0.019194893538951874,
-0.12383019179105759,
-0.023328492417931557,
0.09395769238471985,
0.0833439826965332,
0.0707116350531578,
0.006307366769760847,
0.008323632180690765,
-0.0039132763631641865,
0.1288917362689972,
0.07225736230611801,
-0.00174823182169348,
-0.0061829108744859695,
0.1340922713279724,
0.061612993478775024,
0.002000733045861125,
-0.019237414002418518,
0.06915087252855301,
0.05988088250160217,
0.006859871558845043,
-0.005367955658584833,
-0.027466148138046265,
0.015061220154166222,
-0.09354797005653381,
0.006857078522443771,
-0.018315041437745094,
0.04649561643600464,
0.09276171028614044,
0.0055320304818451405,
-0.010422185994684696,
-0.008695460855960846,
-0.04895353317260742,
0.02047392539680004,
-0.039219509810209274,
-0.02131180837750435,
0.010728532448410988,
-0.06949084997177124,
-0.061453089118003845,
0.016692282631993294,
0.012939904816448689,
-0.04177581146359444,
-1.3061893943310596e-33,
-0.06669596582651138,
0.05944453179836273,
-0.034687772393226624,
0.11138353496789932,
-0.02632254920899868,
-0.04322690889239311,
0.03141707554459572,
-0.012949163094162941,
-0.0250654648989439,
-0.01837046444416046,
-0.06822944432497025,
-0.06640491634607315,
0.09294414520263672,
-0.048841435462236404,
0.06631702184677124,
-0.012547029182314873,
-0.17593155801296234,
-0.03149156644940376,
-0.0061876121908426285,
-0.030532611533999443,
-0.011650987900793552,
0.00838006753474474,
-0.12108602374792099,
0.02506866306066513,
-0.004519835580140352,
0.057640623301267624,
-0.12897895276546478,
0.11942651122808456,
0.0108904168009758,
0.02661518193781376,
-0.022627277299761772,
-0.06385745853185654,
-0.002908548340201378,
-0.07305662333965302,
0.029678532853722572,
0.08445260673761368,
0.027143899351358414,
0.1416867971420288,
0.008062382228672504,
0.04404398426413536,
0.06775613874197006,
-0.07580340653657913,
-0.050524964928627014,
0.05009245499968529,
0.011882217600941658,
0.022903993725776672,
-0.060514647513628006,
-0.08222633600234985,
-0.007227556314319372,
-0.07565553486347198,
-0.03637116402387619,
-0.005173563491553068,
-0.052705105394124985,
-0.004029403440654278,
-0.006329406518489122,
-0.0034889287780970335,
-0.006136794108897448,
-0.010005954653024673,
0.02526593953371048,
-0.0022998808417469263,
0.02235635183751583,
-0.10801618546247482,
0.007836511358618736,
0.037337638437747955,
0.029196958988904953,
-0.02671685628592968,
0.04426579177379608,
-0.04422534629702568,
-0.06493054330348969,
0.05357983335852623,
-0.018684182316064835,
-0.04432157799601555,
0.08806387335062027,
0.05390937626361847,
0.046132542192935944,
0.016564635559916496,
0.05926652252674103,
-0.05449144169688225,
-0.028566239401698112,
-0.012065568938851357,
-0.03973344713449478,
0.036400582641363144,
-0.00989698525518179,
-0.04430036246776581,
0.10358430445194244,
0.06665464490652084,
0.040373172610998154,
0.06348517537117004,
-0.01952812448143959,
0.015479578636586666,
0.017156802117824554,
0.014081288129091263,
0.01169844064861536,
0.031153609976172447,
-0.057110197842121124,
-4.401940856268993e-8,
-0.01975025236606598,
0.061680495738983154,
-0.008516940288245678,
0.03863667696714401,
0.116116963326931,
-0.004474280867725611,
0.008775188587605953,
0.05856996029615402,
-0.014324573799967766,
0.025183159857988358,
0.06492134928703308,
-0.01813558116555214,
-0.07873083651065826,
-0.06382367014884949,
0.07533761858940125,
0.08660118281841278,
0.008451427333056927,
-0.018882961943745613,
-0.03368141129612923,
-0.04658558592200279,
-0.017163679003715515,
0.012057558633387089,
0.02160920947790146,
-0.01445950660854578,
0.03360692411661148,
-0.05501892790198326,
-0.011101869866251945,
0.08675319701433182,
0.03190977871417999,
0.023658886551856995,
-0.0833950787782669,
0.023646658286452293,
0.04194417595863342,
0.0080643305554986,
-0.005207705777138472,
0.11752984672784805,
-0.024792995303869247,
0.016500648111104965,
-0.04942050576210022,
-0.03193579614162445,
0.023334823548793793,
0.012498535215854645,
0.010695712640881538,
0.03679442033171654,
0.030550939962267876,
-0.05436104163527489,
-0.07579142600297928,
-0.05290517583489418,
0.05293192341923714,
0.005537997931241989,
0.008341552689671516,
-0.02582348883152008,
-0.012487707659602165,
0.02551858127117157,
0.0693911612033844,
0.011293941177427769,
0.01665046066045761,
-0.05594293400645256,
0.024076391011476517,
0.0005468866438604891,
0.05522775277495384,
-0.04175810515880585,
-0.076395682990551,
-0.0035197697579860687
] | https://github.com/thu-ml/remoe | 69 | 0 | 0 | 0 |
HyperDAS: Towards Automating Mechanistic Interpretability with Hypernetworks | https://openreview.net/forum?id=6fDjUoEQvm | [
"Jiuding Sun",
"Jing Huang",
"Sidharth Baskaran",
"Karel D'Oosterlinck",
"Christopher Potts",
"Michael Sklar",
"Atticus Geiger"
] | Poster | Mechanistic interpretability has made great strides in identifying neural network features (e.g., directions in hidden activation space) that mediate concepts (e.g., *the birth year of a Nobel laureate*) and enable predictable manipulation. Distributed alignment search (DAS) leverages supervision from counterfactual data to learn concept features within hidden states, but DAS assumes we can afford to conduct a brute force search over potential feature locations. To address this, we present HyperDAS, a transformer-based hypernetwork architecture that (1) automatically locates the token-positions of the residual stream that a concept is realized in and (2) learns features of those residual stream vectors for the concept. In experiments with Llama3-8B, HyperDAS achieves state-of-the-art performance on the RAVEL benchmark for disentangling concepts in hidden states. In addition, we review the design decisions we made to mitigate the concern that HyperDAS (like all powerful interpretabilty methods) might inject new information into the target model rather than faithfully interpreting it. | mechanistic interpretability, causal abstraction, hypernetwork | HyperDAS, a transformer-based hypernetwork, enhances interpretability by pinpointing and learning concept-specific features in neural networks' residual streams. | 13,048 | 2503.10894 | [
-0.08790574967861176,
-0.06507636606693268,
0.005024539306759834,
0.03143218159675598,
0.07347210496664047,
-0.008209099061787128,
-0.03433368355035782,
-0.07121534645557404,
0.01891893893480301,
-0.004154796712100506,
-0.06294438242912292,
-0.04911129176616669,
-0.014105238020420074,
0.060645248740911484,
-0.042279139161109924,
0.03096981719136238,
0.028810836374759674,
0.07509975880384445,
-0.09034948796033859,
-0.06103317812085152,
0.08359235525131226,
0.03139384835958481,
-0.0003464308974798769,
-0.00707511231303215,
0.013917693868279457,
0.07161910831928253,
-0.017045775428414345,
-0.05368024855852127,
0.004903547931462526,
-0.03854788467288017,
0.053499944508075714,
-0.020892353728413582,
-0.114021435379982,
0.05548793077468872,
-0.05258265882730484,
0.05915987491607666,
-0.05823676288127899,
0.038551002740859985,
-0.0002506383170839399,
-0.07889113575220108,
0.05111771821975708,
-0.009831627830862999,
-0.009175312705338001,
0.0033408987801522017,
0.06657478213310242,
-0.008619639091193676,
0.02687043696641922,
-0.027887659147381783,
0.029204009100794792,
0.025624051690101624,
-0.1332811564207077,
-0.07612577825784683,
0.03113221935927868,
0.09596631675958633,
-0.024325639009475708,
0.03004075400531292,
0.048137471079826355,
0.03198269009590149,
-0.015605435706675053,
-0.05639726668596268,
0.02104935236275196,
-0.0798901915550232,
0.011552930809557438,
-0.02825431525707245,
-0.013544701039791107,
-0.05151645094156265,
0.057910844683647156,
0.058735672384500504,
0.008764976635575294,
-0.01602415181696415,
0.02820742130279541,
0.09946966171264648,
-0.05703074112534523,
0.05756581947207451,
0.049946464598178864,
0.06266903877258301,
0.12910254299640656,
0.013435778208076954,
0.07086285203695297,
-0.03799401968717575,
-0.0008410939481109381,
0.010871128179132938,
0.06486951559782028,
0.03699905052781105,
0.08697632700204849,
-0.0177167896181345,
-0.010693001560866833,
0.02076917514204979,
0.0272644255310297,
-0.01874523051083088,
-0.07361507415771484,
-0.16320458054542542,
0.0473611056804657,
-0.04506871849298477,
0.10456573218107224,
0.024518731981515884,
0.030888590961694717,
-0.06521652638912201,
-0.039893630892038345,
0.08600915968418121,
-0.023463385179638863,
0.10142158716917038,
-0.0024380520917475224,
-0.08657380193471909,
0.043767381459474564,
-0.03392176702618599,
-0.01873932220041752,
0.027346476912498474,
0.058978043496608734,
-0.06098584085702896,
-0.0071882386691868305,
0.06757913529872894,
0.047769997268915176,
0.015300100669264793,
0.049670036882162094,
-0.016895193606615067,
0.037094082683324814,
0.016358448192477226,
-0.021570803597569466,
-0.012027754448354244,
-0.03917230665683746,
-0.0021267132833600044,
0.004207977093756199,
0.05194656923413277,
0.043122101575136185,
0.005220214370638132,
-0.10853851586580276,
2.152919281063462e-34,
0.012092565186321735,
0.049377068877220154,
0.014219542033970356,
0.005900263320654631,
0.05957687273621559,
-0.0645333081483841,
0.03131517022848129,
0.048984698951244354,
-0.009182404726743698,
0.012444536201655865,
-0.07695876061916351,
0.047932907938957214,
-0.016723157837986946,
0.08181709796190262,
-0.07150784879922867,
-0.05286254733800888,
-0.04168354719877243,
0.019985251128673553,
-0.07068672776222229,
-0.051789019256830215,
-0.0020629020873457193,
-0.0024374171625822783,
-0.03532716631889343,
-0.082919642329216,
-0.05005364120006561,
-0.029119795188307762,
0.042072754353284836,
-0.01670128107070923,
-0.04742809385061264,
0.033714063465595245,
-0.1479841023683548,
-0.018751636147499084,
0.022356899455189705,
-0.012367179617285728,
0.09232855588197708,
-0.04165828600525856,
-0.03972235694527626,
-0.09608717262744904,
0.03773961216211319,
-0.05750608071684837,
0.004156799521297216,
0.04270835965871811,
-0.08562477678060532,
-0.06383112072944641,
-0.045190855860710144,
-0.04032720997929573,
-0.02025582082569599,
0.01681492291390896,
-0.046990856528282166,
-0.0799199640750885,
0.03706150874495506,
0.1184454932808876,
-0.058525450527668,
-0.08198234438896179,
-0.014233001507818699,
-0.022554730996489525,
0.01778152957558632,
0.021823447197675705,
0.04937104135751724,
0.07341073453426361,
-0.008573499508202076,
0.00473813759163022,
-0.04581216350197792,
0.09947586059570312,
-0.006843579467386007,
0.011859139427542686,
-0.030072832480072975,
0.04440363124012947,
0.07413683086633682,
0.01818677969276905,
-0.034294284880161285,
0.006296548061072826,
0.006891440600156784,
-0.042432695627212524,
-0.005251864902675152,
-0.022044580429792404,
0.07306691259145737,
-0.08635757118463516,
-0.0029310910031199455,
0.01889357902109623,
0.04461503401398659,
0.05446840077638626,
0.003196701407432556,
0.016714727506041527,
-0.09291908144950867,
-0.029843464493751526,
0.03387291729450226,
-0.11502116173505783,
-0.05921415612101555,
-0.04527907073497772,
0.02305573597550392,
-0.033671509474515915,
0.023254720494151115,
0.002379500772804022,
-0.00422405730932951,
-4.7850527114686305e-34,
-0.06037610024213791,
0.020293157547712326,
-0.028214536607265472,
0.0036134999245405197,
-0.013961272314190865,
-0.03426679968833923,
0.011929021216928959,
-0.019787199795246124,
-0.021781345829367638,
-0.035166021436452866,
0.022984128445386887,
-0.06748020648956299,
-0.07766304910182953,
-0.03208550065755844,
0.03987506777048111,
0.05537338927388191,
0.029164886102080345,
0.015275762416422367,
0.04669344797730446,
0.08401640504598618,
0.04964867979288101,
0.09645070880651474,
-0.136208176612854,
0.06935356557369232,
-0.027666103094816208,
0.019663313403725624,
0.03066333569586277,
0.027979444712400436,
0.003967830445617437,
-0.023745514452457428,
-0.041105806827545166,
0.008307535201311111,
-0.07665082812309265,
-0.08488154411315918,
-0.06485339999198914,
-0.0007815478602424264,
-0.012571703642606735,
-0.017781252041459084,
-0.006222859490662813,
0.08348596096038818,
0.07699581980705261,
-0.0033456296660006046,
-0.07057787477970123,
0.06183379143476486,
-0.03543846681714058,
0.011799306608736515,
-0.08300032466650009,
0.0470443032681942,
0.006550023332238197,
0.049362290650606155,
0.0007063406519591808,
0.0030694003216922283,
-0.031073464080691338,
-0.016681278124451637,
-0.009730445221066475,
-0.03661379590630531,
0.0035018050111830235,
0.07095029205083847,
0.060668304562568665,
0.02537889964878559,
-0.06127393618226051,
-0.017008109018206596,
0.03025936335325241,
0.0052686710841953754,
0.02156117931008339,
0.0020955405198037624,
0.017562730237841606,
0.039188385009765625,
-0.00948996003717184,
-0.0472538135945797,
0.08618813753128052,
-0.03295259550213814,
-0.035430118441581726,
0.01018837932497263,
0.03225099667906761,
-0.06068745255470276,
-0.008646531030535698,
-0.05703991279006004,
-0.03334152698516846,
-0.039764322340488434,
0.048577096313238144,
0.008813759312033653,
0.0265972837805748,
0.04424912855029106,
0.061148207634687424,
0.06239180266857147,
0.0700153112411499,
0.06922997534275055,
-0.0005785216926597059,
0.038778334856033325,
-0.026303978636860847,
0.04125853627920151,
-0.10261848568916321,
0.05229978635907173,
0.0023147775791585445,
-4.14727949760163e-8,
-0.0844859927892685,
0.00880538858473301,
-0.026395801454782486,
0.02318231388926506,
-0.05815820023417473,
-0.000450230116257444,
-0.024138981476426125,
0.09229166805744171,
-0.016361786052584648,
-0.003166453680023551,
0.08276069164276123,
-0.012759139761328697,
-0.02976929023861885,
-0.029654979705810547,
-0.03370363265275955,
0.08688252419233322,
-0.01102626696228981,
-0.0002723782672546804,
-0.018815824761986732,
-0.004401536658406258,
0.042797982692718506,
0.08213193714618683,
0.016888871788978577,
0.0038105505518615246,
0.08027751743793488,
-0.06505110114812851,
-0.04464246705174446,
0.1395014226436615,
0.02857627347111702,
0.03986996412277222,
-0.022183604538440704,
0.05086047947406769,
0.016993656754493713,
-0.041960615664720535,
-0.011293446645140648,
0.13734786212444305,
0.041668765246868134,
-0.018510280176997185,
-0.014040556736290455,
-0.010606903582811356,
0.009438351728022099,
0.1352861225605011,
-0.044076018035411835,
0.03851396590471268,
-0.006778270471841097,
-0.020450342446565628,
-0.020512768998742104,
-0.056552015244960785,
0.016231607645750046,
-0.0540219210088253,
0.042091675102710724,
-0.030253520235419273,
0.034146249294281006,
0.0780130997300148,
0.02974446304142475,
0.03580543398857117,
-0.058220699429512024,
-0.05652083829045296,
-0.005370371043682098,
0.1027684435248375,
0.07076030969619751,
0.008745342493057251,
-0.054507408291101456,
0.021027544513344765
] | 0 | 0 | 0 | 0 |
|
GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models | https://openreview.net/forum?id=AjXkRZIvjB | [
"Seyed Iman Mirzadeh",
"Keivan Alizadeh",
"Hooman Shahrokhi",
"Oncel Tuzel",
"Samy Bengio",
"Mehrdad Farajtabar"
] | Poster | Recent advancements in Large Language Models (LLMs) have sparked interest in their mathematical reasoning capabilities. While performance on the widely popular GSM8K benchmark has improved, questions remain about whether reported evaluation metrics are reliable, and reasoning abilities of LLMs have advanced. To overcome the limitations of existing evaluations, we introduce GSM-Symbolic, an improved benchmark created from symbolic templates that allow for the generation of a diverse set of questions. GSM-Symbolic enables more controllable evaluations, providing key insights and more reliable metrics for measuring the reasoning capabilities of models.
Our findings reveal that LLMs exhibit noticeable variance when responding to different instantiations of the same question. Specifically, the performance of models declines when only the numerical values in the question are altered in the GSM-Symbolic benchmark. Furthermore, we investigate the fragility of mathematical reasoning in these models and demonstrate that their performance significantly deteriorates as the number of clauses in a question increases. We hypothesize that this decline is due to the fact that current LLMs are not capable of genuine logical reasoning; instead, they attempt to replicate the reasoning steps observed in their training data. When we add a single clause that appears relevant to the question, we observe significant performance drops (up to 65%) across all state-of-the-art models, even though the added clause does not contribute to the reasoning chain needed to reach the final answer. Overall, our work provides a more nuanced understanding of LLMs' capabilities and limitations in mathematical reasoning. | Reasoning, Large Language Models, Mathematical Reasoning, Datasets | We study the mathematical reasoning capabilities of large language models (LLMs) and introduce a new benchmark called GSM-Symbolic that enables more controlled and nuanced evaluations. | 13,035 | null | [
-0.018468350172042847,
-0.04345907270908356,
0.07054900377988815,
0.044223833829164505,
0.011800556443631649,
-0.0033366733696311712,
-0.013691444881260395,
0.05525504797697067,
0.06454695761203766,
0.0027263157535344362,
-0.06163986399769783,
-0.039187241345644,
0.07290033996105194,
0.0012784054270014167,
0.06861847639083862,
0.00357986381277442,
0.059014614671468735,
-0.0017768240068107843,
-0.10378141701221466,
-0.08074917644262314,
0.010903296060860157,
-0.039267901331186295,
0.02311694249510765,
-0.00928075797855854,
0.049984730780124664,
-0.0680391937494278,
-0.021725356578826904,
-0.05390035733580589,
0.029832495376467705,
0.037781305611133575,
-0.03659128025174141,
0.048645783215761185,
-0.06178135797381401,
0.08741623908281326,
-0.05493074283003807,
0.050908397883176804,
-0.049126919358968735,
0.0162455253303051,
0.030979597941040993,
-0.04262976720929146,
-0.07231640815734863,
-0.062375668436288834,
0.04743736982345581,
-0.03687073662877083,
0.06024966761469841,
-0.06479260325431824,
0.019551293924450874,
0.01964978128671646,
-0.04988507181406021,
0.01105053722858429,
-0.11686407029628754,
-0.022718435153365135,
-0.054854847490787506,
-0.030583886429667473,
-0.0313117578625679,
-0.011110156774520874,
-0.024423224851489067,
0.09722213447093964,
0.02586621791124344,
-0.008885154500603676,
-0.0994020402431488,
-0.09831465780735016,
-0.03855597972869873,
0.022514495998620987,
-0.020802443847060204,
-0.04419105127453804,
0.04457623511552811,
0.016504505649209023,
-0.024079272523522377,
0.10158631950616837,
0.01792537420988083,
0.07474508136510849,
-0.07352975755929947,
0.08059646189212799,
-0.055830612778663635,
0.07363088428974152,
-0.003922698553651571,
0.013335245661437511,
0.030234908685088158,
-0.04918338730931282,
0.00792954582720995,
-0.02774500474333763,
-0.010066037997603416,
0.02159896306693554,
0.09287505596876144,
-0.012013532221317291,
0.023726478219032288,
0.11172322928905487,
0.05568965524435043,
0.020719116553664207,
0.042178649455308914,
-0.009664773009717464,
0.008433662354946136,
0.07109902054071426,
0.040196288377046585,
0.08256139606237411,
-0.024759620428085327,
-0.02914881333708763,
-0.024170802906155586,
0.11188212782144547,
0.04852011427283287,
0.1297520250082016,
0.04733669012784958,
-0.021117307245731354,
0.019021330401301384,
-0.02841692417860031,
0.014267373830080032,
0.009654874913394451,
0.045391011983156204,
-0.0755838006734848,
0.032579001039266586,
0.0873929113149643,
0.03755830228328705,
0.042219128459692,
-0.0757918655872345,
-0.016957996413111687,
0.031123816967010498,
-0.05130807310342789,
0.04633675143122673,
0.036033377051353455,
0.03809608519077301,
0.01950782723724842,
0.011811226606369019,
-0.018687929958105087,
0.015326809138059616,
-0.009816836565732956,
-0.07930180430412292,
3.331928405912456e-34,
0.05801562964916229,
0.0314352884888649,
0.09996683150529861,
0.01477809064090252,
0.03669112175703049,
-0.03184698149561882,
-0.02527742087841034,
0.03172003850340843,
0.014027275145053864,
0.03604026138782501,
-0.05569899454712868,
0.03430468961596489,
-0.048310890793800354,
0.015310029499232769,
0.0714569017291069,
0.04093969985842705,
-0.05766802653670311,
-0.0047334227710962296,
-0.07862928509712219,
-0.044669609516859055,
0.029146170243620872,
-0.05763962119817734,
0.027759969234466553,
-0.012049470096826553,
0.01041675265878439,
0.10962611436843872,
0.07360416650772095,
-0.011421727947890759,
-0.027825431898236275,
0.0011653683613985777,
-0.08411511033773422,
-0.10491177439689636,
-0.018317164853215218,
0.07709427922964096,
0.03471086546778679,
-0.08208715170621872,
-0.00335010327398777,
-0.0923786535859108,
0.04839430749416351,
-0.05679553374648094,
-0.05848683416843414,
0.012321146205067635,
0.017224736511707306,
0.051468946039676666,
-0.042569857090711594,
-0.038133297115564346,
-0.04035975784063339,
-0.010489818640053272,
-0.023897884413599968,
0.002385922009125352,
0.058392804116010666,
0.07535720616579056,
-0.05990970879793167,
0.01005768682807684,
-0.0010523406090214849,
0.00826176442205906,
-0.005800067912787199,
0.01874588429927826,
-0.05234846845269203,
0.08466416597366333,
-0.011419732123613358,
0.052118148654699326,
0.03064730390906334,
0.04290036857128143,
-0.03340228274464607,
0.03303671255707741,
-0.07421858608722687,
0.07159924507141113,
0.1047508493065834,
0.030012669041752815,
-0.005477588623762131,
-0.04085232689976692,
-0.02125207893550396,
-0.03539511188864708,
0.07791971415281296,
-0.06379728019237518,
0.054651983082294464,
-0.03376970440149307,
-0.007092608604580164,
-0.015973029658198357,
0.05719967186450958,
0.012619037181138992,
-0.028866343200206757,
-0.00816389825195074,
-0.023293958976864815,
-0.0815802589058876,
0.02779935859143734,
-0.09014037251472473,
-0.060204848647117615,
-0.03776881471276283,
-0.0794336125254631,
0.0014771840069442987,
-0.03522133454680443,
-0.025017451494932175,
-0.030764784663915634,
-2.638864145754084e-33,
-0.09880108386278152,
-0.07251439988613129,
-0.015614830888807774,
0.10099437087774277,
-0.059625472873449326,
-0.09261266887187958,
0.005744528491050005,
0.033853933215141296,
-0.05974594131112099,
-0.06607621908187866,
-0.03878095746040344,
-0.010213238187134266,
0.06266000121831894,
0.033493708819150925,
0.03171585127711296,
-0.049296557903289795,
-0.01643148995935917,
-0.022415857762098312,
-0.019706791266798973,
0.053575824946165085,
-0.005025925580412149,
0.14931152760982513,
-0.05080731585621834,
0.03790069371461868,
-0.03589210286736488,
0.025177735835313797,
-0.03794049099087715,
0.05921123921871185,
-0.040280282497406006,
0.015610535629093647,
-0.012615317478775978,
-0.020854316651821136,
-0.03870517387986183,
-0.0356312021613121,
-0.01981147937476635,
-0.03049929253757,
0.08845581859350204,
-0.005780555307865143,
-0.018692446872591972,
0.08184255659580231,
0.05211944505572319,
-0.08492796868085861,
0.010704397223889828,
-0.007494955323636532,
-0.022915424779057503,
0.026958512142300606,
-0.03539479151368141,
-0.02308190055191517,
0.03725675493478775,
-0.041049424558877945,
0.025783397257328033,
-0.047867659479379654,
-0.010713712312281132,
0.021335788071155548,
0.0003907040518242866,
-0.056887492537498474,
-0.0933036133646965,
-0.049660574644804,
-0.02383878454566002,
-0.003766345791518688,
-0.0571114718914032,
-0.00010122769162990153,
0.054300278425216675,
-0.09512253850698471,
0.007475976832211018,
-0.016891129314899445,
-0.01241341419517994,
-0.03851092606782913,
0.004178427625447512,
-0.009225518442690372,
0.034251365810632706,
-0.028632832691073418,
-0.05460713803768158,
0.05178893730044365,
-0.0010011183330789208,
0.03504827618598938,
-0.031023159623146057,
-0.056786585599184036,
-0.04899239540100098,
-0.13366732001304626,
0.07555223256349564,
0.01678486168384552,
0.026316462084650993,
0.046283166855573654,
-0.0207674577832222,
-0.009660201147198677,
0.043879199773073196,
0.120850570499897,
0.029135311022400856,
0.08696296066045761,
-0.030114224180579185,
0.051853246986866,
-0.08019682765007019,
0.09798768162727356,
-0.10815411806106567,
-4.004029108273244e-8,
-0.004971465561538935,
0.09344515204429626,
-0.04378778859972954,
0.020209603011608124,
0.0709475576877594,
-0.04032542183995247,
-0.09669389575719833,
0.0414213202893734,
0.004628349095582962,
-0.06000496819615364,
0.10046737641096115,
0.0044991145841777325,
-0.10958714783191681,
0.022321658208966255,
-0.0062709953635931015,
0.03978009894490242,
-0.017877429723739624,
-0.0024741017259657383,
-0.0405859500169754,
0.013948256149888039,
0.025486668571829796,
0.0763143002986908,
-0.04012368246912956,
-0.0019540865905582905,
0.031196855008602142,
-0.060585666447877884,
-0.047261547297239304,
0.15710432827472687,
0.0374518483877182,
0.025587337091565132,
0.003299690317362547,
-0.0021938704885542393,
0.014028706587851048,
-0.029202574864029884,
0.031490739434957504,
0.025677595287561417,
-0.007570074405521154,
-0.002199768088757992,
0.07144966721534729,
-0.07879848033189774,
0.02943001128733158,
0.028325973078608513,
-0.062468696385622025,
0.05975673347711563,
0.0005417391657829285,
-0.057649869471788406,
-0.07241044193506241,
-0.02793884463608265,
-0.02748405560851097,
-0.012591262347996235,
-0.022899556905031204,
-0.04755188152194023,
-0.10453186184167862,
0.016497716307640076,
0.034507960081100464,
-0.051656533032655716,
-0.05264723300933838,
-0.01038079708814621,
-0.029708070680499077,
-0.03394703194499016,
0.07368604838848114,
0.011649672873318195,
-0.029304197058081627,
-0.004921567626297474
] | 0 | 0 | 0 | 0 |
|
SEMDICE: Off-policy State Entropy Maximization via Stationary Distribution Correction Estimation | https://openreview.net/forum?id=rJ5g8ueQaI | [
"Jongmin Lee",
"Meiqi Sun",
"Pieter Abbeel"
] | Poster | In the unsupervised pre-training for reinforcement learning, the agent aims to learn a prior policy for downstream tasks without relying on task-specific reward functions. We focus on state entropy maximization (SEM), where the goal is to learn a policy that maximizes the entropy of the state's stationary distribution. In this paper, we introduce SEMDICE, a principled off-policy algorithm that computes an SEM policy from an arbitrary off-policy dataset, which optimizes the policy directly within the space of stationary distributions. SEMDICE computes a single, stationary Markov state-entropy-maximizing policy from an arbitrary off-policy dataset. Experimental results demonstrate that SEMDICE outperforms baseline algorithms in maximizing state entropy while achieving the best adaptation efficiency for downstream tasks among SEM-based unsupervised RL pre-training methods. | state entropy maximization, unsupervised reinforcement learning | This paper introduces state-entropy maximization method for RL pre-training based on stationary distribution optimization. | 13,026 | null | [
-0.05858851596713066,
-0.026369554921984673,
0.060106657445430756,
0.032042551785707474,
0.017785774543881416,
0.02078380435705185,
-0.023360474035143852,
-0.03354803845286369,
-0.01979757845401764,
-0.0061348513700068,
-0.06027694791555405,
-0.02636624127626419,
0.023516543209552765,
0.03220771998167038,
0.00697920098900795,
-0.01185033842921257,
0.11903952062129974,
0.060735564678907394,
-0.04493635147809982,
-0.04188160225749016,
0.03798970580101013,
0.06086274981498718,
0.08950459212064743,
0.029957104474306107,
0.07652099430561066,
-0.012413665652275085,
0.0022632814943790436,
-0.01456891093403101,
0.043163228780031204,
-0.007976178079843521,
0.07145188748836517,
-0.036064811050891876,
0.0493401475250721,
-0.05287793278694153,
-0.06554903090000153,
0.02715328149497509,
-0.12457893788814545,
-0.0128532275557518,
-0.04401498660445213,
0.07316724956035614,
-0.008483968675136566,
-0.0023261941969394684,
-0.051599081605672836,
0.014794548973441124,
-0.05290519446134567,
-0.011255171149969101,
0.03184739500284195,
-0.10843592882156372,
0.07093727588653564,
0.06930316984653473,
0.04090196639299393,
0.04848095029592514,
0.029897328466176987,
0.09292583167552948,
-0.009206079877912998,
0.10315500944852829,
0.06655561923980713,
0.013016738928854465,
-0.004611849784851074,
-0.07546021044254303,
-0.05651962757110596,
-0.11604432016611099,
-0.01412571407854557,
-0.005388602148741484,
-0.07981739193201065,
-0.04003695771098137,
0.03247808665037155,
0.09773792326450348,
-0.002957114251330495,
-0.023393219336867332,
-0.06905870139598846,
0.04663902148604393,
0.036894604563713074,
0.007393302861601114,
0.013361243531107903,
-0.005853136070072651,
0.035502202808856964,
-0.057232167571783066,
0.028841612860560417,
-0.041425857692956924,
0.004041926469653845,
-0.007933946326375008,
0.0222590621560812,
0.02891516499221325,
0.10260256379842758,
0.01474403589963913,
0.01082434318959713,
0.06233193725347519,
0.13498371839523315,
0.029558969661593437,
-0.01890067383646965,
0.024381021037697792,
0.0651923194527626,
-0.011486788280308247,
-0.01799081824719906,
0.09150251001119614,
-0.05215376988053322,
-0.0620432049036026,
0.017752043902873993,
0.08725526928901672,
0.01616835966706276,
0.09610074013471603,
-0.03663410618901253,
0.009372610598802567,
0.013564285822212696,
-0.09699640423059464,
-0.04588747024536133,
0.07718458026647568,
0.08466220647096634,
-0.05718380957841873,
0.12158741056919098,
0.014737366698682308,
0.05337843671441078,
-0.017598407343029976,
-0.02125691995024681,
0.09427780658006668,
0.0076101250015199184,
0.020032621920108795,
-0.07308099418878555,
0.04223262891173363,
-0.04149564728140831,
-0.06080121919512749,
0.054728101938962936,
0.02954290434718132,
0.000757863512262702,
-0.031763602048158646,
-0.04657265543937683,
5.988375972691674e-34,
0.04844006150960922,
-0.04104483500123024,
0.06370535492897034,
-0.0813790038228035,
-0.05355871096253395,
0.029758280143141747,
0.03455599397420883,
0.024836840108036995,
-0.03428925201296806,
0.012074523605406284,
-0.006662273313850164,
0.05874382704496384,
-0.047317881137132645,
0.09864074736833572,
0.015355060808360577,
-0.052033912390470505,
-0.10447129607200623,
0.05264882370829582,
0.020580969750881195,
-0.016795530915260315,
0.059725116938352585,
-0.06092061474919319,
-0.003612587694078684,
-0.11063840985298157,
0.050991833209991455,
0.045836761593818665,
0.026443446055054665,
-0.021277010440826416,
0.03927679359912872,
0.014396588318049908,
-0.07194719463586807,
0.016649942845106125,
0.012586008757352829,
-0.031372521072626114,
0.03274725005030632,
-0.010592330247163773,
-0.08167965710163116,
-0.0009225261746905744,
-0.043307725340127945,
-0.06056629493832588,
0.023425236344337463,
0.04098990932106972,
0.08165046572685242,
-0.04886019602417946,
-0.12114514410495758,
-0.15040014684200287,
-0.010010944679379463,
-0.07629517465829849,
0.04263240471482277,
-0.04048949107527733,
0.03087082877755165,
-0.009615895338356495,
0.043941088020801544,
-0.1033196896314621,
-0.03367551043629646,
-0.007899172604084015,
-0.0007384773343801498,
0.0564357154071331,
0.003177703358232975,
0.015376502647995949,
0.020612144842743874,
-0.032838206738233566,
0.06813333183526993,
0.04273763298988342,
0.01889103837311268,
0.06519763916730881,
-0.04108448699116707,
0.05679246038198471,
0.015993282198905945,
-0.06172588840126991,
-0.024816827848553658,
0.0357045941054821,
-0.006861864123493433,
-0.02710622549057007,
0.10383708775043488,
-0.0015923233004286885,
0.05562131851911545,
-0.03480290248990059,
-0.05911766365170479,
0.02918119728565216,
0.001753472606651485,
-0.024348318576812744,
-0.03065631166100502,
-0.05887869372963905,
0.04929153621196747,
-0.016765767708420753,
-0.0456957183778286,
-0.040452174842357635,
-0.05678015947341919,
0.0031399165745824575,
-0.03303370252251625,
-0.06583356112241745,
-0.04442333057522774,
0.030881594866514206,
-0.04222050681710243,
-1.0142875147279998e-33,
-0.03553061559796333,
-0.021801702678203583,
-0.03299324959516525,
0.014739351347088814,
-0.05541505664587021,
-0.03025565668940544,
0.04904806613922119,
0.04946210980415344,
-0.05590145289897919,
-0.021334173157811165,
-0.07037699967622757,
0.017273811623454094,
0.03754620999097824,
-0.018532931804656982,
-0.020023176446557045,
0.014167523011565208,
-0.0353170745074749,
0.07176901400089264,
-0.03465815261006355,
-0.012959739193320274,
0.013964072801172733,
0.08429648727178574,
-0.09091057628393173,
0.10048507899045944,
0.017848072573542595,
0.030334360897541046,
0.01914280652999878,
0.05390206724405289,
-0.046197567135095596,
0.05082491412758827,
-0.05701698362827301,
-0.05960696563124657,
-0.05602312460541725,
0.002933854004368186,
-0.09010552614927292,
0.09276439249515533,
-0.01712208241224289,
0.025509901344776154,
-0.08572524785995483,
0.09506223350763321,
0.008516808040440083,
-0.0045391772873699665,
-0.013408980332314968,
-0.01120880525559187,
0.0027675898745656013,
-0.0070588733069598675,
-0.09732311218976974,
0.06277049332857132,
0.02854294329881668,
0.024977967143058777,
-0.11644377559423447,
0.03799199312925339,
-0.06911133974790573,
0.026317723095417023,
-0.028617650270462036,
-0.06650642305612564,
-0.08738405257463455,
-0.010747727937996387,
-0.002001043874770403,
-0.0214263666421175,
-0.016991302371025085,
0.03489653021097183,
0.017608599737286568,
0.0305164847522974,
-0.048323072493076324,
0.025791995227336884,
0.006239692680537701,
0.028679976239800453,
0.040226906538009644,
-0.019705642014741898,
-0.002634369768202305,
0.04130038619041443,
0.003402897622436285,
-0.07063543796539307,
0.07223743200302124,
-0.008322744630277157,
0.0349600575864315,
0.0020464772824198008,
-0.06528922915458679,
0.00120203394908458,
0.022173594683408737,
0.034354016184806824,
0.004298325628042221,
0.09545013308525085,
0.041664764285087585,
0.08333146572113037,
0.042840782552957535,
0.023532425984740257,
0.022295832633972168,
-0.03010808862745762,
-0.04912994056940079,
0.021168654784560204,
0.03998564928770065,
0.06339716911315918,
-0.045868825167417526,
-3.658497504943625e-8,
-0.05082042142748833,
0.020481711253523827,
0.13783346116542816,
0.011875398457050323,
0.08003928512334824,
0.023928336799144745,
-0.06484206020832062,
0.01153789646923542,
-0.06275913864374161,
-0.09054332226514816,
0.06962716579437256,
0.053681373596191406,
0.02080020122230053,
-0.11262873560190201,
-0.03070739656686783,
-0.01690244860947132,
0.012565107084810734,
0.07676596939563751,
-0.0037197601050138474,
-0.04377582669258118,
0.05791483446955681,
-0.01969837211072445,
0.0008135903044603765,
0.005992793943732977,
0.0773833617568016,
-0.05936726927757263,
0.006784538272768259,
0.04981357976794243,
0.11022623628377914,
-0.04486190900206566,
0.005602575372904539,
-0.0496228002011776,
0.02827264554798603,
0.01342033501714468,
-0.014600902795791626,
0.018138661980628967,
0.0413985550403595,
-0.02917931228876114,
0.0027520840521901846,
-0.06930489093065262,
-0.01939241774380207,
-0.0021715331822633743,
0.0005109231569804251,
-0.011773688718676567,
-0.05942676216363907,
0.03715812787413597,
-0.06603238731622696,
-0.05914856493473053,
0.05507662147283554,
-0.05526556447148323,
0.05761934071779251,
-0.0004118495562579483,
-0.0158988106995821,
0.0925585925579071,
0.07442747801542282,
-0.06179572641849518,
-0.023556401953101158,
-0.04985629767179489,
0.020627381280064583,
0.02159879542887211,
-0.00714849540963769,
0.014588393270969391,
-0.07998687028884888,
-0.001232403563335538
] | 0 | 0 | 0 | 0 |
|
Identifiability for Gaussian Processes with Holomorphic Kernels | https://openreview.net/forum?id=FUaDMRVrbS | [
"Ameer Qaqish",
"Didong Li"
] | Poster | Gaussian processes (GPs) are widely recognized for their robustness and flexibility across various domains, including machine learning, time series, spatial statistics, and biomedicine. In addition to their common usage in regression tasks, GP kernel parameters are frequently interpreted in various applications. For example, in spatial transcriptomics, estimated kernel parameters are used to identify spatial variable genes, which exhibit significant expression patterns across different tissue locations. However, before these parameters can be meaningfully interpreted, it is essential to establish their identifiability. Existing studies of GP parameter identifiability have focused primarily on Mat\'ern-type kernels, as their spectral densities allow for more established mathematical tools. In many real-world applications, particuarly in time series analysis, other kernels such as the squared exponential, periodic, and rational quadratic kernels, as well as their combinations, are also widely used. These kernels share the property of being holomorphic around zero, and their parameter identifiability remains underexplored.
In this paper, we bridge this gap by developing a novel theoretical framework for determining kernel parameter identifiability for kernels holomorphic near zero. Our findings enable practitioners to determine which parameters are identifiable in both existing and newly constructed kernels, supporting application-specific interpretation of the identifiable parameters, and highlighting non-identifiable parameters that require careful interpretation. | Equivalence of Gaussian random measure; kernel parameters; periodicity; identifiability; interpretability | We characterize the identifiability of kernel parameters for Gaussian process kernels that are holomorphic around 0. | 13,025 | null | [
-0.10275036096572876,
-0.07105417549610138,
-0.003319438314065337,
0.016780849546194077,
0.09087838232517242,
-0.009516804479062557,
0.04164409637451172,
-0.015291701070964336,
0.011307056993246078,
-0.031908534467220306,
0.02988894097507,
0.003962460905313492,
-0.04221038892865181,
0.05856305733323097,
-0.03690638393163681,
-0.10308877378702164,
-0.07294688373804092,
0.0064890505746006966,
-0.04814540967345238,
0.04638144373893738,
0.06872227787971497,
0.12490681558847427,
-0.055524177849292755,
0.029772788286209106,
0.06121663749217987,
-0.03798333927989006,
0.1276112198829651,
0.0178847573697567,
-0.029072100296616554,
0.013603771105408669,
0.0630095899105072,
0.01190169993788004,
-0.06340155750513077,
-0.04379903897643089,
-0.007691812701523304,
0.041494566947221756,
0.02036329172551632,
-0.012198874726891518,
0.009949108585715294,
-0.04202393442392349,
0.10169023275375366,
0.0066260830499231815,
-0.011932233348488808,
0.06508537381887436,
0.05945458635687828,
0.051851075142621994,
0.03711281716823578,
-0.13222555816173553,
-0.0414767861366272,
-0.008572185412049294,
-0.0655820444226265,
0.07342363148927689,
0.006676740478724241,
0.026618024334311485,
0.03001268394291401,
-0.054502617567777634,
-0.05136800929903984,
-0.06331159919500351,
0.07823532819747925,
0.022922201082110405,
0.016646357253193855,
-0.0990842804312706,
0.00507415272295475,
-0.012810803018510342,
0.01791992038488388,
0.029546452686190605,
0.028271319344639778,
-0.08617974072694778,
-0.07110676914453506,
0.024943189695477486,
-0.09860029816627502,
0.1129184141755104,
-0.11170732229948044,
0.020126765593886375,
-0.02996036224067211,
0.10139481723308563,
-0.08564463257789612,
0.04230831563472748,
0.035397082567214966,
-0.08562476187944412,
0.07571854442358017,
0.03457909822463989,
0.05755319446325302,
-0.0594729483127594,
0.04773035645484924,
0.04545995593070984,
0.001971641555428505,
0.0872814953327179,
0.01711328513920307,
0.005992974154651165,
0.06489884853363037,
0.04111235588788986,
-0.008785163052380085,
-0.03889389708638191,
0.030215643346309662,
-0.014569664373993874,
0.005092444363981485,
-0.0050097377970814705,
0.10962305217981339,
0.0298573337495327,
-0.06403036415576935,
-0.03790298476815224,
0.04581166058778763,
0.10049285739660263,
0.039736129343509674,
-0.0447700135409832,
0.04684608429670334,
0.0010270333150401711,
0.04143117740750313,
0.04130870848894119,
-0.019629912450909615,
-0.012206018902361393,
0.00923429150134325,
0.08746755868196487,
-0.006445493083447218,
0.002141246572136879,
0.03431098163127899,
0.03153889998793602,
0.030513638630509377,
0.014136264100670815,
-0.014962744899094105,
-0.03099147044122219,
-0.048592619597911835,
0.03439437225461006,
0.03386428952217102,
0.038812797516584396,
-0.05780259519815445,
3.0439740291939995e-33,
0.03886457905173302,
0.004246242344379425,
0.00033511940273456275,
-0.019585778936743736,
0.000420200900407508,
0.008978528901934624,
-0.10574833303689957,
-0.07395671308040619,
0.08734412491321564,
0.010838987305760384,
-0.09767619520425797,
0.021863454952836037,
-0.09247727692127228,
0.02216230519115925,
-0.04571404308080673,
0.13918258249759674,
-0.031672727316617966,
0.07560338079929352,
-0.023086434230208397,
-0.04433436319231987,
0.08024754375219345,
-0.08780070394277573,
-0.06357157975435257,
-0.004032149910926819,
-0.027300741523504257,
0.010341092012822628,
-0.006272102706134319,
-0.07663708180189133,
-0.04292210191488266,
-0.0016922763315960765,
-0.028713449835777283,
-0.05375977233052254,
0.022965827956795692,
-0.0243880245834589,
0.10142315179109573,
-0.027699097990989685,
0.03955816477537155,
0.009458109736442566,
-0.003637150628492236,
-0.05890299379825592,
-0.06751279532909393,
-0.013111816719174385,
0.025081422179937363,
-0.0034556707832962275,
-0.04034677892923355,
-0.006205036770552397,
0.056656986474990845,
-0.020732654258608818,
0.011833111755549908,
-0.04811466485261917,
0.008982706815004349,
-0.021584633737802505,
0.01069724839180708,
-0.0959014743566513,
-0.0543396957218647,
0.02017226815223694,
-0.044582951813936234,
0.017052754759788513,
-0.022474778816103935,
0.07838130742311478,
0.02634287066757679,
0.019297225400805473,
0.006869629491120577,
-0.02758059836924076,
0.015936847776174545,
-0.015494543127715588,
-0.01292230561375618,
-0.019424377009272575,
0.05492396280169487,
0.06662862002849579,
0.0997905507683754,
-0.03223327919840813,
0.023587407544255257,
-0.03054438717663288,
-0.02572546899318695,
-0.0060503059066832066,
0.020772883668541908,
0.04849230870604515,
-0.07623094320297241,
0.07935582101345062,
-0.06960734724998474,
-0.010817646980285645,
-0.04587670415639877,
-0.013571011833846569,
-0.016301199793815613,
-0.08104845136404037,
-0.027366871014237404,
-0.018794504925608635,
-0.13918301463127136,
-0.09004519879817963,
0.0022770606447011232,
0.036760080605745316,
-0.0072602820582687855,
-0.008012309670448303,
-0.07732182741165161,
-4.872076211298935e-33,
0.0310770645737648,
0.026429027318954468,
0.05507764592766762,
-0.003982235211879015,
-0.053658705204725266,
0.04530727490782738,
-0.028372416272759438,
0.07077658176422119,
0.04816091060638428,
-0.10103808343410492,
0.02627500519156456,
0.038532767444849014,
0.06755799800157547,
0.03715338185429573,
0.06633627414703369,
-0.004110680427402258,
-0.07709342241287231,
0.011383523233234882,
-0.03142223879694939,
0.023433905094861984,
-0.06308469921350479,
0.0373554453253746,
-0.0033311329316347837,
0.038639917969703674,
-0.006765434052795172,
0.06219851225614548,
-0.021609213203191757,
0.09335668385028839,
-0.09023153781890869,
0.018909400328993797,
-0.13123451173305511,
-0.032512132078409195,
-0.07966652512550354,
-0.0049753389321267605,
-0.04273910075426102,
0.029583264142274857,
0.0328434593975544,
0.029204539954662323,
-0.041703421622514725,
0.08219198137521744,
0.023955971002578735,
0.05112793296575546,
-0.04343799501657486,
0.020349295809864998,
-0.013837588019669056,
-0.03179089352488518,
0.0989709123969078,
0.023292234167456627,
0.05301487445831299,
-0.04932577908039093,
-0.032856374979019165,
0.12080509215593338,
-0.06256131827831268,
0.07234184443950653,
-0.05623378977179527,
0.01110491156578064,
-0.024343717843294144,
-0.0560910589993,
-0.05134601518511772,
-0.035694677382707596,
-0.06334682554006577,
-0.020083267241716385,
0.044625334441661835,
0.03230703994631767,
-0.0231510978192091,
0.038682568818330765,
0.008386559784412384,
-0.05897262319922447,
0.013958141207695007,
0.052000757306814194,
0.014336436055600643,
-0.049374859780073166,
0.015586159192025661,
0.050739411264657974,
-0.02704581245779991,
-0.011474469676613808,
0.023431237787008286,
-0.08328524231910706,
-0.05092373862862587,
-0.005802998319268227,
0.0818907842040062,
0.031210333108901978,
-0.03132663667201996,
0.03720628842711449,
0.04771823063492775,
0.035214681178331375,
0.03311452269554138,
-0.016173575073480606,
0.05098341405391693,
-0.08161600679159164,
-0.004273069091141224,
0.01513227540999651,
-0.1006680354475975,
0.08685468137264252,
-0.054076455533504486,
-4.528592967290024e-8,
0.007978621870279312,
0.016624972224235535,
-0.02916230820119381,
-0.08488528430461884,
0.01018477138131857,
-0.021870652213692665,
-0.026073995977640152,
0.011418529786169529,
-0.0049616205506026745,
0.006227675825357437,
0.04644244536757469,
-0.0868832990527153,
-0.04606596753001213,
-0.08608660846948624,
0.03399496152997017,
-0.005793899297714233,
-0.05031650885939598,
0.048892658203840256,
-0.03728317841887474,
0.008255393244326115,
-0.01923981122672558,
-0.02249050699174404,
-0.07003778219223022,
0.026527471840381622,
0.03981686010956764,
-0.07779742032289505,
-0.01474292017519474,
-0.037623658776283264,
0.014282668009400368,
0.07004307210445404,
0.02338371053338051,
0.0370856411755085,
0.0025233402848243713,
0.07647419720888138,
-0.02893095090985298,
-0.03199486434459686,
0.029664460569620132,
-0.008033111691474915,
-0.08734475821256638,
0.004273170605301857,
-0.035422906279563904,
0.07643232494592667,
-0.09406749159097672,
-0.022224413231015205,
0.026419099420309067,
0.11518899351358414,
0.035023629665374756,
-0.04531668871641159,
0.04515276104211807,
0.07154282182455063,
0.03447825089097023,
0.017133809626102448,
-0.021755900233983994,
0.05221456661820412,
-0.01294007245451212,
-0.015458211302757263,
-0.02895480953156948,
-0.024997510015964508,
-0.004330538213253021,
0.038203272968530655,
0.0301740113645792,
0.06090409308671951,
-0.012344554997980595,
-0.02180047705769539
] | 0 | 0 | 0 | 0 |
|
A Differentiable Metric for Discovering Groups and Unitary Representations | https://openreview.net/forum?id=Tz8Li6G2xU | [
"Dongsung Huh"
] | Poster | Discovering group structures in data poses a fundamental challenge across diverse scientific domains. The primary obstacle lies in the non-differentiable nature of group axioms, impeding their integration into deep learning framework. To overcome this, we present a novel differentiable approach leveraging the representation theory of finite groups. Our method features a unique neural network architecture that models interactions between group elements as multiplications of their matrix representations, coupled with a regularizer that promotes unitarity of these matrices. Crucially, our model implicitly defines a complexity metric that favors the discovery of group structures. Evaluations demonstrate our method's ability to accurately recover group operations and learn their unitary representations from partial observation. Our work lays the foundation for a promising new paradigm in automated algebraic structure discovery, with far-reaching applications across diverse domains, particularly in enabling automatic symmetry discovery for geometric deep learning. | group theory, representation theory, representation learning, symmetry discovery, symbolic relationship | null | 13,021 | null | [
-0.11087661236524582,
-0.05310302972793579,
-0.036905236542224884,
0.048120781779289246,
-0.021265337243676186,
-0.05068720132112503,
-0.07721114158630371,
-0.18211834132671356,
0.037874024361371994,
-0.07644814252853394,
0.03394041210412979,
-0.06676558405160904,
-0.054792240262031555,
0.0935378223657608,
-0.08270975202322006,
0.005039118230342865,
-0.02455613948404789,
0.06501568108797073,
-0.12142422795295715,
-0.015931865200400352,
-0.04590058699250221,
0.002931988099589944,
-0.007241919171065092,
0.02897385135293007,
0.03989435359835625,
0.030890071764588356,
0.04528971388936043,
-0.05524096265435219,
0.04539814963936806,
-0.01610271818935871,
-0.008496200665831566,
0.012191594578325748,
0.023149840533733368,
-0.06171255558729172,
-0.01531689241528511,
0.14019954204559326,
0.006339190062135458,
0.025831881910562515,
0.0053445990197360516,
-0.049406811594963074,
0.04231317713856697,
0.037531767040491104,
0.057327523827552795,
0.10936945676803589,
0.09411625564098358,
0.07415735721588135,
0.02973216399550438,
-0.06945937871932983,
0.03206963092088699,
0.032783348113298416,
0.004162475932389498,
0.015340999700129032,
-0.017743704840540886,
0.044942352920770645,
0.0208013653755188,
-0.00405606534332037,
0.025142481550574303,
-0.03065505251288414,
-0.08013211935758591,
-0.09357964247465134,
0.05557963624596596,
-0.10445059835910797,
0.016716448590159416,
-0.009400641545653343,
-0.05501116067171097,
0.0762498676776886,
0.024263111874461174,
0.049749650061130524,
-0.02623581700026989,
-0.014678344130516052,
0.03597160801291466,
0.04726402834057808,
-0.04282129928469658,
0.08351602405309677,
0.08975173532962799,
0.051599808037281036,
-0.00902409665286541,
-0.04846055060625076,
0.01583406701683998,
-0.07942894846200943,
0.04383073374629021,
0.04702117666602135,
0.04301935061812401,
0.05015241727232933,
0.10886626690626144,
0.010723327286541462,
0.022436922416090965,
0.05406157672405243,
-0.014255573973059654,
-0.034433361142873764,
-0.07074998319149017,
0.025163013488054276,
-0.019765522330999374,
-0.08207649737596512,
0.010023480281233788,
0.06314876675605774,
0.05553754046559334,
-0.01743105612695217,
-0.03024013340473175,
0.12409649789333344,
0.017097488045692444,
0.03512059897184372,
0.02243039943277836,
-0.025637108832597733,
-0.06651502102613449,
0.021187281236052513,
0.041950710117816925,
0.006117012817412615,
0.04013341665267944,
-0.09279860556125641,
0.0563930943608284,
0.023456569761037827,
-0.0174428578466177,
0.030926618725061417,
0.00025552583974786103,
-0.09582147002220154,
0.04033353179693222,
-0.026898734271526337,
0.03824743255972862,
0.0002827484568115324,
0.009471864439547062,
0.03158653527498245,
-0.0007500064093619585,
0.12541140615940094,
0.06922105699777603,
0.09953548014163971,
-0.13666756451129913,
1.1683005320251215e-34,
0.015520631335675716,
0.06149977073073387,
0.03459911793470383,
0.01927793212234974,
0.03304249793291092,
-0.037401147186756134,
-0.028641778975725174,
0.039649803191423416,
0.00007016928429948166,
0.06474489718675613,
-0.06987430900335312,
0.017311828210949898,
0.007436928804963827,
0.02507610060274601,
-0.0037419521249830723,
0.013600410893559456,
-0.016661886125802994,
-0.013806621544063091,
-0.035737521946430206,
-0.11507230997085571,
-0.047905322164297104,
0.01133621484041214,
-0.09636519849300385,
-0.04931372031569481,
-0.04552585259079933,
-0.0065424214117228985,
0.029272885993123055,
-0.019667118787765503,
-0.04180561378598213,
-0.014320959337055683,
-0.015641577541828156,
0.006399552337825298,
-0.03185713291168213,
0.1010027751326561,
0.019854044541716576,
0.004634513054043055,
0.04089108109474182,
-0.04349102824926376,
0.0789531022310257,
-0.02398468367755413,
0.0762980505824089,
0.05006688833236694,
-0.018675832077860832,
-0.02214757911860943,
-0.020912256091833115,
0.028916889801621437,
-0.008974462747573853,
0.0072950329631567,
-0.07325856387615204,
-0.0868052989244461,
0.01335868239402771,
0.0053849476389586926,
-0.05020281299948692,
-0.020762376487255096,
-0.017280619591474533,
0.013551577925682068,
-0.029269475489854813,
0.04380128160119057,
0.044614385813474655,
0.028929157182574272,
-0.005957322660833597,
0.09225285798311234,
-0.039504215121269226,
0.07587955892086029,
-0.10874266177415848,
0.07903285324573517,
-0.01841898262500763,
0.07786791771650314,
0.06542002409696579,
0.020370515063405037,
-0.014864861033856869,
0.028060680255293846,
-0.05137912556529045,
-0.0162826981395483,
0.059446096420288086,
0.004869407974183559,
0.0659632459282875,
-0.12141910195350647,
-0.06874773651361465,
0.05828853324055672,
0.013010682538151741,
0.030381204560399055,
-0.004138912074267864,
-0.030495988205075264,
-0.10830293595790863,
-0.030833324417471886,
0.08243925869464874,
-0.10456334799528122,
-0.033508915454149246,
-0.014809492975473404,
-0.0756935328245163,
-0.0010122888488695025,
-0.0021272115409374237,
0.03872380405664444,
0.00255070673301816,
-1.989302314208004e-33,
-0.015303255058825016,
-0.01197679154574871,
0.0744718685746193,
-0.054570622742176056,
0.03852849453687668,
-0.08028551936149597,
-0.07649367302656174,
-0.014105047099292278,
0.01360483281314373,
-0.03418394550681114,
0.05190090462565422,
-0.016274679452180862,
-0.024641213938593864,
-0.046624112874269485,
0.050825975835323334,
0.014161009341478348,
-0.01519998162984848,
-0.043097421526908875,
0.039453841745853424,
0.009679642505943775,
-0.004053568933159113,
0.0594293512403965,
-0.03760438412427902,
0.04435781389474869,
-0.04843907058238983,
0.01817229762673378,
-0.010766740888357162,
-0.017139768227934837,
0.07740379124879837,
0.08089771866798401,
-0.05639391392469406,
-0.10941635072231293,
-0.06693156808614731,
-0.0023769373074173927,
-0.08244702965021133,
0.08590929955244064,
-0.004728279542177916,
0.039919283241033554,
-0.02142876759171486,
-0.029881220310926437,
-0.03852887824177742,
-0.024052176624536514,
-0.04063394293189049,
0.05702579766511917,
0.022359319031238556,
-0.013147592544555664,
0.08192244917154312,
0.003043557982891798,
-0.09729282557964325,
0.04988455772399902,
0.0277591273188591,
-0.04173707589507103,
-0.020391102880239487,
-0.03946850448846817,
-0.041745640337467194,
0.025000231340527534,
0.03652612119913101,
-0.025562433525919914,
0.037964627146720886,
0.029839379712939262,
-0.05133035406470299,
-0.09080397337675095,
-0.011200622655451298,
0.04951248690485954,
-0.060158345848321915,
-0.03697914630174637,
-0.02705046907067299,
0.0031884796917438507,
0.03928639739751816,
0.1169804111123085,
-0.0059483228251338005,
-0.06644987314939499,
-0.03313884511590004,
0.03807158023118973,
0.0014412460150197148,
-0.03519745171070099,
0.007267765235155821,
-0.029104864224791527,
-0.025590455159544945,
-0.009984397329390049,
0.08376876264810562,
-0.021837176755070686,
0.022735461592674255,
0.06775213778018951,
0.07105448096990585,
0.022486547008156776,
0.07368500530719757,
0.09407014399766922,
0.04438960552215576,
0.005399125628173351,
-0.014361471869051456,
0.0004179771931376308,
-0.041828498244285583,
0.05688329413533211,
-0.006386248394846916,
-3.713962115625691e-8,
-0.10809717327356339,
0.038888655602931976,
0.08173461258411407,
-0.08332458883523941,
0.029965663328766823,
-0.02612757869064808,
-0.01888337731361389,
0.08759313821792603,
-0.04188069328665733,
0.011933334171772003,
0.009162431582808495,
-0.01344135869294405,
-0.050764817744493484,
-0.046077482402324677,
-0.041745301336050034,
0.11454594135284424,
0.02749425359070301,
-0.016579899936914444,
0.005954635329544544,
0.00933306198567152,
-0.014094348065555096,
0.021964801475405693,
-0.006649637594819069,
-0.015118995681405067,
-0.005248924717307091,
-0.1142512708902359,
-0.07787267863750458,
-0.009556923992931843,
0.020627222955226898,
-0.04239365831017494,
0.011056371964514256,
0.02359740436077118,
0.0104753989726305,
0.053442347794771194,
-0.008476673625409603,
0.0703105479478836,
-0.04236684367060661,
-0.02790152095258236,
0.020600132644176483,
-0.0696982890367508,
-0.02278267592191696,
0.06063832342624664,
0.009777117520570755,
0.030302274972200394,
-0.006234270986169577,
0.01183249056339264,
0.02509378269314766,
-0.031584177166223526,
-0.020702896639704704,
0.028527388349175453,
0.10010218620300293,
-0.0467584952712059,
-0.0005126955802552402,
0.006199385970830917,
-0.03290905803442001,
0.021894438192248344,
-0.03058619238436222,
-0.021230265498161316,
0.05844567343592644,
0.008870227262377739,
-0.03630591928958893,
-0.008607259020209312,
-0.05475272983312607,
-0.009690910577774048
] | 0 | 0 | 0 | 0 |
|
Image Watermarks are Removable using Controllable Regeneration from Clean Noise | https://openreview.net/forum?id=mDKxlfraAn | [
"Yepeng Liu",
"Yiren Song",
"Hai Ci",
"Yu Zhang",
"Haofan Wang",
"Mike Zheng Shou",
"Yuheng Bu"
] | Poster | Image watermark techniques provide an effective way to assert ownership, deter misuse, and trace content sources, which has become increasingly essential in the era of large generative models. A critical attribute of watermark techniques is their robustness against various manipulations. In this paper, we introduce a watermark removal approach capable of effectively nullifying state-of-the-art watermarking techniques. Our primary insight involves regenerating the watermarked image starting from a \textbf{clean Gaussian noise} via a controllable diffusion model, utilizing the extracted semantic and spatial features from the watermarked image. The semantic control adapter and the spatial control network are specifically trained to control the denoising process towards ensuring image quality and enhancing consistency between the cleaned image and the original watermarked image. To achieve a smooth trade-off between watermark removal performance and image consistency, we further propose an adjustable and controllable regeneration scheme. This scheme adds varying numbers of noise steps to the latent representation of the watermarked image, followed by a controlled denoising process starting from this noisy latent representation. As the number of noise steps increases, the latent representation progressively approaches clean Gaussian noise, facilitating the desired trade-off. We apply our watermark removal methods across various watermarking techniques, and the results demonstrate that our methods offer superior visual consistency/quality and enhanced watermark removal performance compared to existing regeneration approaches. Our code is available at \url{https://github.com/yepengliu/CtrlRegen}. | Watermark, Detection, Robustness, Diffusion Model | We propose a controllable regeneration method for effective image watermark removal. We aim for our method to serve as a benchmark to evaluate and enhance the robustness of future watermarking techniques. | 13,019 | 2410.05470 | [
-0.0196104533970356,
-0.05684937536716461,
0.06764573603868484,
0.043400853872299194,
0.07550805807113647,
-0.029657971113920212,
-0.000649395864456892,
-0.14844290912151337,
-0.021377919241786003,
-0.02777468040585518,
0.023721856996417046,
0.05006471648812294,
0.09353778511285782,
0.0542420968413353,
-0.04723799601197243,
0.032507702708244324,
-0.013479466550052166,
0.060736674815416336,
-0.07766652852296829,
-0.06739293783903122,
-0.007233587559312582,
-0.05460759624838829,
-0.004308576695621014,
-0.042707934975624084,
0.09690859913825989,
0.06373977661132812,
0.06212243437767029,
-0.018406549468636513,
0.01775771751999855,
-0.033199381083250046,
0.09992188960313797,
-0.09141374379396439,
0.024522390216588974,
-0.02338731661438942,
-0.014510978944599628,
0.01983088068664074,
0.034274522215127945,
0.02239837311208248,
0.0988040566444397,
-0.00023735278227832168,
0.053892072290182114,
-0.010367280803620815,
-0.09304019808769226,
-0.0006100114551372826,
0.00953733455389738,
0.04847656190395355,
0.0267063919454813,
0.019642097875475883,
-0.01563645526766777,
0.016356518492102623,
0.06168022006750107,
0.029137300327420235,
-0.11172766238451004,
0.03308402746915817,
-0.0041455174796283245,
-0.039358075708150864,
0.06342365592718124,
0.01512205507606268,
-0.05759243667125702,
-0.0030794444028288126,
0.0027186647057533264,
-0.0075421021319925785,
0.04688876494765282,
-0.0011362575460225344,
0.058501534163951874,
-0.01616666465997696,
-0.039055027067661285,
-0.0028498717583715916,
0.03373534232378006,
0.00993482768535614,
-0.059145987033843994,
0.06062017008662224,
0.013187413103878498,
-0.009851541370153427,
-0.07825302332639694,
0.0668630450963974,
-0.050217416137456894,
-0.030709993094205856,
0.02337794005870819,
-0.0200858972966671,
0.12309610843658447,
-0.02394767850637436,
0.0719052329659462,
0.018334057182073593,
0.0900803953409195,
-0.048636600375175476,
0.016394613310694695,
-0.024561550468206406,
0.04993562772870064,
-0.005178564228117466,
-0.049190666526556015,
0.05398370325565338,
0.04328925907611847,
0.010945201851427555,
-0.0005764999077655375,
0.01661771349608898,
-0.036603134125471115,
0.026190418750047684,
-0.01987217552959919,
0.0438448004424572,
-0.010523798875510693,
-0.0667717307806015,
-0.024702291935682297,
-0.030895687639713287,
0.12809835374355316,
0.008663344196975231,
0.04557771235704422,
0.03997771814465523,
0.05675460770726204,
-0.024940291419625282,
-0.024385565891861916,
0.03674750402569771,
-0.07223986834287643,
0.025298282504081726,
-0.01629856415092945,
-0.003171413205564022,
-0.0305445846170187,
-0.004141292069107294,
-0.01823393628001213,
-0.05876763164997101,
0.025354083627462387,
-0.05891500040888786,
-0.06394806504249573,
-0.029420355334877968,
-0.020529855042696,
0.056839603930711746,
-0.032145194709300995,
2.4366071183836625e-33,
-0.021094409748911858,
-0.009184131398797035,
-0.03899488225579262,
-0.0607745386660099,
0.04589773714542389,
0.02600858546793461,
-0.05218188092112541,
-0.06629921495914459,
-0.024371594190597534,
-0.09388469904661179,
0.06553803384304047,
-0.09383843094110489,
-0.08581430464982986,
0.08559175580739975,
0.04515199735760689,
-0.014718250371515751,
-0.029477039352059364,
-0.018396619707345963,
0.026483353227376938,
-0.026910919696092606,
0.03705516830086708,
0.032235126942396164,
-0.08457436412572861,
-0.010424906387925148,
-0.012796527706086636,
-0.04858744889497757,
0.08468632400035858,
-0.016115684062242508,
0.04931525141000748,
-0.021017735823988914,
-0.1107257679104805,
-0.003918967675417662,
0.06962694972753525,
0.06353036314249039,
-0.018098000437021255,
-0.020614294335246086,
0.07117750495672226,
-0.0476464107632637,
0.02686277963221073,
0.010014345869421959,
-0.01874529756605625,
0.01621263101696968,
-0.09056214988231659,
-0.0696752592921257,
-0.07092476636171341,
0.026102829724550247,
-0.01138770580291748,
0.04177455976605415,
-0.022863030433654785,
-0.05377272143959999,
0.042829737067222595,
0.06115483492612839,
0.017149820923805237,
-0.019664887338876724,
-0.06871815770864487,
-0.06899431347846985,
0.026549750939011574,
-0.1061498373746872,
0.02609872817993164,
-0.020776327699422836,
0.05818810686469078,
0.03584615886211395,
0.056372418999671936,
0.07133417576551437,
0.07239292562007904,
-0.014963887631893158,
0.11040419340133667,
0.044607002288103104,
-0.02714882604777813,
-0.026428749784827232,
-0.03881564736366272,
0.04853472858667374,
0.001134674996137619,
-0.004180103074759245,
-0.06170213222503662,
-0.047354698181152344,
0.03726467117667198,
0.0036953375674784184,
-0.033853109925985336,
0.016001665964722633,
-0.07555359601974487,
-0.07294031232595444,
0.03216957300901413,
-0.05056580528616905,
-0.03332925960421562,
-0.10492061823606491,
-0.004931719973683357,
-0.03690154477953911,
-0.03883291408419609,
-0.06249523162841797,
0.027235954999923706,
-0.023627394810318947,
0.015501085668802261,
0.007255679927766323,
-0.03437100723385811,
-2.2564093445323753e-33,
0.012964103370904922,
0.019001957029104233,
0.013775612227618694,
0.0662311464548111,
-0.00800381787121296,
-0.027555154636502266,
0.054986950010061264,
0.0807328075170517,
0.061088696122169495,
-0.009985765442252159,
-0.10048535466194153,
-0.0011870809830725193,
-0.051135625690221786,
-0.03335347771644592,
-0.0229880902916193,
-0.04887733981013298,
-0.014740773476660252,
0.012983584776520729,
-0.08532872051000595,
-0.002617613645270467,
0.06678666919469833,
0.05646991357207298,
0.013243013061583042,
0.11366502195596695,
-0.046368829905986786,
0.0663340762257576,
0.02632908895611763,
0.02877158112823963,
-0.02837391197681427,
0.0001920777140185237,
-0.08837003260850906,
0.07219071686267853,
0.07289942353963852,
-0.0841485932469368,
-0.019936202093958855,
0.009523714892566204,
0.041709937155246735,
-0.048243019729852676,
-0.038934506475925446,
0.07037960737943649,
-0.02374088764190674,
-0.02233440801501274,
-0.07061333954334259,
0.03648419305682182,
0.027319103479385376,
-0.02266661450266838,
-0.0017903619445860386,
0.03461568057537079,
0.03801565617322922,
0.06833606958389282,
0.020196735858917236,
-0.02284863032400608,
-0.09322275221347809,
0.06784498691558838,
-0.03081819787621498,
-0.030308686196804047,
-0.06262156367301941,
0.012378917075693607,
0.06517533957958221,
0.016408389434218407,
-0.0004938620259054005,
-0.019853373989462852,
-0.05209757760167122,
-0.026298433542251587,
-0.012957200407981873,
0.012272721156477928,
0.010000207461416721,
0.07563579827547073,
-0.05961260572075844,
0.059748318046331406,
0.024073388427495956,
-0.010865211486816406,
0.006351884454488754,
0.07483159005641937,
0.0874720886349678,
-0.04601454362273216,
-0.017757659777998924,
0.03137010335922241,
0.004554260987788439,
-0.03593974560499191,
0.038150835782289505,
-0.045355796813964844,
0.03987585008144379,
0.033029209822416306,
0.0963168665766716,
0.033890895545482635,
-0.09601472318172455,
-0.10457530617713928,
-0.014655056409537792,
0.01557752676308155,
-0.015727117657661438,
0.0032588031608611345,
-0.009368770755827427,
0.04652068391442299,
-0.04488593339920044,
-4.3076713751588613e-8,
-0.07066656649112701,
-0.01326029747724533,
-0.0050843278877437115,
-0.05848727747797966,
-0.01994379051029682,
-0.057456549257040024,
0.04970209673047066,
0.09294065088033676,
0.0864635556936264,
-0.1678730696439743,
0.05224316939711571,
-0.058177389204502106,
-0.13808950781822205,
-0.10432452708482742,
-0.05290228873491287,
-0.0007023222278803587,
0.005703350994735956,
0.05685970559716225,
-0.011814102530479431,
-0.030475160107016563,
-0.07425422221422195,
0.02454489842057228,
-0.040272925049066544,
-0.028104357421398163,
0.03658999502658844,
0.05239333212375641,
0.06842627376317978,
0.08024357259273529,
0.07119131088256836,
0.038141824305057526,
0.08186771720647812,
0.01695270463824272,
0.08152502030134201,
0.03873573616147041,
-0.005511781666427851,
0.04706257954239845,
0.03403756767511368,
0.007310009561479092,
-0.06646022200584412,
-0.0059162890538573265,
-0.034499965608119965,
0.04129842296242714,
0.004186531528830528,
-0.0745147168636322,
0.012824509292840958,
0.03239794448018074,
0.04709995537996292,
-0.04097943380475044,
0.07909978926181793,
0.03626885265111923,
0.028119761496782303,
0.03106706775724888,
-0.08094289898872375,
0.045847419649362564,
0.041475810110569,
-0.0875849723815918,
0.047951340675354004,
0.06977446377277374,
0.05226338654756546,
0.02233327366411686,
0.04853448271751404,
0.016875192523002625,
0.09539911150932312,
0.02422434836626053
] | https://github.com/yepengliu/ctrlregen | 11 | 1 | 0 | 0 |
Point Cluster: A Compact Message Unit for Communication-Efficient Collaborative Perception | https://openreview.net/forum?id=54XlM8Clkg | [
"Zihan Ding",
"Jiahui Fu",
"Si Liu",
"Hongyu Li",
"Siheng Chen",
"Hongsheng Li",
"Shifeng Zhang",
"Xu Zhou"
] | Poster | The objective of the collaborative perception task is to enhance the individual agent's perception capability through message communication among neighboring agents. A central challenge lies in optimizing the inherent trade-off between perception ability and communication cost. To tackle this bottleneck issue, we argue that a good message unit should encapsulate both semantic and structural information in a sparse format, a feature not present in prior approaches. In this paper, we innovatively propose a compact message unit, namely point cluster, whose core idea is to represent potential objects efficiently with explicitly decoupled low-level structure information and high-level semantic information. Building upon this new message unit, we propose a comprehensive framework CPPC for communication-efficient collaborative perception. The core principle of CPPC is twofold: first, through strategical point sampling, structure information can be well preserved with a few key points, which can significantly reduce communication cost; second, the sequence format of point clusters enables efficient message aggregation by set matching and merging, thereby eliminating unnecessary computation generated when aligning squared BEV maps, especially for long-range collaboration. To handle time latency and pose errors encountered in real-world scenarios, we also carefully design parameter-free solutions that can adapt to different noisy levels without finetuning. Experiments on two widely recognized collaborative perception benchmarks showcase the superior performance of our method compared to the previous state-of-the-art approaches. | Point Cluster, Collaborative Perception | null | 13,011 | null | [
0.05818559601902962,
-0.07295984029769897,
-0.012945164926350117,
-0.0072496673092246056,
0.05182916298508644,
0.014337308704853058,
0.029206130653619766,
-0.042033400386571884,
0.05734458565711975,
-0.03976568579673767,
-0.016394877806305885,
-0.052643779665231705,
0.035028111189603806,
0.097780242562294,
0.07571976631879807,
0.014303796924650669,
0.20148420333862305,
0.04321938008069992,
0.012263746932148933,
0.006525784730911255,
-0.03286042809486389,
0.01878860965371132,
0.006987269036471844,
0.012516188435256481,
-0.00038855199818499386,
0.02712375856935978,
0.01786055602133274,
-0.0661853700876236,
-0.03092842921614647,
-0.005718977190554142,
0.021799633279442787,
-0.0005051078042015433,
0.04321399703621864,
0.021785590797662735,
-0.019361363723874092,
0.11233092099428177,
0.033964820206165314,
-0.016186630353331566,
-0.028411120176315308,
-0.031714461743831635,
0.014468546025454998,
0.03620580956339836,
-0.03746715188026428,
0.09117577224969864,
0.04682587832212448,
0.08308832347393036,
0.015600122511386871,
0.06003294140100479,
0.010963317938148975,
-0.0027106076013296843,
-0.023242758587002754,
-0.001391314435750246,
-0.03210997208952904,
0.041217196732759476,
0.011775689199566841,
0.035519156605005264,
-0.11953514814376831,
-0.019927609711885452,
0.014691264368593693,
-0.0264505036175251,
0.011998241767287254,
-0.05989277735352516,
0.10750449448823929,
-0.011083603836596012,
-0.044581931084394455,
-0.04057058319449425,
0.12753410637378693,
0.025275861844420433,
0.019472798332571983,
-0.012159996666014194,
-0.0012162845814600587,
0.08130980283021927,
0.03540584817528725,
0.014811202883720398,
-0.005364635027945042,
0.05907181650400162,
-0.02708049677312374,
-0.1074061468243599,
-0.06614178419113159,
-0.054225724190473557,
-0.013574972748756409,
0.03934996575117111,
-0.025602884590625763,
0.039272014051675797,
0.07403035461902618,
-0.05160057172179222,
-0.01879257708787918,
0.013390759006142616,
-0.09648299217224121,
0.007427005097270012,
-0.08707644045352936,
-0.047700222581624985,
-0.005120193585753441,
-0.03558645397424698,
-0.0023483014665544033,
0.003050838829949498,
-0.015772556886076927,
-0.05983608216047287,
0.009839690290391445,
0.0912160575389862,
-0.01783013716340065,
0.038215797394514084,
0.03788807615637779,
-0.062966488301754,
-0.02535509504377842,
-0.01786632090806961,
-0.009496660903096199,
0.043926529586315155,
-0.018244512379169464,
-0.010429230518639088,
-0.023020578548312187,
-0.05433041602373123,
-0.05201100930571556,
-0.0033427239395678043,
0.02581467665731907,
-0.024677081033587456,
0.027279581874608994,
-0.027257932350039482,
0.061292361468076706,
-0.025474321097135544,
0.028329063206911087,
-0.11850759387016296,
-0.0008276415755972266,
0.037937480956315994,
0.05114850029349327,
0.05242359638214111,
-0.017746316269040108,
-8.78412949235917e-35,
0.046385258436203,
0.10206472128629684,
0.058234333992004395,
0.009403248317539692,
0.08435099571943283,
0.03347461670637131,
-0.012973918579518795,
0.012245743535459042,
-0.08702993392944336,
-0.05573859438300133,
-0.004658942110836506,
-0.05892264097929001,
0.0466613695025444,
0.06860726326704025,
0.058719243854284286,
-0.06969712674617767,
-0.0063603343442082405,
0.10103407502174377,
-0.01627662405371666,
-0.051150571554899216,
-0.01591922901570797,
0.0344047024846077,
-0.05687671899795532,
-0.006535479333251715,
0.006859395653009415,
0.00828865822404623,
0.02622225694358349,
-0.04421272128820419,
0.017381388694047928,
-0.04054924100637436,
-0.06260357797145844,
0.08076474070549011,
0.04178380221128464,
0.017808139324188232,
0.028689812868833542,
-0.023183779790997505,
0.015235889703035355,
-0.12936365604400635,
-0.018457092344760895,
-0.0493377223610878,
0.017715737223625183,
0.07020371407270432,
-0.04230383783578873,
-0.12361900508403778,
-0.08832082152366638,
-0.053801484405994415,
-0.0018993365811184049,
0.07560072839260101,
-0.060118842869997025,
-0.06845841556787491,
0.05884544923901558,
0.001966560259461403,
-0.011680602096021175,
0.03614984452724457,
-0.019852031022310257,
-0.04907212406396866,
-0.003382297232747078,
-0.02129349112510681,
0.04681868851184845,
0.032450418919324875,
-0.011326005682349205,
0.007806530687958002,
0.010145238600671291,
0.02940957248210907,
0.022222673520445824,
-0.01864255592226982,
-0.0029098782688379288,
0.04127303883433342,
0.07496380805969238,
-0.03910451382398605,
0.020675379782915115,
0.05617059767246246,
-0.007375841028988361,
-0.040436483919620514,
-0.0025406405329704285,
0.0014898477820679545,
0.020155083388090134,
-0.06894957274198532,
0.01269975584000349,
0.09876266866922379,
-0.11893180012702942,
-0.07583054155111313,
-0.04692286252975464,
-0.03427642211318016,
-0.015083299949765205,
0.04510722681879997,
0.05424647033214569,
-0.034305114299058914,
-0.08678509294986725,
-0.0037659069057554007,
-0.0997210368514061,
-0.002758810529485345,
0.01979314349591732,
0.005803293082863092,
-0.0030475114472210407,
-6.374821577234424e-34,
0.02379201538860798,
0.013526104390621185,
0.022721363231539726,
-0.0020907949656248093,
0.011423160322010517,
-0.10066656023263931,
-0.01418718509376049,
-0.09652023017406464,
-0.04856511205434799,
0.02856813371181488,
-0.08875023573637009,
-0.05759135261178017,
0.0830070897936821,
-0.005115520674735308,
0.015921715646982193,
0.07810419052839279,
-0.01212405413389206,
-0.000738164409995079,
0.015912432223558426,
-0.013672317378222942,
0.006012002471834421,
0.05369079113006592,
-0.05826997384428978,
0.0348009392619133,
-0.015638813376426697,
0.050966836512088776,
0.0065832436084747314,
-0.05618590489029884,
0.030848048627376556,
-0.02832924760878086,
-0.046396009624004364,
-0.02144288830459118,
0.01073426567018032,
-0.11225946992635727,
-0.008752882480621338,
0.05750579014420509,
0.012704011052846909,
0.00036774695036001503,
-0.07868305593729019,
0.04628348723053932,
0.05887610465288162,
-0.07665994763374329,
-0.1145828515291214,
0.016323016956448555,
0.012098807841539383,
-0.052229445427656174,
-0.029569653794169426,
-0.015319635160267353,
-0.0916648581624031,
0.08127927035093307,
-0.024954449385404587,
0.025319354608654976,
-0.03381844609975815,
-0.0018609901890158653,
-0.05874870344996452,
-0.026029245927929878,
-0.0633479505777359,
0.09463050961494446,
0.07095180451869965,
-0.015921205282211304,
0.02380918152630329,
-0.04893718287348747,
-0.04643808305263519,
0.09176439046859741,
0.06908270716667175,
-0.011406462639570236,
-0.005490112118422985,
-0.014384008944034576,
0.006947741378098726,
0.015933703631162643,
-0.029690032824873924,
-0.017621006816625595,
-0.0327729731798172,
0.009096862748265266,
0.06839754432439804,
-0.05023359879851341,
0.01051933690905571,
-0.010162082500755787,
0.02162431739270687,
0.02578672580420971,
-0.006725557614117861,
0.0692225992679596,
-0.0042497157119214535,
0.09475631266832352,
0.04866313934326172,
0.036911703646183014,
0.008245808072388172,
0.09610681235790253,
-0.020659860223531723,
0.04557032510638237,
-0.0019308577757328749,
0.0049840896390378475,
0.05007566139101982,
0.0556538850069046,
-0.07850641757249832,
-4.3727393261860925e-8,
-0.20772802829742432,
0.023178797215223312,
0.02725539728999138,
-0.04094715043902397,
0.010203665122389793,
-0.05281407758593559,
-0.008956747129559517,
0.08037000894546509,
0.03843662142753601,
0.01794183999300003,
0.033903736621141434,
-0.07124627381563187,
-0.043887097388505936,
-0.03469828888773918,
0.02999979816377163,
0.08170245587825775,
-0.05953940749168396,
-0.08400320261716843,
-0.016129788011312485,
0.006852964870631695,
-0.017659682780504227,
0.05133666843175888,
-0.056499965488910675,
0.07637842744588852,
-0.014338923618197441,
-0.03312491998076439,
0.04562746360898018,
0.1204226016998291,
0.039910316467285156,
-0.05897986888885498,
-0.0698281079530716,
0.006851466838270426,
0.011355253867805004,
-0.004889807663857937,
0.02170700766146183,
0.012806291691958904,
-0.013378755189478397,
0.03043261356651783,
-0.015259601175785065,
0.003152107587084174,
0.026726368814706802,
0.02800324745476246,
-0.03046542964875698,
0.02499658800661564,
0.07324962317943573,
0.06281450390815735,
-0.08790681511163712,
-0.02628694474697113,
-0.05598635971546173,
-0.016735536977648735,
-0.03412162512540817,
0.05818016454577446,
-0.05033344402909279,
0.09840520471334457,
0.016388101503252983,
0.01596561260521412,
0.022051332518458366,
-0.017764026299118996,
0.13943937420845032,
0.018822167068719864,
0.00972340814769268,
0.06441844999790192,
-0.13005591928958893,
-0.020837314426898956
] | 0 | 0 | 0 | 0 |
|
A Theoretical Analysis of Self-Supervised Learning for Vision Transformers | https://openreview.net/forum?id=Antib6Uovh | [
"Yu Huang",
"Zixin Wen",
"Yuejie Chi",
"Yingbin Liang"
] | Poster | Self-supervised learning has become a cornerstone in computer vision, primarily divided into reconstruction-based methods like masked autoencoders (MAE) and discriminative methods such as contrastive learning (CL). Recent empirical observations reveal that MAE and CL capture different types of representations: CL tends to focus on global patterns, while MAE adeptly captures **both global and subtle local** information simultaneously. Despite a flurry of recent empirical investigations to shed light on this difference, theoretical understanding remains limited, especially on the dominant architecture **vision transformers** (ViTs). In this paper, to provide rigorous insights, we model the visual data distribution by considering two types of spatial features: dominant global features and comparatively minuscule local features, and study the impact of imbalance among these features. We analyze the training dynamics of one-layer softmax-based ViTs on both MAE and CL objectives using gradient descent. Our analysis shows that as the degree of feature imbalance varies, ViTs trained with the MAE objective effectively learn both global and local features to achieve near-optimal reconstruction, while the CL-trained ViTs favor predominantly global features, even under mild imbalance. These results provide a theoretical explanation for distinct behaviors of MAE and CL observed in empirical studies. | Theory of transformers, Convergence analysis, Nonconvex optimization, Theory of self-supervised learning | null | 13,009 | 2403.02233 | [
0.016309795901179314,
-0.10657665878534317,
0.04796971008181572,
-0.013082600198686123,
0.09853868931531906,
0.03527407720685005,
-0.006257213652133942,
-0.0025556820910423994,
0.038342997431755066,
-0.05742029845714569,
-0.06377460062503815,
-0.078933484852314,
-0.009785451926290989,
0.09390974044799805,
-0.11734774708747864,
-0.029533982276916504,
0.09552833437919617,
0.09926574677228928,
-0.11569646745920181,
-0.019725076854228973,
0.05794237554073334,
0.0251011960208416,
-0.006434848066419363,
0.04662672430276871,
0.08475944399833679,
-0.02199302613735199,
0.04490166902542114,
-0.0495402067899704,
0.03212443366646767,
-0.0514102503657341,
-0.025439942255616188,
-0.05061430484056473,
-0.016327401623129845,
0.06259208172559738,
-0.01889609731733799,
0.00962433498352766,
-0.03248092159628868,
0.0790807455778122,
-0.061518266797065735,
-0.026766981929540634,
-0.0005962246214039624,
0.04840829223394394,
0.020764824002981186,
-0.01838572882115841,
0.051175281405448914,
0.04868369549512863,
0.021282240748405457,
-0.03263584151864052,
-0.001960733672603965,
-0.00745545607060194,
0.006154809147119522,
-0.002361241029575467,
-0.06983763724565506,
0.03688326105475426,
-0.029425017535686493,
0.07492892444133759,
0.05149364471435547,
-0.01877867802977562,
0.03050239011645317,
0.05812208727002144,
0.02811359241604805,
-0.023082919418811798,
0.046810880303382874,
0.008589312434196472,
-0.028444984927773476,
-0.01694895513355732,
0.038864538073539734,
0.009956859983503819,
-0.040108609944581985,
-0.0439763218164444,
-0.0045644198544323444,
0.016192203387618065,
-0.05270623043179512,
0.0214245542883873,
0.08343473821878433,
0.019535761326551437,
0.05857490748167038,
-0.036943286657333374,
0.09717775881290436,
-0.056838568300008774,
0.04108856990933418,
0.054458294063806534,
0.02101181074976921,
0.02141689509153366,
0.11305153369903564,
-0.015405366197228432,
-0.06554991751909256,
0.05973100662231445,
-0.0026730576064437628,
0.03589228168129921,
-0.016531862318515778,
-0.06448756158351898,
-0.0344146303832531,
-0.01732051745057106,
0.047815773636102676,
-0.08195938169956207,
0.025414995849132538,
-0.0896906703710556,
0.03575760871171951,
0.046513572335243225,
-0.0613134428858757,
-0.031423356384038925,
0.054154470562934875,
0.0028850638773292303,
0.06680046766996384,
-0.053351305425167084,
0.0761025920510292,
0.03972718119621277,
0.05550958961248398,
-0.08813125640153885,
0.06231137737631798,
0.01538881566375494,
-0.02265825681388378,
-0.04701702296733856,
0.020974930375814438,
0.005750240292400122,
0.0011981180869042873,
0.031107591465115547,
-0.022567352280020714,
0.059646446257829666,
-0.06349192559719086,
-0.053313806653022766,
0.0066189891658723354,
0.05621993541717529,
0.06820575147867203,
-0.020300861448049545,
-0.04933750629425049,
1.1149079944484644e-33,
-0.0034224262926727533,
0.0013002350460737944,
0.004578087013214827,
-0.012877092696726322,
-0.014003941789269447,
-0.04415755346417427,
0.014483108185231686,
0.022981557995080948,
0.012522542849183083,
0.04096957668662071,
-0.0242927223443985,
0.0348074696958065,
-0.0368662066757679,
0.14992350339889526,
0.018911052495241165,
0.007778057362884283,
0.0012561783660203218,
-0.0597294457256794,
-0.04136801138520241,
-0.03779779374599457,
-0.004860700108110905,
0.05902764946222305,
0.020708169788122177,
-0.07398631423711777,
-0.05318758264183998,
-0.054297029972076416,
0.011659035459160805,
-0.005633072927594185,
-0.027239318937063217,
0.021301688626408577,
-0.09140485525131226,
0.08888489753007889,
-0.021990500390529633,
-0.00916224904358387,
0.06101524457335472,
0.02904769778251648,
0.017258420586586,
-0.012668093666434288,
0.021730812266469002,
-0.00161588448099792,
-0.023977788165211678,
0.08275774121284485,
-0.07923612743616104,
-0.08379067480564117,
0.023158056661486626,
-0.04064542055130005,
0.0703611895442009,
0.048958320170640945,
-0.11573880910873413,
-0.037232909351587296,
0.020073266699910164,
0.032838623970746994,
-0.09695335477590561,
-0.09361797571182251,
0.0007614727946929634,
0.045377325266599655,
0.02019594796001911,
0.037429265677928925,
0.02627319097518921,
-0.0011981650022789836,
0.021264148876070976,
-0.009957081638276577,
0.019291946664452553,
0.07999356836080551,
0.024545200169086456,
-0.016114501282572746,
0.09710205346345901,
0.01284915953874588,
-0.007963533513247967,
-0.019502826035022736,
0.0157176461070776,
0.05355548486113548,
-0.07299897074699402,
-0.11866853386163712,
0.04382513463497162,
0.011748749762773514,
0.11289935559034348,
0.008442576974630356,
-0.0465768538415432,
0.030257362872362137,
-0.030550159513950348,
0.014768116176128387,
0.022765493020415306,
-0.06712302565574646,
-0.05751330405473709,
0.07304887473583221,
0.05991935357451439,
-0.06702324748039246,
-0.015041911043226719,
-0.0025686833541840315,
-0.04078097268939018,
-0.08790218085050583,
0.010592270642518997,
-0.040910277515649796,
-0.031791772693395615,
-2.491228116433551e-33,
-0.04467124491930008,
0.055998172610998154,
-0.048206545412540436,
-0.009705903008580208,
-0.045783087611198425,
-0.07362394779920578,
-0.0008622587192803621,
0.07502321898937225,
-0.06082926318049431,
-0.063950315117836,
-0.018750041723251343,
-0.02917708083987236,
-0.02587251551449299,
0.012940777465701103,
-0.010146652348339558,
0.03130052983760834,
-0.01202397234737873,
0.03466782718896866,
-0.051248155534267426,
-0.023692673072218895,
-0.024842217564582825,
0.08549163490533829,
-0.00043439114233478904,
0.060517434030771255,
-0.08569865673780441,
0.061163514852523804,
-0.08638506382703781,
0.07811819016933441,
-0.0024821339175105095,
-0.019517341628670692,
0.03383153676986694,
0.03601646050810814,
-0.023173801600933075,
-0.02977360039949417,
-0.0958106592297554,
0.05181884765625,
-0.048600614070892334,
-0.06873350590467453,
-0.04866255447268486,
0.12038476765155792,
-0.004559761378914118,
-0.05221095681190491,
-0.019853772595524788,
-0.02076863870024681,
-0.0656777173280716,
-0.1018308699131012,
-0.008294257335364819,
-0.020625783130526543,
0.06867363303899765,
0.020887285470962524,
-0.05403899773955345,
0.03667523339390755,
-0.04676741734147072,
0.04695790633559227,
-0.09387586265802383,
-0.03758476674556732,
0.01904059201478958,
0.02983861044049263,
0.11297876387834549,
0.04911300912499428,
-0.02849995531141758,
-0.04285073280334473,
-0.03928850218653679,
-0.027187833562493324,
-0.05847007408738136,
-0.04636047035455704,
-0.008718150667846203,
-0.03123287670314312,
0.05558000132441521,
0.043178681284189224,
0.04846767336130142,
0.022769922390580177,
0.021398654207587242,
-0.03012349084019661,
-0.012828433886170387,
-0.02179693803191185,
0.033270757645368576,
-0.013740887865424156,
-0.02301136590540409,
-0.016554400324821472,
0.04524251073598862,
-0.061497848480939865,
0.043011751025915146,
0.08481992036104202,
0.048953697085380554,
0.043217699974775314,
0.030446043238043785,
-0.044996608048677444,
0.061014026403427124,
-0.007345977239310741,
-0.04724445194005966,
0.006621796637773514,
-0.11358637362718582,
-0.051828350871801376,
-0.07738428562879562,
-4.778138062988546e-8,
-0.0847383588552475,
0.03555649518966675,
0.007316200528293848,
-0.02376239188015461,
-0.002196528483182192,
-0.02193608507514,
-0.011814621277153492,
0.1268034279346466,
-0.09867831319570541,
0.03718280419707298,
0.043517667800188065,
-0.04602765664458275,
-0.02463732473552227,
-0.05485993251204491,
0.034363627433776855,
0.07125931978225708,
0.04516781121492386,
0.016888508573174477,
0.02651311829686165,
0.06452387571334839,
-0.029441364109516144,
0.012169196270406246,
0.02201143279671669,
-0.07375417649745941,
0.003540217410773039,
-0.0624241828918457,
-0.05598508194088936,
0.08444178849458694,
0.04756123200058937,
0.03715012967586517,
0.03524074703454971,
0.0814136490225792,
0.056556761264801025,
-0.06026501953601837,
-0.014680279418826103,
0.07381496578454971,
0.0578417144715786,
-0.01799829863011837,
-0.07980424165725708,
-0.007548829074949026,
-0.014451626688241959,
0.030324462801218033,
-0.009993440471589565,
0.028155764564871788,
-0.009389663115143776,
0.044950637966394424,
0.07415798306465149,
-0.09192078560590744,
-0.01886346936225891,
0.00572599284350872,
0.024043094366788864,
-0.03986340016126633,
0.002591158030554652,
0.14693215489387512,
0.08634316921234131,
-0.06291293352842331,
0.04028140380978584,
-0.08006834983825684,
0.01910417340695858,
0.09244103729724884,
-0.0205061137676239,
0.0424792543053627,
-0.06738916784524918,
-0.027970870956778526
] | 0 | 0 | 0 | 0 |
|
ProteinBench: A Holistic Evaluation of Protein Foundation Models | https://openreview.net/forum?id=BksqWM8737 | [
"Fei YE",
"Zaixiang Zheng",
"Dongyu Xue",
"Yuning Shen",
"Lihao Wang",
"Yiming Ma",
"Yan Wang",
"Xinyou Wang",
"Xiangxin Zhou",
"Quanquan Gu"
] | Poster | Recent years have witnessed a surge in the development of protein foundation models, significantly improving performance in protein prediction and generative tasks ranging from 3D structure prediction and protein design to conformational dynamics. However, the capabilities and limitations associated with these models remain poorly understood due to the absence of a unified evaluation framework. To fill this gap, we introduce ProteinBench, a holistic evaluation framework designed to enhance the transparency of protein foundation models. Our approach consists of three key components: (i) A taxonomic classification of tasks that broadly encompass the main challenges in the protein domain, based on the relationships between different protein modalities; (ii) A multi-metric evaluation approach that assesses performance across four key dimensions: quality, novelty, diversity, and robustness; and (iii) In-depth analyses from various user objectives, providing a holistic view of model performance. Our comprehensive evaluation of protein foundation models reveals several key findings that shed light on their current capabilities and limitations. To promote transparency and facilitate further research, we release the evaluation dataset, code, and a public leaderboard publicly for further analysis and a general modular toolkit. We intend for ProteinBench to be a living benchmark for establishing a standardized, in-depth evaluation framework for protein foundation models, driving their development and application while fostering collaboration within the field. | Protein foundation model, benchmark, protein design, protein conformation prediction | This work provides a holistic evaluation of protein foundation models. | 13,006 | 2409.06744 | [
-0.08387114852666855,
-0.15394090116024017,
-0.045475900173187256,
-0.018332697451114655,
0.09521092474460602,
-0.0067182257771492004,
-0.07513581961393356,
0.036784227937459946,
0.020267607644200325,
-0.019571268931031227,
-0.06203276291489601,
-0.09180446714162827,
-0.016387205570936203,
0.07559262216091156,
-0.030262043699622154,
-0.009083885699510574,
0.0318804495036602,
0.0021675261668860912,
-0.08802425116300583,
-0.056988272815942764,
0.011446070857346058,
0.04434652253985405,
0.017498377710580826,
0.02495422214269638,
0.015521968714892864,
-0.030886013060808182,
-0.04742444306612015,
-0.012771420180797577,
-0.0183913204818964,
-0.026374250650405884,
0.014893908053636551,
-0.013052619993686676,
0.005774204153567553,
-0.026111146435141563,
-0.007745249196887016,
0.10681837797164917,
0.0011064561549574137,
-0.06959055364131927,
-0.07559442520141602,
-0.042547207325696945,
0.011082473210990429,
0.044147152453660965,
0.02114729955792427,
0.025701625272631645,
0.05525178089737892,
0.009045671671628952,
-0.03167169913649559,
-0.06393478810787201,
0.03138340637087822,
-0.02334107644855976,
-0.031170109286904335,
-0.12241299450397491,
-0.014513222500681877,
-0.0004451932618394494,
0.0003676513733807951,
0.06148916482925415,
-0.027087662369012833,
-0.08844462037086487,
-0.03511393442749977,
-0.02787252515554428,
-0.02027699537575245,
-0.03222888708114624,
-0.029398489743471146,
-0.0181998573243618,
0.10704932361841202,
-0.019008120521903038,
0.0020464088302105665,
-0.001276251976378262,
0.036917876452207565,
0.009631450287997723,
-0.03297136351466179,
0.04966670647263527,
-0.046347375959157944,
0.07966684550046921,
0.054033637046813965,
0.022152256220579147,
-0.017428232356905937,
0.08540072292089462,
0.010999825783073902,
-0.061421722173690796,
-0.009275280870497227,
0.06310746818780899,
0.08680650591850281,
0.06901953369379044,
0.05817319080233574,
-0.015386011451482773,
0.02593567594885826,
0.03804954141378403,
-0.04625622555613518,
0.06620372831821442,
0.038452547043561935,
0.03720054402947426,
0.08876729011535645,
-0.011433533392846584,
-0.05195898562669754,
0.023578930646181107,
0.05006725713610649,
-0.06688780337572098,
0.028787223622202873,
0.06776653230190277,
-0.05235832557082176,
0.062147341668605804,
0.06554403156042099,
-0.14654254913330078,
0.07100895792245865,
0.032813142985105515,
0.04495887830853462,
0.08275838941335678,
0.04287205636501312,
0.01455781888216734,
0.035737816244363785,
0.03030645102262497,
0.05397365614771843,
-0.03619608283042908,
0.0373019203543663,
0.003946448676288128,
0.013389511965215206,
-0.07229113578796387,
0.025381997227668762,
0.043491192162036896,
0.0061759501695632935,
0.034501563757658005,
0.08167675882577896,
-0.022006532177329063,
-0.04664963111281395,
0.056850772351026535,
-0.12201839685440063,
-1.1280668415357508e-35,
0.03882964327931404,
-0.006863993592560291,
0.08946514129638672,
0.13150355219841003,
0.01174499187618494,
-0.07941477745771408,
-0.015892699360847473,
-0.015803247690200806,
-0.07087200880050659,
0.04057000204920769,
-0.03545418381690979,
0.03600426763296127,
-0.017804423347115517,
0.07925428450107574,
0.04588489606976509,
-0.0222080759704113,
-0.004459155723452568,
0.11992965638637543,
-0.06535907834768295,
0.05414154753088951,
0.02092805877327919,
-0.007267291191965342,
-0.011263849213719368,
-0.022999482229351997,
0.06548192352056503,
-0.0008276383741758764,
-0.007444596383720636,
-0.04644402116537094,
-0.1161227822303772,
0.026935860514640808,
-0.08353316783905029,
-0.024548104032874107,
-0.05235886946320534,
0.014628185890614986,
0.008822273463010788,
-0.011791763827204704,
-0.008110566064715385,
-0.06618574261665344,
0.020820852369070053,
-0.02543945424258709,
0.025577303022146225,
-0.039980512112379074,
-0.06663574278354645,
-0.0457262322306633,
-0.012441751547157764,
-0.010178150609135628,
0.06552252918481827,
0.016248296946287155,
-0.03794679790735245,
-0.046849992126226425,
0.08256056159734726,
0.020649265497922897,
0.07327895611524582,
0.05058844015002251,
-0.13110843300819397,
0.05466974899172783,
-0.016080960631370544,
-0.010878557339310646,
0.04694126546382904,
0.054095324128866196,
-0.007313405629247427,
0.0747089609503746,
-0.09715229272842407,
0.04124818742275238,
-0.011597520671784878,
-0.0050993748009204865,
-0.0634148120880127,
-0.004446042701601982,
0.010015364736318588,
0.10561001300811768,
-0.05541389435529709,
-0.003059414913877845,
-0.0014272466069087386,
-0.030869290232658386,
0.08598815649747849,
-0.0794348195195198,
-0.016223954036831856,
-0.04466979578137398,
0.028325114399194717,
0.04279966652393341,
-0.0472065769135952,
0.04623621329665184,
-0.03976107016205788,
-0.055081117898225784,
-0.08439666777849197,
-0.051605429500341415,
0.1239151731133461,
-0.016901200637221336,
-0.08427701890468597,
-0.00964350625872612,
-0.02065313793718815,
-0.03822062909603119,
0.025187186896800995,
0.023959463462233543,
-0.02964569814503193,
-1.253804406941626e-33,
0.010877947323024273,
-0.01560654491186142,
0.01447164174169302,
0.018959244713187218,
-0.008622841909527779,
-0.042664285749197006,
-0.03797376900911331,
0.021568791940808296,
0.00800316222012043,
-0.0693855807185173,
0.10043852776288986,
-0.005413166247308254,
0.01270972564816475,
-0.03758905455470085,
0.00836429838091135,
0.05508395656943321,
-0.12054809927940369,
-0.011574105359613895,
0.02531857043504715,
0.048077214509248734,
-0.014458535239100456,
0.019262855872511864,
-0.02816144749522209,
0.0050344159826636314,
-0.032596394419670105,
0.06015228480100632,
-0.04938340559601784,
-0.011296545155346394,
0.050179433077573776,
-0.003824173705652356,
-0.0910641998052597,
-0.01073797419667244,
-0.03447548672556877,
-0.006778469309210777,
-0.07079201191663742,
0.03308892250061035,
0.01453916635364294,
-0.10005241632461548,
0.0704202950000763,
0.012953178025782108,
-0.0514889732003212,
0.06769898533821106,
-0.10575273633003235,
-0.0070953369140625,
0.03789512813091278,
0.03886670246720314,
-0.04447430744767189,
0.02687663584947586,
-0.012573677115142345,
-0.01781725138425827,
0.014509993605315685,
0.022500406950712204,
0.01439716387540102,
-0.0094190938398242,
-0.07368116825819016,
-0.05458088964223862,
-0.0352778285741806,
-0.010686837136745453,
0.008307678624987602,
0.06311801075935364,
-0.039933107793331146,
0.029381511732935905,
0.01838516630232334,
0.08735630661249161,
0.014733768068253994,
-0.026806699112057686,
-0.028924912214279175,
-0.00664012273773551,
-0.030911589041352272,
0.0006992346607148647,
-0.03633100166916847,
-0.027039293199777603,
0.06796873360872269,
0.06472238898277283,
0.08051243424415588,
-0.04790426790714264,
0.03764951974153519,
-0.03095855563879013,
-0.03486860543489456,
-0.006597608793526888,
0.004868756048381329,
0.036883093416690826,
0.03127655014395714,
-0.007568462751805782,
0.09583594650030136,
0.07878818362951279,
0.09464453160762787,
0.08838897198438644,
-0.010507732629776001,
0.04824446886777878,
-0.027535898610949516,
-0.022315023466944695,
-0.07977065443992615,
0.12801054120063782,
0.05590015649795532,
-4.342749093666498e-8,
0.036116596311330795,
0.09029287844896317,
0.014401613734662533,
-0.007307729218155146,
-0.0923079401254654,
0.015100671909749508,
-0.025438079610466957,
0.013030429370701313,
-0.044798512011766434,
0.01649298146367073,
0.06172625347971916,
-0.0036196024157106876,
-0.09288056939840317,
0.027351368218660355,
0.02960881032049656,
0.025534095242619514,
-0.030628567561507225,
0.04631641507148743,
-0.05651445314288139,
0.011519157327711582,
0.015908831730484962,
0.08995901793241501,
0.012462490238249302,
-0.056330833584070206,
0.046175505965948105,
-0.08378928154706955,
-0.06325215101242065,
0.07574371248483658,
-0.028818456456065178,
-0.03460288420319557,
0.013715755194425583,
0.07448751479387283,
-0.040927059948444366,
0.043188437819480896,
-0.0033884120639413595,
0.05088428780436516,
-0.0789770781993866,
-0.053801048547029495,
0.010400161147117615,
0.06792692095041275,
0.006419734563678503,
0.07804228365421295,
-0.08652765303850174,
-0.00488600879907608,
0.027113057672977448,
-0.019243771210312843,
-0.021382732316851616,
0.04952939599752426,
-0.015480784699320793,
-0.016468457877635956,
-0.0023668634239584208,
-0.03022836707532406,
-0.05339963361620903,
0.013395930640399456,
0.001740884268656373,
0.05684150382876396,
-0.07633233070373535,
-0.047778937965631485,
0.05326414108276367,
-0.05744690075516701,
0.04571367800235748,
-0.11303147673606873,
-0.014639830216765404,
-0.013418387621641159
] | 0 | 0 | 0 | 1 |
|
LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid | https://openreview.net/forum?id=ISqx8giekS | [
"Tianyi Zhang",
"Anshumali Shrivastava"
] | Poster | Large language models (LLMs) have shown immense potential across various domains, but their high memory requirements and inference costs remain critical challenges for deployment. Post-training quantization (PTQ) has emerged as a promising technique to reduce memory requirements and decoding latency. However, recent accurate quantization methods often depend on specialized computations or custom data formats to achieve better model quality, which limits their compatibility with popular frameworks, as they require dedicated inference kernels tailored to specific hardware and software platforms, hindering wider adoption. Furthermore, many competitive methods have high resource requirements and computational overhead for quantizing models, making it challenging to scale them to hundreds of billions of parameters. In response to these challenges, we propose LeanQuant (Loss-error-aware network Quantization), a novel quantization method that is accurate, versatile, and scalable. In the existing popular iterative loss-error-based quantization framework, we identify a critical limitation in prior methods: the min-max affine quantization grid fails to preserve model quality due to outliers in inverse Hessian diagonals. To overcome this fundamental issue, we propose learning loss-error-aware grids, instead of using non-adaptive min-max affine grids. Our approach not only produces quantized models that are more accurate but also generalizes to a wider range of quantization types, including affine and non-uniform quantization, enhancing compatibility with more frameworks. Extensive experiments with recent LLMs demonstrate that LeanQuant is highly accurate, comparing favorably against competitive baselines in model quality, and scalable, achieving very accurate quantization of Llama-3.1 405B, one of the largest open-source LLMs to date, using two Quadro RTX 8000-48GB GPUs in 21 hours. Our code is available at https://github.com/LeanModels/LeanQuant. | large language model, quantization | null | 12,988 | 2407.10032 | [
-0.004550582729279995,
-0.09812557697296143,
-0.023002592846751213,
0.028740152716636658,
0.023406492546200752,
0.037655699998140335,
-0.05818835273385048,
-0.014369407668709755,
0.032427944242954254,
-0.05787880718708038,
-0.0002576420083642006,
-0.017235388979315758,
-0.044059645384550095,
0.02151070348918438,
-0.007217936683446169,
-0.015072561800479889,
0.05647195503115654,
0.11791615188121796,
-0.06376270949840546,
-0.09441731870174408,
0.02977273054420948,
0.025362659245729446,
-0.054925113916397095,
0.030945589765906334,
0.08983448147773743,
-0.0069890134036540985,
-0.005595372524112463,
0.01062527671456337,
0.013318438082933426,
0.012261872179806232,
0.011464065872132778,
0.03979146480560303,
-0.0005251435213722289,
-0.014657896012067795,
0.012678963132202625,
0.07358163595199585,
0.00022091819846536964,
-0.05530828982591629,
0.014757358469069004,
-0.0792677104473114,
0.013412104919552803,
-0.012990754097700119,
0.04495140165090561,
0.011381275951862335,
0.036872200667858124,
-0.012437062337994576,
0.019965361803770065,
0.004546894691884518,
0.05126592889428139,
0.025737155228853226,
-0.023660996928811073,
-0.005704913754016161,
-0.029055073857307434,
0.03443323075771332,
-0.004208628088235855,
-0.06513214111328125,
0.030342167243361473,
0.02238788641989231,
-0.028485845774412155,
0.009090462699532509,
-0.07933102548122406,
-0.12458402663469315,
0.017050735652446747,
0.032440491020679474,
-0.04124390706419945,
-0.01434352621436119,
0.055072981864213943,
0.019494399428367615,
-0.022863484919071198,
0.040968868881464005,
-0.0490250363945961,
0.0023201846051961184,
-0.08225438743829727,
0.1095825806260109,
-0.028768304735422134,
0.11517174541950226,
0.06627152860164642,
-0.04002099856734276,
0.0785624161362648,
-0.03386056423187256,
-0.003775876946747303,
0.065837562084198,
0.05749231204390526,
-0.00036342532257549465,
0.13845428824424744,
-0.02550586126744747,
-0.02931777574121952,
0.04648549109697342,
0.016143834218382835,
-0.01590311899781227,
-0.05279698595404625,
0.035016778856515884,
0.020331937819719315,
0.014756695367395878,
0.04306153208017349,
0.02222973108291626,
0.06165485456585884,
-0.005978184286504984,
-0.027781998738646507,
0.07181566208600998,
0.04030326381325722,
0.07565362006425858,
0.0305794607847929,
-0.020684190094470978,
0.04693879187107086,
-0.033379361033439636,
0.04905064031481743,
0.09502384066581726,
0.03156140074133873,
-0.0863099992275238,
-0.021061928942799568,
-0.009730957448482513,
0.013957351446151733,
-0.016384610906243324,
0.08618827164173126,
0.011723315343260765,
0.022332455962896347,
-0.06280327588319778,
0.02095804549753666,
-0.010746733285486698,
-0.12044354528188705,
0.013457941822707653,
-0.012690078467130661,
0.007872309535741806,
0.03600763529539108,
-0.017393548041582108,
-0.13858939707279205,
8.991085374726011e-34,
0.07046351581811905,
0.09137141704559326,
0.06066351383924484,
-0.008731485344469547,
0.048456475138664246,
-0.01736859232187271,
0.048756279051303864,
-0.0014288965612649918,
-0.04080025851726532,
0.039299145340919495,
-0.024310320615768433,
-0.04686186835169792,
-0.08252377808094025,
0.005675274878740311,
0.06607436388731003,
0.00555858900770545,
-0.048348136246204376,
0.03701255843043327,
-0.019509868696331978,
-0.05493123084306717,
0.09667252749204636,
-0.10717121511697769,
-0.013889383524656296,
0.02374209091067314,
0.03677170351147652,
0.02419682778418064,
0.117877297103405,
0.012830128893256187,
-0.04782097041606903,
-0.0034499517641961575,
-0.09239017963409424,
-0.07111907005310059,
0.030287347733974457,
-0.03507605567574501,
0.06541607528924942,
-0.09036919474601746,
0.019672991707921028,
-0.03538336977362633,
0.0290258526802063,
-0.06684266030788422,
-0.049396954476833344,
0.006755547132343054,
-0.007197595201432705,
-0.06730591505765915,
-0.09754182398319244,
-0.05646126717329025,
-0.014299093745648861,
-0.04524233937263489,
-0.015185839496552944,
-0.08951627463102341,
0.09292341768741608,
-0.019072314724326134,
-0.03485279157757759,
0.05545123293995857,
0.011700785718858242,
-0.0389985665678978,
0.0516788586974144,
0.015650702640414238,
0.05012594908475876,
0.10715815424919128,
-0.034705840051174164,
-0.04738643392920494,
-0.0279369056224823,
0.05444483086466789,
0.05527333542704582,
-0.00007264792657224461,
-0.01167357712984085,
0.038369596004486084,
-0.03619009256362915,
0.03583819419145584,
-0.029616504907608032,
-0.04310107231140137,
-0.014540771953761578,
-0.03155208379030228,
0.1034909188747406,
-0.05225200206041336,
0.03877510502934456,
-0.08232946693897247,
0.023786025121808052,
-0.01650550775229931,
-0.001421127119101584,
0.037280064076185226,
-0.03317345306277275,
0.004557894077152014,
-0.045445483177900314,
-0.07797770947217941,
0.033020343631505966,
-0.053617507219314575,
-0.06908228993415833,
0.007267479784786701,
-0.10296406596899033,
-0.06271825730800629,
0.00617602001875639,
-0.021427420899271965,
-0.10039226710796356,
-1.6232912095863607e-33,
-0.09493628144264221,
0.04886367917060852,
-0.02467423304915428,
0.079533152282238,
-0.05866540968418121,
-0.11416573822498322,
0.02716413512825966,
0.06974077224731445,
0.028115760535001755,
-0.1273508071899414,
-0.008928117342293262,
-0.07870551198720932,
0.04098937660455704,
0.05179750546813011,
0.04983191937208176,
0.02220090478658676,
-0.035217367112636566,
0.010517432354390621,
-0.020607588812708855,
0.05929451808333397,
-0.013799124397337437,
0.01213444210588932,
-0.020305391401052475,
-0.0380919985473156,
0.02446793206036091,
0.02935957722365856,
-0.05324927717447281,
0.05494129657745361,
0.012231112457811832,
-0.01653614081442356,
-0.05485401302576065,
-0.015319102443754673,
-0.028534643352031708,
-0.058428503572940826,
-0.03549709916114807,
0.04196035489439964,
0.05909484997391701,
0.0012404622975736856,
-0.03290972858667374,
0.0754295364022255,
0.11291155964136124,
-0.09483637660741806,
-0.11346696317195892,
0.023719606921076775,
0.0071509587578475475,
-0.007764569018036127,
-0.0988178476691246,
0.027374550700187683,
0.0072750127874314785,
0.0015378713142126799,
-0.012641397304832935,
0.0317058227956295,
0.04408865049481392,
0.1225259006023407,
-0.011411173269152641,
-0.06151396036148071,
-0.028238998726010323,
0.009205816313624382,
0.01717968098819256,
-0.026899123564362526,
-0.09038709849119186,
-0.07304669171571732,
0.015550882555544376,
-0.04924028366804123,
0.019363898783922195,
-0.053861960768699646,
-0.01752972975373268,
-0.00627503078430891,
-0.0013312773080542684,
0.026924697682261467,
0.029889540746808052,
-0.00023173289082478732,
-0.026349185034632683,
0.053081758320331573,
-0.03458897024393082,
-0.09003215283155441,
-0.015012189745903015,
-0.014759473502635956,
0.01360694132745266,
0.04133932664990425,
0.07006017863750458,
0.09122473001480103,
0.05583830922842026,
0.0011454764753580093,
0.12496551126241684,
0.0320553295314312,
0.06779307872056961,
0.06672321259975433,
-0.021272072568535805,
0.021841060370206833,
-0.053962647914886475,
0.010238975286483765,
0.01635112799704075,
0.05834218114614487,
-0.10849866271018982,
-4.025801203511037e-8,
-0.11263163387775421,
0.13011443614959717,
0.0031853928230702877,
-0.04348580539226532,
0.03915393725037575,
-0.0406038798391819,
-0.009760629385709763,
0.07039083540439606,
0.035217929631471634,
0.0012472642119973898,
0.05219360068440437,
-0.012228517793118954,
-0.07379040122032166,
-0.05216461047530174,
-0.07516712695360184,
0.07442289590835571,
-0.017701050266623497,
0.006497581955045462,
0.03209591656923294,
-0.02626008912920952,
-0.03776327148079872,
0.06962045282125473,
-0.02931126020848751,
0.015753742307424545,
0.06755929440259933,
-0.06463903933763504,
0.04663421958684921,
0.10951923578977585,
0.1264050155878067,
0.02659895084798336,
-0.05377225950360298,
0.006857466883957386,
0.058732226490974426,
-0.019742203876376152,
0.0018264757236465812,
0.03824895992875099,
-0.010044243186712265,
-0.021584540605545044,
0.016442999243736267,
-0.05090724304318428,
-0.028993459418416023,
0.003340478055179119,
-0.08829791098833084,
0.02107408456504345,
0.05931350216269493,
0.016963589936494827,
0.009302031248807907,
0.01944895088672638,
0.017418665811419487,
-0.02704666554927826,
0.019547177478671074,
0.0255773663520813,
-0.013531919568777084,
0.02176610566675663,
-0.023653779178857803,
0.018387453630566597,
0.001842767815105617,
-0.02981642261147499,
0.041726723313331604,
0.02129252813756466,
-0.007014710456132889,
-0.033029165118932724,
-0.021179962903261185,
0.012229708954691887
] | 0 | 5 | 0 | 0 |
|
Strategist: Self-improvement of LLM Decision Making via Bi-Level Tree Search | https://openreview.net/forum?id=gfI9v7AbFg | [
"Jonathan Light",
"Min Cai",
"Weiqin Chen",
"Guanzhi Wang",
"Xiusi Chen",
"Wei Cheng",
"Yisong Yue",
"Ziniu Hu"
] | Poster | Traditional reinforcement learning and planning require a lot of data and training to develop effective strategies. On the other hand, large language models (LLMs) can generalize well and perform tasks without prior training but struggle with complex planning and decision-making. We introduce **STRATEGIST**, a new approach that combines the strengths of both methods. It uses LLMs to generate and update high-level strategies in text form, while a Monte Carlo Tree Search (MCTS) algorithm refines and executes them. STRATEGIST is a general framework that optimizes strategies through self-play simulations without requiring any training data. We test STRATEGIST in competitive, multi-turn games with partial information, such as **Game of Pure Strategy (GOPS)** and **The Resistance: Avalon**, a multi-agent hidden-identity discussion game. Our results show that STRATEGIST-based agents outperform traditional reinforcement learning models, other LLM-based methods, and existing LLM agents while achieving performance levels comparable to human players. | LLMs, games, search, self-improvement, self-play, RL, agent, multi-agent, planning, decision-making, Monte Carlo Tree Search, hierarchical learning, strategy optimization, reinforcement learning, game AI, strategic reasoning, population-based training, adversarial learning, policy refinement, social deduction games, partially observable environments, heuristic learning, evolutionary strategies. | We introduce STRATEGIST, a bi-level framework that enables LLMs to self-improve decision-making in multi-agent games by iteratively refining high-level strategies through self-play and optimizing execution with Monte Carlo Tree Search (MCTS). | 12,980 | null | [
-0.017594628036022186,
-0.012052174657583237,
0.009545773267745972,
0.004756659269332886,
-0.002111929701641202,
0.0026016782503575087,
-0.008588235825300217,
0.03003772348165512,
0.04893866553902626,
-0.00420203385874629,
-0.0709756463766098,
-0.05365141108632088,
0.0323634035885334,
-0.0017640212317928672,
0.10451871156692505,
-0.02977759949862957,
0.09571028500795364,
0.07345155626535416,
0.031330693513154984,
-0.08616412431001663,
0.041183825582265854,
0.01986929029226303,
0.08295675367116928,
0.014406819827854633,
-0.0018724144902080297,
-0.029587555676698685,
0.0012675421312451363,
-0.0010059972992166877,
-0.0184226855635643,
-0.004090600647032261,
0.07632657885551453,
-0.01342943124473095,
0.05181825906038284,
-0.016346227377653122,
-0.13384409248828888,
0.02311297506093979,
-0.14695273339748383,
-0.035678815096616745,
-0.02580808661878109,
-0.041373129934072495,
-0.056747470051050186,
0.018970921635627747,
0.01230828370898962,
0.008180909790098667,
0.10274778306484222,
0.022449109703302383,
-0.020215362310409546,
-0.004031097516417503,
0.004126652609556913,
0.006076042074710131,
-0.07261358201503754,
-0.06476804614067078,
-0.008276316337287426,
-0.038942210376262665,
0.011223133653402328,
0.08628308027982712,
0.02733728475868702,
0.028237227350473404,
0.036859285086393356,
-0.04621042683720589,
0.015300607308745384,
-0.1104096844792366,
-0.008116813376545906,
-0.033808864653110504,
-0.07446977496147156,
-0.013879972510039806,
0.04001878947019577,
0.09202081710100174,
0.014610830694437027,
-0.017199868336319923,
0.026454638689756393,
0.055046647787094116,
-0.004405017476528883,
0.013318013399839401,
0.010059291496872902,
0.038768965750932693,
-0.040913812816143036,
-0.016249103471636772,
0.06707436591386795,
0.003572179703041911,
-0.06838811933994293,
-0.002837154781445861,
-0.01960938610136509,
0.05419610068202019,
0.04932650551199913,
-0.02627096325159073,
-0.03321604058146477,
0.09685030579566956,
0.07677491754293442,
-0.024038001894950867,
-0.010300702415406704,
-0.025218158960342407,
0.009329376742243767,
-0.018990231677889824,
0.000026960537070408463,
0.06020526960492134,
-0.060622941702604294,
-0.08137743920087814,
-0.03057192824780941,
0.033453330397605896,
0.017081258818507195,
0.0835246592760086,
0.03812152147293091,
-0.02909942902624607,
-0.02144785225391388,
0.016367485746741295,
0.03733464702963829,
0.021903008222579956,
0.055820077657699585,
-0.03491489961743355,
-0.01864142343401909,
-0.0011232306715101004,
-0.011994005180895329,
0.032151397317647934,
0.02610720321536064,
0.062249086797237396,
0.05960574746131897,
-0.03586442023515701,
-0.06180243939161301,
0.12587960064411163,
0.02681196853518486,
-0.015015268698334694,
0.022693311795592308,
0.05763496458530426,
0.04710640385746956,
0.014791618101298809,
-0.01611289009451866,
3.594704772703379e-35,
0.07397323101758957,
0.04834776371717453,
0.078243188560009,
0.09780696034431458,
0.02334565669298172,
0.01670471578836441,
0.08262817561626434,
-0.016877368092536926,
-0.02356334589421749,
-0.004115636460483074,
-0.03289679437875748,
0.03112809732556343,
-0.06664053350687027,
0.033900100737810135,
0.08805805444717407,
0.0028046879451721907,
-0.07225938886404037,
0.036847617477178574,
0.010608071461319923,
-0.056064605712890625,
0.0905587449669838,
0.05846920609474182,
0.0667896419763565,
-0.11418437212705612,
0.10543651133775711,
0.07358129322528839,
0.005159562453627586,
-0.08689378947019577,
0.003551952075213194,
-0.014939728192985058,
-0.0747617706656456,
-0.02009398117661476,
-0.0653839185833931,
0.04762990400195122,
0.03657227382063866,
0.03314738720655441,
-0.11011612415313721,
-0.05919865891337395,
-0.012653552927076817,
0.030733622610569,
-0.08129093796014786,
-0.033765003085136414,
0.08429642766714096,
0.018959183245897293,
-0.09995649009943008,
-0.08946073055267334,
-0.06144154071807861,
-0.031457383185625076,
-0.03602174296975136,
-0.025446880608797073,
0.021726172417402267,
0.030808277428150177,
-0.05307049676775932,
-0.04227520897984505,
-0.014612103812396526,
0.047889359295368195,
0.013855447061359882,
0.04227888584136963,
-0.05354449525475502,
0.04601955786347389,
0.04977254569530487,
0.01799881085753441,
-0.01464416366070509,
0.1512216478586197,
-0.017030762508511543,
0.0835917517542839,
-0.02131451852619648,
0.04126523807644844,
0.08300407975912094,
-0.07450500875711441,
0.0597708523273468,
-0.03238994628190994,
0.005741948261857033,
-0.059386905282735825,
0.03592127934098244,
-0.0288996584713459,
0.06060129404067993,
-0.07595346868038177,
0.01620992086827755,
-0.009105518460273743,
-0.006617848761379719,
-0.002935524797067046,
-0.05229276418685913,
-0.06202924996614456,
0.03459411486983299,
-0.06624943763017654,
0.03235765919089317,
-0.08415113389492035,
0.04188600927591324,
0.0459861122071743,
-0.08779669553041458,
-0.014547668397426605,
-0.05347462370991707,
0.04522784426808357,
0.023234840482473373,
-3.9277968238375915e-34,
-0.06157239153981209,
-0.09848114103078842,
-0.027995089069008827,
0.0478365533053875,
0.0030850626062601805,
-0.024869849905371666,
-0.034962065517902374,
-0.05199253931641579,
0.029931435361504555,
-0.058706413954496384,
-0.13083580136299133,
0.05043457821011543,
0.05482533201575279,
-0.007624026387929916,
0.017883507534861565,
-0.07110273092985153,
0.04320715367794037,
0.040306732058525085,
0.02975296415388584,
0.03803993761539459,
0.0010016643209382892,
0.11236225813627243,
-0.16020652651786804,
0.005292900837957859,
-0.004987709224224091,
0.003359853057190776,
0.021346595138311386,
0.010631403885781765,
0.04649597406387329,
0.03426015004515648,
-0.00416681170463562,
-0.009721703827381134,
-0.008692624047398567,
-0.02725941874086857,
-0.04937436431646347,
0.04136907309293747,
0.02973027154803276,
0.029414353892207146,
-0.03958655521273613,
0.0635605901479721,
0.03858940303325653,
-0.08786989003419876,
-0.06313895434141159,
-0.00893247127532959,
-0.056563783437013626,
0.006834770552814007,
-0.060582127422094345,
-0.007406014483422041,
0.02702050469815731,
0.04121313616633415,
-0.022764986380934715,
-0.025235872715711594,
-0.07470116764307022,
0.002492684405297041,
-0.07311293482780457,
-0.09632095694541931,
-0.04489855840802193,
0.023302115499973297,
0.006038486026227474,
0.004038655664771795,
-0.041825246065855026,
0.052203528583049774,
0.020945146679878235,
-0.010472742840647697,
0.022404981777071953,
-0.06813348084688187,
-0.07198264449834824,
0.028099529445171356,
0.03312024846673012,
-0.08914586156606674,
-0.0008126043248921633,
-0.008516130968928337,
-0.032855547964572906,
-0.022867994382977486,
0.045953914523124695,
0.03918980434536934,
-0.0636078342795372,
-0.008014739491045475,
-0.013739071786403656,
-0.03146879002451897,
0.05525585636496544,
0.0729280635714531,
0.009906901977956295,
0.07593198865652084,
0.050159379839897156,
0.039963144809007645,
0.015821194276213646,
0.12988033890724182,
-0.009514260105788708,
-0.024467458948493004,
0.0063262456096708775,
0.013577821664512157,
0.03298812359571457,
0.04380214214324951,
-0.07642409950494766,
-4.2166981017999206e-8,
-0.08646658807992935,
0.0348280668258667,
0.07343029230833054,
0.06030156463384628,
0.04727013036608696,
-0.026981549337506294,
-0.03861004486680031,
0.0570729598402977,
-0.03261486068367958,
0.007221986074000597,
0.07392899692058563,
-0.029077308252453804,
0.00450749509036541,
-0.10229083150625229,
0.04113911837339401,
0.014046970754861832,
0.004331935662776232,
-0.05135372281074524,
-0.017244232818484306,
-0.006606792099773884,
0.04559721052646637,
0.007193503901362419,
-0.04120294749736786,
-0.03265255317091942,
-0.005137236788868904,
-0.035931967198848724,
-0.04250985383987427,
0.031120669096708298,
0.06900721043348312,
0.027769314125180244,
0.05542350560426712,
-0.06342461705207825,
-0.004146967548877001,
0.0916631743311882,
0.03885558247566223,
0.017706086859107018,
0.10556326806545258,
0.008437921293079853,
0.014508195221424103,
-0.01447213627398014,
-0.0042667509987950325,
0.01680426113307476,
-0.004205481614917517,
0.03701373562216759,
-0.03867948055267334,
0.03721622750163078,
-0.09714920818805695,
-0.12386218458414078,
0.05832499638199806,
-0.05946815013885498,
0.013407336547970772,
-0.017700381577014923,
-0.04335852712392807,
0.011851775459945202,
0.09924396872520447,
0.053612567484378815,
0.03422529622912407,
-0.008422519080340862,
0.04481489211320877,
0.07965438067913055,
0.03593018651008606,
-0.07360947877168655,
-0.05108402296900749,
0.03593318164348602
] | 0 | 0 | 0 | 0 |
|
DataMan: Data Manager for Pre-training Large Language Models | https://openreview.net/forum?id=eNbA8Fqir4 | [
"Ru Peng",
"Kexin Yang",
"Yawen Zeng",
"Junyang Lin",
"Dayiheng Liu",
"Junbo Zhao"
] | Poster | The performance emergence of large language models (LLMs) driven by data scaling laws makes the selection of pre-training data increasingly important.
However, existing methods rely on limited heuristics and human intuition, lacking comprehensive and clear guidelines.
To address this, we are inspired by *``reverse thinking''* -- prompting LLMs to self-identify which criteria benefit its performance.
As its pre-training capabilities are related to perplexity (PPL), we derive 14 quality criteria from the causes of text perplexity anomalies and introduce 15 common application domains to support domain mixing.
In this paper, we train a **Data** **Man**ager (**DataMan**) to learn quality ratings and domain recognition from pointwise rating, and use it to annotate a 447B token pre-training corpus with 14 quality ratings and domain type.
Our experiments validate our approach, using DataMan to select 30B tokens to train a 1.3B-parameter language model, demonstrating significant improvements in in-context learning (ICL), perplexity, and instruction-following ability over the state-of-the-art baseline.
The best-performing model, based on the *Overall Score l=5* surpasses a model trained with 50% more data using uniform sampling.
We continue pre-training with high-rated, domain-specific data annotated by DataMan to enhance domain-specific ICL performance and thus verify DataMan's domain mixing ability.
Our findings emphasize the importance of quality ranking, the complementary nature of quality criteria, and their low correlation with perplexity, analyzing misalignment between PPL and ICL performance.
We also thoroughly analyzed our pre-training dataset, examining its composition, the distribution of quality ratings, and the original document sources. | Large Language Models, Data Selection, Pre-Training | null | 12,977 | 2502.19363 | [
-0.018338369205594063,
-0.08400990068912506,
0.031158775091171265,
0.02216804213821888,
-0.04060843214392662,
-0.04619528725743294,
0.01846068538725376,
0.0023044515401124954,
-0.004707971587777138,
-0.01247850339859724,
-0.018078098073601723,
-0.06124115362763405,
0.07638914883136749,
0.016576332971453667,
0.03438173606991768,
0.02119595743715763,
0.12240785360336304,
0.006663348991423845,
-0.09854721277952194,
-0.04140901193022728,
0.03047187440097332,
0.055289264768362045,
-0.01777331903576851,
-0.03110291063785553,
-0.0033722934313118458,
0.008574355393648148,
0.006810041610151529,
-0.013919612392783165,
0.021002231165766716,
0.01057632640004158,
0.03328625112771988,
0.07899926602840424,
0.024146977812051773,
0.038646895438432693,
-0.010380315594375134,
0.03893495351076126,
-0.05381383001804352,
0.027103042230010033,
0.019867945462465286,
-0.009316845797002316,
0.02054586447775364,
-0.04899873584508896,
0.04945378750562668,
0.025873946025967598,
0.02281167544424534,
-0.013244584202766418,
-0.024700883775949478,
-0.03201799467206001,
-0.0007593132904730737,
0.027602769434452057,
-0.07680386304855347,
0.02150026708841324,
0.022068051621317863,
0.04927345737814903,
-0.08759840577840805,
-0.0016458005411550403,
-0.007814288139343262,
0.09460459649562836,
-0.05586383864283562,
-0.032101064920425415,
-0.13484138250350952,
-0.13142526149749756,
-0.05477423220872879,
-0.018428394570946693,
-0.0015998601447790861,
-0.0531838983297348,
-0.010541057214140892,
0.051614295691251755,
-0.026330914348363876,
0.0349542535841465,
0.011084001511335373,
0.06502138078212738,
-0.022091791033744812,
0.1456729918718338,
0.0210768710821867,
0.029599551111459732,
-0.023867182433605194,
-0.0398431159555912,
0.07224074006080627,
-0.08911044150590897,
-0.014309229329228401,
0.04918481037020683,
0.09379326552152634,
0.015029816888272762,
0.09823929518461227,
0.004256556276232004,
0.0797225832939148,
0.012291712686419487,
-0.011416124179959297,
0.07633689790964127,
-0.04082310199737549,
-0.04854578152298927,
0.07144557684659958,
0.0119505375623703,
0.007711953949183226,
0.06265988945960999,
-0.03875069320201874,
-0.07824279367923737,
-0.013506293296813965,
0.03844187408685684,
0.011751465499401093,
0.08782784640789032,
0.004667692817747593,
0.0019181995885446668,
-0.0034467654768377542,
-0.04840201511979103,
0.02876114659011364,
0.0704818069934845,
0.04641590639948845,
-0.06984622776508331,
0.04974955692887306,
0.08146132528781891,
0.027956748381257057,
-0.06793469935655594,
0.06305035203695297,
-0.023060301318764687,
0.0029865754768252373,
-0.04041069746017456,
0.06517624109983444,
0.051025617867708206,
-0.06405913084745407,
0.003202610183507204,
0.000751927203964442,
-0.016157835721969604,
0.008643165230751038,
-0.07592988759279251,
-0.02556760050356388,
-6.5853971556823325e-34,
0.020330093801021576,
0.06377461552619934,
0.00016922500799410045,
0.05726311355829239,
-0.01372386422008276,
-0.005708080716431141,
-0.01772242598235607,
0.046310558915138245,
-0.018646294251084328,
0.017130127176642418,
0.008876760490238667,
0.08630058914422989,
-0.028657574206590652,
0.02728327177464962,
0.017853934317827225,
0.02568204514682293,
-0.021853605285286903,
0.01791827753186226,
-0.04039047658443451,
-0.014054356142878532,
0.08151872456073761,
0.03160816803574562,
0.003991476260125637,
-0.05387042462825775,
0.05936595797538757,
0.012142829596996307,
0.055267348885536194,
-0.04306532070040703,
-0.01049833931028843,
0.021493008360266685,
-0.13571827113628387,
-0.05251549556851387,
0.014259874820709229,
0.026615263894200325,
0.06173485144972801,
-0.015505674295127392,
0.004473676439374685,
-0.026251114904880524,
0.02275671251118183,
-0.03176229074597359,
-0.05851009860634804,
0.033767975866794586,
-0.016550319269299507,
0.027762893587350845,
-0.08136141300201416,
-0.0594174787402153,
0.027651403099298477,
-0.0039172060787677765,
-0.030102331191301346,
-0.04063961282372475,
0.0974070131778717,
-0.004589966963976622,
-0.023351242765784264,
-0.001930517377331853,
-0.04581538960337639,
0.02280462719500065,
0.04435613751411438,
0.0013621727703139186,
0.03140827640891075,
0.07580260932445526,
0.033064235001802444,
0.02386038564145565,
-0.0279069896787405,
0.10888291150331497,
0.004346200730651617,
-0.020559227094054222,
0.010526005178689957,
0.024614838883280754,
0.01932949386537075,
-0.07619751989841461,
-0.02949717454612255,
0.01402795035392046,
0.00706050731241703,
-0.06564533710479736,
0.07038667798042297,
-0.011508303694427013,
0.05562824010848999,
-0.05967208370566368,
0.019905250519514084,
0.010692009702324867,
0.026260916143655777,
-0.04087543115019798,
-0.008958077058196068,
-0.0922476053237915,
-0.0627257376909256,
-0.04956076294183731,
0.03591911122202873,
-0.08710458874702454,
0.02058195136487484,
-0.0399707593023777,
-0.005648783873766661,
-0.0015454029198735952,
0.0023916177451610565,
0.045086123049259186,
-0.03437458723783493,
1.1697058676639532e-33,
-0.028644446283578873,
-0.04939660057425499,
-0.02872529998421669,
0.1242588609457016,
-0.04289703816175461,
-0.05086657032370567,
0.05127280578017235,
0.08213461935520172,
0.024243967607617378,
-0.11413650214672089,
-0.04509901627898216,
-0.06330637633800507,
0.06512618809938431,
-0.017520401626825333,
0.05849991738796234,
0.04176075756549835,
-0.04536685720086098,
0.09130443632602692,
-0.013423942029476166,
0.07176939398050308,
-0.028093848377466202,
0.05917628854513168,
-0.15600870549678802,
0.0736585184931755,
-0.051192112267017365,
0.02820335514843464,
-0.04148746654391289,
-0.0007632022607140243,
0.022875919938087463,
0.04644570127129555,
-0.010817552916705608,
-0.01858087070286274,
0.04421726241707802,
-0.02245192416012287,
-0.11150585114955902,
-0.026726530864834785,
0.053282927721738815,
-0.023013893514871597,
-0.03250043839216232,
0.0789499506354332,
0.0877939835190773,
-0.03616642579436302,
-0.11053025722503662,
-0.006725076120346785,
-0.05902262404561043,
-0.06297598779201508,
-0.07685644924640656,
0.01467336155474186,
0.028856763616204262,
-0.020254142582416534,
-0.053695619106292725,
0.04425846040248871,
0.0009061748278327286,
0.0021686688996851444,
-0.06375986337661743,
-0.11536457389593124,
0.04171580448746681,
-0.05600256100296974,
0.062427818775177,
0.013490106910467148,
-0.03702250495553017,
0.035783953964710236,
0.12535998225212097,
-0.014371675439178944,
-0.009667307138442993,
-0.055826976895332336,
0.0017453898908570409,
0.03102167695760727,
-0.07214654237031937,
-0.006797169800847769,
-0.028733717277646065,
-0.02971820719540119,
-0.04017471894621849,
-0.04949147626757622,
-0.0162209440022707,
0.002848477102816105,
0.07244130969047546,
-0.07186111062765121,
-0.012648429721593857,
-0.027820296585559845,
0.004758104216307402,
0.017715133726596832,
0.015791555866599083,
0.08443304151296616,
0.06560559570789337,
0.1301063448190689,
0.0703498125076294,
0.05172397568821907,
-0.025045672431588173,
0.04029618203639984,
-0.002079060999676585,
0.00837680697441101,
-0.016890866681933403,
0.045981038361787796,
-0.05057622864842415,
-4.614159365701198e-8,
-0.10178320109844208,
-0.0016376640414819121,
-0.05999447777867317,
0.06377676129341125,
-0.016038214787840843,
-0.011121553368866444,
-0.09232132881879807,
0.08845271170139313,
0.05242634937167168,
0.006207899656146765,
0.04980562627315521,
0.030422771349549294,
-0.1469617486000061,
-0.11400161683559418,
0.024587826803326607,
0.057781416922807693,
0.04909037426114082,
0.022295014932751656,
0.012808308005332947,
0.0011829239083454013,
0.06348670274019241,
0.05804413557052612,
0.0658944696187973,
-0.06919794529676437,
0.04504667967557907,
-0.031516920775175095,
-0.014478707686066628,
0.11856380105018616,
0.030411599203944206,
-0.043233804404735565,
0.005093065090477467,
-0.013358143158257008,
-0.03518364951014519,
-0.05064334347844124,
0.04381038621068001,
0.08043048530817032,
0.0766921415925026,
-0.010449630208313465,
0.002397614996880293,
-0.04894459247589111,
0.042931411415338516,
0.077088363468647,
-0.09098341315984726,
0.014004993252456188,
0.013526641763746738,
0.0015387102030217648,
-0.005555749870836735,
-0.0853487178683281,
0.049185849726200104,
-0.05572381988167763,
0.00782835390418768,
-0.0062745800241827965,
-0.001589911407791078,
0.040167514234781265,
0.09064720571041107,
0.06137658655643463,
-0.05825470760464668,
-0.005952139850705862,
-0.017621276900172234,
0.05474898964166641,
0.04709245264530182,
-0.020950183272361755,
-0.04497591406106949,
0.016826624050736427
] | 0 | 0 | 0 | 0 |
|
EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing | https://openreview.net/forum?id=Y2Dh8rWwlb | [
"Kaizhi Zheng",
"Xiaotong Chen",
"Xuehai He",
"Jing Gu",
"Linjie Li",
"Zhengyuan Yang",
"Kevin Lin",
"Jianfeng Wang",
"Lijuan Wang",
"Xin Eric Wang"
] | Poster | Given the steep learning curve of professional 3D software and the time-
consuming process of managing large 3D assets, language-guided 3D scene editing has significant potential in fields such as virtual reality, augmented reality, and
gaming. However, recent approaches to language-guided 3D scene editing either
require manual interventions or focus only on appearance modifications without
supporting comprehensive scene layout changes. In response, we propose EditRoom, a unified framework capable of executing a variety of layout edits through
natural language commands, without requiring manual intervention. Specifically,
EditRoom leverages Large Language Models (LLMs) for command planning and
generates target scenes using a diffusion-based method, enabling six types of edits: rotate, translate, scale, replace, add, and remove. To address
the lack of data for language-guided 3D scene editing, we have developed an automatic pipeline to augment existing 3D scene synthesis datasets and introduced
EditRoom-DB, a large-scale dataset with 83k editing pairs, for training and evaluation. Our experiments demonstrate that our approach consistently outperforms
other baselines across all metrics, indicating higher accuracy and coherence in
language-guided scene layout editing. | 3D Scene Editing, Large Language Model, Diffusion-based Models | We propose a graph diffusion-based method fo language-guided 3D scene layout editing | 12,972 | 2410.12836 | [
0.004386999644339085,
-0.06857126951217651,
0.04052584618330002,
0.024280134588479996,
0.013981753960251808,
0.008167413994669914,
-0.12203701585531235,
-0.07202962785959244,
0.016567448154091835,
0.010715424083173275,
-0.027491865679621696,
-0.04660680145025253,
0.025034621357917786,
0.0554969385266304,
0.0672868937253952,
-0.0015552211552858353,
0.06944694370031357,
0.06074000149965286,
-0.07622668147087097,
0.000366770604159683,
-0.03796757757663727,
0.05470959097146988,
0.005889380816370249,
-0.008430601097643375,
0.06740391254425049,
0.02497372217476368,
0.04750977084040642,
-0.05948377773165703,
0.026416726410388947,
-0.04708479344844818,
0.031841836869716644,
0.05034215375781059,
-0.004745685029774904,
0.06007220968604088,
-0.01534118503332138,
0.0830056220293045,
-0.03601555526256561,
-0.0003703804686665535,
-0.06906042993068695,
-0.0792548730969429,
-0.056448712944984436,
0.04040016606450081,
-0.0016551074804738164,
0.021128302440047264,
0.11245322972536087,
0.009002931416034698,
-0.07722455263137817,
-0.06809346377849579,
-0.02145334519445896,
0.016539182513952255,
-0.06245935335755348,
-0.10299619287252426,
-0.05155491456389427,
0.04527398571372032,
-0.06382321566343307,
0.07386790961027145,
0.1230827197432518,
0.0009668238344602287,
0.0670393779873848,
-0.04170023649930954,
-0.008231060579419136,
-0.03750209882855415,
0.000990400556474924,
0.0031014662235975266,
-0.06321858614683151,
-0.016455240547657013,
0.031867384910583496,
-0.009411978535354137,
-0.05282629281282425,
-0.007563075982034206,
-0.02436576783657074,
-0.020277030766010284,
0.0097710732370615,
-0.0030509873759001493,
-0.023935750126838684,
0.00690129678696394,
-0.040375400334596634,
-0.09755820780992508,
-0.01381901279091835,
-0.09424389153718948,
0.06727610528469086,
-0.016684411093592644,
0.0013205797877162695,
0.06258842349052429,
0.0006472241948358715,
-0.0008868015138432384,
0.00624682754278183,
0.04213257506489754,
0.0010318098356947303,
0.04207292199134827,
-0.045912209898233414,
-0.029281537979841232,
-0.024454157799482346,
-0.03914111107587814,
-0.030254244804382324,
-0.01001615822315216,
0.00484246201813221,
-0.08730556815862656,
-0.026780402287840843,
0.0468968041241169,
0.003863175632432103,
-0.03300816938281059,
0.015390337444841862,
-0.04161025583744049,
0.003367466852068901,
-0.022675488144159317,
0.09908569604158401,
-0.02461010217666626,
-0.047060370445251465,
-0.03436756879091263,
0.049840047955513,
-0.010402644984424114,
-0.03515790402889252,
-0.027292953804135323,
0.08048490434885025,
-0.04451572895050049,
0.09320897608995438,
-0.0263488981872797,
0.03247259184718132,
0.056389134377241135,
0.022311875596642494,
0.031131327152252197,
0.02723216824233532,
0.04868829622864723,
-0.043271079659461975,
-0.02170734852552414,
0.0032488342840224504,
9.557779262642935e-34,
0.0859191045165062,
0.017115605995059013,
-0.014739199541509151,
0.04999847710132599,
0.14676930010318756,
0.02182108536362648,
0.01963968388736248,
0.051440902054309845,
-0.0430472157895565,
-0.057050611823797226,
0.023546939715743065,
-0.05488405004143715,
-0.03165347874164581,
0.12315547466278076,
0.05548793077468872,
0.0005325232050381601,
0.01754596456885338,
0.09457923471927643,
-0.05026297643780708,
0.0603344663977623,
0.0008646885980851948,
-0.01895197667181492,
-0.02401503175497055,
0.002935639349743724,
-0.03850729763507843,
0.025513041764497757,
0.048438381403684616,
0.00795477069914341,
0.01829228550195694,
-0.03081464022397995,
-0.11507228016853333,
0.0060438122600317,
0.019466008991003036,
0.02819828875362873,
-0.013729389756917953,
-0.03881177306175232,
-0.0537082701921463,
-0.07824045419692993,
0.028644856065511703,
-0.006972374860197306,
0.010229270905256271,
0.08279161900281906,
-0.02200126275420189,
-0.03308970108628273,
-0.10997574031352997,
0.08561588078737259,
0.033580146729946136,
0.0604754202067852,
-0.0029154347721487284,
-0.03245091438293457,
0.06658569723367691,
0.07112928479909897,
-0.012119795195758343,
-0.0224009919911623,
-0.02230006456375122,
-0.06961628049612045,
0.06196637079119682,
-0.04927271977066994,
0.07201137393712997,
0.013967888429760933,
0.04634256660938263,
0.10000166296958923,
-0.03194089233875275,
0.12094961106777191,
0.039024341851472855,
0.009503294713795185,
0.06375496089458466,
0.06927211582660675,
0.06044671684503555,
-0.040906865149736404,
-0.12804710865020752,
-0.012778283096849918,
0.004231042228639126,
-0.0029422470834106207,
0.0549873486161232,
-0.09209729731082916,
-0.032901011407375336,
-0.06851304322481155,
0.02840839885175228,
0.010579473339021206,
-0.06169744208455086,
-0.01207044068723917,
-0.0010000901529565454,
-0.018671253696084023,
0.007394515909254551,
-0.09580403566360474,
0.00385979819111526,
-0.016712969169020653,
0.010270318016409874,
-0.08801886439323425,
-0.020995836704969406,
-0.10690474510192871,
-0.03485746309161186,
0.008892184123396873,
0.029923105612397194,
-4.804757853879251e-34,
0.021436650305986404,
-0.01649397611618042,
-0.015807580202817917,
0.06072669476270676,
-0.024315765127539635,
-0.06824523955583572,
-0.05563785880804062,
0.014751285314559937,
0.10786977410316467,
-0.08306524902582169,
-0.026560300961136818,
0.0025584364775568247,
0.026937944814562798,
-0.041500385850667953,
0.013170842081308365,
-0.04349544271826744,
-0.011082921177148819,
-0.03612096607685089,
-0.04553650692105293,
0.0015076030977070332,
0.019777115434408188,
-0.02009117417037487,
-0.06773378700017929,
0.032009758055210114,
-0.08799722790718079,
0.05013786256313324,
-0.026669194921851158,
0.026541462168097496,
0.010539964772760868,
0.047982871532440186,
-0.05136609449982643,
0.004330235533416271,
0.022693436592817307,
0.004418832715600729,
-0.06017500162124634,
0.02234463207423687,
0.056536659598350525,
-0.0101308673620224,
-0.0745481625199318,
0.02524580992758274,
0.04287636652588844,
-0.01813497021794319,
-0.058863479644060135,
0.08475115895271301,
-0.05228037387132645,
0.06062425300478935,
-0.06515615433454514,
0.022213274613022804,
-0.03758886829018593,
0.02713550254702568,
-0.10329070687294006,
0.05740763247013092,
-0.055256664752960205,
-0.08128905296325684,
0.032233886420726776,
-0.13836108148097992,
-0.06444928050041199,
-0.04128233343362808,
0.09064094722270966,
-0.060922931879758835,
-0.06420121341943741,
-0.0025582751259207726,
0.008373157121241093,
-0.05764342471957207,
-0.0029749134555459023,
0.03451498970389366,
-0.0804731622338295,
-0.06542126089334488,
0.009211528114974499,
0.049668315798044205,
0.04836568236351013,
-0.017417458817362785,
-0.009593089111149311,
0.02419980801641941,
0.007319897413253784,
-0.06276524066925049,
0.0703863725066185,
-0.0488533116877079,
0.044130053371191025,
-0.043166983872652054,
0.04541032388806343,
-0.0354173518717289,
0.04080028086900711,
0.06639491021633148,
0.03196388855576515,
0.08915018290281296,
-0.036387525498867035,
0.10441149771213531,
0.010408436879515648,
0.07374458760023117,
-0.0639287605881691,
0.03916935622692108,
0.03812984377145767,
0.07277639210224152,
-0.06889372318983078,
-3.710205831453095e-8,
-0.07166723906993866,
0.044736411422491074,
0.010447360575199127,
0.012073328718543053,
-0.035617951303720474,
-0.030687445774674416,
0.05286345258355141,
0.1533006876707077,
0.061366960406303406,
-0.01876498945057392,
0.034337397664785385,
-0.0076455469243228436,
-0.0031691808253526688,
-0.005487569607794285,
-0.04173896089196205,
0.10601693391799927,
-0.018274003639817238,
0.05514182522892952,
-0.008673577569425106,
0.03821181505918503,
0.0034201168455183506,
0.0951714888215065,
0.05089845508337021,
-0.03236128017306328,
0.06526850163936615,
-0.01533126924186945,
-0.0534382201731205,
0.08165392279624939,
0.041985079646110535,
-0.04038099944591522,
0.0725630670785904,
-0.04449472948908806,
0.05858680605888367,
0.10605993866920471,
-0.019069910049438477,
-0.00340943387709558,
0.028480621054768562,
0.011084520258009434,
0.015569149516522884,
-0.04258551821112633,
0.009432286024093628,
0.03233471140265465,
-0.01243777945637703,
0.02195807173848152,
-0.005245096515864134,
-0.0036827484145760536,
0.0432504341006279,
-0.04753294959664345,
0.013763672672212124,
-0.09191473573446274,
-0.0845613107085228,
-0.011225434951484203,
-0.0330086424946785,
0.05158352106809616,
0.03730885684490204,
0.09570738673210144,
0.008831607177853584,
-0.006800143979489803,
0.09925186634063721,
0.04043680056929588,
-0.03659649193286896,
-0.002342347288504243,
-0.03247088938951492,
-0.0003962407063227147
] | 0 | 0 | 0 | 0 |
|
LocoVR: Multiuser Indoor Locomotion Dataset in Virtual Reality | https://openreview.net/forum?id=9mBodivRIo | [
"Kojiro Takeyama",
"Yimeng Liu",
"Misha Sra"
] | Poster | Understanding human locomotion is crucial for AI agents such as robots, particularly in complex indoor home environments. Modeling human trajectories in these spaces requires insight into how individuals maneuver around physical obstacles and manage social navigation dynamics. These dynamics include subtle behaviors influenced by proxemics - the social use of space, such as stepping aside to allow others to pass or choosing longer routes to avoid collisions. Previous research has developed datasets of human motion in indoor scenes, but these are often limited in scale and lack the nuanced social navigation dynamics common in home environments.
To address this, we present LocoVR, a dataset of 7000+ two-person trajectories captured in virtual reality from over 130 different indoor home environments. LocoVR provides accurate trajectory and precise spatial information, along with rich examples of socially-motivated movement behaviors.
For example, the dataset captures instances of individuals navigating around each other in narrow spaces, adjusting paths to respect personal boundaries in living areas, and coordinating movements in high-traffic zones like entryways and kitchens. Our evaluation shows that LocoVR significantly enhances model performance in three practical indoor tasks utilizing human trajectories, and demonstrates predicting socially-aware navigation patterns in home environments. | Dataset, Human trajectory, Indoor locomotion, Virtual reality, Social motion behavior | We present a dataset of two-person trajectories across 130+ home environments, capturing geometrically and socially-aware motion behaviors. | 12,967 | 2410.06437 | [
0.04846208542585373,
-0.10116000473499298,
0.019174128770828247,
-0.020341338589787483,
0.06199437752366066,
-0.025700867176055908,
-0.004783071577548981,
-0.07246049493551254,
0.0009386349702253938,
-0.034600649029016495,
0.005932767875492573,
-0.08052309602499008,
-0.021129587665200233,
0.056955303996801376,
0.018211988732218742,
-0.016760295256972313,
0.07022714614868164,
0.07419305294752121,
-0.068303681910038,
0.0635981485247612,
-0.006010914221405983,
0.020649870857596397,
0.02550816349685192,
0.03077486902475357,
0.0018711892189458013,
0.017604272812604904,
0.046361155807971954,
-0.06045662611722946,
-0.0032087014988064766,
-0.04361902177333832,
0.10277111083269119,
-0.04434848204255104,
-0.02854994684457779,
-0.007085680030286312,
-0.03270136937499046,
0.012748856097459793,
0.0156259685754776,
-0.027983536943793297,
-0.04604880511760712,
-0.004231462720781565,
-0.00960585754364729,
-0.0038891404401510954,
0.06175592169165611,
0.026255406439304352,
0.05554429441690445,
0.11119266599416733,
-0.015350169502198696,
-0.08349649608135223,
0.08763469755649567,
0.02667098306119442,
-0.06042659282684326,
0.0036218317691236734,
0.10135658085346222,
0.11828868091106415,
-0.05200060456991196,
0.013576536439359188,
-0.024354491382837296,
-0.007531929761171341,
0.09671083837747574,
-0.03553760424256325,
0.08422922343015671,
-0.054579898715019226,
0.030551984906196594,
-0.004463115707039833,
-0.030388519167900085,
-0.07661375403404236,
0.025737714022397995,
0.032267581671476364,
0.050593990832567215,
0.007205440662801266,
-0.008849765174090862,
0.03944515064358711,
0.03395110368728638,
0.033769141882658005,
0.04819198697805405,
-0.007023899350315332,
-0.07576809823513031,
-0.09632527828216553,
-0.004013365134596825,
-0.03048221953213215,
0.05724266543984413,
-0.08512886613607407,
-0.03247362747788429,
0.06465330719947815,
0.07664377242326736,
0.023198485374450684,
-0.008616489358246326,
0.061041273176670074,
-0.001585203455761075,
0.010526961646974087,
-0.08102985471487045,
0.032943472266197205,
0.05550268664956093,
-0.07513003796339035,
-0.03566160053014755,
0.019026968628168106,
-0.09428777545690536,
-0.09057488292455673,
0.03344788774847984,
0.06383001059293747,
-0.08056194335222244,
0.011357339099049568,
0.050904229283332825,
0.06483474373817444,
0.09392250329256058,
0.01133374311029911,
0.07510475069284439,
0.026354443281888962,
0.009125014767050743,
0.08164478838443756,
0.0009550984250381589,
-0.03235943987965584,
0.014097629114985466,
0.013112729415297508,
0.06629515439271927,
0.03337973356246948,
-0.026084089651703835,
0.07960580289363861,
0.018259087577462196,
0.04052206501364708,
0.006297181826084852,
-0.04574096202850342,
0.07188747078180313,
0.006941054482012987,
0.049220453947782516,
-0.06893175840377808,
-0.03708866983652115,
7.132623295946205e-34,
-0.015889998525381088,
0.038868196308612823,
0.023199276998639107,
0.07504861801862717,
0.06773572415113449,
-0.049462269991636276,
-0.06479312479496002,
-0.03414927050471306,
-0.02115859091281891,
0.03808197006583214,
-0.00021416597883217037,
-0.044218894094228745,
-0.031700875610113144,
0.06491958349943161,
0.04670262709259987,
0.00341001502238214,
-0.023075535893440247,
-0.01825251616537571,
-0.0662197545170784,
0.111640065908432,
0.10335065424442291,
-0.08683477342128754,
-0.03292251378297806,
0.03398517519235611,
-0.022649412974715233,
0.005430134478956461,
0.046177834272384644,
0.033324386924505234,
0.04421156272292137,
-0.021810581907629967,
-0.010561536997556686,
0.05907737836241722,
-0.05160698667168617,
0.056797027587890625,
0.05233720690011978,
0.02553839050233364,
-0.04929104447364807,
-0.05651375651359558,
-0.0007660659030079842,
0.061945509165525436,
-0.06216229498386383,
0.013027144595980644,
-0.027666602283716202,
-0.012949598021805286,
-0.05568854510784149,
-0.06302370876073837,
0.07089202105998993,
0.0465821772813797,
-0.07941585779190063,
0.015271448530256748,
-0.029043281450867653,
0.06893748044967651,
-0.042228084057569504,
-0.057173777371644974,
-0.0700034499168396,
-0.0849774181842804,
0.06028694659471512,
-0.016626544296741486,
0.010594408959150314,
0.019336257129907608,
0.014945191331207752,
0.009207002818584442,
0.03244715556502342,
0.0540330596268177,
-0.01657148450613022,
-0.06513683497905731,
0.009729370474815369,
0.13459135591983795,
0.09769963473081589,
0.04164813831448555,
-0.04185931757092476,
0.08320699632167816,
0.02295239269733429,
-0.0585017167031765,
0.07788171619176865,
-0.017632845789194107,
-0.025986336171627045,
-0.09680745750665665,
-0.040074385702610016,
-0.023303555324673653,
-0.05800238624215126,
-0.03504277393221855,
-0.008749867789447308,
0.04962896928191185,
0.04611703008413315,
-0.036198634654283524,
-0.047476254403591156,
-0.04500212520360947,
-0.10614724457263947,
-0.0009342462290078402,
-0.022963400930166245,
-0.0382799357175827,
-0.06729968637228012,
0.021785803139209747,
-0.10111784934997559,
-1.946885886255528e-33,
-0.02054572105407715,
0.023099709302186966,
-0.0019235035870224237,
-0.008263158611953259,
-0.028983693569898605,
-0.02735920064151287,
-0.015854550525546074,
-0.015055880881845951,
0.03409355506300926,
-0.015356379561126232,
-0.05144056677818298,
0.03155004605650902,
0.1123366728425026,
0.04302153363823891,
0.08446647971868515,
-0.025839390233159065,
0.10251642763614655,
-0.05352567136287689,
-0.005641234572976828,
-0.03040509857237339,
0.019448649138212204,
0.0500672273337841,
-0.042570751160383224,
-0.04108057916164398,
0.015618976205587387,
-0.011367211118340492,
0.03739939630031586,
0.035352420061826706,
-0.007807515561580658,
0.04777209460735321,
-0.01673542708158493,
-0.057685788720846176,
0.008502882905304432,
-0.06613150984048843,
0.006643143482506275,
0.04073932021856308,
0.024349667131900787,
0.023253582417964935,
-0.01300795003771782,
-0.042383823543787,
0.05841692537069321,
-0.055642951279878616,
-0.019617123529314995,
-0.020081408321857452,
-0.05236513167619705,
0.041469424962997437,
-0.09738202393054962,
0.03515246883034706,
-0.06386051326990128,
0.034895241260528564,
0.021802367642521858,
0.07968278229236603,
-0.1340569257736206,
-0.08401424437761307,
-0.02863076701760292,
-0.053123291581869125,
-0.07471676170825958,
-0.0018011833308264613,
0.050829868763685226,
-0.03646516054868698,
-0.03581623733043671,
0.057976242154836655,
-0.08270174264907837,
0.11770794540643692,
-0.025186199694871902,
0.04471634700894356,
-0.08092714846134186,
-0.0239583570510149,
-0.017555803060531616,
0.01760590448975563,
-0.021958623081445694,
-0.051862671971321106,
0.01053532212972641,
-0.007960253395140171,
-0.017657769843935966,
-0.046220459043979645,
0.06654511392116547,
-0.031363677233457565,
0.06618465483188629,
-0.023459360003471375,
-0.007512720767408609,
-0.06487134099006653,
0.037098228931427,
0.005636232439428568,
-0.01293994765728712,
0.09409292042255402,
-0.07437188923358917,
0.03754241764545441,
-0.07255576550960541,
0.1090242862701416,
0.05043395236134529,
0.07598290592432022,
-0.08774417638778687,
-0.044720832258462906,
-0.08894761651754379,
-3.6725683827398825e-8,
-0.12669254839420319,
0.0475635789334774,
0.01854659803211689,
-0.02308753877878189,
-0.017612222582101822,
0.0023568884935230017,
0.06795200705528259,
0.053852830082178116,
-0.02159043215215206,
0.0385860800743103,
0.045589957386255264,
0.022229058668017387,
0.030288496986031532,
0.018368858844041824,
0.008199125528335571,
0.038132477551698685,
-0.01433588843792677,
0.02267867885529995,
-0.06071513518691063,
0.07906971126794815,
0.011535300873219967,
-0.0011082730488851666,
-0.011149427853524685,
0.00859819445759058,
0.008733750320971012,
-0.016674930229783058,
-0.08631981909275055,
0.011695724911987782,
0.036331675946712494,
-0.04773467034101486,
0.05001964792609215,
-0.019504407420754433,
-0.0279719065874815,
-0.03091387264430523,
-0.06650407612323761,
0.04325440898537636,
-0.028112562373280525,
0.0031386991031467915,
0.015941794961690903,
-0.044355303049087524,
-0.01887865923345089,
0.002373570343479514,
-0.05856519564986229,
0.03395025432109833,
0.01171592902392149,
0.02890988253057003,
0.0017359363846480846,
-0.10781076550483704,
0.050676245242357254,
-0.008774477057158947,
-0.06445419043302536,
-0.05225630849599838,
0.046983033418655396,
0.07065699994564056,
0.060350071638822556,
0.03514096885919571,
-0.031848929822444916,
-0.062364425510168076,
0.09701930731534958,
-0.007329603191465139,
-0.0030417165253311396,
0.0216522254049778,
-0.115598164498806,
-0.03302919492125511
] | https://github.com/kojirotakeyama/locovr | 9 | 0 | 0 | 0 |
Prompting Fairness: Integrating Causality to Debias Large Language Models | https://openreview.net/forum?id=7GKbQ1WT1C | [
"Jingling Li",
"Zeyu Tang",
"Xiaoyu Liu",
"Peter Spirtes",
"Kun Zhang",
"Liu Leqi",
"Yang Liu"
] | Poster | Large language models (LLMs), despite their remarkable capabilities, are susceptible to generating biased and discriminatory responses. As LLMs increasingly influence high-stakes decision-making (e.g., hiring and healthcare), mitigating these biases becomes critical. In this work, we propose a causality-guided debiasing framework to tackle social biases, aiming to reduce the objectionable dependence between LLMs' decisions and the social information in the input. Our framework introduces a novel perspective to identify how social information can affect an LLM's decision through different causal pathways. Leveraging these causal insights, we outline principled prompting strategies that regulate these pathways through selection mechanisms. This framework not only unifies existing prompting-based debiasing techniques, but also opens up new directions for reducing bias by encouraging the model to prioritize fact-based reasoning over reliance on biased social cues. We validate our framework through extensive experiments on real-world datasets across multiple domains, demonstrating its effectiveness in debiasing LLM decisions, even with only black-box access to the model. | Large Language Model, Prompting, Social Bias, Causality, Debias, Selection Mechanism | We propose a causality-guided framework to mitigate social biases in large language models by regulating the flow of social information through different causal pathways to influence the model’s decisions. | 12,960 | 2403.08743 | [
-0.016001898795366287,
-0.004264491144567728,
0.01945260912179947,
0.03265490382909775,
0.13739177584648132,
0.05151699110865593,
0.03876907005906105,
-0.02768506295979023,
0.04959363862872124,
-0.013963364064693451,
-0.08714821934700012,
-0.020969992503523827,
0.05440255627036095,
-0.002762804040685296,
-0.0003228384302929044,
-0.020471928641200066,
0.14462603628635406,
-0.005687600001692772,
-0.09750965982675552,
-0.08768066763877869,
-0.008059367537498474,
0.003764741588383913,
0.018006961792707443,
-0.003048989921808243,
-0.011982124298810959,
-0.11549647897481918,
0.04647880420088768,
-0.026108626276254654,
-0.021858764812350273,
0.030440766364336014,
0.009145224466919899,
-0.0019609318114817142,
0.0024855216033756733,
-0.005147003103047609,
-0.04226237162947655,
-0.02017981931567192,
-0.03242781385779381,
-0.00781018752604723,
-0.03040478564798832,
-0.015368795022368431,
0.022011345252394676,
-0.04722638428211212,
0.00536931399255991,
0.020212028175592422,
0.03627892583608627,
-0.01901850476861,
0.04945666715502739,
0.04571795463562012,
-0.05457991734147072,
0.0032972057815641165,
-0.07361248880624771,
-0.040541402995586395,
0.07259230315685272,
0.04618452861905098,
-0.030390622094273567,
-0.0023003320675343275,
0.01424277015030384,
0.08675408363342285,
-0.07127095013856888,
-0.004770409781485796,
-0.057079609483480453,
-0.1380387246608734,
-0.09000486135482788,
0.04780275747179985,
-0.05721662566065788,
0.03078235499560833,
-0.03700154274702072,
0.094969242811203,
-0.08067997545003891,
0.04109478369355202,
0.0782298743724823,
0.07069870084524155,
0.022353684529662132,
0.04335395246744156,
0.06668678671121597,
0.019581088796257973,
-0.013668745756149292,
-0.02393493801355362,
0.07240204513072968,
-0.07652243226766586,
-0.0034113505389541388,
-0.02736411616206169,
0.027852920815348625,
0.03327410668134689,
0.12399985641241074,
0.015930477529764175,
0.007809550501406193,
0.006979444529861212,
0.0030573252588510513,
0.06384904682636261,
-0.026941930875182152,
-0.07827925682067871,
0.04354920610785484,
0.04015512019395828,
0.012376525439321995,
0.053298577666282654,
-0.047478944063186646,
-0.01686408929526806,
0.011832642368972301,
0.0682109072804451,
-0.026382004842162132,
0.012984107248485088,
-0.0435277596116066,
0.03188537806272507,
0.02547791413962841,
-0.050518106669187546,
-0.0375056229531765,
0.051241856068372726,
-0.013054135255515575,
-0.009751003235578537,
-0.03702176734805107,
0.054287683218717575,
0.06207545846700668,
0.005139590706676245,
0.008455517701804638,
-0.05804803594946861,
0.05450386554002762,
0.04165278747677803,
-0.006852158345282078,
-0.02199418842792511,
-0.02134106308221817,
0.04541325941681862,
0.023751871660351753,
-0.025235962122678757,
0.047284938395023346,
-0.021958721801638603,
-0.050487637519836426,
-8.977853716439568e-34,
0.0703059583902359,
0.034444402903318405,
0.04880974441766739,
0.001801341655664146,
0.04121819883584976,
-0.011302630417048931,
-0.08687464147806168,
0.034456636756658554,
-0.031371958553791046,
0.0845969021320343,
-0.03306451439857483,
0.02159418910741806,
0.03482380509376526,
0.016389576718211174,
0.015487326309084892,
0.0550541914999485,
-0.08808685839176178,
0.09546513110399246,
-0.03861912712454796,
0.024623945355415344,
0.09151238203048706,
-0.009168961085379124,
-0.050909373909235,
0.006010283250361681,
-0.04082605242729187,
-0.0013779915170744061,
0.014452900737524033,
-0.0096582705155015,
0.03085239790380001,
0.032723575830459595,
-0.09190373867750168,
-0.05484655499458313,
0.08473242819309235,
-0.011763736605644226,
0.01294084545224905,
-0.05411454290151596,
-0.09679204970598221,
-0.024766534566879272,
0.016069315373897552,
-0.008338783867657185,
0.03471982479095459,
-0.007036582101136446,
0.05684620141983032,
0.024085931479930878,
-0.04173223301768303,
-0.03819315508008003,
-0.008211633190512657,
-0.03910558670759201,
-0.020050352439284325,
0.01529240608215332,
0.08659384399652481,
0.059475768357515335,
0.06522195041179657,
0.030254818499088287,
-0.05267864838242531,
-0.014521625824272633,
-0.031275056302547455,
0.0058974334970116615,
-0.023698942735791206,
-0.04933314025402069,
0.011961081065237522,
0.007844923995435238,
-0.012580154463648796,
0.02767883986234665,
0.04207761213183403,
-0.006006311625242233,
-0.07283351570367813,
-0.027956271544098854,
0.047422144562006,
-0.04214015603065491,
0.008757814764976501,
0.061075568199157715,
-0.08525819331407547,
-0.033085897564888,
-0.04376307874917984,
0.06981639564037323,
0.00006372208736138418,
0.037796054035425186,
0.11219731718301773,
-0.020969437435269356,
0.07247929275035858,
-0.10447423160076141,
0.0004428272368386388,
-0.041077662259340286,
0.011522158980369568,
0.00013335295079741627,
0.006666356697678566,
-0.06337469816207886,
0.053743958473205566,
-0.02985271066427231,
-0.022922463715076447,
-0.02392616495490074,
-0.04127003997564316,
0.03337393328547478,
-0.018026228994131088,
-1.4687416204440619e-33,
-0.09698574244976044,
-0.048701800405979156,
0.027134064584970474,
-0.0011455878848209977,
-0.03562943637371063,
-0.00011253345292061567,
0.02984689548611641,
-0.06790449470281601,
0.08743954449892044,
-0.021716343238949776,
-0.011988149955868721,
-0.06131967529654503,
0.1115521714091301,
-0.013151651248335838,
0.02195926569402218,
-0.037982892245054245,
0.03793662041425705,
0.025876587256789207,
-0.005826212000101805,
0.015770087018609047,
0.053580354899168015,
0.06636340171098709,
-0.042645759880542755,
0.028850985690951347,
0.02235262095928192,
-0.003776501165702939,
-0.013372554443776608,
0.008349129930138588,
0.04117773473262787,
0.027834296226501465,
-0.14629961550235748,
0.0149308480322361,
-0.0319739393889904,
-0.04719133675098419,
-0.1059362068772316,
0.04631301760673523,
0.047145646065473557,
-0.023283781483769417,
-0.09072383493185043,
0.06706912815570831,
0.025111118331551552,
-0.06262760609388351,
-0.1371067315340042,
0.00010353773541282862,
-0.07556647807359695,
-0.016460834071040154,
-0.06419963389635086,
-0.0027572300750762224,
0.02281882055103779,
-0.010265573859214783,
-0.06464780867099762,
-0.0006161665660329163,
0.039736710488796234,
0.15744620561599731,
-0.056651771068573,
-0.1301865130662918,
0.015434160828590393,
-0.06489738821983337,
0.0553157739341259,
0.05765447020530701,
-0.061658989638090134,
0.037615351378917694,
0.006296040024608374,
-0.04099518433213234,
-0.03337084874510765,
-0.12262515723705292,
0.01287102047353983,
-0.025525320321321487,
0.025443214923143387,
-0.05658366531133652,
0.07114030420780182,
0.00904417596757412,
-0.04888257756829262,
-0.037917040288448334,
-0.03092855215072632,
0.028212472796440125,
0.03622639551758766,
-0.011744643561542034,
-0.09003801643848419,
0.010487793013453484,
0.12594516575336456,
0.0001239334378624335,
0.09760591387748718,
-0.01819252036511898,
0.06002869829535484,
0.046225715428590775,
0.009190567769110203,
0.07381071150302887,
-0.019236695021390915,
0.019948195666074753,
-0.05611548572778702,
-0.024868305772542953,
-0.018272550776600838,
0.04780947044491768,
-0.1187572106719017,
-4.151863208790019e-8,
-0.03102778270840645,
0.025185326114296913,
-0.02172950468957424,
0.127427875995636,
-0.01952608861029148,
-0.09625034779310226,
-0.1188613548874855,
0.06311874836683273,
-0.013262760825455189,
-0.05376796796917915,
0.003676416352391243,
0.05553290992975235,
-0.04912842810153961,
-0.053997959941625595,
0.056713689118623734,
0.05662842094898224,
-0.03688347339630127,
0.015160628594458103,
0.01899591274559498,
-0.05595901235938072,
0.07220809161663055,
-0.010624546557664871,
-0.003667579498142004,
-0.006674510892480612,
0.07800783962011337,
-0.016142556443810463,
-0.06605842709541321,
0.040481600910425186,
0.06236597150564194,
-0.0062081837095320225,
-0.03074645809829235,
-0.008484001271426678,
-0.09088728576898575,
-0.03903086110949516,
0.03623203933238983,
0.02099326066672802,
-0.06018180400133133,
-0.046350765973329544,
0.05416909232735634,
-0.030461741611361504,
0.06171591579914093,
0.11969548463821411,
-0.04603605717420578,
0.02663491666316986,
0.019792545586824417,
-0.011222624219954014,
-0.03091205470263958,
-0.004285554401576519,
-0.033584222197532654,
-0.009491401724517345,
-0.016835680231451988,
-0.04930073395371437,
0.047936052083969116,
0.01171165145933628,
0.06954437494277954,
0.016491327434778214,
0.010046257637441158,
-0.007320560049265623,
-0.001632241765037179,
0.0458822026848793,
0.11605475842952728,
0.02548709139227867,
-0.03278117999434471,
-0.0389840193092823
] | 0 | 0 | 0 | 0 |
|
GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers | https://openreview.net/forum?id=fWRBheSJth | [
"Sarkar Snigdha Sarathi Das",
"Ryo Kamoi",
"Bo Pang",
"Yusen Zhang",
"Caiming Xiong",
"Rui Zhang"
] | Poster | The effectiveness of large language models (LLMs) is closely tied to the design of prompts, making prompt optimization essential for enhancing their performance across a wide range of tasks. Although recent advancements have focused on automating prompt engineering, many existing approaches rely exclusively on textual feedback, refining prompts based solely on inference errors identified by large, computationally expensive LLMs. Unfortunately, smaller models struggle to generate high-quality feedback, resulting in complete dependence on large LLM judgment. Moreover, these methods fail to leverage more direct and finer-grained information, such as gradients, due to operating purely in text space. To this end, we introduce, we introduce *GReaTer*, a novel prompt optimization technique that directly incorporates *gradient information over task-specific reasoning*. By utilizing task loss gradients, *GReaTer* enables self-optimization of prompts for smaller, lightweight language models (LM) without the need for costly closed-source LLMs, while maintaining reasonable prompt structures. This allows high-performance prompt optimization without dependence on massive LLMs, closing the gap between smaller models and the sophisticated reasoning often needed for prompt refinement. Extensive evaluations across diverse tasks demonstrate that \ours consistently outperforms previous methods, even those reliant on powerful LLMs. Additionally, *GReaTer*-optimized prompts frequently exhibit better transferability and, in some cases, boost task performance to levels comparable to or surpassing those achieved by larger language models, highlighting the effectiveness of *"gradient over reasoning"*-based prompt optimization. Code of *GReaTer* is available at: https://github.com/psunlpgroup/GreaTer | Large Language Model, Prompt Optimization | We introduce a prompt optimization method using gradient over reasoning to boost performance on open-source, smaller language models. | 12,959 | 2412.09722 | [
-0.006537850946187973,
-0.003935391083359718,
0.02771664410829544,
0.037083763629198074,
0.09364989399909973,
-0.03388099744915962,
0.0007698572590015829,
0.054874952882528305,
0.0014479951933026314,
0.011013252660632133,
-0.10518437623977661,
-0.08493972569704056,
0.060548435896635056,
0.03519158810377121,
0.041902635246515274,
0.015419862233102322,
0.0953531488776207,
0.01908203214406967,
-0.08069298416376114,
-0.07831918448209763,
-0.02410101145505905,
0.018204718828201294,
0.034949302673339844,
0.01936565525829792,
-0.015706121921539307,
-0.019706500694155693,
-0.027088450267910957,
-0.04036356136202812,
0.006802735850214958,
0.00764906220138073,
0.005412448663264513,
-0.0077516729943454266,
0.016999660059809685,
0.04850254952907562,
-0.038707464933395386,
0.055075641721487045,
-0.057301755994558334,
0.0465644896030426,
-0.024359606206417084,
-0.08270201086997986,
-0.10513465851545334,
-0.03463641554117203,
0.022854046896100044,
0.041399210691452026,
0.06328728049993515,
-0.0589820072054863,
-0.006984252482652664,
-0.012287801131606102,
0.008441368117928505,
0.03586013615131378,
-0.08784125745296478,
-0.05950450524687767,
0.007876326330006123,
0.009749380871653557,
-0.03741145506501198,
0.03550443798303604,
0.054910313338041306,
0.03842499479651451,
-0.016851335763931274,
-0.032990410923957825,
-0.11617616564035416,
-0.13186995685100555,
-0.023047497496008873,
0.03005211241543293,
-0.018206624314188957,
-0.021301066502928734,
-0.016034739091992378,
0.0250015277415514,
-0.024844862520694733,
0.08638694882392883,
0.00009902174497256055,
0.07702823728322983,
-0.031204935163259506,
0.08608042448759079,
-0.009076115675270557,
0.056392524391412735,
-0.04649609699845314,
-0.02566811814904213,
0.0033015969675034285,
-0.039172347635030746,
-0.02799852006137371,
0.0017087695887312293,
0.01085576694458723,
0.023034919053316116,
0.11184010654687881,
-0.011552577838301659,
0.03525375574827194,
0.02198529802262783,
0.06244425103068352,
0.024610299617052078,
-0.003149104304611683,
-0.0911187157034874,
0.030489696189761162,
0.07266419380903244,
-0.017258351668715477,
0.07705633342266083,
-0.007677807006984949,
-0.14593848586082458,
-0.04182170331478119,
0.08716357499361038,
0.027336247265338898,
0.07842598855495453,
0.021143287420272827,
-0.09851372241973877,
-0.027074456214904785,
-0.025892212986946106,
0.03722243010997772,
0.07367772608995438,
-0.006291449535638094,
-0.1140318289399147,
0.04069603607058525,
0.07482808828353882,
0.06784317642450333,
-0.01330831740051508,
0.023778878152370453,
-0.031639214605093,
0.06085675582289696,
-0.07112453877925873,
0.032069478183984756,
0.03372160717844963,
0.049062665551900864,
0.020763922482728958,
-0.019785553216934204,
0.0035587858874350786,
0.06211776286363602,
-0.05059722810983658,
-0.0175975002348423,
1.4206564628202415e-33,
0.06014132872223854,
0.08008785545825958,
0.046100880950689316,
0.10281865298748016,
0.0693446695804596,
0.011811818927526474,
0.03337903693318367,
0.07705292105674744,
-0.046434011310338974,
-0.047859907150268555,
0.01087907887995243,
-0.042729176580905914,
-0.031149694696068764,
0.03892318904399872,
0.04877907782793045,
-0.007178345695137978,
-0.016436615958809853,
0.08340682089328766,
-0.04549349099397659,
-0.05961274728178978,
-0.008535395376384258,
-0.030397383496165276,
-0.018508827313780785,
0.013195541687309742,
0.059225622564554214,
0.0045093875378370285,
0.09970077872276306,
-0.037183813750743866,
-0.03847531974315643,
-0.002509456593543291,
-0.108929343521595,
-0.1163964718580246,
0.022757703438401222,
0.04574330151081085,
-0.01782962866127491,
0.025263264775276184,
-0.038099564611911774,
-0.07690367102622986,
0.06135406717658043,
0.014052224345505238,
0.02426890656352043,
0.01006084494292736,
0.018619410693645477,
0.06574814766645432,
-0.09001409262418747,
-0.08287757635116577,
-0.04978862777352333,
0.018871665000915527,
-0.049133721739053726,
-0.01692744717001915,
0.05835527181625366,
0.06235165148973465,
-0.008041409775614738,
0.015459545888006687,
-0.005048938561230898,
-0.0035721161402761936,
0.023584773764014244,
0.021875405684113503,
0.016328254714608192,
0.01604279689490795,
0.011600269004702568,
0.05897676944732666,
-0.012914392165839672,
0.06501082330942154,
-0.006368671078234911,
-0.010403248481452465,
-0.02694246545433998,
0.06734715402126312,
0.05943240225315094,
0.024035142734646797,
-0.07655102014541626,
-0.029781240969896317,
0.040822695940732956,
-0.066789910197258,
0.007738459389656782,
0.030856668949127197,
0.019941529259085655,
-0.09166432917118073,
0.08586473762989044,
-0.048222966492176056,
0.040760964155197144,
-0.005392994731664658,
-0.027696825563907623,
-0.008163721300661564,
-0.0013307338813319802,
-0.06895425915718079,
0.05438099056482315,
-0.061067916452884674,
-0.0359504297375679,
-0.034220028668642044,
-0.04142557457089424,
-0.026932718232274055,
-0.02539852075278759,
-0.006997630000114441,
-0.01815410703420639,
-1.3620137128871062e-33,
-0.013275117613375187,
-0.046896956861019135,
-0.05890251323580742,
0.038594115525484085,
-0.07353847473859787,
-0.04056285694241524,
-0.005806359928101301,
-0.018141066655516624,
-0.04235270991921425,
-0.0846349224448204,
-0.053223952651023865,
-0.015880433842539787,
0.04501701146364212,
0.014181062579154968,
0.025922071188688278,
-0.013142251409590244,
0.013624846003949642,
0.01985038071870804,
0.03641689941287041,
0.04388671740889549,
-0.009314484894275665,
0.1193106546998024,
-0.1206299215555191,
-0.014101303182542324,
-0.05246957018971443,
-0.016624068841338158,
-0.054095201194286346,
0.08257137984037399,
0.011840528808534145,
0.019301321357488632,
-0.07296780496835709,
0.007845447398722172,
-0.03651408106088638,
-0.02123206853866577,
-0.05546049401164055,
0.03809032216668129,
0.033292561769485474,
-0.06808818876743317,
-0.0479578822851181,
0.10903788357973099,
0.07938588410615921,
-0.09586019068956375,
-0.008368410170078278,
-0.030225809663534164,
-0.07559176534414291,
0.009440097957849503,
-0.06920564919710159,
-0.021959355100989342,
0.010272693820297718,
0.03374536335468292,
-0.053892284631729126,
-0.029973074793815613,
-0.015896501019597054,
0.0425674244761467,
-0.040782272815704346,
-0.11634847521781921,
-0.09489817917346954,
-0.060068629682064056,
0.047390613704919815,
-0.019881030544638634,
-0.058971308171749115,
0.014903085306286812,
0.07204292714595795,
-0.07532086223363876,
0.02468164637684822,
-0.06649713218212128,
-0.02799277938902378,
-0.01867782138288021,
0.0230655986815691,
0.012822086922824383,
0.07785184681415558,
-0.026298843324184418,
0.01328696496784687,
0.06520286947488785,
0.06073196604847908,
0.05338718742132187,
0.03969331458210945,
-0.07523614913225174,
-0.051399752497673035,
-0.052329063415527344,
0.0690748393535614,
0.04770340770483017,
0.05608964338898659,
0.03944913297891617,
-0.0009266350534744561,
0.03663048893213272,
-0.017481837421655655,
0.17722712457180023,
0.02238716371357441,
0.058025818318128586,
-0.028636226430535316,
0.010429632849991322,
0.07088524103164673,
0.035730477422475815,
-0.10386701673269272,
-3.6447563189767607e-8,
-0.07969343662261963,
0.03896414116024971,
0.030192380771040916,
0.07067444175481796,
0.032295167446136475,
-0.04497909918427467,
-0.030672814697027206,
0.05054636299610138,
0.026803022250533104,
-0.05064365640282631,
0.02789033204317093,
-0.005877705756574869,
-0.014634108170866966,
-0.049434807151556015,
-0.02552792616188526,
0.08958038687705994,
-0.03988475352525711,
-0.0078748082742095,
-0.025533292442560196,
-0.05497295409440994,
0.01906747929751873,
0.055396463721990585,
-0.053064294159412384,
0.016071515157818794,
0.06774943321943283,
-0.03869068622589111,
0.006449335720390081,
0.11365053802728653,
0.023922095075249672,
-0.04944708198308945,
-0.05176983401179314,
-0.0066510532051324844,
-0.02934327721595764,
0.02626311406493187,
0.08268250524997711,
0.10449124872684479,
0.046064700931310654,
0.021034540608525276,
-0.011033455841243267,
-0.05139562115073204,
0.01144801452755928,
0.017456544563174248,
-0.07021071761846542,
0.031869493424892426,
0.002051539486274123,
-0.022943200543522835,
-0.03237871080636978,
-0.06315978616476059,
-0.0056747435592114925,
-0.0649593323469162,
0.0033040232956409454,
-0.09088271856307983,
-0.02228112518787384,
0.03444209322333336,
-0.030588144436478615,
0.06639203429222107,
-0.028345821425318718,
0.04027518257498741,
-0.015342174097895622,
0.053361766040325165,
0.03899770602583885,
0.10445112735033035,
0.004403932485729456,
0.0018481684383004904
] | https://github.com/psunlpgroup/greater | 21 | 0 | 0 | 0 |
Efficient Causal Decision Making with One-sided Feedback | https://openreview.net/forum?id=UWdPsY7agk | [
"Jianing Chu",
"Shu Yang",
"Wenbin Lu",
"PULAK GHOSH"
] | Poster | We study a class of decision-making problems with one-sided feedback, where outcomes are only observable for specific actions. A typical example is bank loans, where the repayment status is known only if a loan is approved and remains undefined if rejected. In such scenarios, conventional approaches to causal decision evaluation and learning from observational data are not directly applicable. In this paper, we introduce a novel value function to evaluate decision rules that addresses the issue of undefined counterfactual outcomes. Without assuming no unmeasured confounders, we establish the identification of the value function using shadow variables. Furthermore, leveraging semiparametric theory, we derive the efficiency bound for the proposed value function and develop efficient methods for decision evaluation and learning. Numerical experiments and a real-world data application demonstrate the empirical performance of our proposed methods. | semiparametric efficiency, one-sided feedback, causal decision making | null | 12,956 | null | [
-0.0879334807395935,
0.017100196331739426,
-0.0005536630633287132,
0.021762331947684288,
0.026468271389603615,
-0.023360690101981163,
0.009365866892039776,
-0.059223346412181854,
0.028648488223552704,
0.07383361458778381,
-0.003198845311999321,
0.013728034682571888,
0.06829163432121277,
-0.015113542787730694,
-0.04995185509324074,
-0.0004103882238268852,
0.07350513339042664,
0.006245485041290522,
-0.059082455933094025,
-0.012932821176946163,
-0.02081621065735817,
-0.036321692168712616,
-0.020575346425175667,
0.00939916167408228,
0.09572403877973557,
-0.07042467594146729,
0.02635692059993744,
0.020527038723230362,
-0.03967824578285217,
-0.03857456520199776,
-0.02655092254281044,
-0.012438367120921612,
-0.03856467828154564,
-0.11666777729988098,
-0.051742225885391235,
0.03973497450351715,
0.02756972797214985,
-0.01063223835080862,
-0.05464359000325203,
0.02477904036641121,
0.009800763800740242,
-0.07625411450862885,
-0.04074572026729584,
-0.027420982718467712,
0.042122047394514084,
-0.030710959807038307,
0.03266231343150139,
0.0053176493383944035,
-0.014641892164945602,
0.049440156668424606,
0.007868589833378792,
0.049361806362867355,
-0.004806513898074627,
0.0012828942853957415,
0.0510079599916935,
0.055480390787124634,
0.024502329528331757,
-0.01629350334405899,
-0.06257274001836777,
0.007129993289709091,
-0.018319763243198395,
-0.07993240654468536,
-0.06121479347348213,
-0.038878895342350006,
0.028510665521025658,
0.05069161206483841,
-0.03288901597261429,
-0.003107630182057619,
-0.04645979776978493,
0.051687054336071014,
0.017791880294680595,
-0.004135167226195335,
-0.05079985409975052,
0.0001541954989079386,
0.03935025632381439,
0.005978716071695089,
0.043099965900182724,
0.09138219803571701,
-0.022715572267770767,
-0.0432220883667469,
0.014458627440035343,
-0.007752723526209593,
-0.03660661354660988,
-0.025936931371688843,
0.08135474473237991,
-0.021500099450349808,
0.05354390665888786,
0.12102948874235153,
0.0853343978524208,
0.08547475934028625,
0.050311099737882614,
0.08568203449249268,
0.043956924229860306,
-0.11745396256446838,
0.003105178941041231,
-0.072454534471035,
-0.03377090394496918,
-0.02942664548754692,
0.01785181276500225,
0.0631697028875351,
-0.015246585942804813,
0.07476001977920532,
0.004249362275004387,
0.001612881082110107,
0.056140363216400146,
-0.02085752598941326,
0.008833048865199089,
0.06400591135025024,
0.05970621854066849,
-0.09481368213891983,
-0.06675794720649719,
0.013063685968518257,
0.03103114292025566,
0.013832787051796913,
-0.02410920523107052,
-0.016992248594760895,
0.10105758160352707,
0.01314166933298111,
-0.043405309319496155,
-0.035198453813791275,
-0.004970753565430641,
-0.012765990570187569,
0.07453121989965439,
0.0038809801917523146,
-0.018350377678871155,
-0.0514109767973423,
-0.07912160456180573,
4.75443108362971e-34,
0.00893012061715126,
0.01805523969233036,
0.05470869690179825,
-0.10577445477247238,
-0.06469464302062988,
0.001992129720747471,
-0.04157843440771103,
0.015611485578119755,
-0.04535742104053497,
0.036610156297683716,
0.04800548404455185,
-0.008445559069514275,
-0.05592765659093857,
-0.03643422573804855,
0.057091377675533295,
0.03448323905467987,
-0.05812074989080429,
0.08429615199565887,
-0.012240801937878132,
-0.03868262469768524,
0.03178920969367027,
-0.07406911998987198,
0.008773958310484886,
-0.05349411442875862,
0.029686057940125465,
-0.08940664678812027,
0.01920822262763977,
-0.011170008219778538,
0.006217682734131813,
-0.015408421866595745,
0.012074174359440804,
-0.030512021854519844,
0.1497374325990677,
-0.026804471388459206,
-0.03795129433274269,
-0.017075909301638603,
-0.12152482569217682,
0.002836626023054123,
-0.05287047475576401,
0.024237824603915215,
-0.04760091379284859,
0.011053729802370071,
-0.04058363661170006,
-0.012065678834915161,
-0.1013752743601799,
-0.04807952791452408,
0.03678988292813301,
-0.04109678789973259,
-0.0187213197350502,
-0.061019785702228546,
0.028285937383770943,
0.025354374200105667,
0.02771826460957527,
-0.07454575598239899,
-0.07393073290586472,
-0.015600617974996567,
0.00483580119907856,
0.016078416258096695,
-0.13622984290122986,
-0.03477710112929344,
-0.01551664900034666,
-0.0069846478290855885,
-0.04607221484184265,
-0.004118847195059061,
-0.09275461733341217,
0.08148425072431564,
-0.0478217750787735,
-0.020614035427570343,
0.07165534049272537,
-0.05936992168426514,
-0.06154885143041611,
0.017057480290532112,
-0.01320868730545044,
-0.04983026534318924,
0.0470106303691864,
-0.02427772618830204,
-0.057418301701545715,
-0.018870020285248756,
0.056029755622148514,
-0.008419468067586422,
-0.010070828720927238,
-0.11063610017299652,
-0.004387734457850456,
0.036960020661354065,
0.05772896483540535,
0.04487739875912666,
0.03572594374418259,
-0.09531709551811218,
-0.045211054384708405,
-0.062170520424842834,
-0.1195276603102684,
0.007066446356475353,
-0.02399960160255432,
-0.009351354092359543,
0.11658379435539246,
-3.1742156808582396e-33,
-0.05139855667948723,
-0.04703530669212341,
-0.01017476711422205,
0.025423824787139893,
-0.020150071009993553,
0.06694401800632477,
-0.07272716611623764,
-0.047364357858896255,
0.044658735394477844,
-0.014861532486975193,
0.008774644695222378,
0.006927158683538437,
-0.005894924513995647,
0.027946792542934418,
0.02059202268719673,
0.009812888689339161,
-0.001503947889432311,
-0.04531189054250717,
0.04568077251315117,
0.06445278227329254,
-0.009538550861179829,
0.08750658482313156,
0.06599992513656616,
0.04033677652478218,
-0.06417326629161835,
0.007660721894353628,
0.01225556805729866,
0.08051479607820511,
-0.004705552011728287,
0.01921265386044979,
-0.06093103438615799,
0.026525380089879036,
0.02247510850429535,
0.011643652804195881,
-0.030149148777127266,
0.007498739752918482,
0.08798671513795853,
-0.015778465196490288,
-0.051681242883205414,
0.057721298187971115,
0.05731000751256943,
-0.02154158428311348,
-0.09196420758962631,
-0.08165761828422546,
-0.009094742126762867,
-0.0407746322453022,
0.018537703901529312,
0.09344712644815445,
0.002033253200352192,
-0.029241928830742836,
0.008045142516493797,
0.06965751200914383,
-0.03459721431136131,
0.1457725167274475,
-0.0685139000415802,
-0.06850222498178482,
0.06849417090415955,
-0.05143987759947777,
0.12939569354057312,
0.12441090494394302,
-0.07647157460451126,
0.08278943598270416,
0.013652435503900051,
0.05162786692380905,
0.01438107155263424,
-0.03991568088531494,
0.04480552300810814,
0.012629899196326733,
0.08465535193681717,
-0.053413331508636475,
0.04350073263049126,
0.04765934497117996,
0.0007107730489224195,
0.005442833993583918,
0.07297000288963318,
0.09155265241861343,
-0.020772606134414673,
-0.043802693486213684,
-0.07302633672952652,
0.05919690802693367,
0.059217698872089386,
0.00684716971591115,
0.12443208694458008,
-0.016239453107118607,
-0.036678917706012726,
0.024211855605244637,
0.0881606712937355,
0.027986103668808937,
0.03834451735019684,
-0.03411761298775673,
-0.028274409472942352,
0.04223598539829254,
0.013835386373102665,
0.03892860189080238,
-0.028980081900954247,
-4.234118833323919e-8,
0.011219467036426067,
-0.045501019805669785,
0.06753664463758469,
0.027775652706623077,
0.012634129263460636,
-0.02116510644555092,
-0.044897548854351044,
0.015116159804165363,
-0.03717222809791565,
-0.06652609258890152,
-0.007958018220961094,
0.046887289732694626,
-0.020239630714058876,
-0.050094593316316605,
-0.03691596910357475,
-0.012412009760737419,
-0.0024855013471096754,
0.025330310687422752,
-0.027368959039449692,
0.06457972526550293,
0.06642653048038483,
-0.034848593175411224,
-0.017666326835751534,
-0.04710450395941734,
0.07842198759317398,
-0.09070678800344467,
-0.017000142484903336,
0.09408808499574661,
0.0726536214351654,
-0.0017870146548375487,
0.04347659647464752,
0.02754211612045765,
0.04736388474702835,
0.011297931894659996,
0.09581001102924347,
0.01720757782459259,
0.08896201103925705,
-0.026138145476579666,
-0.027366502210497856,
-0.028507666662335396,
-0.0038221837021410465,
0.04912557452917099,
-0.10073427855968475,
0.001318924012593925,
0.06323738396167755,
0.033518850803375244,
-0.014192781411111355,
-0.10689739137887955,
-0.013500208035111427,
-0.05696243792772293,
0.047265395522117615,
-0.01832280494272709,
0.07291852682828903,
-0.01467135176062584,
0.09242001920938492,
-0.023043416440486908,
-0.01742440089583397,
-0.04684137552976608,
0.0012271328596398234,
0.0073798466473817825,
0.11760840564966202,
-0.004913279786705971,
-0.014721805229783058,
0.011727558448910713
] | 0 | 0 | 0 | 0 |
|
LANTERN: Accelerating Visual Autoregressive Models with Relaxed Speculative Decoding | https://openreview.net/forum?id=98d7DLMGdt | [
"Doohyuk Jang",
"Sihwan Park",
"June Yong Yang",
"Yeonsung Jung",
"Jihun Yun",
"Souvik Kundu",
"Sung-Yub Kim",
"Eunho Yang"
] | Poster | Auto-Regressive (AR) models have recently gained prominence in image generation, often matching or even surpassing the performance of diffusion models. However, one major limitation of AR models is their sequential nature, which processes tokens one at a time, slowing down generation compared to models like GANs or diffusion-based methods that operate more efficiently. While speculative decoding has proven effective for accelerating LLMs by generating multiple tokens in a single forward, its application in visual AR models remains largely unexplored. In this work, we identify a challenge in this setting, which we term \textit{token selection ambiguity}, wherein visual AR models frequently assign uniformly low probabilities to tokens, hampering the performance of speculative decoding. To overcome this challenge, we propose a relaxed acceptance condition referred to as LANTERN that leverages the interchangeability of tokens in latent space. This relaxation restores the effectiveness of speculative decoding in visual AR models by enabling more flexible use of candidate tokens that would otherwise be prematurely rejected. Furthermore, by incorporating a total variation distance bound, we ensure that these speed gains are achieved without significantly compromising image quality or semantic coherence. Experimental results demonstrate the efficacy of our method in providing a substantial speed-up over speculative decoding. In specific, compared to a na\"ive application of the state-of-the-art speculative decoding, LANTERN increases speed-ups by $\mathbf{1.75}\times$ and $\mathbf{1.82}\times$, as compared to greedy decoding and random sampling, respectively, when applied to LlamaGen, a contemporary visual AR model. The code is publicly available at \url{https://github.com/jadohu/LANTERN}. | Speculative decoding, Visual Autoregressive Models | null | 12,935 | 2410.03355 | [
-0.029785389080643654,
-0.1072721779346466,
0.04373091831803322,
0.06814004480838776,
0.1331751048564911,
-0.05051710456609726,
-0.03286377713084221,
-0.09727462381124496,
0.11323990672826767,
-0.07167817652225494,
-0.019946500658988953,
-0.09566745162010193,
0.007651688531041145,
0.01615508645772934,
-0.010947082191705704,
-0.025886211544275284,
0.06867216527462006,
0.08412308245897293,
-0.039247605949640274,
-0.030228564515709877,
0.014529570937156677,
-0.022832583636045456,
-0.0074291592463850975,
-0.04308763146400452,
0.11794460564851761,
0.0194806270301342,
0.009498462080955505,
-0.022609321400523186,
0.02684801258146763,
-0.025328248739242554,
0.02577322907745838,
0.0334407240152359,
-0.079898402094841,
0.04337587580084801,
-0.06261078268289566,
0.0717604011297226,
-0.03842811658978462,
0.0220261849462986,
0.0666310042142868,
-0.014233814552426338,
-0.0030446622986346483,
0.017762616276741028,
0.023937828838825226,
-0.018123304471373558,
0.06543654203414917,
-0.002098234137520194,
0.01812734641134739,
-0.04500921443104744,
-0.04666348174214363,
-0.03788206726312637,
-0.043258216232061386,
0.0063025010749697685,
-0.06953197717666626,
-0.005643794313073158,
-0.03161293640732765,
0.014815988019108772,
0.05423564463853836,
-0.04304179176688194,
0.04710119217634201,
-0.009368588216602802,
-0.03390047699213028,
-0.043605998158454895,
-0.015361107885837555,
-0.00232295086607337,
0.014424419961869717,
-0.05018145591020584,
0.04040130600333214,
-0.010852831415832043,
0.021312935277819633,
0.06198485195636749,
-0.03407920524477959,
0.08123570680618286,
-0.04473290592432022,
0.0030624298378825188,
-0.005254190415143967,
0.08709560334682465,
0.07431572675704956,
-0.048919398337602615,
0.07926695048809052,
-0.1279047429561615,
0.06027299165725708,
-0.015615364536643028,
0.06593453139066696,
0.019899506121873856,
0.08338196575641632,
0.030371423810720444,
-0.03638243302702904,
0.04236596077680588,
0.052027832716703415,
0.014746645465493202,
-0.08818992972373962,
-0.08246617019176483,
0.024121275171637535,
0.06327225267887115,
0.08256405591964722,
0.003305061487480998,
0.003554751630872488,
-0.078884556889534,
-0.022836724296212196,
0.08607665449380875,
0.03514297306537628,
0.020637620240449905,
0.0333995446562767,
-0.009837882593274117,
0.07118169218301773,
-0.07234150916337967,
0.08723487704992294,
0.03612864390015602,
0.014914514496922493,
-0.03363141790032387,
0.053014107048511505,
0.032556820660829544,
0.017612924799323082,
-0.008503192104399204,
0.0015200695488601923,
-0.04538891091942787,
-0.005109555087983608,
-0.05145176127552986,
0.03173317760229111,
0.04497530311346054,
-0.004191263113170862,
0.020961547270417213,
-0.033690162003040314,
0.006334687117487192,
0.029686439782381058,
-0.07390847057104111,
-0.0024989487137645483,
1.3629536491902832e-33,
-0.030013011768460274,
0.006660378072410822,
0.031308818608522415,
0.038449548184871674,
0.049679823219776154,
0.02164047211408615,
-0.04841163009405136,
-0.0030041125137358904,
-0.00697354506701231,
-0.04580342397093773,
-0.02917413040995598,
-0.021163828670978546,
-0.07017619162797928,
0.118264339864254,
0.026062965393066406,
-0.0180362481623888,
-0.03433367609977722,
0.029545528814196587,
-0.034572798758745193,
0.012183157727122307,
-0.013996550813317299,
-0.021527668461203575,
-0.09482984244823456,
-0.04970362037420273,
-0.03874173387885094,
0.018873197957873344,
0.07204999029636383,
-0.08930322527885437,
-0.01091728638857603,
0.03002425655722618,
-0.11964051425457001,
0.028338242322206497,
0.07023460417985916,
0.025555212050676346,
0.05291671305894852,
-0.016790935769677162,
-0.0012515437556430697,
-0.04332711175084114,
-0.010853706859052181,
-0.011342261917889118,
-0.03457643836736679,
0.07638950645923615,
-0.07775399833917618,
-0.0005432140314951539,
-0.09199025481939316,
-0.021908123046159744,
0.02269330620765686,
0.06848713755607605,
-0.12738610804080963,
0.0300477035343647,
0.07156597822904587,
0.04268889129161835,
-0.08513916283845901,
-0.0684572234749794,
-0.05237000808119774,
-0.11048197001218796,
0.0016968080308288336,
-0.06267393380403519,
0.04926227033138275,
-0.012280681170523167,
-0.035763632506132126,
0.04037580266594887,
0.007441963534802198,
0.0698663666844368,
-0.02840128168463707,
0.04202338308095932,
0.005498078186064959,
0.02433898113667965,
-0.03963315859436989,
0.0650053471326828,
-0.07929586619138718,
0.012402839958667755,
-0.05916550010442734,
-0.1416761875152588,
0.02413606457412243,
-0.020533697679638863,
0.09942063689231873,
0.036404937505722046,
-0.0010693403892219067,
0.03551385551691055,
-0.06572852283716202,
-0.053824376314878464,
-0.003979947417974472,
-0.04582984372973442,
-0.05584060028195381,
-0.05333314090967178,
0.019574573263525963,
-0.09088091552257538,
-0.019255200400948524,
-0.10225318372249603,
-0.011179646477103233,
-0.01890699937939644,
-0.018582232296466827,
0.016710665076971054,
0.015617809258401394,
-2.7269540393657752e-33,
0.020527729764580727,
0.04068738594651222,
-0.015393040142953396,
0.031161576509475708,
-0.01730264350771904,
-0.06276886910200119,
0.02716980315744877,
-0.01570950262248516,
0.021646009758114815,
-0.12992854416370392,
-0.0914720743894577,
0.019441628828644753,
-0.016440998762845993,
-0.021697482094168663,
0.02111196331679821,
-0.0745977908372879,
0.01522612851113081,
0.02087966538965702,
-0.017958348616957664,
0.08195598423480988,
0.0486728735268116,
0.0349179208278656,
-0.08631134033203125,
0.02691708318889141,
0.0009925616905093193,
0.09957519918680191,
-0.03319887816905975,
0.049967486411333084,
-0.056309014558792114,
-0.07931560277938843,
0.043293390423059464,
0.0017711432883515954,
0.01996013894677162,
-0.0011279472382739186,
-0.06799835711717606,
0.0743723213672638,
0.07127901166677475,
0.037595488131046295,
-0.05436256155371666,
0.11276090145111084,
0.07597228139638901,
-0.04101061448454857,
-0.04282725229859352,
0.05754123628139496,
-0.045978765934705734,
-0.033742353320121765,
0.04952426627278328,
-0.025246620178222656,
0.04825577512383461,
0.04408528655767441,
0.01640435680747032,
0.008615761995315552,
-0.08369866013526917,
0.03409880772233009,
-0.0761546641588211,
-0.0769050195813179,
-0.10010573267936707,
0.008227006532251835,
0.07409113645553589,
0.027694402262568474,
-0.03406636789441109,
-0.025089411064982414,
0.02054191567003727,
-0.08110301941633224,
-0.004033563658595085,
-0.08076657354831696,
-0.03373151272535324,
-0.007477660197764635,
0.01650628075003624,
0.0037575247697532177,
0.06550485640764236,
-0.05039328709244728,
-0.023069385439157486,
0.06356916576623917,
0.07171261310577393,
-0.02454477734863758,
0.11430328339338303,
-0.0038422616198658943,
0.06160690262913704,
-0.03035097010433674,
-0.007227080874145031,
-0.00620754761621356,
0.0788799598813057,
0.06072212755680084,
-0.002717181108891964,
0.08207355439662933,
-0.012115621007978916,
0.009744875133037567,
0.002341077197343111,
0.01672298274934292,
-0.01172301359474659,
-0.004842228256165981,
0.06324668973684311,
0.03996574133634567,
-0.005409317556768656,
-4.014865595536321e-8,
-0.06727445125579834,
0.0663883239030838,
0.0017093666829168797,
-0.024339087307453156,
0.018769126385450363,
-0.030640972778201103,
0.0031146707478910685,
0.03287125378847122,
0.014560227282345295,
-0.06441082805395126,
0.08539832383394241,
-0.040188029408454895,
-0.033074717968702316,
-0.08109032362699509,
-0.03134191036224365,
0.04270985350012779,
-0.033068858087062836,
-0.0469280369579792,
-0.024544183164834976,
0.02399969846010208,
-0.06591839343309402,
0.019798772409558296,
-0.0038433584850281477,
-0.07795096933841705,
0.0196138396859169,
-0.011791247874498367,
-0.03287007659673691,
0.027903931215405464,
0.07316865026950836,
0.009955724701285362,
0.010606399737298489,
0.07337875664234161,
0.03333397954702377,
-0.016733385622501373,
-0.05677922070026398,
0.03904123976826668,
0.02156335860490799,
-0.035048868507146835,
-0.007345276419073343,
-0.008429309353232384,
0.02791925147175789,
0.018366850912570953,
-0.022640472277998924,
-0.005058870185166597,
-0.028576618060469627,
0.03809752315282822,
0.09187944233417511,
-0.10905308276414871,
0.015557495877146721,
0.007680938579142094,
-0.0003078200970776379,
-0.07766873389482498,
0.00697162514552474,
0.12076594680547714,
0.06023712828755379,
-0.02866082824766636,
0.05377870798110962,
-0.03602194786071777,
0.04723231494426727,
0.07579802721738815,
0.07916218042373657,
-0.0086019616574049,
-0.07153963297605515,
0.028409238904714584
] | 0 | 3 | 0 | 0 |
|
LICO: Large Language Models for In-Context Molecular Optimization | https://openreview.net/forum?id=yu1vqQqKkx | [
"Tung Nguyen",
"Aditya Grover"
] | Poster | Optimizing black-box functions is a fundamental problem in science and engineering. To solve this problem, many approaches learn a surrogate function that estimates the underlying objective from limited historical evaluations. Large Language Models (LLMs), with their strong pattern-matching capabilities via pretraining on vast amounts of data, stand out as a potential candidate for surrogate modeling. However, directly prompting a pretrained language model to produce predictions is not feasible in many scientific domains due to the scarcity of domain-specific data in the pretraining corpora and the challenges of articulating complex problems in natural language. In this work, we introduce LICO, a general-purpose model that extends arbitrary base LLMs for black-box optimization, with a particular application to the molecular domain. To achieve this, we equip the language model with a separate embedding layer and prediction layer, and train the model to perform in-context predictions on a diverse set of functions defined over the domain. Once trained, LICO can generalize to unseen molecule properties simply via in-context prompting. LICO performs competitively on PMO, a challenging molecular optimization benchmark comprising 23 objective functions, and achieves state-of-the-art performance on its low-budget version PMO-1K. | large language models, molecular optimization, black-box optimization, foundation models, in-context learning | We introduce LICO, a general-purpose model that extends arbitrary base LLMs for black-box optimization, with a particular application to the molecular domain. | 12,932 | 2406.18851 | [
-0.06921843439340591,
-0.08215147256851196,
0.01887047104537487,
0.01708689145743847,
0.03931925445795059,
-0.01996692456305027,
-0.03392873331904411,
0.0038695167750120163,
-0.007937216199934483,
-0.049584705382585526,
-0.05529498681426048,
-0.12738800048828125,
0.015333250164985657,
0.05380949005484581,
-0.027786606922745705,
0.07309537380933762,
0.06267533451318741,
0.10479115694761276,
-0.10440288484096527,
-0.038569290190935135,
0.07388795167207718,
0.070840023458004,
0.04945050925016403,
-0.04043544828891754,
0.006141358520835638,
-0.003932330757379532,
0.0016972869634628296,
0.0054388330318033695,
0.024169059470295906,
0.006620212458074093,
0.07586593925952911,
0.025541573762893677,
0.035353660583496094,
0.024429408833384514,
0.002369001042097807,
0.047693297266960144,
-0.17423152923583984,
-0.0006013938109390438,
0.005888241808861494,
-0.042965613305568695,
-0.008165735751390457,
-0.03309547156095505,
0.019060879945755005,
0.044312383979558945,
0.049726784229278564,
-0.04251415655016899,
0.04012710973620415,
-0.0014503522543236613,
0.02125629037618637,
-0.023114565759897232,
-0.044771697372198105,
-0.06970521807670593,
-0.07808203250169754,
0.021848473697900772,
0.00587774533778429,
0.03273606300354004,
0.0017267680959776044,
-0.02463575266301632,
0.001172184944152832,
-0.07203298062086105,
-0.024791866540908813,
-0.09670107066631317,
-0.03174690902233124,
0.00829216931015253,
-0.006286639720201492,
-0.016794580966234207,
0.08824142813682556,
0.10188494622707367,
-0.013467129319906235,
0.08710859715938568,
-0.025331638753414154,
0.09305328130722046,
-0.009088066406548023,
0.056371692568063736,
0.040018174797296524,
0.11587442457675934,
0.08037474006414413,
-0.023816967383027077,
0.04925457015633583,
-0.05958898738026619,
0.026349004358053207,
0.03269563615322113,
0.06611446291208267,
0.04134877026081085,
0.12247178703546524,
0.009912004694342613,
0.034499332308769226,
0.05948358029127121,
0.05176554620265961,
0.02718336321413517,
-0.0337597094476223,
-0.09107720851898193,
0.05073815584182739,
-0.043443430215120316,
-0.04755612090229988,
0.044414740055799484,
0.04408181831240654,
-0.08303919434547424,
0.007404687814414501,
0.058023545891046524,
0.018209602683782578,
0.06991292536258698,
0.023963728919625282,
-0.04946892336010933,
-0.05932486429810524,
-0.05018267780542374,
0.017223523929715157,
0.07867148518562317,
0.024193597957491875,
-0.021118180826306343,
0.029605785384774208,
0.07825187593698502,
0.028465501964092255,
0.003438027109950781,
0.0423823781311512,
0.0043741026893258095,
0.0771993100643158,
-0.0823131650686264,
0.044366106390953064,
0.1030663475394249,
-0.0912104919552803,
0.014970578253269196,
-0.03887735679745674,
0.004885702393949032,
-0.03342761471867561,
0.029420729726552963,
-0.07162975519895554,
1.268069352313091e-34,
0.06000466272234917,
0.02323540672659874,
0.045822832733392715,
0.036524832248687744,
0.05870649591088295,
-0.03644883632659912,
0.008773998357355595,
0.04108916223049164,
-0.13945171236991882,
0.025852741673588753,
-0.024821851402521133,
0.01972908526659012,
-0.058947306126356125,
0.020968841388821602,
0.01876187138259411,
0.010179397650063038,
-0.017441000789403915,
0.0453246608376503,
-0.03287598490715027,
0.0009495129925198853,
-0.02717270888388157,
-0.01117867324501276,
-0.018113771453499794,
-0.030444080010056496,
0.022979864850640297,
0.025453859940171242,
0.09810830652713776,
-0.07645036280155182,
-0.04103895276784897,
-0.004948846995830536,
-0.1300603449344635,
-0.014703251421451569,
-0.006642433349043131,
0.05929942801594734,
0.07607321441173553,
-0.01691000536084175,
0.01348030474036932,
-0.053782179951667786,
-0.006608540657907724,
-0.002055823104456067,
-0.0017687688814476132,
0.036037627607584,
0.02572438307106495,
0.0012962758773937821,
-0.0778169259428978,
-0.07173750549554825,
0.010831374675035477,
0.024423949420452118,
-0.02820592373609543,
-0.03546290844678879,
0.030132753774523735,
0.020492790266871452,
-0.013782555237412453,
-0.0245343167334795,
-0.018549485132098198,
0.01240859366953373,
-0.02406524121761322,
-0.001751711592078209,
0.03656560927629471,
0.053021159023046494,
-0.003498189151287079,
0.055278047919273376,
-0.0058942134492099285,
0.06855570524930954,
0.007650602608919144,
-0.016675665974617004,
-0.10586220026016235,
-0.021715868264436722,
0.04377337172627449,
0.057913195341825485,
-0.007523923646658659,
0.037476930767297745,
0.005343675147742033,
-0.05463845282793045,
0.10157565027475357,
0.036606162786483765,
0.037971410900354385,
-0.10333073884248734,
0.038028087466955185,
0.03481365740299225,
0.0758628249168396,
0.021988937631249428,
-0.02503146417438984,
-0.05416398495435715,
-0.07429618388414383,
-0.12083525955677032,
0.015168766491115093,
-0.06487981975078583,
-0.08902806043624878,
-0.08276229351758957,
-0.055538080632686615,
-0.10168465971946716,
-0.022047853097319603,
-0.04958651214838028,
0.005303371697664261,
-7.279431060926889e-34,
-0.006351239047944546,
-0.022558603435754776,
0.0319875068962574,
0.0193120576441288,
-0.005256790667772293,
-0.05238467454910278,
-0.026730859652161598,
-0.04537608101963997,
0.0004694037197623402,
-0.023086506873369217,
0.04920106753706932,
-0.06163293123245239,
0.06553778052330017,
-0.04912496358156204,
-0.002884655725210905,
-0.016547519713640213,
-0.05291294306516647,
0.008668459951877594,
0.0627826675772667,
0.08309066295623779,
-0.02380528301000595,
0.09820960462093353,
-0.05393737927079201,
0.07002110779285431,
-0.07296469807624817,
0.017252644523978233,
-0.017583277076482773,
0.03981553018093109,
-0.008594720624387264,
0.014981796033680439,
-0.05036298558115959,
0.019731057807803154,
-0.095351941883564,
0.009039600379765034,
-0.11101052165031433,
0.08953608572483063,
0.05061237886548042,
-0.008907074108719826,
0.001202064217068255,
0.025641681626439095,
0.0914597138762474,
-0.011234644800424576,
-0.1055145263671875,
-0.009399618953466415,
0.043762996792793274,
-0.007418068591505289,
-0.09031020849943161,
-0.025890693068504333,
0.10376957803964615,
-0.03597542271018028,
-0.03313653543591499,
0.008512909524142742,
-0.01619865745306015,
-0.036205366253852844,
-0.03350388631224632,
-0.0971606895327568,
-0.07652153074741364,
-0.06288918852806091,
0.08417496085166931,
-0.016321832314133644,
-0.04320213198661804,
-0.015121019445359707,
0.023266740143299103,
-0.00940615776926279,
-0.043377432972192764,
-0.08487755060195923,
-0.014209814369678497,
0.055586256086826324,
0.012111685238778591,
-0.06333763152360916,
-0.0034738744143396616,
-0.017186874523758888,
0.026302022859454155,
0.05325429141521454,
-0.006567566655576229,
-0.022466016933321953,
0.03597153350710869,
-0.09005813300609589,
-0.022880079224705696,
-0.003938327077776194,
0.08624503761529922,
0.03519061952829361,
0.03286353126168251,
0.04251675680279732,
0.07766768336296082,
0.03692122548818588,
0.08132899552583694,
0.11012759059667587,
0.039920128881931305,
-0.04116907715797424,
-0.06152210384607315,
0.02178107015788555,
-0.019034944474697113,
0.10653314739465714,
-0.03401060402393341,
-3.895735289916047e-8,
0.01177757978439331,
0.02187175676226616,
0.042718879878520966,
-0.01306324265897274,
-0.007129468955099583,
0.00783673394471407,
0.004065206740051508,
0.054550863802433014,
0.01500026136636734,
-0.025284547358751297,
0.05921676382422447,
0.041398245841264725,
-0.058970313519239426,
-0.04919622093439102,
-0.03571951761841774,
0.10362065583467484,
-0.03467930108308792,
-0.017215417698025703,
0.013813245110213757,
-0.02733088657259941,
0.0007679251721128821,
0.08676977455615997,
0.024804625660181046,
0.0009710410377010703,
0.08188755065202713,
-0.1175922080874443,
0.01109290961176157,
0.030797580257058144,
0.07681789994239807,
-0.013889303430914879,
-0.08050408214330673,
-0.00729316147044301,
0.003327864222228527,
0.006726131774485111,
-0.00978159997612238,
0.05387359485030174,
0.027690062299370766,
-0.03987143188714981,
-0.05214325711131096,
0.023122547194361687,
0.015195640735328197,
0.08674929291009903,
-0.09716172516345978,
-0.03258068859577179,
-0.0022665653377771378,
-0.009345818310976028,
-0.020395353436470032,
-0.012445163913071156,
0.04101858288049698,
-0.03413417190313339,
0.0697883889079094,
0.015046950429677963,
-0.054785579442977905,
0.030283695086836815,
0.027814330533146858,
0.03237949684262276,
-0.059057071805000305,
-0.043109405785799026,
0.029383068904280663,
0.004465353675186634,
-0.02228926680982113,
-0.02549203485250473,
-0.03935427591204643,
0.007036864757537842
] | 0 | 0 | 0 | 0 |
|
Forking Paths in Neural Text Generation | https://openreview.net/forum?id=8RCmNLeeXx | [
"Eric J Bigelow",
"Ari Holtzman",
"Hidenori Tanaka",
"Tomer Ullman"
] | Poster | Estimating uncertainty in Large Language Models (LLMs) is important for properly evaluating LLMs, and ensuring safety for users. However, prior approaches to uncertainty estimation focus on the final answer in generated text, ignoring intermediate steps that might dramatically impact the outcome. We hypothesize that there exist key forking tokens, such that re-sampling the system at those specific tokens, but not others, leads to very different outcomes. To test this empirically, we develop a novel approach to representing uncertainty dynamics across individual tokens of text generation, and applying statistical models to test our hypothesis. Our approach is highly flexible: it can be applied to any dataset and any LLM, without fine tuning or accessing model weights. We use our method to analyze LLM responses on 7 different tasks across 4 domains, spanning a wide range of typical use cases. We find many examples of forking tokens, including surprising ones such as a space character instead of a colon, suggesting that LLMs are often just a single token away from saying something very different. | Large Language Models, Uncertainty Estimation, Interpretability | null | 12,926 | 2412.07961 | [
-0.04740535095334053,
-0.07901635766029358,
0.09013693779706955,
0.07440869510173798,
0.01856350153684616,
-0.04698340967297554,
0.0007475007441826165,
0.03565060347318649,
0.11143945902585983,
-0.09300778061151505,
-0.041243892163038254,
-0.0924849584698677,
0.09555090218782425,
-0.049648262560367584,
-0.07399298995733261,
0.0030541231390088797,
0.051926158368587494,
0.06831470876932144,
-0.09685252606868744,
-0.034120965749025345,
0.05119633674621582,
0.07590021193027496,
0.03747081384062767,
0.012867150828242302,
0.05208905413746834,
-0.008172035217285156,
-0.03585072234272957,
0.018788695335388184,
0.023449117317795753,
0.010484328493475914,
0.01724361628293991,
0.010667725466191769,
-0.0897698625922203,
-0.00010977294732583687,
0.05526599660515785,
0.03410874307155609,
-0.061370767652988434,
0.005175819154828787,
0.11878418922424316,
-0.07697232067584991,
-0.037454817444086075,
-0.019352663308382034,
0.0663776844739914,
0.036827556788921356,
0.11642208695411682,
-0.028210092335939407,
-0.014996884390711784,
-0.024137532338500023,
-0.10845114290714264,
0.05598282814025879,
-0.04352075979113579,
-0.04134444147348404,
0.03730683773756027,
-0.03471508249640465,
-0.07406769692897797,
0.0035645635798573494,
0.023386744782328606,
0.07656911015510559,
0.0034546437673270702,
-0.06184801086783409,
-0.07277362048625946,
-0.10839482396841049,
-0.051767006516456604,
-0.017503196373581886,
0.019536297768354416,
-0.01096060499548912,
0.008188433945178986,
0.09491456300020218,
-0.03396608680486679,
0.0931675136089325,
0.0118831442669034,
0.06588946282863617,
-0.06782008707523346,
0.11606545001268387,
-0.033601872622966766,
0.09185217320919037,
0.06206836923956871,
-0.030521532520651817,
0.016924381256103516,
-0.034286223351955414,
-0.011774101294577122,
-0.033559370785951614,
0.08983254432678223,
0.027566177770495415,
0.08948672562837601,
0.034695401787757874,
0.02460782788693905,
0.06803612411022186,
-0.013388696126639843,
0.0417095422744751,
-0.08964872360229492,
-0.08313938230276108,
0.14786742627620697,
0.03144730255007744,
-0.03665183112025261,
0.07737564295530319,
-0.03549492359161377,
-0.0958767905831337,
0.01998690515756607,
0.08291573822498322,
-0.012055988423526287,
0.10794442147016525,
0.05079690366983414,
-0.08556722104549408,
-0.013753600418567657,
-0.002889568917453289,
0.00425717094913125,
-0.037547893822193146,
-0.017661157995462418,
-0.07691284269094467,
0.041864730417728424,
0.03282480686903,
-0.009855262003839016,
0.007106369361281395,
0.06533819437026978,
-0.08506130427122116,
-0.0017943857237696648,
-0.054077841341495514,
0.014332490041851997,
0.08379211276769638,
-0.04683533310890198,
0.044498324394226074,
0.03654571995139122,
0.04129013419151306,
-0.04514482617378235,
-0.023449787870049477,
0.02478201873600483,
8.796494223561882e-34,
0.05911695584654808,
-0.012117772363126278,
-0.019624436274170876,
0.058036159723997116,
0.06316914409399033,
0.03625490888953209,
-0.023296421393752098,
0.013619893230497837,
-0.031034374609589577,
-0.0381932407617569,
-0.020015593618154526,
-0.021176369860768318,
-0.02673063427209854,
-0.028998101130127907,
0.07207319140434265,
-0.009722462855279446,
-0.034476138651371,
0.020023729652166367,
-0.050159964710474014,
0.02330150082707405,
-0.018773656338453293,
-0.03025796078145504,
-0.0214142594486475,
-0.03725038096308708,
0.04614090919494629,
0.018789028748869896,
0.029119303449988365,
-0.015072389505803585,
-0.0544108971953392,
0.00735026178881526,
-0.1429213136434555,
-0.032165996730327606,
0.03529613837599754,
0.009238369762897491,
0.03536316007375717,
-0.04085586220026016,
-0.022037772461771965,
-0.02161257714033127,
-0.04325954243540764,
0.010995804332196712,
0.01373314019292593,
0.015769504010677338,
0.05865294113755226,
0.06135622039437294,
-0.06332261860370636,
-0.018388640135526657,
-0.025122446939349174,
-0.04698353633284569,
-0.027842115610837936,
-0.0722714439034462,
0.010960182175040245,
0.07476279884576797,
0.04119512066245079,
0.004825830925256014,
0.020477181300520897,
-0.02025248296558857,
-0.06759563088417053,
-0.01176851149648428,
0.015350586734712124,
0.01652282476425171,
-0.0038913851603865623,
0.04062221199274063,
-0.005316116847097874,
0.0571279413998127,
0.08106095343828201,
0.09108809381723404,
-0.034133221954107285,
0.01800033636391163,
0.009560560807585716,
-0.02775505557656288,
-0.018490450456738472,
-0.03051951341331005,
-0.07733455300331116,
-0.04146795719861984,
0.0035129948519170284,
0.02191527560353279,
0.03290780261158943,
0.000813405669759959,
-0.012883706949651241,
-0.0258869007229805,
0.062208641320466995,
-0.06972438097000122,
0.0016847011866047978,
-0.029756035655736923,
-0.023228416219353676,
-0.03986741974949837,
0.06368353217840195,
-0.07839178293943405,
-0.0677245482802391,
0.020115774124860764,
0.006790727376937866,
-0.06160971149802208,
0.012499582022428513,
0.03645600378513336,
-0.015254802070558071,
-2.0108700641517005e-33,
-0.09638027101755142,
-0.013168985955417156,
-0.004364644177258015,
0.07166962325572968,
-0.058683130890131,
-0.04047449678182602,
-0.008134137839078903,
0.00959189422428608,
-0.039490070194005966,
-0.09806878119707108,
-0.10462145507335663,
0.008356014266610146,
0.02993512898683548,
0.0018439703853800893,
0.007604931015521288,
-0.050903335213661194,
0.0631927028298378,
0.029009832069277763,
0.053503282368183136,
0.06731454282999039,
0.0016020727343857288,
0.025065993890166283,
-0.12580110132694244,
0.035872265696525574,
0.0008814425091259181,
0.061680082231760025,
-0.0284526739269495,
0.009216011501848698,
-0.04345320537686348,
-0.03394299000501633,
-0.041616491973400116,
0.017005300149321556,
0.028242018073797226,
-0.041311152279376984,
-0.051061663776636124,
0.01510627195239067,
0.05306827649474144,
0.004442234989255667,
-0.004263182170689106,
0.03257950767874718,
0.046197809278964996,
-0.04012360796332359,
-0.04493965581059456,
-0.003230748698115349,
-0.0349637046456337,
0.019879033789038658,
-0.06558814644813538,
0.0073328022845089436,
-0.02903137169778347,
-0.006074079778045416,
0.011557500809431076,
0.010531575419008732,
-0.09929165244102478,
0.02484462969005108,
-0.03586726635694504,
-0.11565055698156357,
-0.025027276948094368,
-0.09502041339874268,
-0.008830898441374302,
0.027935227379202843,
-0.10207656025886536,
0.014064958319067955,
0.019864631816744804,
-0.05513407662510872,
0.03121350146830082,
-0.043946098536252975,
-0.05367244780063629,
-0.02709057182073593,
-0.0008798282942734659,
-0.024105791002511978,
0.11052331328392029,
-0.07846648246049881,
-0.028141461312770844,
0.016151847317814827,
0.028090041130781174,
-0.03341582417488098,
0.06771238893270493,
-0.071586474776268,
0.005587122868746519,
-0.048591360449790955,
0.032929595559835434,
0.023519432172179222,
0.05542254075407982,
0.06448686867952347,
0.04200277477502823,
0.05178999528288841,
0.02313326857984066,
0.06976344436407089,
0.029543491080403328,
0.06668595224618912,
-0.005102758761495352,
-0.004972292110323906,
-0.04167282208800316,
0.028159743174910545,
-0.058209292590618134,
-4.3964011098296396e-8,
0.013951941393315792,
0.05551426112651825,
-0.03233537822961807,
0.11165168881416321,
0.015396400354802608,
0.017365556210279465,
-0.0312013141810894,
-0.01419864408671856,
-0.0349501296877861,
-0.053375061601400375,
0.05042315647006035,
0.02842770144343376,
-0.07862021774053574,
-0.058609314262866974,
-0.03971555456519127,
0.07313255220651627,
-0.011508346535265446,
0.08069600909948349,
-0.024296509101986885,
-0.015865756198763847,
0.0521235354244709,
0.09551004320383072,
-0.04084281250834465,
-0.030715223401784897,
0.045972272753715515,
-0.039878323674201965,
-0.042024943977594376,
0.08574198931455612,
0.06813464313745499,
-0.012278907932341099,
0.049254804849624634,
-0.04427085444331169,
-0.02667303755879402,
0.019792508333921432,
0.014873931184411049,
0.09433621913194656,
0.018306057900190353,
-0.05994477495551109,
0.05101325735449791,
-0.026819314807653427,
0.043763015419244766,
0.05365407466888428,
-0.05568952485918999,
0.068680040538311,
0.003017482813447714,
-0.05046427622437477,
-0.009674970991909504,
-0.07343962043523788,
0.07473733276128769,
-0.0799984559416771,
-0.03713719919323921,
-0.04396488144993782,
-0.037477146834135056,
0.06130850315093994,
0.05478106439113617,
0.05641744285821915,
-0.0110646802932024,
-0.01903095282614231,
-0.03923249617218971,
0.049667418003082275,
0.1096210852265358,
0.021758586168289185,
-0.02953319251537323,
-0.03557980805635452
] | 0 | 0 | 0 | 0 |
|
BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data | https://openreview.net/forum?id=3usdM1AuI3 | [
"Parsa Vahidi",
"Omid G. Sani",
"Maryam Shanechi"
] | Poster | Neural populations exhibit complex recurrent structures that drive behavior, while continuously receiving and integrating external inputs from sensory stimuli, upstream regions, and neurostimulation. However, neural populations are often modeled as autonomous dynamical systems, with little consideration given to the influence of external inputs that shape the population activity and behavioral outcomes. Here, we introduce BRAID, a deep learning framework that models nonlinear neural dynamics underlying behavior while explicitly incorporating any measured external inputs. Our method disentangles intrinsic recurrent neural population dynamics from the effects of inputs by including a forecasting objective within input-driven recurrent neural networks. BRAID further prioritizes the learning of intrinsic dynamics that are related to a behavior of interest by using a multi-stage optimization scheme. We validate BRAID with nonlinear simulations, showing that it can accurately learn the intrinsic dynamics shared between neural and behavioral modalities. We then apply BRAID to motor cortical activity recorded during a motor task and demonstrate that our method more accurately fits the neural-behavioral data by incorporating measured sensory stimuli into the model and improves the forecasting of neural-behavioral data compared with various baseline methods, whether input-driven or not. | Deep learning, Dynamic modeling, Sensory stimuli, RNN, Intrinsic, Behavior | null | 12,909 | null | [
-0.07876605540513992,
-0.1623329371213913,
0.0788402184844017,
0.05162583291530609,
-0.1345866471529007,
0.019952498376369476,
-0.0006888139760121703,
-0.06170497089624405,
0.008409632369875908,
-0.06036698445677757,
-0.04622538015246391,
-0.04029175266623497,
-0.02059033140540123,
0.0614120252430439,
-0.08144534379243851,
0.042308587580919266,
0.041812121868133545,
0.08509691059589386,
-0.0821312740445137,
0.011345244944095612,
0.04736949875950813,
0.023524323478341103,
0.03892328217625618,
0.06282658129930496,
0.02899574115872383,
0.041399311274290085,
-0.0252715814858675,
-0.012075172737240791,
0.018618980422616005,
-0.03878660872578621,
0.1065647155046463,
-0.027952158823609352,
-0.015360300429165363,
-0.02923600748181343,
-0.043263066560029984,
0.0422058068215847,
-0.09395245462656021,
-0.006643554661422968,
-0.05390792712569237,
-0.08608338981866837,
0.04666418954730034,
0.017809417098760605,
0.011932882480323315,
0.03283010050654411,
0.03737836703658104,
0.030193615704774857,
0.024382205680012703,
-0.09441067278385162,
-0.041397664695978165,
0.04509866610169411,
-0.0512981116771698,
-0.002039617858827114,
0.059428006410598755,
0.05455506220459938,
0.08617464452981949,
0.04263310506939888,
0.07015623152256012,
0.037784282118082047,
-0.03426443412899971,
-0.03843086212873459,
0.011953636072576046,
-0.06426338851451874,
0.009424460120499134,
-0.033807553350925446,
-0.05600511282682419,
0.056150928139686584,
-0.004047027789056301,
0.11700316518545151,
-0.023489270359277725,
-0.0013910033740103245,
-0.0075818272307515144,
0.05238271877169609,
-0.049059148877859116,
-0.007365928962826729,
0.041439421474933624,
0.0008753250585868955,
0.03809947520494461,
0.01339979749172926,
0.06266097724437714,
0.009430821985006332,
0.018121175467967987,
0.04617809131741524,
0.057695500552654266,
0.011667231097817421,
0.07409697026014328,
0.07407734543085098,
-0.007270061410963535,
0.014745998196303844,
-0.007376718334853649,
-0.048449207097291946,
-0.04311670735478401,
0.006998008117079735,
-0.013984293676912785,
-0.047358691692352295,
0.0016567085403949022,
0.043568000197410583,
-0.05703753978013992,
-0.05614715814590454,
0.03691769018769264,
0.04163205623626709,
0.009723321534693241,
0.04589097201824188,
-0.002505709184333682,
0.05141380429267883,
-0.014779153279960155,
0.00411982974037528,
0.0511694960296154,
0.03660234063863754,
0.04129556566476822,
-0.06336569786071777,
0.05490982532501221,
0.013552184216678143,
0.030100366100668907,
0.06705183535814285,
0.00033381229150108993,
-0.03517601266503334,
0.04726467281579971,
-0.011647640727460384,
0.08846144378185272,
0.15195460617542267,
-0.06545450538396835,
0.014593815430998802,
-0.03926971182227135,
0.010436583310365677,
-0.02748202532529831,
-0.03167338669300079,
-0.08265955746173859,
7.159466811473186e-34,
-0.022567735984921455,
-0.03217242285609245,
0.08786074817180634,
-0.031188108026981354,
0.027185281738638878,
-0.0691787451505661,
-0.030439557507634163,
-0.0284406840801239,
0.026479680091142654,
0.030615653842687607,
-0.04772964492440224,
0.037901848554611206,
-0.028330236673355103,
0.08954757452011108,
0.030643755570054054,
-0.0030847026500850916,
-0.053860120475292206,
0.01977454498410225,
0.013484600000083447,
-0.019399067386984825,
0.053144440054893494,
0.00519747706130147,
-0.03188365325331688,
-0.027048654854297638,
-0.03661561757326126,
-0.002495725406333804,
0.03956767916679382,
0.031797319650650024,
-0.02405792474746704,
0.03325066715478897,
-0.0292279664427042,
-0.0010046821553260088,
-0.09179770201444626,
-0.002178465947508812,
0.04766731336712837,
-0.008830425329506397,
0.045685309916734695,
0.03222270309925079,
0.000476479617645964,
0.03777293488383293,
0.028171397745609283,
-0.0009168880642391741,
-0.011618600226938725,
0.018864735960960388,
-0.045653264969587326,
-0.11135733872652054,
0.041445061564445496,
0.009263691492378712,
-0.07408424466848373,
-0.06764023005962372,
0.008579525165259838,
0.007543636951595545,
0.045430488884449005,
-0.10599717497825623,
-0.02782432921230793,
-0.04321122169494629,
-0.009441099129617214,
0.012285846285521984,
-0.0131796570494771,
0.048926256597042084,
-0.02079065702855587,
0.018247921019792557,
-0.00020935772045049816,
-0.033349666744470596,
-0.017735833302140236,
0.06882358342409134,
0.004656428005546331,
0.1115674301981926,
0.02768704481422901,
-0.04973607510328293,
0.005196333397179842,
0.0580655001103878,
0.0072576673701405525,
-0.06776276975870132,
0.08050164580345154,
-0.00896473042666912,
0.06785501539707184,
-0.12188773602247238,
-0.10762450098991394,
0.0561380572617054,
0.0467519536614418,
0.05307013541460037,
-0.0029721600003540516,
0.06087348610162735,
0.008005348965525627,
-0.0819234549999237,
0.09446794539690018,
-0.07585155218839645,
-0.08057355880737305,
0.020827163010835648,
-0.09709329158067703,
-0.0758901908993721,
-0.029198121279478073,
0.017849838361144066,
-0.01571880653500557,
-2.475214209955505e-33,
-0.018501821905374527,
0.0620170533657074,
-0.024714088067412376,
0.000653817318379879,
-0.037288323044776917,
-0.010820314288139343,
0.009967060759663582,
0.01901722513139248,
0.004905601963400841,
-0.04075388237833977,
0.020783988758921623,
-0.048808809369802475,
0.05410132557153702,
0.031058931723237038,
0.048811811953783035,
-0.010577748529613018,
-0.09127801656723022,
-0.0074323625303804874,
0.08199037611484528,
-0.01953677274286747,
-0.0077410247176885605,
0.053861092776060104,
-0.06840073317289352,
0.02596934512257576,
-0.06788017600774765,
0.021888311952352524,
-0.04710528627038002,
0.07399279624223709,
0.047707121819257736,
0.03650694712996483,
-0.06558860838413239,
-0.07371287792921066,
-0.024057170376181602,
0.016193417832255363,
-0.11023758351802826,
0.08851821720600128,
0.022700607776641846,
0.003155825426802039,
-0.010205693542957306,
-0.048760101199150085,
0.041960399597883224,
-0.05375281721353531,
0.010337203741073608,
0.04252991825342178,
0.013097538612782955,
-0.05949801579117775,
-0.10061590373516083,
-0.006008554250001907,
-0.039334412664175034,
0.07352178543806076,
0.024084819480776787,
0.07145056873559952,
-0.0959104523062706,
-0.01966378092765808,
-0.08507935702800751,
-0.037352435290813446,
0.030304113402962685,
-0.0007999418885447085,
0.020169824361801147,
0.05299178883433342,
-0.08833978325128555,
-0.029758792370557785,
-0.043029580265283585,
-0.02691044844686985,
-0.03821033239364624,
0.024479560554027557,
-0.023369288071990013,
-0.07222148776054382,
0.09654704481363297,
-0.060926828533411026,
0.004390303045511246,
0.06829392910003662,
0.012464219704270363,
0.02663777396082878,
-0.011123738251626492,
-0.020927494391798973,
-0.006514092441648245,
-0.06823021918535233,
-0.03382888436317444,
-0.0005239094025455415,
0.04705719277262688,
0.027573423460125923,
0.009671201929450035,
-0.008327599614858627,
0.03301306441426277,
0.007741620298475027,
-0.027618303894996643,
0.04410324990749359,
-0.018793053925037384,
-0.003933630883693695,
-0.010605521500110626,
-0.01885378547012806,
-0.08534908294677734,
-0.00552536454051733,
-0.036223284900188446,
-4.1733486000339326e-8,
-0.03830145299434662,
0.11370302736759186,
0.10462749004364014,
0.08016680181026459,
0.02789267525076866,
0.015293687582015991,
0.0827326700091362,
-0.03916662931442261,
-0.011352242901921272,
-0.04452455788850784,
0.15017853677272797,
-0.022118547931313515,
0.028434472158551216,
-0.043495576828718185,
-0.029800143092870712,
0.09160696715116501,
0.020140837877988815,
-0.02780734747648239,
0.0014833103632554412,
-0.015417905524373055,
0.06888483464717865,
0.011679067276418209,
-0.00802539475262165,
-0.029503311961889267,
0.06176290288567543,
-0.10642720013856888,
-0.09692034125328064,
0.07287883013486862,
0.012719376012682915,
-0.0324040986597538,
0.042803406715393066,
0.02799045667052269,
-0.004130247514694929,
-0.018285343423485756,
-0.11535204201936722,
0.008707737550139427,
0.009121044538915157,
-0.02859964407980442,
-0.03422626480460167,
-0.07415387779474258,
-0.04001297056674957,
0.07233928143978119,
-0.004538704175502062,
-0.009318728931248188,
-0.008837291970849037,
-0.04396054521203041,
0.0733189806342125,
-0.13814233243465424,
0.06773995608091354,
0.06206250935792923,
0.019083689898252487,
0.020378148183226585,
0.0469217486679554,
0.06541521102190018,
0.05695434659719467,
0.016013657674193382,
-0.02581680193543434,
-0.0501229353249073,
-0.0004991548485122621,
0.07028280198574066,
-0.04061071202158928,
0.0633992999792099,
-0.07157619297504425,
-0.059484343975782394
] | 0 | 0 | 0 | 0 |
|
LLM-based Typed Hyperresolution for Commonsense Reasoning with Knowledge Bases | https://openreview.net/forum?id=wNobG8bV5Q | [
"Armin Toroghi",
"Ali Pesaranghader",
"Tanmana Sadhu",
"Scott Sanner"
] | Poster | Large language models (LLM) are being increasingly applied to tasks requiring commonsense reasoning. Despite their outstanding potential, the reasoning process of LLMs is prone to errors and hallucinations that hinder their applicability, especially in high-stakes scenarios. Several works have attempted to enhance commonsense reasoning performance of LLMs by (i) using prompting styles that elicit more accurate reasoning, (ii) utilizing the LLM as a semantic parser for a symbolic reasoner, or (iii) enforcing the LLM to simulate a logical inference rule. However, all these solutions have critical limitations: they are unable to leverage the internal commonsense knowledge of the LLM in tandem with an axiomatic knowledge base, they lack a mechanism to reliably repair erroneous inference steps, and their application is restricted to small knowledge bases that fit the context limit of the LLM. In this work, we present LLM-based Typed Hyperresolution (LLM-TH), a logical commonsense reasoning framework that leverages "theory resolution", a concept from classical logical inference which enables integrating LLMs into the "resolution" inference rule, thus mitigating reasoning errors and hallucinations and enabling verification of the reasoning procedure. LLM-TH is also equipped with a mechanism for repairing erroneous inference steps supported by theoretical guarantees. Using "Hyperresolution" and "Typed inference" schemes, we show that LLM-TH can efficiently reason over large knowledge bases consisting of tens of thousands of rules with arbitrary predicate arities. Our experiments on three diverse language-based reasoning tasks—preference reasoning, multi-domain deductive reasoning, and geographical question answering—showcase that LLM-TH, using merely a BART 406M parameter NLI entailment model, significantly reduces reasoning errors compared to baselines using Llama3-70B, Gemini1.5-Flash, GPT-3.5-Turbo, and Mixtral-46.7B. | Large Language Models, Commonsense reasoning, Logical inference | null | 12,891 | null | [
0.027230165898799896,
-0.05654493719339371,
0.05675046145915985,
-0.029508326202630997,
0.080458864569664,
-0.04129236191511154,
0.016147544607520103,
-0.007326443679630756,
-0.0497661754488945,
0.049407847225666046,
-0.03695131093263626,
-0.046563200652599335,
0.06421475857496262,
0.05642656609416008,
0.08975639194250107,
0.06870061159133911,
0.06891434639692307,
0.09107861667871475,
-0.0740623027086258,
0.011409909464418888,
0.05369865521788597,
-0.013766632415354252,
-0.023390106856822968,
0.002922814106568694,
-0.058716099709272385,
0.028117747977375984,
0.04107068106532097,
-0.0034621250815689564,
0.03464873507618904,
-0.0016215924406424165,
0.027001796290278435,
0.09523359686136246,
-0.04738033190369606,
0.049814239144325256,
0.06157004088163376,
0.08567489683628082,
-0.004459228366613388,
0.036605775356292725,
-0.00569266639649868,
-0.013926043175160885,
-0.03798885643482208,
-0.0984835997223854,
-0.003960739821195602,
0.017230911180377007,
0.04581385850906372,
0.05352706462144852,
0.03963300958275795,
0.017111750319600105,
-0.06176137179136276,
-0.005884009413421154,
-0.015795569866895676,
-0.04362795501947403,
-0.041462983936071396,
-0.06291192024946213,
0.05565907433629036,
0.005067744757980108,
-0.02848546765744686,
0.006088786292821169,
-0.053295787423849106,
-0.10433672368526459,
-0.06973744928836823,
-0.07809428870677948,
-0.07400543987751007,
0.03309771791100502,
-0.013383384793996811,
0.004572276957333088,
0.03354840725660324,
0.06778118759393692,
-0.016351131722331047,
0.037738069891929626,
-0.026448912918567657,
0.059544969350099564,
-0.04412202537059784,
0.07097972929477692,
-0.06871844083070755,
0.040733762085437775,
0.030023301020264626,
-0.01857108622789383,
0.0803612470626831,
-0.03375765308737755,
0.00814368948340416,
0.062496595084667206,
-0.014514857903122902,
0.03978259861469269,
0.059301652014255524,
-0.0067486087791621685,
0.037151090800762177,
0.008951110765337944,
0.05337826907634735,
0.06323762238025665,
-0.01612935960292816,
-0.08975224196910858,
0.07006058096885681,
0.09673681110143661,
0.07599108666181564,
0.05411196127533913,
0.020116308704018593,
-0.0875808447599411,
-0.011055640876293182,
0.09361065179109573,
0.049511078745126724,
0.12169340252876282,
-0.0040876660495996475,
-0.1329759657382965,
0.0497274175286293,
-0.04650171473622322,
-0.06117710843682289,
0.08387772738933563,
0.04710523411631584,
-0.08629242330789566,
0.06667910516262054,
0.05361980199813843,
0.07311959564685822,
-0.012768471613526344,
0.003682867856696248,
-0.021129688248038292,
0.043065499514341354,
-0.01380226295441389,
0.0476275272667408,
-0.10949558764696121,
0.00006382165156537667,
0.03465702012181282,
0.025562705472111702,
-0.041083548218011856,
0.10867226868867874,
-0.042141880840063095,
-0.08460459113121033,
2.3438040434381507e-33,
0.030419787392020226,
0.07769922912120819,
0.01613168604671955,
0.05253811925649643,
0.08806231617927551,
0.007232096046209335,
0.0474235825240612,
0.05323897302150726,
-0.04405321553349495,
0.0007879406912252307,
0.016705188900232315,
-0.04298131912946701,
-0.046331748366355896,
-0.02040715329349041,
-0.03317505493760109,
0.019786186516284943,
-0.03690527752041817,
0.05974805727601051,
-0.11261153966188431,
-0.014031182043254375,
0.008607042953372002,
-0.0053389426320791245,
0.000594779325183481,
0.038702014833688736,
-0.06018596142530441,
-0.013709892518818378,
0.08006259053945541,
-0.016173716634511948,
0.09915316849946976,
0.0009179332409985363,
-0.06327735632658005,
-0.057615891098976135,
-0.03332434594631195,
0.07175388932228088,
-0.0330723375082016,
-0.024210125207901,
-0.03690246120095253,
-0.11807575076818466,
0.030676385387778282,
-0.08289017528295517,
-0.028632085770368576,
0.038659799844026566,
0.0319054014980793,
0.012020241469144821,
-0.07247010618448257,
-0.09500348567962646,
-0.03349388390779495,
0.04767630994319916,
-0.010928650386631489,
-0.013066405430436134,
0.09913402795791626,
0.03296973928809166,
0.01725289411842823,
-0.05106475204229355,
-0.013223502784967422,
-0.028478914871811867,
-0.02374974824488163,
0.05356850475072861,
-0.030807746574282646,
0.017025716602802277,
-0.006941244471818209,
-0.00853039138019085,
0.028135959059000015,
0.09517297893762589,
0.040961168706417084,
0.05451555922627449,
-0.0987137109041214,
-0.059526823461055756,
0.05160318315029144,
-0.05666001886129379,
0.020498257130384445,
0.01756051555275917,
0.029718192294239998,
0.03715890273451805,
0.01619047485291958,
-0.02915317565202713,
0.011905800551176071,
-0.04936124384403229,
0.05963614955544472,
-0.03586677834391594,
-0.006629245821386576,
0.06806239485740662,
-0.04111003875732422,
0.013180714100599289,
0.003058850299566984,
-0.041871100664138794,
0.00448857806622982,
-0.06596039235591888,
0.02485605701804161,
-0.11244147270917892,
0.01049470528960228,
-0.05353394150733948,
-0.02524571307003498,
-0.07769306004047394,
-0.02755400352180004,
-2.3149530039651562e-33,
-0.021950924769043922,
-0.03214738890528679,
-0.03175682574510574,
0.07329371571540833,
-0.030656971037387848,
-0.054442379623651505,
-0.00494597340002656,
-0.028485007584095,
-0.049049343913793564,
-0.10947586596012115,
-0.0294960904866457,
-0.0229706559330225,
0.021025491878390312,
-0.005174748133867979,
-0.02126375027000904,
-0.029752008616924286,
0.02790556102991104,
-0.09034117311239243,
-0.06100797280669212,
0.12201213836669922,
0.006797329522669315,
0.026158250868320465,
-0.09304196387529373,
0.027431359514594078,
-0.013469938188791275,
0.04553867131471634,
-0.04984680190682411,
0.08041101694107056,
0.031119562685489655,
0.05591640993952751,
0.019095472991466522,
-0.031838566064834595,
-0.08197974413633347,
-0.11912141740322113,
-0.05308438464999199,
0.01057554129511118,
0.03354969248175621,
0.014997195452451706,
-0.06873364001512527,
0.02548946626484394,
0.013472533784806728,
-0.07833215594291687,
0.01973707228899002,
-0.03291501849889755,
-0.04285114258527756,
0.013751275837421417,
-0.047939904034137726,
-0.0339558906853199,
0.07255025207996368,
-0.007562910672277212,
-0.02259312942624092,
-0.06383262574672699,
0.040582794696092606,
0.025645805522799492,
0.0034478255547583103,
-0.02234678529202938,
-0.034548986703157425,
-0.024331767112016678,
-0.061053115874528885,
-0.019400371238589287,
-0.05527898296713829,
0.003870931454002857,
0.015082347206771374,
0.012286975979804993,
-0.007024239748716354,
0.014819384552538395,
-0.0884544774889946,
0.003991304896771908,
-0.04681149125099182,
-0.05216825380921364,
0.05336783826351166,
0.013673846609890461,
-0.07068563997745514,
-0.03289199247956276,
0.10950793325901031,
-0.0066850860603153706,
0.06027815118432045,
0.008801640011370182,
-0.027428558096289635,
-0.05645817890763283,
0.039369333535432816,
-0.0321049690246582,
0.05352376028895378,
0.08281305432319641,
-0.050713323056697845,
-0.05413927137851715,
0.09273790568113327,
0.1003125011920929,
-0.012744615785777569,
0.040001336485147476,
-0.021068591624498367,
0.06679258495569229,
-0.002617662074044347,
0.07460352778434753,
-0.023477310314774513,
-4.461829661295269e-8,
-0.11893397569656372,
0.018975306302309036,
-0.09074974805116653,
0.0280144140124321,
0.02040271833539009,
0.02539859525859356,
-0.05327600985765457,
0.03889480605721474,
-0.020277755334973335,
-0.05689239129424095,
0.03315076231956482,
-0.007962951436638832,
-0.03349434584379196,
-0.019879810512065887,
0.02799043618142605,
0.06541437655687332,
0.03848627954721451,
-0.0348362997174263,
-0.08322396874427795,
0.025850484147667885,
-0.003416066989302635,
-0.0016479181358590722,
-0.04135902225971222,
0.02762899175286293,
0.0948590636253357,
-0.016031939536333084,
0.02035093680024147,
0.09957253932952881,
0.04930891469120979,
0.008662517182528973,
-0.019642481580376625,
0.01436651311814785,
-0.005474301055073738,
0.003736854065209627,
0.06994654983282089,
0.04632970318198204,
0.003130996134132147,
0.012179314158856869,
0.056423116475343704,
-0.054935306310653687,
0.03496168553829193,
-0.0011969762854278088,
-0.09027890115976334,
0.08673175424337387,
0.08075088262557983,
-0.06281223893165588,
-0.09365055710077286,
-0.019563594833016396,
-0.02268505096435547,
-0.006694091949611902,
-0.024883853271603584,
0.04061194881796837,
-0.02377132512629032,
-0.008436356671154499,
-0.020396912470459938,
0.07560402899980545,
-0.030307527631521225,
0.06808623671531677,
0.0046323323622345924,
0.036318521946668625,
0.05795460566878319,
0.016340894624590874,
-0.0019853629637509584,
0.010063719935715199
] | 0 | 0 | 0 | 0 |
|
What's the Move? Hybrid Imitation Learning via Salient Points | https://openreview.net/forum?id=r0pLGGcuY6 | [
"Priya Sundaresan",
"Hengyuan Hu",
"Quan Vuong",
"Jeannette Bohg",
"Dorsa Sadigh"
] | Poster | While imitation learning (IL) offers a promising framework for teaching robots various behaviors, learning complex tasks remains challenging. Existing IL policies struggle to generalize effectively across visual and spatial variations even for simple tasks. In this work, we introduce **SPHINX**: **S**alient **P**oint-based **H**ybrid **I**mitatio**N** and e**X**ecution, a flexible IL policy that leverages multimodal observations (point clouds and wrist images), along with a hybrid action space of low-frequency, sparse waypoints and high-frequency, dense end effector movements. Given 3D point cloud observations, SPHINX learns to infer task-relevant points within a point cloud, or *salient points*, which support spatial generalization by focusing on semantically meaningful features. These salient points serve as anchor points to predict waypoints for long-range movement, such as reaching target poses in free-space. Once near a salient point, SPHINX learns to switch to predicting dense end-effector movements given close-up wrist images for precise phases of a task. By exploiting the strengths of different input modalities and action representations for different manipulation phases, SPHINX tackles complex tasks in a sample-efficient, generalizable manner. Our method achieves **86.7%** success across 4 real-world and 2 simulated tasks, outperforming the next best state-of-the-art IL baseline by **41.1%** on average across **440** real world trials. SPHINX additionally generalizes to novel viewpoints, visual distractors, spatial arrangements, and execution speeds with a **1.7x** speedup over the most competitive baseline. Our website (http://sphinx-manip.github.io) provides open-sourced code for data collection, training, and evaluation, along with supplementary videos. | Imitation Learning, Robot Learning, Robot Manipulation, Robotics | We propose an imitation learning algorithm for complex robot manipulation with visuospatial generalization; it substantially outperforms SOTA existing methods across 4 real-world tasks and 2 simulated benchmarks. | 12,887 | 2412.05426 | [
-0.001985738519579172,
-0.09888243675231934,
0.09795715659856796,
-0.02503194287419319,
0.06841558963060379,
0.015601934865117073,
0.035235270857810974,
-0.03528761491179466,
0.026901360601186752,
0.0717577263712883,
0.04679252579808235,
-0.03701562434434891,
0.022649720311164856,
0.030197281390428543,
0.08308649808168411,
0.017236042767763138,
0.0962928980588913,
0.015600751154124737,
-0.032394278794527054,
0.07479238510131836,
0.08564376085996628,
-0.03448978811502457,
0.07591725885868073,
0.02702476643025875,
-0.0037385243922472,
0.06794346868991852,
-0.01993192732334137,
-0.08463245630264282,
-0.006082545034587383,
-0.11027467995882034,
0.007770034950226545,
0.02383359521627426,
-0.03837965056300163,
-0.03609893471002579,
-0.06530474126338959,
0.10436191409826279,
-0.014544648118317127,
-0.050978850573301315,
-0.04948557913303375,
-0.02459757775068283,
0.005714010912925005,
-0.004883245099335909,
-0.0009676439221948385,
0.015123783610761166,
0.11877907067537308,
0.08380846679210663,
0.027982555329799652,
-0.060688383877277374,
0.07673180103302002,
0.027847228571772575,
-0.08835913985967636,
-0.06514537334442139,
0.07043223828077316,
-0.025337250903248787,
-0.01749187335371971,
0.030524088069796562,
0.024934915825724602,
-0.03684001788496971,
0.10312507301568985,
-0.04474484547972679,
0.02416416071355343,
-0.034966349601745605,
0.018181761726737022,
-0.00786029826849699,
-0.04062269255518913,
-0.11511777341365814,
-0.02396124228835106,
-0.05285615473985672,
0.0013134449254721403,
0.0009389693732373416,
-0.04633048176765442,
-0.0077231768518686295,
0.0127224400639534,
-0.029005251824855804,
0.05216038599610329,
-0.03274957463145256,
0.024308525025844574,
-0.0639948770403862,
-0.021549975499510765,
-0.08129612356424332,
0.03127547353506088,
-0.022366447374224663,
-0.01910260319709778,
0.07226981222629547,
0.0306248739361763,
0.017517145723104477,
-0.02483469434082508,
0.06437899172306061,
0.010798340663313866,
-0.01087924838066101,
0.004287051036953926,
-0.07728968560695648,
-0.019480636343359947,
-0.09474605321884155,
0.036193083971738815,
-0.03857377544045448,
-0.0196611937135458,
-0.07465305179357529,
-0.0111708277836442,
0.06161566451191902,
0.059602975845336914,
0.019177952781319618,
-0.007127989549189806,
0.027163192629814148,
0.018290890380740166,
-0.03066219948232174,
-0.0039017396047711372,
-0.047394152730703354,
0.027253763750195503,
-0.008025896735489368,
0.0021886038593947887,
-0.1286640763282776,
0.0366964265704155,
0.027608288452029228,
0.008575200103223324,
0.029588188976049423,
-0.05447818711400032,
-0.023329710587859154,
0.01569763571023941,
0.008182906545698643,
0.0013012744020670652,
-0.07247645407915115,
-0.03147874400019646,
0.08665622025728226,
0.02696726657450199,
-0.029472457244992256,
-0.058651819825172424,
3.0502425364907517e-34,
0.06990209221839905,
0.0036231407430022955,
0.06488781422376633,
-0.019245855510234833,
0.019253892824053764,
0.0011775704333558679,
-0.025517776608467102,
0.012862871401011944,
-0.035937681794166565,
0.003653707215562463,
-0.026278795674443245,
-0.017983905971050262,
-0.01407971978187561,
0.1560676097869873,
0.08413079380989075,
-0.016379691660404205,
-0.05155688896775246,
0.04166467487812042,
-0.05440746992826462,
0.06743120402097702,
0.08545014262199402,
-0.0299227274954319,
-0.047824133187532425,
-0.04710943251848221,
0.03766385465860367,
0.02729658968746662,
-0.0376858226954937,
-0.047791074961423874,
0.012371913529932499,
-0.009858778677880764,
-0.05267997831106186,
0.03979083150625229,
-0.040776316076517105,
0.03546201437711716,
-0.006565502379089594,
-0.01890198513865471,
-0.07335551083087921,
-0.06751582771539688,
-0.015602621249854565,
-0.04094073548913002,
0.008047888986766338,
0.03234950080513954,
-0.025790536776185036,
-0.04058380052447319,
-0.07260336726903915,
-0.022478876635432243,
0.05863518640398979,
0.08311568200588226,
-0.028086373582482338,
-0.021795369684696198,
0.040508728474378586,
0.04156329110264778,
-0.045303016901016235,
-0.12877047061920166,
-0.007236420642584562,
-0.0049897246062755585,
-0.0008493088535033166,
0.012413760647177696,
0.007771152537316084,
-0.0471295528113842,
0.007991146296262741,
-0.08300546556711197,
-0.02181887812912464,
0.08064469695091248,
-0.010940320789813995,
0.0602918341755867,
0.0025805749464780092,
0.13533326983451843,
0.11085010319948196,
0.05415507033467293,
-0.028778985142707825,
0.016900481656193733,
0.008639569394290447,
-0.06635348498821259,
0.07612504810094833,
-0.03954477235674858,
-0.006558122579008341,
-0.10756611078977585,
0.03675808757543564,
0.052971530705690384,
-0.05763539671897888,
-0.019146230071783066,
-0.06208597496151924,
-0.035219285637140274,
0.004877970088273287,
-0.05943724513053894,
0.03953075781464577,
-0.10253551602363586,
-0.0891183391213417,
-0.0075131431221961975,
-0.0437266007065773,
-0.07323458045721054,
-0.10165891796350479,
0.04536872357130051,
-0.09345808625221252,
-1.3776627569389164e-33,
0.036337751895189285,
-0.01157400943338871,
0.017329495400190353,
0.021531611680984497,
-0.07137339562177658,
-0.01274832058697939,
0.02177671529352665,
0.048470932990312576,
0.012913892045617104,
-0.006445872597396374,
-0.058468468487262726,
0.01851644366979599,
0.04081409052014351,
-0.012643393129110336,
0.042067766189575195,
0.030313752591609955,
0.0075771501287817955,
0.07457937300205231,
-0.024545148015022278,
0.032966598868370056,
0.0426102876663208,
0.03230806812644005,
0.022091934457421303,
0.006994992028921843,
-0.017183056101202965,
0.05463331565260887,
-0.025523940101265907,
0.03182099387049675,
-0.03648985177278519,
0.01462346501648426,
-0.034146301448345184,
-0.04475101828575134,
0.007295211777091026,
-0.012227159924805164,
-0.017206311225891113,
0.05794311687350273,
-0.03684999421238899,
0.04248267039656639,
-0.04385662078857422,
0.06406797468662262,
0.06446708738803864,
-0.04201594367623329,
0.029312659054994583,
-0.004358185455203056,
0.053155869245529175,
-0.019592134281992912,
-0.08934901654720306,
0.06981463730335236,
-0.08316457271575928,
0.04157865047454834,
-0.050769273191690445,
0.012660647742450237,
-0.0696277841925621,
-0.074137382209301,
-0.0340479351580143,
-0.041051968932151794,
-0.02715643309056759,
-0.000986196449957788,
0.02664824016392231,
-0.0728347897529602,
-0.03185582906007767,
0.016980554908514023,
-0.05067417025566101,
0.08617036789655685,
0.007643291261047125,
0.035847362130880356,
-0.0601922981441021,
-0.007663298398256302,
-0.008352624252438545,
-0.013940500095486641,
0.03578588739037514,
0.0032419476192444563,
0.0389370396733284,
-0.034390639513731,
0.03976011648774147,
0.0015509547665715218,
0.034972772002220154,
-0.04670555517077446,
0.001360520371235907,
-0.031173842027783394,
0.009883031249046326,
-0.05010860785841942,
0.08876673877239227,
0.11707243323326111,
0.018963541835546494,
0.19013893604278564,
-0.06302750110626221,
0.08147614449262619,
0.014163237065076828,
0.057999372482299805,
-0.0068330783396959305,
0.08043508231639862,
0.013455419801175594,
0.07354635745286942,
-0.000404457125114277,
-4.015001309198851e-8,
-0.12602455914020538,
0.0738552138209343,
0.11166030168533325,
0.03153083845973015,
-0.0003568124957382679,
0.0032003349624574184,
0.019073430448770523,
-0.01160643994808197,
-0.03655901923775673,
-0.011507286690175533,
0.0590219721198082,
-0.02390923537313938,
0.03818204253911972,
-0.044697362929582596,
0.06578390300273895,
0.0644502341747284,
0.061556894332170486,
-0.03234230726957321,
-0.023334620520472527,
0.052368685603141785,
-0.00824646931141615,
-0.04779193177819252,
0.058067407459020615,
0.0018878737464547157,
-0.051623180508613586,
-0.06397540122270584,
-0.0608057864010334,
0.04806370288133621,
0.03561243787407875,
-0.03328726068139076,
0.034857600927352905,
-0.02653508074581623,
0.026599161326885223,
-0.015924550592899323,
-0.006514195818454027,
0.013995413668453693,
0.013656746596097946,
-0.04055873304605484,
0.015235845930874348,
0.017278872430324554,
-0.042897630482912064,
0.05063838139176369,
-0.003071724670007825,
-0.007956575602293015,
0.026792334392666817,
0.05112896487116814,
-0.017799776047468185,
-0.18982332944869995,
0.045621637254953384,
-0.03803344815969467,
-0.008137627504765987,
-0.031303536146879196,
0.006379834376275539,
0.08925627171993256,
0.07189981639385223,
0.052628353238105774,
-0.026893222704529762,
-0.027131816372275352,
0.017713608220219612,
0.09743153303861618,
0.07509975880384445,
0.002264060778543353,
-0.07876085489988327,
0.020906025543808937
] | 0 | 0 | 0 | 0 |
|
PEARL: Towards Permutation-Resilient LLMs | https://openreview.net/forum?id=txoJvjfI9w | [
"Liang CHEN",
"Li Shen",
"Yang Deng",
"Xiaoyan Zhao",
"Bin Liang",
"Kam-Fai Wong"
] | Poster | The in-context learning (ICL) capability of large language models (LLMs) enables them to perform challenging tasks using provided demonstrations. However, ICL is highly sensitive to the ordering of demonstrations, leading to instability in predictions. This paper shows that this vulnerability can be exploited to design a natural attack—difficult for model providers to detect—that achieves nearly 80% success rate on LLaMA-3 by simply permuting the demonstrations. Existing mitigation methods primarily rely on post-processing and fail to enhance the model's inherent robustness to input permutations, raising concerns about safety and reliability of LLMs. To address this issue, we propose Permutation-resilient learning (PEARL), a novel framework based on distributionally robust optimization (DRO), which optimizes model performance against the worst-case input permutation. Specifically, PEARL consists of a permutation-proposal network (P-Net) and the LLM. The P-Net generates the most challenging permutations by treating it as an optimal transport problem, which is solved using an entropy-constrained Sinkhorn algorithm. Through minimax optimization, the P-Net and the LLM iteratively optimize against each other, progressively improving the LLM's robustness. Experiments on synthetic pre-training and real-world instruction tuning tasks demonstrate that PEARL effectively mitigates permutation attacks and enhances performance. Notably, despite being trained on fewer shots and shorter contexts, PEARL achieves performance gains of up to 40% when scaled to many-shot and long-context scenarios, highlighting its efficiency and generalization capabilities. | In-Context Learning, Large Language Models, Instruction Tuning, Robustness, Attack, Distributionally Robust Optimization, Optimal Transport, Sinkhorn Algorithm | null | 12,884 | 2502.14628 | [
-0.04830208048224449,
-0.05858469381928444,
0.04343424364924431,
-0.013192616403102875,
0.02509145811200142,
0.046372007578611374,
0.01474499050527811,
-0.010440177284181118,
0.07984094321727753,
0.018212346360087395,
-0.04683414101600647,
0.022907178848981857,
0.11487387865781784,
-0.007252455689013004,
-0.05395001918077469,
0.023340102285146713,
0.1173640713095665,
0.11893461644649506,
-0.10613960027694702,
-0.004376180935651064,
0.013121116906404495,
-0.007491010241210461,
-0.03590394929051399,
-0.050357695668935776,
0.03723575919866562,
0.0393810048699379,
-0.055608946830034256,
0.012540231458842754,
0.006520950701087713,
-0.033536940813064575,
0.004984771367162466,
0.02729148417711258,
-0.00041688515921123326,
-0.008489889092743397,
-0.05329950153827667,
0.03583797439932823,
-0.13218927383422852,
-0.0349077545106411,
0.01605186052620411,
-0.028899313881993294,
-0.06568840891122818,
0.022809192538261414,
0.048548374325037,
-0.016400333493947983,
0.026559574529528618,
-0.0602538175880909,
0.06037183478474617,
-0.0017751873238012195,
-0.01030019111931324,
0.0006249980651773512,
-0.017650224268436432,
-0.025679290294647217,
-0.013353241607546806,
-0.0027144919149577618,
-0.03795357421040535,
-0.06651784479618073,
0.05602838844060898,
-0.020748328417539597,
-0.021683786064386368,
0.0003137871390208602,
-0.021869782358407974,
-0.08972984552383423,
-0.027039090171456337,
0.0349055640399456,
-0.04728073626756668,
0.018561504781246185,
0.016087835654616356,
0.06152503564953804,
0.021528396755456924,
0.03582220524549484,
0.010265097953379154,
0.041086163371801376,
-0.040867358446121216,
0.06889884918928146,
0.0712289810180664,
0.10158269107341766,
0.01661774143576622,
-0.06686048954725266,
0.04022398963570595,
-0.019227875396609306,
-0.013123913668096066,
-0.007793216034770012,
0.040696363896131516,
-0.028250450268387794,
0.1357974261045456,
0.020668597891926765,
-0.034736283123493195,
0.052535202354192734,
0.051899220794439316,
0.031004928052425385,
-0.0436164066195488,
-0.05775679647922516,
0.07319847494363785,
0.04373868182301521,
0.014641037210822105,
0.030234474688768387,
0.006027248688042164,
-0.07238974422216415,
-0.021732894703745842,
0.09258200973272324,
-0.05514521524310112,
0.08462250977754593,
0.03928668424487114,
-0.10617255419492722,
0.0069255889393389225,
-0.028785334900021553,
-0.026916448026895523,
-0.05873364210128784,
0.03803844004869461,
-0.06813928484916687,
0.02133866772055626,
0.08542542159557343,
0.046179547905921936,
0.05624353513121605,
-0.03006681241095066,
-0.005943678319454193,
-0.0120993722230196,
-0.0026966286823153496,
0.05120181664824486,
0.12772513926029205,
-0.03915053606033325,
-0.007051697000861168,
-0.011344078928232193,
-0.015246650204062462,
0.01885703019797802,
-0.045804399996995926,
-0.023842215538024902,
-1.1122713055182724e-33,
0.0018830634653568268,
0.06399708986282349,
0.0218253992497921,
0.00330211129039526,
0.09751663357019424,
-0.04698599502444267,
0.02988898940384388,
-0.012678355909883976,
-0.01950170285999775,
-0.05694076418876648,
0.04471609741449356,
-0.007412009872496128,
0.016700560227036476,
-0.03914773091673851,
0.05857965350151062,
0.022812049835920334,
0.026049451902508736,
-0.0018338855588808656,
-0.028186587616801262,
-0.09124686568975449,
0.058509353548288345,
-0.06363222002983093,
0.007606297265738249,
-0.12367265671491623,
-0.02187994308769703,
0.06829991936683655,
0.05846124142408371,
0.02314569801092148,
0.02085874229669571,
0.02966446988284588,
-0.09620236605405807,
0.019608192145824432,
-0.037770919501781464,
0.0563812181353569,
0.021840495988726616,
0.0009028021595440805,
-0.05726126208901405,
-0.01818118430674076,
0.008961258456110954,
0.010277651250362396,
-0.04371044039726257,
-0.03215259313583374,
-0.0032046432606875896,
0.02627757005393505,
-0.010953949764370918,
-0.004060981795191765,
-0.009274437092244625,
0.05482562631368637,
-0.02015506848692894,
-0.03148918226361275,
0.016579359769821167,
0.006539216265082359,
-0.03376086801290512,
-0.01663319580256939,
0.01319295633584261,
-0.05059681460261345,
0.015476332977414131,
0.03815852105617523,
0.02387966401875019,
0.14040787518024445,
0.0017163786105811596,
0.03058537282049656,
0.0002992853696923703,
0.08464380353689194,
0.1018536388874054,
-0.023584259673953056,
-0.07288380712270737,
-0.01754765957593918,
0.07028119266033173,
-0.028532439842820168,
-0.025720341131091118,
-0.04345635324716568,
-0.022906983271241188,
0.006849354133009911,
0.026267999783158302,
-0.038467876613140106,
0.03649686649441719,
-0.011363351717591286,
0.013160520233213902,
-0.06021842360496521,
0.0018362636910751462,
0.05865653231739998,
0.028120696544647217,
0.02967616356909275,
-0.08045970648527145,
-0.007164916023612022,
-0.04117007553577423,
-0.04146091267466545,
-0.07491610199213028,
-0.03698958456516266,
-0.0012268644059076905,
-0.038545407354831696,
0.010061917826533318,
0.025680264458060265,
0.02957814373075962,
-7.270596715785231e-35,
-0.08350205421447754,
-0.022608838975429535,
-0.041774433106184006,
0.07094011455774307,
-0.016324182972311974,
-0.10538313537836075,
0.04790500923991203,
0.052140410989522934,
-0.05414465442299843,
-0.044448476284742355,
-0.12695789337158203,
-0.03621769696474075,
0.07606983929872513,
-0.024351254105567932,
0.053881969302892685,
-0.015593874268233776,
0.006772513035684824,
0.027166983112692833,
0.03425852581858635,
0.03641074150800705,
0.008227798156440258,
0.11682574450969696,
-0.1341754049062729,
0.10381712764501572,
-0.07898028939962387,
-0.01107893418520689,
-0.019354697316884995,
0.004366206005215645,
0.06054505333304405,
0.06176193803548813,
-0.04846316948533058,
0.017878683283925056,
-0.030814725905656815,
-0.030678899958729744,
-0.053122781217098236,
0.07476653903722763,
-0.03940025717020035,
0.1060984805226326,
-0.007905912585556507,
0.13618607819080353,
0.09184171259403229,
-0.009164910763502121,
-0.05842836573719978,
-0.009350230917334557,
0.007783765438944101,
0.02385181188583374,
-0.038192521780729294,
-0.08848042041063309,
0.09021952003240585,
-0.020978432148694992,
-0.02183421514928341,
0.0048791044391691685,
0.025123724713921547,
-0.02898266166448593,
-0.022194335237145424,
-0.03334544971585274,
0.04940950125455856,
-0.0751466304063797,
0.1355537623167038,
0.0017997195245698094,
-0.08890395611524582,
-0.03565113618969917,
0.09971319139003754,
-0.019423333927989006,
-0.0032551318872720003,
0.026689141988754272,
-0.054540541023015976,
-0.0063230618834495544,
-0.027840327471494675,
-0.04556213691830635,
0.022216973826289177,
0.012297391891479492,
-0.014905015006661415,
-0.0016686208546161652,
0.041075579822063446,
-0.02145015075802803,
-0.02090637758374214,
-0.03703294321894646,
-0.02950199693441391,
-0.021491503342986107,
0.0015852790093049407,
0.014286824502050877,
-0.018915345892310143,
0.09679238498210907,
0.05348081886768341,
0.024713879451155663,
0.07815643399953842,
0.06406497955322266,
0.04209955409169197,
-0.008561043068766594,
-0.01783670112490654,
0.05524788424372673,
0.04398348927497864,
0.03240973502397537,
-0.00436502555385232,
-4.773968953486474e-8,
-0.07269378751516342,
-0.004533406812697649,
0.06291640549898148,
0.046662647277116776,
0.06370732933282852,
0.05397998169064522,
-0.03197300806641579,
0.02132546342909336,
0.031552139669656754,
0.00047996462672017515,
0.03403609246015549,
0.01136845350265503,
-0.021687917411327362,
-0.06242832541465759,
-0.07919872552156448,
0.13825537264347076,
0.004022376611828804,
-0.012445193715393543,
-0.031214796006679535,
-0.06955058127641678,
-0.0039570992812514305,
0.024222176522016525,
0.06473258882761002,
-0.012610836885869503,
0.058059804141521454,
-0.05718445032835007,
-0.01153011154383421,
0.02453356236219406,
0.07568337768316269,
0.007613509893417358,
-0.08366115391254425,
-0.07054225355386734,
-0.010420573875308037,
-0.02127770520746708,
-0.06709655374288559,
0.1498311161994934,
-0.025639675557613373,
-0.07597753405570984,
0.018600720912218094,
-0.08995616436004639,
-0.013318737037479877,
0.0068079084157943726,
-0.02354400046169758,
0.007049684412777424,
0.005106098484247923,
-0.027115600183606148,
-0.008859468623995781,
-0.14868496358394623,
0.002075499389320612,
-0.0834459513425827,
0.034088052809238434,
-0.06666652858257294,
-0.040626004338264465,
0.03216231241822243,
0.050641946494579315,
0.07272060215473175,
-0.02142941579222679,
-0.07058461010456085,
0.027552220970392227,
0.10231224447488785,
-0.02396284230053425,
0.005654239561408758,
-0.0661478266119957,
-0.027849189937114716
] | https://github.com/chanliang/pearl | 0 | 0 | 0 | 0 |
SWE-bench Multimodal: Do AI Systems Generalize to Visual Software Domains? | https://openreview.net/forum?id=riTiq3i21b | [
"John Yang",
"Carlos E Jimenez",
"Alex L Zhang",
"Kilian Lieret",
"Joyce Yang",
"Xindi Wu",
"Ori Press",
"Niklas Muennighoff",
"Gabriel Synnaeve",
"Karthik R Narasimhan",
"Diyi Yang",
"Sida Wang",
"Ofir Press"
] | Poster | Autonomous systems for software engineering are now capable of fixing bugs and developing features. These systems are commonly evaluated on SWE-bench (Jimenez et al., 2024a), which assesses their ability to solve software issues from GitHub repositories. However, SWE-bench uses only Python repositories, with problem statements presented predominantly as text and lacking visual elements such as images. This limited coverage motivates our inquiry into how existing systems might perform on unrepresented software engineering domains
(e.g., front-end, game development, DevOps), which use different programming languages and paradigms. Therefore, we propose SWE-bench Multimodal (SWE-bench M), to evaluate systems on their ability to fix bugs in visual, user-facing JavaScript software. SWE-bench M features 617 task instances collected from 17 JavaScript libraries used for web interface design, diagramming, data visualization, syntax highlighting, and interactive mapping. Each SWE-bench M task instance contains at least one image in its problem statement or unit tests. Our analysis finds that top-performing SWE-bench systems struggle with SWE-bench M, revealing limitations in visual problem-solving and cross-language generalization. Lastly, we show that SWE-agent’s flexible language-agnostic features enable it to substantially outperform alternatives on SWE-bench M, resolving 12% of task
instances compared to 6% for the next best system. | Language models, Natural language processing, Software engineering | Our new multimodal software engineering benchmark reveals that current LM agent systems do not generalize to other programming languages. | 12,867 | 2410.03859 | [
-0.05872379615902901,
-0.0953618586063385,
0.059812530875205994,
0.028295597061514854,
0.10863204300403595,
-0.1386880725622177,
-0.029620669782161713,
0.03743999823927879,
-0.027006102725863457,
-0.03097817488014698,
-0.09977536648511887,
-0.07811545580625534,
0.07027913630008698,
0.0001470174902351573,
0.0677269697189331,
0.006392736453562975,
0.023296814411878586,
0.0021552047692239285,
-0.007965276949107647,
-0.07264914363622665,
0.006402985658496618,
-0.021634016185998917,
0.001605205237865448,
-0.05724272131919861,
0.023114265874028206,
0.061580222100019455,
-0.03434527665376663,
-0.06799394637346268,
-0.02525242045521736,
-0.06347793340682983,
-0.026131004095077515,
-0.003439463209360838,
-0.08806294947862625,
0.05764403194189072,
0.0045630354434251785,
0.13419200479984283,
-0.03456860035657883,
-0.04116382822394371,
-0.026510337367653847,
-0.028234604746103287,
-0.10147316008806229,
0.020187878981232643,
0.021434495225548744,
-0.039803583174943924,
0.06921257823705673,
-0.07881227135658264,
-0.058464277535676956,
-0.08979742974042892,
0.01671573519706726,
-0.0039710174314677715,
-0.10363840311765671,
-0.11849623918533325,
-0.0010752723319455981,
-0.015183784998953342,
0.029347041621804237,
0.02407621033489704,
0.08248719573020935,
-0.003562817582860589,
0.04187501594424248,
0.05057350918650627,
0.025433562695980072,
-0.010064049623906612,
0.05856423079967499,
0.0280735082924366,
0.03015010431408882,
-0.015236450359225273,
0.041191790252923965,
-0.006374916527420282,
0.03248903900384903,
-0.047616537660360336,
-0.06784603744745255,
0.04637587442994118,
-0.002871946431696415,
0.020326580852270126,
-0.036242809146642685,
0.00322672794573009,
-0.04165361076593399,
-0.0011053585913032293,
0.03817001357674599,
-0.09050890803337097,
0.020372889935970306,
-0.034980520606040955,
-0.009187716990709305,
0.10159850120544434,
0.10028841346502304,
0.0492677204310894,
0.044138453900814056,
0.04288523644208908,
0.057183630764484406,
0.101981982588768,
-0.04123895987868309,
-0.08211967349052429,
0.10695531219244003,
0.031205883249640465,
0.0271180160343647,
-0.002026250818744302,
0.11661501228809357,
-0.10395260900259018,
-0.07683605700731277,
0.08997141569852829,
-0.04333099350333214,
0.00852087140083313,
0.0959056168794632,
-0.041637029498815536,
-0.022757841274142265,
0.03303929418325424,
0.086472287774086,
-0.011743573471903801,
-0.004640813916921616,
-0.0570666529238224,
-0.016766805201768875,
0.04186880216002464,
0.02629086747765541,
-0.04053042456507683,
-0.00410578353330493,
-0.0362517312169075,
0.023786280304193497,
-0.002456157235428691,
0.044276896864175797,
0.04820522293448448,
0.09777174144983292,
0.09896842390298843,
-0.0075207664631307125,
-0.042079661041498184,
0.11721260845661163,
0.04028846323490143,
-0.024595674127340317,
2.3344827405781146e-33,
0.0531219057738781,
-0.004729338921606541,
0.020431572571396828,
0.033019453287124634,
0.09815651923418045,
-0.07465816289186478,
0.0011363596422597766,
0.07063154876232147,
-0.0956135168671608,
-0.035245753824710846,
0.002609835471957922,
0.0459938570857048,
-0.05900229141116142,
0.10030315071344376,
0.045500498265028,
0.013601281680166721,
-0.024318015202879906,
0.03373873978853226,
-0.0469060093164444,
0.0025440226309001446,
-0.058975700289011,
-0.06602845340967178,
0.01417563483119011,
0.01290520466864109,
0.009077695198357105,
0.02419373020529747,
0.030968571081757545,
-0.012736757285892963,
-0.01650952734053135,
0.02617666870355606,
-0.13161639869213104,
0.012942994013428688,
0.007710270117968321,
0.05219721421599388,
0.0203070230782032,
-0.06605470925569534,
-0.06530175358057022,
-0.0647842213511467,
-0.00565014174208045,
0.06907301396131516,
-0.05982866883277893,
-0.01210756879299879,
-0.11039535701274872,
-0.011521132662892342,
0.05741787329316139,
-0.052603065967559814,
0.018963193520903587,
-0.03907877206802368,
-0.03325296938419342,
-0.029370730742812157,
-0.020171141251921654,
0.07719533890485764,
0.03511633723974228,
-0.008852741681039333,
0.001393443439155817,
-0.0531797893345356,
0.041752588003873825,
-0.011185828596353531,
0.03759966790676117,
0.019404424354434013,
-0.012768526561558247,
0.05397344008088112,
-0.0023795648012310266,
0.018505889922380447,
-0.03060264140367508,
0.02266279235482216,
-0.01626116968691349,
0.040454160422086716,
0.05664635822176933,
0.03508027642965317,
-0.003306139027699828,
0.04009315371513367,
0.010261276736855507,
0.004021157044917345,
-0.006808815989643335,
-0.027215922251343727,
0.042593855410814285,
-0.04816808924078941,
0.06225118041038513,
-0.048620644956827164,
0.02232666127383709,
-0.030655961483716965,
-0.00231569935567677,
-0.06974323838949203,
0.00700234342366457,
-0.018482573330402374,
0.007813474163413048,
-0.025352513417601585,
-0.04697345942258835,
-0.01963752694427967,
-0.01586134172976017,
-0.006775615271180868,
-0.05046360567212105,
-0.057505398988723755,
0.003949195146560669,
-2.883156116425941e-33,
-0.06337974965572357,
-0.031848058104515076,
-0.07753712683916092,
0.09269620478153229,
-0.048722878098487854,
-0.007272549904882908,
0.01009045448154211,
-0.011632425710558891,
0.006071402691304684,
-0.0347227118909359,
-0.017093036323785782,
-0.019937429577112198,
-0.07755151391029358,
0.010210058651864529,
0.0380294993519783,
-0.007769282441586256,
-0.02945060469210148,
-0.07205002009868622,
-0.02037731744349003,
0.04582623392343521,
0.05507267266511917,
0.08817712217569351,
0.019897187128663063,
-0.026900436729192734,
-0.047875575721263885,
0.022488532587885857,
-0.055586472153663635,
-0.05048133432865143,
-0.03175835683941841,
-0.0346955806016922,
0.08226200193166733,
0.01447882130742073,
-0.079696424305439,
-0.018578266724944115,
0.09797680377960205,
-0.011046269908547401,
0.05344684422016144,
-0.008131553418934345,
-0.01977946236729622,
0.08114898949861526,
0.10546953976154327,
-0.07919556647539139,
-0.02098057232797146,
0.013850744813680649,
0.01184546947479248,
0.0672035738825798,
-0.11960811167955399,
-0.0007381265168078244,
0.026612192392349243,
-0.05246538668870926,
-0.05316803976893425,
-0.03750051185488701,
-0.0254836268723011,
-0.12625281512737274,
-0.004166977945715189,
-0.10266932845115662,
-0.02754545584321022,
0.034675367176532745,
-0.019554750993847847,
0.048222560435533524,
-0.05778281390666962,
-0.05530044436454773,
0.024599583819508553,
-0.045764658600091934,
-0.038577716797590256,
-0.015824714675545692,
-0.05379750207066536,
0.022408390417695045,
-0.02586667239665985,
-0.05412193015217781,
0.00515526719391346,
-0.07265017926692963,
-0.03875609114766121,
-0.011621453799307346,
0.02333003468811512,
0.05387387424707413,
0.03473030775785446,
-0.035860612988471985,
0.005151430610567331,
-0.042275287210941315,
0.052574291825294495,
0.01557924970984459,
0.0448489785194397,
0.02317061834037304,
-0.015721743926405907,
0.12961256504058838,
-0.04734213650226593,
0.07621642202138901,
0.06494998186826706,
-0.02478666976094246,
-0.03451180458068848,
0.06858668476343155,
-0.030777491629123688,
0.008561789058148861,
0.019560758024454117,
-3.862982111968449e-8,
-0.0056762914173305035,
0.05006156116724014,
0.018809465691447258,
-0.026422692462801933,
0.00720225740224123,
-0.009797221049666405,
-0.029106587171554565,
0.09518369287252426,
0.03552565723657608,
0.012015549466013908,
0.018491191789507866,
0.019434785470366478,
-0.014778981916606426,
0.039382148534059525,
0.027993768453598022,
0.08234624564647675,
-0.06752686947584152,
0.07648911327123642,
-0.009959344752132893,
-0.0734308511018753,
0.05781000107526779,
0.03698411583900452,
0.01441641803830862,
0.02819700352847576,
-0.039642296731472015,
-0.09679566323757172,
0.012958805076777935,
0.0796571671962738,
0.01708504743874073,
-0.03295522555708885,
0.06442420184612274,
0.03716333955526352,
0.03896833583712578,
-0.006493046879768372,
0.06223466619849205,
-0.027653243392705917,
0.04927818104624748,
0.022575151175260544,
0.018533239141106606,
-0.032853350043296814,
-0.017216302454471588,
0.06322291493415833,
0.01868024282157421,
0.0060781012289226055,
-0.011143863201141357,
0.005700007546693087,
-0.05090472474694252,
-0.04001937061548233,
-0.02988174371421337,
-0.11530926823616028,
-0.043124426156282425,
-0.061379943042993546,
-0.04532347992062569,
0.14617770910263062,
0.04305814579129219,
0.012699849903583527,
0.021202564239501953,
-0.07339572161436081,
0.04484924301505089,
0.07331427186727524,
0.02141740545630455,
0.018675435334444046,
0.004252552054822445,
-0.0011962502030655742
] | 0 | 0 | 0 | 0 |
|
Follow My Instruction and Spill the Beans: Scalable Data Extraction from Retrieval-Augmented Generation Systems | https://openreview.net/forum?id=Y4aWwRh25b | [
"Zhenting Qi",
"Hanlin Zhang",
"Eric P. Xing",
"Sham M. Kakade",
"Himabindu Lakkaraju"
] | Poster | Retrieval-Augmented Generation (RAG) improves pre-trained models by incorporating external knowledge at test time to enable customized adaptation.
We study the risk of datastore leakage in Retrieval-In-Context RAG Language Models (LMs). We show that an adversary can exploit LMs' instruction-following capabilities to easily extract text data verbatim from the datastore of RAG systems built with instruction-tuned LMs via prompt injection.
The vulnerability exists for a wide range of modern LMs that span Llama2, Mistral/Mixtral, Vicuna, SOLAR, WizardLM, Qwen1.5, and Platypus2, and the exploitability exacerbates as the model size scales up.
We also study multiple effects of RAG setup on the extractability of data, indicating that following unexpected instructions to regurgitate data can be an outcome of failure in effectively utilizing contexts for modern LMs, and further show that such vulnerability can be greatly mitigated by position bias elimination strategies.
Extending our study to production RAG models, GPTs, we design an attack that can cause datastore leakage with a near-perfect success rate on 25 randomly selected customized GPTs with at most 2 queries, and we extract text data verbatim at a rate of 41\% from a book of 77,000 words and 3\% from a corpus of 1,569,000 words by prompting the GPTs with only 100 queries generated by themselves. | Retrieval-Augmented Generation, Security, Privacy | null | 12,866 | 2402.17840 | [
-0.12206767499446869,
-0.001485571963712573,
0.04975016042590141,
0.0633038580417633,
0.02725687250494957,
-0.005896095186471939,
-0.011185445822775364,
-0.01982720009982586,
0.004384043160825968,
0.017457790672779083,
0.01032839436084032,
-0.004941027145832777,
0.12193254381418228,
-0.07804147154092789,
0.05107814446091652,
-0.010496494360268116,
0.08669334650039673,
0.0853106752038002,
-0.07209118455648422,
-0.0128908921033144,
0.04145476594567299,
0.08446074277162552,
-0.036736246198415756,
0.005276131443679333,
0.00022586123668588698,
-0.009138226509094238,
-0.06073109805583954,
0.009138423949480057,
0.08516770601272583,
-0.08249867707490921,
0.00044804843491874635,
0.04657749831676483,
-0.09003745764493942,
0.06985528767108917,
0.01819099299609661,
0.03617018088698387,
-0.0661744773387909,
0.00519994692876935,
0.001993022160604596,
-0.01820991560816765,
0.010130888782441616,
0.018433284014463425,
-0.018922796472907066,
0.03646567091345787,
0.023273414000868797,
-0.06665191054344177,
0.0025743097066879272,
-0.005524148233234882,
-0.0576314777135849,
-0.027814172208309174,
-0.031284675002098083,
-0.03465946763753891,
0.04858342185616493,
0.019657209515571594,
-0.10400847345590591,
-0.0544012226164341,
0.0770433321595192,
0.035145070403814316,
-0.06801653653383255,
0.046937450766563416,
-0.05889501795172691,
-0.06907297670841217,
-0.09646996855735779,
-0.007926568388938904,
-0.028212685137987137,
-0.04170454666018486,
0.061487581580877304,
0.07871732860803604,
0.006549544166773558,
0.00883918534964323,
-0.08356738090515137,
0.003129061544314027,
-0.10194412618875504,
0.0720820501446724,
0.023142259567975998,
0.10997738689184189,
0.0011569021735340357,
-0.07981812208890915,
0.019438719376921654,
-0.03958595171570778,
0.03767736256122589,
-0.020007023587822914,
0.06234343722462654,
0.029957642778754234,
0.04403603449463844,
0.013627054169774055,
0.0709979236125946,
0.03628898039460182,
0.14803211390972137,
-0.0017758839530870318,
0.007615589536726475,
-0.07764643430709839,
0.08167928457260132,
0.052874818444252014,
0.08673720806837082,
0.057609278708696365,
0.038535021245479584,
-0.035476963967084885,
0.021798232570290565,
0.06117147207260132,
0.009847528301179409,
0.08109618723392487,
0.005264609586447477,
-0.020350288599729538,
-0.02687695622444153,
-0.047140758484601974,
0.027728058397769928,
0.0005468965973705053,
0.01317027397453785,
-0.07906579226255417,
0.06469658017158508,
0.11078603565692902,
0.023966621607542038,
-0.04409614950418472,
-0.030833257362246513,
-0.024484463036060333,
-0.05979647487401962,
-0.10620592534542084,
-0.1027754694223404,
0.03673626109957695,
0.0031888545490801334,
0.029156433418393135,
0.027248451486229897,
0.009223180823028088,
-0.020506886765360832,
-0.026742199435830116,
-0.021067701280117035,
5.110358834444198e-34,
0.019650518894195557,
0.01608305796980858,
-0.023206936195492744,
0.03913269191980362,
0.037987854331731796,
-0.032217226922512054,
0.06326565146446228,
0.06634890288114548,
-0.03596287593245506,
-0.005174863617867231,
-0.001795737654902041,
0.012346194125711918,
-0.02682565152645111,
-0.00296326563693583,
0.07542674243450165,
0.09662479907274246,
0.02072463184595108,
0.05654970929026604,
-0.011968035250902176,
0.025611983612179756,
0.03469870984554291,
-0.04728378728032112,
0.01242239959537983,
-0.1013689935207367,
-0.026493264362215996,
0.08203378319740295,
0.0014081960543990135,
-0.0042266808450222015,
-0.020064959302544594,
0.029040271416306496,
-0.10037683695554733,
-0.04940579831600189,
0.03336917608976364,
0.07156138122081757,
-0.00044885094393976033,
-0.05391924828290939,
-0.03912285342812538,
-0.05132024362683296,
0.0008668615482747555,
0.01294667087495327,
0.03969702497124672,
0.027209436520934105,
0.049666035920381546,
-0.000583054032176733,
-0.027057137340307236,
-0.0950796976685524,
-0.020876778289675713,
0.005797981284558773,
-0.04918161407113075,
-0.017047366127371788,
0.0580938383936882,
0.11896559596061707,
-0.041844405233860016,
0.018590671941637993,
-0.01403438113629818,
0.0034309797920286655,
0.030686181038618088,
-0.03462308645248413,
0.0413399264216423,
0.12195996195077896,
-0.023148953914642334,
0.10788794606924057,
0.05807573348283768,
0.07104181498289108,
0.030548738315701485,
-0.01952584646642208,
-0.005893298890441656,
-0.0340375155210495,
0.04242448881268501,
0.0219479501247406,
-0.05361412093043327,
-0.10670556873083115,
-0.038087766617536545,
0.00005749826232204214,
0.05367353558540344,
-0.05713009461760521,
0.042820241302251816,
-0.02097165212035179,
0.059615347534418106,
-0.04960322752594948,
0.12351354211568832,
-0.07902595400810242,
0.05207442492246628,
-0.01000705175101757,
-0.12034980952739716,
-0.09738380461931229,
-0.021621793508529663,
-0.1353158801794052,
-0.024779092520475388,
-0.022559702396392822,
0.001685743685811758,
-0.019569512456655502,
-0.033857472240924835,
-0.013536900281906128,
0.01227698102593422,
-1.574335175952474e-33,
0.023141536861658096,
-0.048372797667980194,
-0.03142636641860008,
0.03132297471165657,
-0.04040784016251564,
-0.07688266038894653,
-0.0044925943948328495,
0.010746357031166553,
-0.07051621377468109,
-0.07248618453741074,
-0.1045176237821579,
0.021968897432088852,
0.01809076778590679,
-0.01649145595729351,
0.038187213242053986,
0.0050682793371379375,
0.009135116823017597,
-0.022079017013311386,
-0.013520991429686546,
0.06870388239622116,
0.04967048391699791,
0.05354684218764305,
-0.031537991017103195,
0.03707713261246681,
0.016258565708994865,
-0.030385814607143402,
-0.01754237525165081,
0.01654892973601818,
0.04766528308391571,
0.026614002883434296,
-0.03917904570698738,
0.00213521346449852,
0.030967891216278076,
-0.041661687195301056,
-0.050657548010349274,
0.022146277129650116,
0.10803483426570892,
0.006224513985216618,
-0.06496630609035492,
0.06908782571554184,
0.06554364413022995,
0.0005668196245096624,
-0.08976238965988159,
0.00826027151197195,
-0.014529034495353699,
-0.02885514125227928,
-0.0853130966424942,
-0.020839761942625046,
0.07938645780086517,
0.01043378934264183,
0.03941032662987709,
-0.005034126341342926,
0.039597202092409134,
-0.027286795899271965,
-0.028649603947997093,
-0.10605598241090775,
-0.023465652018785477,
-0.040157269686460495,
-0.030660174787044525,
0.0520559623837471,
0.00418137339875102,
-0.02255469560623169,
-0.024372637271881104,
-0.05616426095366478,
0.02807745151221752,
-0.040262818336486816,
-0.018669968470931053,
0.012148967944085598,
-0.0796123594045639,
-0.02253323420882225,
0.05812167376279831,
-0.06644508987665176,
0.016803717240691185,
-0.03369998559355736,
0.06647743284702301,
0.0625605583190918,
-0.041215118020772934,
-0.10425197333097458,
-0.01811668835580349,
-0.007174537982791662,
0.0026501843240112066,
0.005782711319625378,
0.04429503530263901,
0.06547415256500244,
0.029656831175088882,
0.009122289717197418,
0.005866379011422396,
0.014445386826992035,
-0.021169845014810562,
-0.011328155174851418,
-0.06623722612857819,
-0.04894157871603966,
-0.026135757565498352,
0.0889105275273323,
0.02504802495241165,
-4.760064697961752e-8,
-0.01259591430425644,
0.000654448929708451,
0.0335397906601429,
0.1407407969236374,
-0.0016110563883557916,
-0.008366133086383343,
-0.12730096280574799,
0.05579362064599991,
0.023679308593273163,
0.041177619248628616,
0.048197608441114426,
-0.010939800180494785,
-0.02330085262656212,
-0.027621006593108177,
0.03053954243659973,
0.06463181972503662,
-0.027696555480360985,
-0.006414413917809725,
-0.08581136167049408,
-0.02983107976615429,
-0.002120035933330655,
0.0869830846786499,
0.06039668992161751,
-0.06584249436855316,
0.03907174989581108,
-0.033466145396232605,
0.02788367308676243,
0.11584888398647308,
0.08724678307771683,
0.021668672561645508,
0.030316423624753952,
-0.076555997133255,
-0.0140888886526227,
0.0062239146791398525,
-0.03732537478208542,
0.09835874289274216,
-0.039561133831739426,
-0.05007017031311989,
0.03483910858631134,
-0.005456201266497374,
0.01829221285879612,
0.06098250672221184,
-0.024686183780431747,
0.023621097207069397,
-0.03458656743168831,
0.0007926188409328461,
-0.006769493222236633,
-0.09453260898590088,
0.030064977705478668,
-0.05117494612932205,
-0.008973559364676476,
-0.07021176815032959,
-0.04092191532254219,
0.07522156089544296,
0.025471875444054604,
-0.004914384335279465,
0.0016853893175721169,
-0.01429650280624628,
0.014846430160105228,
0.04706525057554245,
0.028707752004265785,
-0.03490978106856346,
0.013898706994950771,
-0.017239999026060104
] | https://github.com/zhentingqi/rag-privacy | 6 | 0 | 0 | 0 |
Collab: Controlled Decoding using Mixture of Agents for LLM Alignment | https://openreview.net/forum?id=7ohlQUbTpp | [
"Souradip Chakraborty",
"Sujay Bhatt",
"Udari Madhushani Sehwag",
"Soumya Suvra Ghosal",
"Jiahao Qiu",
"Mengdi Wang",
"Dinesh Manocha",
"Furong Huang",
"Alec Koppel",
"Sumitra Ganesh"
] | Poster | Alignment of Large Language models (LLMs) is crucial for safe and trustworthy deployment in applications. Reinforcement learning from human feedback (RLHF) has emerged as an effective technique to align LLMs to human preferences, and broader utilities, but it requires updating billions of model parameters which is computationally expensive. Controlled Decoding, by contrast, provides a mechanism for aligning a model at inference time without retraining. However, single-agent decoding approaches often struggle to adapt to diverse tasks due to the complexity and variability inherent in these tasks. To strengthen the test-time performance w.r.t the target task, we propose a mixture of agents-based decoding strategies leveraging the existing off-the-shelf aligned LLM policies. Treating each prior policy as an agent in the spirit of mixture of agent collaboration, we develop a decoding method that allows for inference-time alignment through a token-level selection strategy among multiple agents. For each token, the most suitable LLM is dynamically chosen from a pool of models based on a long-term utility metric. This policy-switching mechanism ensures optimal model selection at each step, enabling efficient collaboration and alignment among LLMs during decoding. Theoretical analysis of our proposed algorithm establishes optimal performance with respect to the target task represented via a target reward, for the given off-the-shelf models. We conduct comprehensive empirical evaluations with open-source aligned models on diverse tasks and preferences, which demonstrates the merits of this approach over single-agent decoding baselines. Notably, COLLAB surpasses the current SoTA decoding strategy, achieving an improvement of {up to 1.56x} in average reward and $71.89\%$ in GPT-4 based win-tie rate. | Alignment, Decoding, RLHF, Transfer Decoding, LLM | Decoding with Mixture of LLM Agents via policy switching | 12,855 | null | [
-0.02653699554502964,
-0.1210142970085144,
-0.01810390315949917,
0.009801780804991722,
0.05213179439306259,
0.032560497522354126,
-0.019650835543870926,
-0.03231023997068405,
0.08278799057006836,
-0.018029669299721718,
-0.039030395448207855,
-0.09031842648983002,
0.0017299484461545944,
0.03341713175177574,
0.020850148051977158,
-0.038093119859695435,
0.10764788836240768,
0.0921829491853714,
-0.09880079329013824,
-0.05114118009805679,
-0.001095556770451367,
0.040388353168964386,
0.03676048666238785,
0.005657305009663105,
0.007225983776152134,
0.003959541209042072,
-0.003148643998429179,
-0.02538810484111309,
0.004882908426225185,
-0.022948333993554115,
0.08675653487443924,
-0.0010876802261918783,
0.01256936602294445,
-0.010339171625673771,
-0.056688062846660614,
0.07329927384853363,
-0.12225867062807083,
-0.03265657648444176,
0.03527849540114403,
-0.03718800097703934,
-0.054415278136730194,
-0.012572807259857655,
0.0001269056083401665,
-0.006797492038458586,
0.029448064044117928,
-0.004603034816682339,
0.01357862539589405,
-0.011433111503720284,
0.0004461581411305815,
0.03446535766124725,
-0.01537724956870079,
-0.013273918069899082,
0.03545517101883888,
0.03778228163719177,
-0.06574341654777527,
0.0068346350453794,
0.040857598185539246,
0.042681265622377396,
0.023321140557527542,
-0.09342110902070999,
-0.09366922825574875,
-0.13484707474708557,
-0.002883634530007839,
0.04184506833553314,
-0.06384072452783585,
-0.02097940817475319,
-0.035468000918626785,
0.07587692141532898,
-0.050028473138809204,
0.02979147620499134,
-0.07878528535366058,
0.017028601840138435,
0.046406056731939316,
0.007943431846797466,
-0.01140834391117096,
0.0634135827422142,
0.04194486886262894,
-0.044273264706134796,
0.14384573698043823,
-0.05279232934117317,
0.003904694924131036,
0.046005625277757645,
0.08815816044807434,
0.0004176860093139112,
0.09559398144483566,
0.004097737371921539,
0.07014277577400208,
0.0674019604921341,
0.07453460991382599,
-0.006028494331985712,
-0.04207945242524147,
-0.026503371074795723,
0.09833622723817825,
0.0320674329996109,
-0.005507101304829121,
0.07699207216501236,
-0.029620002955198288,
-0.04844167083501816,
-0.04428170621395111,
0.08844099193811417,
-0.05926767364144325,
0.07221079617738724,
0.00625685416162014,
-0.0716123953461647,
-0.05988452583551407,
-0.04285310208797455,
0.00625034561380744,
0.026610439643263817,
0.04887325316667557,
-0.06972501426935196,
0.030841542407870293,
0.03252226859331131,
-0.010752462781965733,
0.10023622214794159,
0.01891525276005268,
-0.01140977256000042,
0.02644612453877926,
-0.04105944558978081,
-0.0001259112177649513,
0.013886386528611183,
-0.009224502369761467,
-0.027914805337786674,
-0.0019449078245088458,
-0.019871672615408897,
0.04335109889507294,
0.016334256157279015,
0.0598849356174469,
-3.6563944636869185e-34,
0.056008294224739075,
0.10910964012145996,
0.02680458500981331,
0.07568445056676865,
0.02200615219771862,
-0.04042353108525276,
0.05470984801650047,
0.09355917572975159,
-0.045891545712947845,
0.028105081990361214,
0.01916602812707424,
-0.005528975278139114,
-0.0640321895480156,
0.07412069290876389,
-0.030841201543807983,
-0.0019187376601621509,
-0.08217023313045502,
0.10277511924505234,
-0.09850206971168518,
-0.03945362940430641,
0.07913961261510849,
0.02494986541569233,
0.01948487013578415,
-0.04375527426600456,
0.019329747185111046,
-0.048064105212688446,
0.04351470246911049,
-0.05082833021879196,
0.0023589711636304855,
-0.005290670786052942,
-0.12309518456459045,
-0.019373904913663864,
-0.021449293941259384,
0.05152014642953873,
0.019214939326047897,
-0.031498104333877563,
-0.0606350339949131,
-0.043944407254457474,
-0.004942670464515686,
-0.0004395860596559942,
0.04124976322054863,
0.026611242443323135,
0.052848707884550095,
-0.00877298042178154,
-0.05393736809492111,
-0.11833233386278152,
-0.07970735430717468,
-0.026684634387493134,
-0.017494898289442062,
-0.062482867389917374,
0.03593655303120613,
0.06031408905982971,
-0.011468694545328617,
-0.039230577647686005,
0.009229805320501328,
0.011160836555063725,
-0.008416497148573399,
0.04514351487159729,
-0.017862342298030853,
0.05504024401307106,
-0.008380222134292126,
0.006849770434200764,
0.0013829870149493217,
0.08981289714574814,
0.0809038057923317,
0.013888828456401825,
-0.04549679160118103,
0.06957707554101944,
0.015465352684259415,
-0.021756984293460846,
-0.012617646716535091,
-0.0009227402042597532,
-0.001867301994934678,
-0.043994635343551636,
0.02193567529320717,
0.013868344947695732,
0.09238234162330627,
-0.0515589639544487,
-0.0012711180606856942,
0.002645641565322876,
-0.020266208797693253,
-0.03792709857225418,
-0.06697947531938553,
-0.029388824477791786,
-0.032512400299310684,
-0.1257346272468567,
0.018477272242307663,
-0.05751420557498932,
-0.022327838465571404,
0.0014152154326438904,
-0.06386759132146835,
-0.008457256481051445,
-0.005272014997899532,
0.03385802358388901,
0.03325790539383888,
-4.740702595309052e-34,
-0.01911081001162529,
-0.058807868510484695,
0.03853501379489899,
0.06613332033157349,
-0.02084583416581154,
-0.07265458256006241,
-0.015158839523792267,
-0.01346102450042963,
-0.03452439606189728,
-0.015459722839295864,
-0.10227736085653305,
-0.07532337307929993,
0.06329167634248734,
-0.04571238160133362,
0.025490326806902885,
-0.001387373311445117,
0.040669407695531845,
0.029517987743020058,
0.05969936028122902,
0.042381830513477325,
0.030191602185368538,
0.06817150861024857,
-0.10044945031404495,
0.02915939688682556,
0.004137557465583086,
0.024280767887830734,
0.031109532341361046,
0.042767349630594254,
0.06063012406229973,
0.06469465047121048,
-0.08572545647621155,
0.0524451807141304,
-0.04092918336391449,
-0.02306641824543476,
-0.09870177507400513,
0.1083051785826683,
0.007406227756291628,
0.030224069952964783,
-0.00004944025931763463,
0.12374736368656158,
0.04608660191297531,
-0.09328749030828476,
-0.09973941743373871,
0.004948489833623171,
-0.006873671431094408,
0.015982994809746742,
-0.08020564168691635,
-0.04481523483991623,
-0.03525819629430771,
-0.030311472713947296,
-0.015554497018456459,
0.042986657470464706,
-0.07588469237089157,
-0.0024256391916424036,
-0.057338789105415344,
-0.030471058562397957,
-0.043613456189632416,
-0.019326793029904366,
0.015315398573875427,
-0.021916838362812996,
0.0032916306518018246,
0.04815302789211273,
0.08704014122486115,
-0.02238324284553528,
-0.005176078528165817,
-0.00781994964927435,
0.033122722059488297,
-0.010281185619533062,
0.013444083742797375,
0.010215163230895996,
0.08171515166759491,
-0.10266568511724472,
0.0013296843972057104,
0.01786484755575657,
0.09056949615478516,
-0.058800265192985535,
0.00417923042550683,
-0.06251814216375351,
0.013091381639242172,
-0.06521132588386536,
0.05256962403655052,
0.06544540822505951,
-0.023666393011808395,
0.06624066084623337,
0.0319940485060215,
0.07887199521064758,
0.04643309861421585,
0.14591725170612335,
0.02259560488164425,
0.013740443624556065,
-0.02670726552605629,
0.004523251671344042,
0.05081506446003914,
0.02921631932258606,
-0.0895344614982605,
-4.653889007499856e-8,
-0.033204346895217896,
0.006366638466715813,
0.03348889201879501,
0.04940599575638771,
-0.08291810005903244,
-0.056355081498622894,
-0.053305238485336304,
-0.026791898533701897,
0.041434578597545624,
0.0014427641872316599,
0.07099849730730057,
0.061936307698488235,
0.00018420918786432594,
-0.0874830037355423,
-0.031550511717796326,
0.0794387087225914,
-0.08016151934862137,
-0.005550324451178312,
-0.01729811355471611,
-0.000029070255550323054,
0.017650246620178223,
0.05691882595419884,
0.01582314632833004,
0.00689317099750042,
0.054296135902404785,
-0.039437536150217056,
-0.009736848063766956,
0.06269604712724686,
0.06131194904446602,
-0.03350670635700226,
0.016432596370577812,
-0.004252897575497627,
-0.012585382908582687,
-0.022606324404478073,
-0.018456967547535896,
0.0688437819480896,
-0.012143627740442753,
-0.02449898049235344,
0.05665186420083046,
-0.025615889579057693,
0.006685934960842133,
0.05442308261990547,
-0.06609451770782471,
0.0020648795180022717,
-0.02938493713736534,
-0.029379285871982574,
-0.06786008924245834,
-0.08351565152406693,
0.0518270805478096,
-0.08894681930541992,
-0.005504341330379248,
-0.05319041758775711,
-0.01313060149550438,
0.05092519149184227,
0.026948174461722374,
0.010350923053920269,
0.014779491350054741,
-0.07342441380023956,
0.13801811635494232,
0.004754908382892609,
0.010494427755475044,
-0.006497628521174192,
-0.05642514303326607,
-0.05849580094218254
] | 0 | 0 | 0 | 0 |
|
Finally Rank-Breaking Conquers MNL Bandits: Optimal and Efficient Algorithms for MNL Assortment | https://openreview.net/forum?id=kx8i1yfkRX | [
"Aadirupa Saha",
"Pierre Gaillard"
] | Poster | We address the problem of active online assortment optimization problem with preference feedback, which is a framework for modeling user choices and subsetwise utility maximization. The framework is useful in various real-world applications including ad placement, online retail, recommender systems, and fine-tuning language models, amongst many others. The problem, although has been studied in the past, lacks an intuitive and practical solution approach with simultaneously efficient algorithm and optimal regret guarantee. E.g., popularly used assortment selection algorithms often require the presence of a ``strong reference" which is always included in the choice sets, further they are also designed to offer the same assortments repeatedly until the reference item gets selected---all such requirements are quite unrealistic for practical applications. In this paper, we designed efficient algorithms for the problem of regret minimization in assortment selection with \emph{Plackett Luce} (PL) based user choices. We designed a novel concentration guarantee for estimating the score parameters of the PL model using `\emph{Pairwise Rank-Breaking}', which builds the foundation of our proposed algorithms. Moreover, our methods are practical, provably optimal, and devoid of the aforementioned limitations of the existing methods. | Active online assortment optimization, Preference feedback, Subsetwise utility maximization, Assortment selection algorithms, Plackett Luce model, Regret minimization, Pairwise Rank-Breaking, Concentration guarantee, Practical algorithms, Empirical evaluations | null | 12,846 | null | [
-0.10453345626592636,
-0.05317026376724243,
0.01797662302851677,
-0.0016962814843282104,
0.02693992853164673,
0.026859285309910774,
0.08645455539226532,
0.07853057980537415,
0.006353755947202444,
-0.03343702852725983,
0.03453727066516876,
0.012512757442891598,
0.09034435451030731,
-0.023592645302414894,
0.02790195681154728,
0.0473460778594017,
0.10933859646320343,
0.02695246785879135,
-0.019840555265545845,
-0.042587704956531525,
0.04107692837715149,
-0.025678634643554688,
0.01857510767877102,
-0.07657856494188309,
-0.013995128683745861,
-0.0910053476691246,
-0.03946202993392944,
0.02634587325155735,
-0.0439402274787426,
-0.03272103890776634,
-0.008368542417883873,
0.06702426820993423,
0.04296541213989258,
-0.0485217347741127,
-0.039976052939891815,
-0.04775969311594963,
-0.08507555723190308,
-0.08349292725324631,
0.0017814140301197767,
0.027306359261274338,
0.0011006476124748588,
0.028257250785827637,
-0.0833454355597496,
0.019243502989411354,
-0.022549400106072426,
-0.017267487943172455,
0.0038484828546643257,
0.04331890121102333,
0.013937191106379032,
0.09144074469804764,
-0.024768400937318802,
0.050472307950258255,
-0.03332329913973808,
0.02591577358543873,
0.0059369527734816074,
0.0016416802536696196,
-0.021655837073922157,
0.007911448366940022,
-0.015488398261368275,
-0.031096210703253746,
-0.045057907700538635,
-0.12167832255363464,
-0.04523902013897896,
-0.04478183016180992,
0.005572057329118252,
-0.021220814436674118,
-0.028309747576713562,
-0.0005635912530124187,
-0.07221411913633347,
0.09260355681180954,
-0.0165870264172554,
0.03446946293115616,
-0.037022508680820465,
0.07540150731801987,
0.05216654762625694,
0.04353752359747887,
0.08090028911828995,
-0.11136437207460403,
-0.0067141917534172535,
0.011592152528464794,
-0.11831779778003693,
0.009384727105498314,
-0.015008433721959591,
-0.0382077731192112,
0.043247126042842865,
-0.022020041942596436,
0.07202254235744476,
-0.056043464690446854,
0.14734116196632385,
-0.02530771680176258,
-0.031921666115522385,
0.011325675062835217,
0.01663965918123722,
0.004510532133281231,
0.012047904543578625,
0.04959043860435486,
-0.024221990257501602,
-0.02881956845521927,
0.022603580728173256,
0.047672856599092484,
-0.0032099331729114056,
0.06784258782863617,
0.04132035747170448,
-0.079128697514534,
-0.03623305633664131,
-0.09793918579816818,
-0.03584735095500946,
0.11320701986551285,
0.036858610808849335,
-0.09196341782808304,
-0.04301723465323448,
0.06182651221752167,
0.0372399166226387,
0.06804373115301132,
-0.03550703451037407,
-0.006903282832354307,
0.057190902531147,
-0.013947280123829842,
0.06167937442660332,
0.022232696413993835,
0.005971461534500122,
0.060955602675676346,
0.1047416627407074,
-0.039176955819129944,
-0.01650521345436573,
0.06446622312068939,
0.024942671880126,
1.1195338402254386e-33,
-0.03460049256682396,
-0.009079603478312492,
0.019993199035525322,
-0.07157431542873383,
-0.005613798275589943,
-0.004726821556687355,
0.06656365096569061,
0.03487614542245865,
-0.0671180710196495,
-0.000818216591142118,
0.0030637194868177176,
0.0043471986427903175,
-0.05979614332318306,
0.10438070446252823,
0.13022683560848236,
-0.05030346289277077,
0.016856195405125618,
0.044093307107686996,
0.003514534095302224,
-0.08195051550865173,
0.07456927001476288,
0.0332266166806221,
0.02061043679714203,
-0.04813109338283539,
-0.041823454201221466,
-0.05673199146986008,
-0.01496297400444746,
0.010507297702133656,
0.027893850579857826,
0.0024371626786887646,
-0.060379788279533386,
-0.04386267811059952,
0.027017513290047646,
-0.014285307377576828,
-0.031907666474580765,
0.015668198466300964,
-0.10628882050514221,
0.009433089755475521,
0.007202972657978535,
-0.022403476759791374,
0.00028754849336110055,
0.016998371109366417,
0.05076771229505539,
0.08387148380279541,
-0.09750938415527344,
-0.013152766972780228,
-0.06413183361291885,
0.06378180533647537,
-0.01441151276230812,
0.026533622294664383,
-0.014658229425549507,
-0.029047662392258644,
0.021132580935955048,
0.11061199009418488,
-0.09737672656774521,
-0.022820577025413513,
-0.04698354750871658,
-0.008167984895408154,
0.013023113831877708,
-0.005950983613729477,
0.045579444617033005,
-0.02797938697040081,
0.06437937170267105,
-0.0013430551625788212,
-0.030861997976899147,
-0.026488222181797028,
-0.014844680204987526,
-0.0795503556728363,
0.05043284222483635,
-0.007174378726631403,
0.024320317432284355,
0.04436742514371872,
0.04806177318096161,
-0.06587286293506622,
0.007195130921900272,
0.007835707627236843,
0.052916403859853745,
0.07228419929742813,
0.05372822657227516,
-0.09697990864515305,
0.0433577299118042,
-0.006533314473927021,
0.011027188040316105,
-0.09432335942983627,
-0.042976077646017075,
-0.041996896266937256,
0.00430401973426342,
-0.06144342198967934,
-0.01686432957649231,
-0.044203635305166245,
-0.007529980503022671,
0.0946112722158432,
-0.0645269975066185,
-0.01916702836751938,
0.061275843530893326,
-1.1040160290976382e-33,
-0.008579561486840248,
-0.07830292731523514,
0.17139825224876404,
0.040795110166072845,
0.0357116237282753,
-0.03876803070306778,
-0.05433070287108421,
-0.020146425813436508,
-0.04328388720750809,
-0.06347639113664627,
-0.08836561441421509,
0.034588586539030075,
0.08070363104343414,
0.01333759818226099,
0.030122127383947372,
0.047855194658041,
-0.04452976956963539,
0.02970937080681324,
0.04676472768187523,
0.02312677539885044,
0.002243678318336606,
0.044947125017642975,
-0.07433368265628815,
0.05860510468482971,
-0.03553997725248337,
-0.0022548427805304527,
0.00915217213332653,
0.045532915741205215,
-0.05005122721195221,
-0.08575877547264099,
-0.009766736999154091,
-0.025006385520100594,
-0.0864461362361908,
0.06499172002077103,
0.05654843896627426,
0.008473603054881096,
-0.0004312570672482252,
0.04575628042221069,
-0.01297720056027174,
0.13607366383075714,
0.0340699627995491,
-0.007101336028426886,
-0.061569150537252426,
-0.022571012377738953,
0.05174955725669861,
-0.01805999130010605,
-0.05206822603940964,
-0.0015046679181978106,
0.06809313595294952,
-0.038784097880125046,
-0.046817950904369354,
0.07865644246339798,
-0.011708991602063179,
0.0683308020234108,
-0.05505857244133949,
-0.021010421216487885,
-0.019561901688575745,
-0.015056127682328224,
0.10862374305725098,
0.013990775682032108,
-0.05197041481733322,
0.04135706275701523,
0.0005828768480569124,
0.009265069849789143,
0.0278937891125679,
-0.08398155868053436,
0.07355491071939468,
-0.03770746290683746,
-0.05542689189314842,
0.014485613442957401,
-0.011836329475045204,
0.007321187760680914,
0.038497038185596466,
-0.08673172444105148,
-0.06426666676998138,
0.018874501809477806,
-0.0018482737941667438,
0.03868213668465614,
-0.0004891477874480188,
-0.032743290066719055,
0.013951546512544155,
0.03590827062726021,
0.025383859872817993,
0.013296631164848804,
0.008185851387679577,
0.011005247943103313,
0.08359424024820328,
0.05424198508262634,
0.001611116575077176,
-0.06299837678670883,
0.0010746236657723784,
-0.07320693135261536,
0.049562521278858185,
0.023403553292155266,
-0.017867915332317352,
-4.615284510123274e-8,
-0.004825308918952942,
-0.02749583125114441,
0.001750678988173604,
0.12338337302207947,
0.08149741590023041,
-0.017524924129247665,
-0.0076662227511405945,
0.06948179751634598,
-0.043173156678676605,
0.030782241374254227,
0.012358181178569794,
0.05767141655087471,
-0.025802716612815857,
-0.004444994498044252,
-0.04583576321601868,
0.011219717562198639,
0.00657220184803009,
0.04236802086234093,
0.005238679703325033,
-0.020531265065073967,
-0.00859276857227087,
0.029750920832157135,
0.11267021298408508,
0.027411069720983505,
0.0007032976718619466,
0.05525260418653488,
0.05060921981930733,
0.01673353649675846,
0.09270116686820984,
0.09504415839910507,
0.04910334572196007,
-0.013398478738963604,
0.035138994455337524,
-0.03243459016084671,
0.0304863303899765,
0.03738698735833168,
-0.11017239093780518,
-0.046650972217321396,
-0.07354342937469482,
-0.06709890067577362,
0.04990169405937195,
-0.0438259094953537,
-0.010255865752696991,
-0.028656158596277237,
0.08445358276367188,
-0.02541208639740944,
-0.035465363413095474,
-0.06415338069200516,
0.051025014370679855,
0.04018353670835495,
-0.04042579233646393,
-0.07401043176651001,
0.018900880590081215,
-0.04529871419072151,
0.035995010286569595,
-0.05260622873902321,
-0.07754427194595337,
-0.036176230758428574,
0.0910431519150734,
0.02556237205862999,
0.05670258402824402,
-0.050758544355630875,
-0.05120362713932991,
0.03208069130778313
] | 0 | 0 | 0 | 0 |
|
Efficient Top-m Data Values Identification for Data Selection | https://openreview.net/forum?id=lOfuvmi2HT | [
"Xiaoqiang Lin",
"Xinyi Xu",
"See-Kiong Ng",
"Bryan Kian Hsiang Low"
] | Poster | Data valuation has found many real-world applications, e.g., data pricing and data selection. However, the most adopted approach -- Shapley value (SV) -- is computationally expensive due to the large number of model trainings required. Fortunately, most applications (e.g., data selection) require only knowing the $m$ data points with the highest data values (i.e., top-$m$ data values), which implies the potential for fewer model trainings as exact data values are not required. Existing work formulates top-$m$ Shapley value identification as top-$m$ arms identification in multi-armed bandits (MAB). However, the proposed approach falls short because it does not utilize data features to predict data values, a method that has been shown empirically to be effective. A recent top-$m$ arms identification work does consider the use of arm features while assuming a linear relationship between arm features and rewards, which is often not satisfied in data valuation. To this end, we propose the GPGapE algorithm that uses the Gaussian process to model the \emph{non-linear} mapping from data features to data values, removing the linear assumption. We theoretically analyze the correctness and stopping iteration of GPGapE in finding an $(\epsilon, \delta)$-approximation to the top-$m$ data values. We further improve the computational efficiency, by calculating data values using small data subsets to reduce the computation cost of model training. We empirically demonstrate that GPGapE outperforms other baselines in top-$m$ data values identification, noisy data detection, and data subset selection on real-world datasets. We also demonstrate the efficiency of our GPGapE in data selection for large language model fine-tuning. | data selection, data valuation, top-m arms identification | We propose an efficient top-m data values identification algorithm for data selection with both theoretical results and empirical efficiency | 12,845 | null | [
-0.09759826213121414,
-0.031133335083723068,
-0.024323290213942528,
-0.015778684988617897,
0.04587056487798691,
-0.003803134895861149,
0.08510356396436691,
0.011005646549165249,
0.007380607072263956,
0.002624242100864649,
-0.06661651283502579,
0.0018633228028193116,
0.1076488345861435,
-0.014150301925837994,
0.011044124141335487,
0.04605036973953247,
0.1213836818933487,
0.06130259484052658,
-0.005280194338411093,
-0.023005923256278038,
-0.07032101601362228,
0.037241436541080475,
0.04571682959794998,
-0.006014806218445301,
0.025490229949355125,
-0.07075347751379013,
0.00930743757635355,
0.043398160487413406,
-0.02019856870174408,
-0.02743416465818882,
0.04365122318267822,
-0.005703945178538561,
0.0054364632815122604,
-0.04797414317727089,
-0.046328842639923096,
0.012414410710334778,
-0.07240951061248779,
-0.02914494462311268,
-0.026532109826803207,
-0.007799812126904726,
0.00878197606652975,
0.013094935566186905,
-0.02681027539074421,
0.007039637304842472,
0.023394808173179626,
0.023470349609851837,
-0.021615976467728615,
-0.0357915535569191,
0.022487351670861244,
0.008276308886706829,
-0.07288314402103424,
0.012040503323078156,
-0.015258533880114555,
0.01930840313434601,
-0.024514373391866684,
-0.01993013732135296,
0.0013127014972269535,
-0.08827996253967285,
0.046392764896154404,
0.057967204600572586,
-0.04275109991431236,
-0.06066371500492096,
-0.0033725877292454243,
-0.021306945011019707,
0.023316752165555954,
-0.010061298497021198,
0.017560947686433792,
0.01733390800654888,
-0.08996491879224777,
0.016683675348758698,
0.07151349633932114,
0.027132973074913025,
-0.11558163911104202,
-0.0026708426885306835,
0.008773209527134895,
0.06699242442846298,
0.040428098291158676,
0.03605515509843826,
0.08245725929737091,
0.010378866456449032,
-0.10386419296264648,
-0.0002682977938093245,
-0.011861825361847878,
-0.027741892263293266,
0.07568155974149704,
0.00014082645066082478,
0.0502835251390934,
0.06164645776152611,
0.11203578859567642,
0.019053205847740173,
0.00016291257634293288,
0.01753127947449684,
-0.04331066086888313,
-0.033371564000844955,
0.050351306796073914,
0.027633657678961754,
-0.01649492420256138,
-0.02981170266866684,
-0.019504176452755928,
0.05623026564717293,
-0.009537003934383392,
0.05837634950876236,
0.04497635364532471,
-0.0456840954720974,
0.03644559159874916,
-0.05181228369474411,
0.047669295221567154,
0.061346814036369324,
0.044322434812784195,
-0.08186385780572891,
-0.005890474189072847,
-0.021805409342050552,
-0.001206147950142622,
0.05623874068260193,
-0.00872922595590353,
0.04034231975674629,
-0.018954474478960037,
-0.0270259790122509,
-0.05143264681100845,
0.05489121377468109,
0.0189628005027771,
-0.02848025970160961,
0.11703221499919891,
0.030550731346011162,
-0.02429085038602352,
0.05422144755721092,
-0.08597937226295471,
1.553141380149133e-33,
0.03995024040341377,
0.011313410475850105,
0.06338826566934586,
-0.10211247205734253,
-0.02324581891298294,
-0.03732931241393089,
-0.006192223634570837,
-0.027256282046437263,
-0.026126766577363014,
0.1187998354434967,
-0.06225597485899925,
0.11092030256986618,
-0.032837603241205215,
0.05712093040347099,
0.08485186845064163,
0.040859561413526535,
-0.015437068417668343,
0.036756522953510284,
-0.02020697295665741,
-0.03501012176275253,
0.06070737913250923,
-0.01299066562205553,
0.039721451699733734,
-0.06939736008644104,
-0.0024630334228277206,
-0.0010038757463917136,
-0.08205563575029373,
0.012141634710133076,
0.009146362543106079,
0.026504576206207275,
-0.038805752992630005,
-0.016879744827747345,
-0.022912155836820602,
0.00020086928270757198,
0.02096698433160782,
-0.008340488187968731,
-0.060890309512615204,
-0.03061787225306034,
-0.009008505381643772,
0.06760673969984055,
0.028420690447092056,
0.016581540927290916,
0.03658667206764221,
-0.03474752604961395,
-0.12332411855459213,
-0.046088870614767075,
-0.03460180386900902,
-0.07455694675445557,
-0.059566840529441833,
-0.046751998364925385,
-0.021473340690135956,
-0.02481287531554699,
-0.02642042003571987,
-0.0032629813067615032,
-0.1321876049041748,
0.0337848961353302,
-0.03598608449101448,
0.02855127863585949,
-0.043292541056871414,
0.01715615950524807,
-0.003273629117757082,
-0.03524061292409897,
0.022580476477742195,
-0.009598448872566223,
-0.07493405044078827,
-0.014778322540223598,
-0.005645178724080324,
0.07645141333341599,
0.06390940397977829,
-0.0007729975623078644,
0.05695052444934845,
-0.04831056669354439,
0.03185166418552399,
-0.030604416504502296,
0.08328943699598312,
-0.0460980087518692,
0.12200873345136642,
-0.022763852030038834,
0.025594277307391167,
-0.02256474643945694,
0.018690932542085648,
0.024880046024918556,
0.03671440854668617,
-0.03937399014830589,
0.006132589187473059,
0.012879330664873123,
0.015094650909304619,
-0.12491849809885025,
-0.0881761983036995,
-0.0464903824031353,
-0.1373978555202484,
0.08864440023899078,
-0.12696339190006256,
-0.02692994847893715,
-0.04205656051635742,
-2.1448598897561328e-33,
0.0070416247472167015,
-0.000644553336314857,
0.09332257509231567,
0.0616929866373539,
0.010095121338963509,
-0.06650277972221375,
0.05615744739770889,
-0.028646748512983322,
-0.011043432168662548,
-0.03643948584794998,
-0.026610039174556732,
-0.023546185344457626,
0.04073592647910118,
0.008764130994677544,
0.1052657887339592,
-0.0052881306037306786,
-0.030726537108421326,
-0.04020853713154793,
0.014379260130226612,
-0.017699778079986572,
0.002456534653902054,
0.04192144796252251,
-0.027342434972524643,
0.007501410320401192,
-0.028919002041220665,
-0.006148553919047117,
-0.04464581981301308,
-0.00426090694963932,
0.00822014082223177,
-0.013971996493637562,
-0.04101210832595825,
-0.019507352262735367,
-0.04290067031979561,
-0.015994843095541,
-0.08929914981126785,
0.017603524029254913,
0.11758537590503693,
0.029188591986894608,
-0.037144675850868225,
0.18135803937911987,
0.0637463629245758,
0.0205055121332407,
-0.004447492770850658,
0.033707451075315475,
0.0232516098767519,
0.02645660750567913,
-0.06999804079532623,
0.08187418431043625,
0.07680468261241913,
-0.012274017557501793,
-0.0026014549657702446,
0.061427559703588486,
-0.09328935295343399,
0.12193938344717026,
-0.04326382651925087,
-0.05499594658613205,
0.010057237930595875,
0.007805920206010342,
0.008659368380904198,
0.04645472764968872,
-0.036647696048021317,
0.08423463255167007,
-0.028039176017045975,
0.027448788285255432,
0.014797370880842209,
0.0059273215010762215,
-0.02222004346549511,
-0.030155088752508163,
-0.10092882066965103,
-0.03700967878103256,
-0.04343724623322487,
-0.02597092278301716,
0.0738685131072998,
-0.04342523589730263,
-0.0748646929860115,
0.08752742409706116,
-0.013923483900725842,
-0.023717379197478294,
0.026519451290369034,
-0.06690247356891632,
0.02657247520983219,
0.03996264189481735,
0.013986156322062016,
0.02850249595940113,
0.009634355083107948,
0.03379197418689728,
0.05684799700975418,
0.0015788997989147902,
-0.018085259944200516,
-0.1274159699678421,
-0.05615737661719322,
-0.04121646285057068,
-0.018507400527596474,
0.06940983980894089,
-0.07782609760761261,
-4.697096400718692e-8,
-0.03161022439599037,
0.06810034066438675,
0.04219042882323265,
0.059531308710575104,
0.054372530430555344,
0.07697655260562897,
-0.04096008464694023,
0.028484975919127464,
0.014099448919296265,
-0.03522748872637749,
0.10099371522665024,
-0.011234156787395477,
-0.05087306350469589,
-0.01797802746295929,
-0.01920893043279648,
0.04850569739937782,
-0.016500180587172508,
-0.030987145379185677,
0.004187009762972593,
0.014453109353780746,
-0.035847511142492294,
0.04285890609025955,
0.06116701290011406,
-0.05160633847117424,
0.04921116307377815,
-0.06743458658456802,
-0.08657114952802658,
0.06162671372294426,
0.062195539474487305,
0.09929921478033066,
-0.021282901987433434,
-0.0013179874513298273,
0.07848306745290756,
-0.03215567395091057,
0.1612222045660019,
0.026969581842422485,
-0.06757863610982895,
-0.005442488007247448,
-0.0033623899798840284,
-0.020297015085816383,
-0.009077909402549267,
0.06472203135490417,
-0.020551413297653198,
-0.05945262312889099,
0.04040851816534996,
0.0713859498500824,
-0.011053051799535751,
-0.042243592441082,
0.015148486942052841,
-0.0026910637971013784,
0.046594709157943726,
-0.03414047136902809,
0.04331006854772568,
0.0919867679476738,
0.05206925421953201,
-0.02609935589134693,
-0.005596808157861233,
-0.06747662276029587,
0.012840799055993557,
-0.010493073612451553,
0.03833555057644844,
-0.12387388944625854,
-0.07571934908628464,
0.030515020713210106
] | 0 | 0 | 0 | 0 |
|
ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery | https://openreview.net/forum?id=6z4YKr0GK6 | [
"Ziru Chen",
"Shijie Chen",
"Yuting Ning",
"Qianheng Zhang",
"Boshi Wang",
"Botao Yu",
"Yifei Li",
"Zeyi Liao",
"Chen Wei",
"Zitong Lu",
"Vishal Dey",
"Mingyi Xue",
"Frazier N. Baker",
"Benjamin Burns",
"Daniel Adu-Ampratwum",
"Xuhui Huang",
"Xia Ning",
"Song Gao",
"Yu Su",
"Huan Sun"
] | Poster | The advancements of language language models (LLMs) have piqued growing interest in developing LLM-based language agents to automate scientific discovery end-to-end, which has sparked both excitement and skepticism about the true capabilities of such agents. In this work, we argue that for an agent to fully automate scientific discovery, it must be able to complete all essential tasks in the workflow. Thus, we call for rigorous assessment of agents on individual tasks in a scientific workflow before making bold claims on end-to-end automation. To this end, we present ScienceAgentBench, a new benchmark for evaluating language agents for data-driven scientific discovery. To ensure the scientific authenticity and real-world relevance of our benchmark, we extract 102 tasks from 44 peer-reviewed publications in four disciplines and engage nine subject matter experts to validate them. We unify the target output for every task to a self-contained Python program file and employ an array of evaluation metrics to examine the generated programs, execution results, and costs. Each task goes through multiple rounds of manual validation by annotators and subject matter experts to ensure its annotation quality and scientific plausibility. We also propose two effective strategies to mitigate data contamination concerns. Using our benchmark, we evaluate five open-weight and proprietary LLMs, each with three frameworks: direct prompting, OpenHands, and self-debug. Given three attempts for each task, the best-performing agent can only solve 32.4% of the tasks independently and 34.3% with expert-provided knowledge. These results underscore the limited capacities of current language agents in generating code for data-driven discovery, let alone end-to-end automation for scientific research. | Benchmark, Evaluation, Large Language Model, Language Agent, AI for Science, Code Generation, Task Automation | We present ScienceAgentBench, a new benchmark for rigorously measuring progress towards developing language agents to assist human scientists in data-driven scientific discovery. | 12,844 | 2410.05080 | [
-0.10281448066234589,
-0.004290874116122723,
-0.05446537584066391,
0.027763739228248596,
0.05923333391547203,
-0.04053396359086037,
-0.039508163928985596,
-0.006333488039672375,
0.010545169934630394,
0.03484760969877243,
-0.09858813881874084,
-0.12705032527446747,
0.0360947884619236,
0.057692065834999084,
-0.03638216480612755,
0.039825085550546646,
0.05045992508530617,
-0.007976260036230087,
-0.028697198256850243,
-0.07109595835208893,
-0.0074042147025465965,
0.08524766564369202,
0.08468230813741684,
0.017725571990013123,
0.02020360715687275,
0.014891598373651505,
-0.0505920946598053,
-0.07687665522098541,
-0.017961649224162102,
0.047788891941308975,
0.02742711454629898,
0.06775417923927307,
0.0040141684003174305,
0.048007749021053314,
0.0394410602748394,
0.03225209563970566,
-0.021234536543488503,
0.00043718493543565273,
-0.01475675217807293,
0.01802244782447815,
0.024748098105192184,
-0.057252660393714905,
-0.019562916830182076,
-0.0010282930452376604,
0.030796833336353302,
-0.0170365609228611,
-0.03652488812804222,
-0.03535599634051323,
-0.01841653324663639,
0.026256805285811424,
-0.10781292617321014,
-0.10960331559181213,
0.06244255229830742,
0.012600825168192387,
-0.03912461921572685,
0.013390833511948586,
0.11184999346733093,
-0.03789687156677246,
-0.008053632453083992,
-0.10221627354621887,
0.005243350751698017,
-0.0985468253493309,
-0.048978738486766815,
-0.0064406199380755424,
0.0006954430718906224,
-0.021317528560757637,
-0.04858192428946495,
0.01893123984336853,
0.004357193596661091,
0.017194272950291634,
-0.0631910115480423,
0.07351846992969513,
0.014004800468683243,
0.05943398177623749,
0.016409873962402344,
-0.02011019177734852,
0.004019503481686115,
-0.002126323990523815,
0.11018051207065582,
-0.08617764711380005,
-0.06048831343650818,
-0.047716591507196426,
-0.009542904794216156,
0.07302886992692947,
0.06622157245874405,
-0.0022029802203178406,
0.07738194614648819,
0.10291139036417007,
0.011213085614144802,
0.06398138403892517,
0.003953366074711084,
-0.08768105506896973,
0.06894085556268692,
0.01583237387239933,
0.02318698540329933,
0.12077458947896957,
-0.0021076316479593515,
-0.12670843303203583,
0.015293911099433899,
0.027478154748678207,
-0.042967841029167175,
0.06494707614183426,
0.00023052659526001662,
-0.09174524992704391,
-0.05092667415738106,
-0.030642181634902954,
0.015056894160807133,
0.005003074649721384,
0.03860611096024513,
-0.0696997120976448,
0.018877701833844185,
-0.006670909933745861,
0.06389091908931732,
0.0368698351085186,
0.07683904469013214,
0.01871446892619133,
0.06887201964855194,
0.0580517016351223,
-0.012089269235730171,
0.011636621318757534,
-0.0478014200925827,
0.032361626625061035,
0.05524410307407379,
-0.017975665628910065,
0.08889705687761307,
0.03775019571185112,
-0.014113015495240688,
1.1071115281666356e-33,
0.09006392955780029,
0.06351500749588013,
0.040077779442071915,
0.09182772785425186,
0.025685947388410568,
-0.03202756121754646,
-0.001961276400834322,
0.011512774974107742,
-0.08447765558958054,
-0.06166520714759827,
0.0056283664889633656,
0.11682764440774918,
-0.056923165917396545,
-0.000258634943747893,
-0.017886264249682426,
-0.010349440388381481,
-0.013396144844591618,
0.07720500230789185,
-0.059681009501218796,
-0.031172217801213264,
0.10228948295116425,
-0.04650633782148361,
0.03793598711490631,
-0.049405407160520554,
0.05500290170311928,
-0.008876509964466095,
0.012428917922079563,
-0.06145937368273735,
-0.00020248509827069938,
0.003853358095511794,
-0.11215703934431076,
-0.028636299073696136,
0.06543746590614319,
0.10039839148521423,
-0.05208173394203186,
-0.003940701484680176,
-0.05518224090337753,
-0.06804009526968002,
0.0607326403260231,
0.01509944163262844,
-0.00047501642256975174,
0.05166309326887131,
0.028351539745926857,
-0.02153402753174305,
-0.03397226333618164,
-0.05769776552915573,
-0.0034831061493605375,
0.040415551513433456,
0.06884174048900604,
-0.01201761607080698,
0.05380586162209511,
0.00656975619494915,
0.07533576339483261,
0.029442131519317627,
0.007473531179130077,
0.05819961428642273,
-0.017771024256944656,
0.010390046052634716,
0.0366327241063118,
-0.017379654571413994,
0.005775902885943651,
0.06128600612282753,
-0.020147396251559258,
0.0610293410718441,
0.05821628123521805,
0.018810883164405823,
-0.006324558984488249,
0.13098597526550293,
0.020061438903212547,
0.021699409931898117,
-0.04295461252331734,
0.021195555105805397,
0.02411745674908161,
-0.033034246414899826,
0.10549993813037872,
-0.04775848612189293,
0.038801178336143494,
-0.1470557153224945,
0.008709554560482502,
-0.037199486047029495,
-0.004078928381204605,
-0.03866928070783615,
-0.0047784592024981976,
-0.04202330857515335,
-0.01523263193666935,
-0.06730107963085175,
-0.06270556151866913,
-0.0015929077053442597,
-0.049595095217227936,
-0.05801898613572121,
-0.06258906424045563,
-0.02736019156873226,
-0.020758233964443207,
0.028443148359656334,
0.00018682122754398733,
-8.740990686393915e-34,
-0.06182600557804108,
-0.09256387501955032,
-0.054475706070661545,
0.12753689289093018,
0.0007477567996829748,
-0.07526783645153046,
-0.1078338623046875,
-0.04524397850036621,
0.04000544548034668,
-0.05266173928976059,
-0.026478158310055733,
-0.08571325987577438,
0.027402691543102264,
-0.030394429340958595,
-0.002142558339983225,
-0.06286713480949402,
-0.037144750356674194,
-0.0003445125767029822,
0.0064903791062533855,
0.04831957444548607,
-0.050674598664045334,
0.06337722390890121,
-0.09903024882078171,
0.001590664847753942,
0.011891284957528114,
0.0390038825571537,
0.06011776626110077,
-0.043053749948740005,
0.006884652189910412,
0.0267327968031168,
0.0008855857304297388,
0.028458615764975548,
-0.045871686190366745,
-0.04669330641627312,
-0.045965053141117096,
0.07208601385354996,
0.07322109490633011,
0.009974604472517967,
0.0058846366591751575,
0.07470956444740295,
0.05422328785061836,
0.024434484541416168,
-0.11049575358629227,
-0.03239717334508896,
-0.00863615795969963,
-0.019602162763476372,
-0.1018487736582756,
0.07657456398010254,
-0.022688204422593117,
-0.012104334309697151,
-0.026810631155967712,
-0.03393268957734108,
0.0026461679954081774,
-0.05304638668894768,
-0.006377889774739742,
-0.05274253711104393,
-0.014333533123135567,
0.007288950029760599,
-0.06332145631313324,
0.005727218464016914,
-0.07766357064247131,
0.06822772324085236,
0.025325272232294083,
0.04937174543738365,
-0.007550413720309734,
-0.050402332097291946,
-0.030371850356459618,
0.052965372800827026,
0.01929861679673195,
-0.0348094217479229,
0.09138765931129456,
-0.029705533757805824,
-0.03697197884321213,
0.029308032244443893,
0.08886652439832687,
0.017310917377471924,
0.04344196617603302,
-0.05570681393146515,
0.06760479509830475,
0.016248594969511032,
0.021870143711566925,
0.06695031374692917,
0.061080146580934525,
-0.032584309577941895,
-0.008511710911989212,
-0.0037954349536448717,
-0.009620534256100655,
0.07688623666763306,
-0.050345610827207565,
0.05041259527206421,
-0.003903544507920742,
-0.021325625479221344,
-0.04466833546757698,
0.021203164011240005,
-0.039739660918712616,
-3.614700716525476e-8,
0.02267439290881157,
0.042353127151727676,
0.008068799041211605,
0.019088849425315857,
-0.01303783804178238,
-0.023439908400177956,
-0.029536405578255653,
0.056052517145872116,
-0.00881991907954216,
-0.03715401142835617,
0.07120684534311295,
-0.05043141916394234,
-0.030997874215245247,
-0.023649239912629128,
-0.01886053942143917,
0.07996302098035812,
0.021560706198215485,
0.0004729707434307784,
-0.060239870101213455,
-0.025501945987343788,
0.05771931633353233,
0.029739011079072952,
0.03357807174324989,
-0.01981515809893608,
0.04494364187121391,
-0.07161331921815872,
-0.01539562363177538,
0.07146691530942917,
0.021100010722875595,
-0.03476507589221001,
-0.005629252176731825,
-0.06732745468616486,
-0.019785773009061813,
0.01864163763821125,
0.06885139644145966,
0.08052922785282135,
-0.018508102744817734,
-0.07458455115556717,
0.05236443877220154,
0.042842429131269455,
0.01160935778170824,
0.1017700582742691,
-0.03509185090661049,
0.0051833330653607845,
-0.0043215504847466946,
-0.04811178147792816,
-0.08406620472669601,
-0.014725638553500175,
0.050744879990816116,
-0.05078966170549393,
-0.02245183102786541,
-0.07392901927232742,
-0.025962993502616882,
0.027108782902359962,
0.0064545245841145515,
0.1453913152217865,
-0.08072072267532349,
-0.007764396723359823,
0.05683792009949684,
-0.003792172996327281,
0.027047643437981606,
0.0049214609898626804,
0.05721667781472206,
0.026194743812084198
] | 0 | 0 | 1 | 1 |
|
RazorAttention: Efficient KV Cache Compression Through Retrieval Heads | https://openreview.net/forum?id=tkiZQlL04w | [
"Hanlin Tang",
"Yang Lin",
"Jing Lin",
"Qingsen Han",
"Danning Ke",
"Shikuan Hong",
"Yiwu Yao",
"Gongyi Wang"
] | Poster | The memory and computational demands of Key-Value (KV) cache present significant challenges for deploying long-context language models. Previous approaches attempt to mitigate this issue by selectively dropping tokens, which irreversibly erases critical information that might be needed for future queries. In this paper, we propose a novel compression technique for KV cache that preserves all token information. Our investigation reveals that: i) Most attention heads primarily focus on the local context; ii) Only a few heads, denoted as retrieval heads, can essentially pay attention to all input tokens. These key observations motivate us to use separate caching strategy for attention heads.Therefore, we propose RazorAttention, a training-free KV cache compression algorithm, which maintains a full cache for these crucial retrieval heads and discards the remote tokens in non-retrieval heads. Furthermore, we introduce a novel mechanism involving a “compensation token” to further recover the information in the dropped tokens. Extensive evaluations across a diverse set of large language models (LLMs) demonstrate that RazorAttention achieves a reduction in KV cache size by over 70% without noticeable impacts on performance. Additionally, RazorAttention is compatible with FlashAttention, rendering it an efficient and plug-and-play solution that enhances LLM inference efficiency without overhead or retraining of the original model. | LLMs, KV cache compression, LLM inference acceleration | A novel KV cache compression algorithm that achieves 3X reduction under general cases, compatible with FlashAttention | 12,841 | 2407.15891 | [
-0.03781360760331154,
0.05091915279626846,
0.004121103789657354,
0.032353490591049194,
0.012805921956896782,
0.0024217707104980946,
0.05850738659501076,
-0.017514802515506744,
0.06467544287443161,
-0.011579702608287334,
0.02046908624470234,
-0.030293673276901245,
0.042576905339956284,
-0.013109909370541573,
0.02837531641125679,
-0.040557585656642914,
0.049930643290281296,
0.10859094560146332,
-0.07848060876131058,
-0.02959582395851612,
0.028583791106939316,
0.017908573150634766,
0.05608142912387848,
-0.011250833049416542,
0.05092572420835495,
0.011133014224469662,
-0.07126368582248688,
-0.07293830066919327,
0.03947560489177704,
-0.011237096972763538,
0.046216025948524475,
0.03575184568762779,
-0.018257316201925278,
0.0575193427503109,
-0.015178494155406952,
0.07275721430778503,
-0.08279555290937424,
-0.014189044944941998,
0.015428469516336918,
-0.09943068772554398,
-0.05576230585575104,
0.03874720260500908,
-0.0640367642045021,
0.09061606973409653,
0.07170398533344269,
0.03962798789143562,
0.0668134018778801,
0.005146549083292484,
-0.08114288002252579,
-0.05508024990558624,
-0.037444692105054855,
-0.059535346925258636,
-0.030579090118408203,
0.03830793499946594,
-0.07823776453733444,
0.010338996537029743,
-0.007419671397656202,
0.0489518977701664,
-0.057139504700899124,
0.05891183018684387,
-0.028910648077726364,
-0.10489274561405182,
-0.02631559409201145,
-0.011941527016460896,
-0.035511430352926254,
-0.055396199226379395,
0.1072777509689331,
-0.008785759098827839,
0.003679986810311675,
0.005646687000989914,
-0.04630259424448013,
0.04598149284720421,
-0.019557571038603783,
0.04194878041744232,
0.002186953090131283,
0.13130198419094086,
0.00982404500246048,
-0.06525631248950958,
0.04551662504673004,
-0.060532812029123306,
-0.0033909969497472048,
-0.051134977489709854,
0.07066069543361664,
0.006891281343996525,
0.11231906712055206,
0.01850079372525215,
0.10474520921707153,
-0.06576859205961227,
0.017246155068278313,
0.010542809963226318,
-0.02638109028339386,
-0.034389834851026535,
0.1609542816877365,
-0.06540829688310623,
0.007989488542079926,
0.01825289987027645,
0.03086848556995392,
0.0756273865699768,
-0.0309167318046093,
0.08557209372520447,
0.04064858704805374,
0.06698662042617798,
-0.052036140114068985,
-0.014662946574389935,
-0.01579049415886402,
0.008095728233456612,
0.003075400833040476,
0.05003378540277481,
-0.004262262023985386,
-0.03435068577528,
-0.012761539779603481,
0.08098986744880676,
-0.06576559692621231,
-0.06843413412570953,
-0.012490962632000446,
0.02676408737897873,
-0.038678672164678574,
-0.04876675829291344,
0.0383196696639061,
0.06220864877104759,
0.04106534272432327,
-0.0024082039017230272,
-0.06742817908525467,
-0.06065182760357857,
-0.042574796825647354,
-0.04217219725251198,
0.005225646309554577,
3.0170526371662767e-33,
0.036010466516017914,
0.07056224346160889,
-0.09253628551959991,
-0.01086196769028902,
0.0026604903396219015,
-0.014713374897837639,
0.03188415244221687,
0.0546242892742157,
-0.037828948348760605,
-0.03695349767804146,
0.04195740818977356,
0.02074865624308586,
0.01699233241379261,
0.03031090646982193,
0.013409027829766273,
0.0038678739219903946,
-0.0006411437643691897,
0.0879613608121872,
-0.026402050629258156,
-0.059944745153188705,
0.06509754061698914,
-0.01197057031095028,
0.038319360464811325,
-0.06570614129304886,
0.044032078236341476,
0.03066260740160942,
0.04063097760081291,
-0.061697717756032944,
-0.04343681409955025,
0.018346741795539856,
-0.07343510538339615,
-0.045367658138275146,
-0.017835192382335663,
0.003494022646918893,
0.025458620861172676,
-0.015949387103319168,
0.046826861798763275,
-0.07070629298686981,
0.05844489485025406,
-0.012087934650480747,
0.017727429047226906,
0.07369118183851242,
-0.007518000900745392,
-0.040661849081516266,
-0.13601405918598175,
-0.0712568536400795,
-0.04961518943309784,
0.010401075705885887,
-0.026657002046704292,
0.0024330164305865765,
0.06291036307811737,
0.0023953737691044807,
-0.020745307207107544,
0.0231739841401577,
0.015684595331549644,
-0.01059412769973278,
0.05140320584177971,
-0.0011625186307355762,
0.028717834502458572,
0.09231828153133392,
-0.011141287162899971,
0.011052248068153858,
0.017261981964111328,
0.09291981905698776,
0.0060324338264763355,
0.023153165355324745,
0.04360409826040268,
0.09043174237012863,
0.01693972572684288,
-0.011535967700183392,
-0.005251291673630476,
-0.0084812818095088,
0.04160250350832939,
-0.06307800114154816,
-0.02079661563038826,
-0.023391520604491234,
-0.019164761528372765,
-0.03822101652622223,
-0.0287662073969841,
-0.026944007724523544,
0.049918223172426224,
0.008505097590386868,
-0.026472562924027443,
0.00852635782212019,
-0.002782902680337429,
-0.08054877817630768,
0.03790947422385216,
-0.09740626811981201,
-0.024439582601189613,
-0.039026688784360886,
0.0020725936628878117,
-0.034799885004758835,
0.018418969586491585,
-0.07194516062736511,
-0.06947211921215057,
-2.429781536967216e-33,
0.015864256769418716,
-0.05152649059891701,
-0.058317504823207855,
0.07529307156801224,
-0.04291902109980583,
-0.010693134739995003,
0.005433906335383654,
-0.015450989827513695,
-0.13520434498786926,
-0.10403475910425186,
-0.08673787117004395,
-0.06685085594654083,
0.030565204098820686,
0.05162976682186127,
0.013520457781851292,
0.03760511055588722,
0.0014031960163265467,
-0.07496726512908936,
-0.05653873831033707,
0.06333576142787933,
-0.018047936260700226,
0.034964606165885925,
-0.07318704575300217,
0.048744913190603256,
-0.00936945155262947,
-0.004484475590288639,
-0.0708942860364914,
0.009791223332285881,
0.04663776233792305,
-0.040091775357723236,
-0.08677234500646591,
-0.006499965209513903,
-0.061216216534376144,
-0.07141000777482986,
-0.05719662085175514,
0.04447474703192711,
0.03610355034470558,
0.05591663718223572,
-0.041731759905815125,
0.11295568197965622,
0.1008836030960083,
-0.057233069092035294,
-0.017603181302547455,
0.04892620071768761,
-0.016527386382222176,
-0.03852012753486633,
-0.12395498901605606,
0.01258897129446268,
0.020530978217720985,
0.05098399892449379,
0.03271196782588959,
-0.03740764781832695,
-0.08980631828308105,
-0.026015860959887505,
-0.055259820073843,
-0.028490277007222176,
-0.08622029423713684,
-0.01573927514255047,
0.01408880203962326,
-0.06476349383592606,
0.023043056949973106,
-0.08208806067705154,
-0.021687636151909828,
0.006436166353523731,
0.026219002902507782,
-0.04101239889860153,
-0.0021998523734509945,
-0.036578018218278885,
-0.047985274344682693,
-0.07738424092531204,
0.06766533106565475,
-0.02839038148522377,
-0.012784634716808796,
-0.03375496715307236,
0.0029183542355895042,
0.05995732173323631,
0.06113244220614433,
-0.07644420117139816,
0.012171333655714989,
0.023299185559153557,
-0.01757064089179039,
0.08160056918859482,
-0.011339513584971428,
0.016046658158302307,
0.05859968811273575,
0.0703342854976654,
-0.03580741956830025,
0.07420138269662857,
-0.01055710669606924,
-0.025207795202732086,
-0.007751730270683765,
0.03557930514216423,
0.013754690997302532,
0.08591997623443604,
0.011793157085776329,
-4.057294233916764e-8,
-0.0656236857175827,
0.051863282918930054,
-0.02642160840332508,
0.12159813940525055,
0.030284376814961433,
-0.014474482275545597,
-0.035798877477645874,
0.12555190920829773,
0.02566596493124962,
-0.01259984914213419,
0.07842807471752167,
-0.04171803221106529,
-0.07950405031442642,
-0.03502337634563446,
0.01374219823628664,
0.05317788943648338,
-0.010483765043318272,
0.04565167799592018,
-0.0373833104968071,
-0.05544266477227211,
-0.04429382458329201,
0.11646896600723267,
0.049033958464860916,
-0.01085138414055109,
0.009436346590518951,
0.007195256184786558,
0.0545109324157238,
0.13502725958824158,
0.06983266025781631,
-0.055615123361349106,
-0.012171431444585323,
0.03164530172944069,
-0.0007621584227308631,
-0.09637250006198883,
0.02964525856077671,
0.0496051162481308,
-0.01078883558511734,
-0.012816778384149075,
0.012891968712210655,
0.05826016142964363,
0.058605849742889404,
-0.004311679396778345,
-0.042338307946920395,
0.08834176510572433,
-0.0009388364851474762,
0.047797080129384995,
0.008333476260304451,
-0.03517243638634682,
0.014614821411669254,
-0.043278712779283524,
0.0035382932983338833,
-0.07310476899147034,
-0.04552686959505081,
0.07174152135848999,
-0.03896660730242729,
0.048157136887311935,
0.04483665153384209,
0.022644007578492165,
0.06016640365123749,
0.07656367123126984,
0.05326879024505615,
-0.04590151831507683,
-0.07004988938570023,
-0.017593415454030037
] | 0 | 0 | 0 | 0 |
|
Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation | https://openreview.net/forum?id=JvkuZZ04O7 | [
"Mufei Li",
"Siqi Miao",
"Pan Li"
] | Poster | Large Language Models (LLMs) demonstrate strong reasoning abilities but face limitations such as hallucinations and outdated knowledge. Knowledge Graph (KG)-based Retrieval-Augmented Generation (RAG) addresses these issues by grounding LLM outputs in structured external knowledge from KGs. However, current KG-based RAG frameworks still struggle to optimize the trade-off between retrieval effectiveness and efficiency in identifying a suitable amount of relevant graph information for the LLM to digest. We introduce SubgraphRAG, extending the KG-based RAG framework that retrieves subgraphs and leverages LLMs for reasoning and answer prediction. Our approach innovatively integrates a lightweight multilayer perceptron (MLP) with a parallel triple-scoring mechanism for efficient and flexible subgraph retrieval while encoding directional structural distances to enhance retrieval effectiveness. The size of retrieved subgraphs can be flexibly adjusted to match the query's needs and the downstream LLM's capabilities. This design strikes a balance between model complexity and reasoning power, enabling scalable and generalizable retrieval processes. Notably, based on our retrieved subgraphs, smaller LLMs like Llama3.1-8B-Instruct deliver competitive results with explainable reasoning, while larger models like GPT-4o achieve state-of-the-art accuracy compared with previous baselines—all without fine-tuning. Extensive evaluations on the WebQSP and CWQ benchmarks highlight SubgraphRAG's strengths in efficiency, accuracy, and reliability by reducing hallucinations and improving response grounding. | Knowledge Graphs, Large Language Models, Retrieval-Augmented Generation, Retrieval | null | 12,839 | 2410.20724 | [
-0.05743762478232384,
-0.08668555319309235,
0.009120380505919456,
0.02912084385752678,
0.04345335811376572,
-0.0035083494149148464,
-0.032738424837589264,
0.023666514083743095,
-0.0017083714483305812,
-0.006219490431249142,
-0.04647108167409897,
-0.06353764981031418,
0.07104730606079102,
0.08384370803833008,
-0.007840869016945362,
0.09111465513706207,
0.09546791762113571,
0.04751235246658325,
-0.034775517880916595,
-0.0812382772564888,
0.05290770158171654,
0.018333587795495987,
0.07286297529935837,
-0.019198963418602943,
0.02794514037668705,
-0.03563545644283295,
-0.03668801486492157,
-0.018159760162234306,
0.03498189151287079,
0.02011818438768387,
0.05690860003232956,
0.048148371279239655,
-0.02424253150820732,
0.06299086660146713,
-0.034342966973781586,
0.055372193455696106,
-0.05417504161596298,
0.07281505316495895,
0.006036028265953064,
0.020137213170528412,
-0.018160201609134674,
0.017794040963053703,
0.04955272749066353,
0.04166850447654724,
0.11018197238445282,
0.04483441263437271,
-0.1048337072134018,
-0.011579679325222969,
0.007481095381081104,
0.05402495339512825,
-0.04080066829919815,
0.02563355676829815,
-0.033387791365385056,
0.0030188318341970444,
-0.000529096694663167,
-0.02271716482937336,
-0.04119495302438736,
0.025586239993572235,
-0.03312816098332405,
-0.061097823083400726,
-0.04844559356570244,
-0.14194998145103455,
-0.03682241216301918,
-0.044650670140981674,
-0.026156075298786163,
-0.029073039069771767,
-0.0025332870427519083,
0.04190194979310036,
-0.06455016881227493,
0.04912734776735306,
-0.006268822122365236,
0.016618190333247185,
-0.043058037757873535,
0.07488427311182022,
0.04004885256290436,
0.1109490692615509,
0.05802434682846069,
-0.01480997633188963,
-0.003688675817102194,
-0.05223451554775238,
-0.009388451464474201,
0.04778849706053734,
0.03846175596117973,
0.04088076949119568,
0.09223390370607376,
0.020738933235406876,
0.021410057321190834,
0.001458521350286901,
-0.010764379985630512,
0.008357799611985683,
-0.04914942383766174,
-0.09300341457128525,
0.01335603091865778,
-0.01698232628405094,
0.057357706129550934,
0.10513672977685928,
0.038632363080978394,
-0.03284693881869316,
-0.05953577533364296,
0.0985211655497551,
0.02240770310163498,
0.13901479542255402,
0.011266396380960941,
-0.07853661477565765,
-0.005810083821415901,
0.0019088182598352432,
-0.030107051134109497,
0.06998150795698166,
-0.0010710690403357148,
-0.1146286204457283,
-0.0010556544875726104,
0.14830806851387024,
0.07106431573629379,
0.02305958978831768,
0.03523089736700058,
-0.025047749280929565,
0.030373310670256615,
-0.04524781554937363,
0.04526427015662193,
0.0022427665535360575,
-0.060925208032131195,
0.04278455302119255,
0.022308379411697388,
-0.011165495961904526,
-0.018887650221586227,
-0.02020876109600067,
0.009193730540573597,
1.157402574408171e-33,
0.0541590191423893,
0.05579305440187454,
0.09277670830488205,
0.052024275064468384,
-0.005854041315615177,
0.018379271030426025,
-0.011839252896606922,
0.0968688502907753,
-0.037684496492147446,
0.004238559398800135,
-0.06982165575027466,
0.029123323038220406,
-0.003402060829102993,
0.005907554179430008,
0.022608032450079918,
0.048952192068099976,
-0.019042162224650383,
-0.039120517671108246,
-0.04294401407241821,
-0.06853296607732773,
0.06254073977470398,
-0.026850739493966103,
0.051948681473731995,
0.004210506100207567,
0.03789012134075165,
0.008948713541030884,
0.0667661800980568,
-0.06185736879706383,
-0.061670128256082535,
-0.019047414883971214,
-0.0855608806014061,
0.0102196354418993,
0.008703255094587803,
-0.014074995182454586,
-0.01741514727473259,
-0.09150046110153198,
-0.07210243493318558,
-0.08942187577486038,
0.0014607353368774056,
-0.0011914287460967898,
-0.05055558681488037,
0.043198730796575546,
0.030165381729602814,
-0.00038505735574290156,
-0.07116535305976868,
-0.04741949960589409,
-0.05569568648934364,
-0.015412898734211922,
0.015575659461319447,
-0.01315118744969368,
0.06626258790493011,
0.015094916336238384,
-0.056805722415447235,
-0.03364365175366402,
0.059848614037036896,
-0.028827188536524773,
-0.01159672811627388,
0.05799594894051552,
0.05014410614967346,
0.04884754493832588,
0.02510201744735241,
0.018490152433514595,
0.03390120714902878,
0.09223367273807526,
0.047693829983472824,
0.044478658586740494,
-0.09433606266975403,
0.09688500314950943,
0.09052583575248718,
-0.041145164519548416,
0.02828449010848999,
-0.006299089174717665,
0.03770763427019119,
-0.07789122313261032,
-0.024461545050144196,
0.0006060276064090431,
-0.010910910554230213,
-0.09610573202371597,
0.027913346886634827,
-0.022721968591213226,
0.00817410834133625,
-0.09322687238454819,
-0.02441067434847355,
-0.052792470902204514,
-0.01694810949265957,
-0.04006194695830345,
0.08862760663032532,
-0.14155294001102448,
0.043435871601104736,
-0.04212100803852081,
-0.028579814359545708,
-0.023168101906776428,
-0.03671223297715187,
-0.07051350176334381,
0.007781195919960737,
1.1624308662881835e-34,
0.044287022203207016,
-0.0068075209856033325,
-0.0462765246629715,
-0.03641265258193016,
0.022835012525320053,
-0.09759053587913513,
-0.00021620267943944782,
0.0702819898724556,
-0.13607561588287354,
0.006193181034177542,
-0.047174036502838135,
-0.05945281684398651,
0.055307213217020035,
-0.036044541746377945,
0.04640769585967064,
0.017649896442890167,
-0.011443670839071274,
0.012220906093716621,
-0.05509471148252487,
0.10345451533794403,
-0.05960327759385109,
0.06982847303152084,
-0.09610747545957565,
0.052757252007722855,
-0.007374994922429323,
0.019992824643850327,
-0.00359003571793437,
-0.00848823506385088,
-0.05271178111433983,
0.05315449833869934,
-0.025320500135421753,
-0.017287563532590866,
0.03612932190299034,
-0.028937416151165962,
-0.03475530818104744,
-0.002512566512450576,
0.05036202445626259,
-0.08308983594179153,
-0.074269600212574,
0.0014412103919312358,
0.002600928535684943,
0.02956625446677208,
-0.023029480129480362,
0.0045256344601511955,
-0.0049486784264445305,
0.06451558321714401,
-0.09788253903388977,
0.001616989029571414,
-0.00021405053848866373,
0.002722380217164755,
-0.0012832990614697337,
-0.045172449201345444,
0.02553977631032467,
-0.029488377273082733,
0.012524351477622986,
-0.10168957710266113,
-0.07495369762182236,
-0.008123076520860195,
0.049191202968358994,
-0.0008332739816978574,
-0.036592211574316025,
-0.01669243350625038,
0.041593655943870544,
-0.03520137071609497,
0.039800919592380524,
0.0007168612792156637,
-0.03719029203057289,
0.008360996842384338,
-0.05271771922707558,
0.04021748527884483,
-0.03571802377700806,
-0.024946095421910286,
0.01294404361397028,
-0.027093471959233284,
0.06962008774280548,
0.04160762205719948,
0.03734995052218437,
0.012070857919752598,
-0.0585334338247776,
-0.08159160614013672,
-0.05825372785329819,
-0.039959151297807693,
0.03464730829000473,
0.10397431999444962,
0.04845341667532921,
-0.04916418716311455,
0.022395608946681023,
0.13674518465995789,
0.04317878931760788,
0.025081373751163483,
0.016088809818029404,
0.004957939963787794,
-0.022843942046165466,
0.07753057032823563,
0.05394406616687775,
-4.098033556942937e-8,
-0.116982601583004,
0.04853028804063797,
-0.049179933965206146,
0.06990541517734528,
0.025582890957593918,
-0.02504965476691723,
-0.03949083015322685,
0.10514556616544724,
-0.009695704095065594,
0.06271351873874664,
0.06680101156234741,
-0.008058330975472927,
-0.01902724802494049,
-0.04364354908466339,
0.023662356659770012,
0.06356655061244965,
0.04152372106909752,
0.006368777249008417,
0.001180086052045226,
-0.014652918092906475,
0.04913382977247238,
0.03423721715807915,
0.00649301940575242,
0.09689044207334518,
0.0340636745095253,
-0.03794538974761963,
-0.08124469220638275,
0.07917241007089615,
0.031779754906892776,
-0.017627691850066185,
-0.008830469101667404,
-0.05740043893456459,
0.02864600531756878,
-0.0345354862511158,
0.05337117612361908,
0.05575203150510788,
-0.03380375728011131,
-0.0393783263862133,
0.0473276786506176,
-0.025670286267995834,
0.046749014407396317,
0.07522299885749817,
-0.017135892063379288,
0.060490529984235764,
0.016666200011968613,
-0.02095475234091282,
-0.0694098249077797,
-0.02698567882180214,
-0.01255856268107891,
-0.012782186269760132,
-0.053308021277189255,
-0.11817868798971176,
0.040407512336969376,
0.0028957040049135685,
0.0498870313167572,
0.017148589715361595,
-0.08236799389123917,
0.007944656535983086,
0.09133516252040863,
-0.0404142290353775,
0.04906408116221428,
0.011858709156513214,
-0.029611559584736824,
0.056264977902173996
] | https://github.com/graph-com/subgraphrag | 70 | 0 | 0 | 0 |
MrT5: Dynamic Token Merging for Efficient Byte-level Language Models | https://openreview.net/forum?id=VYWBMq1L7H | [
"Julie Kallini",
"Shikhar Murty",
"Christopher D Manning",
"Christopher Potts",
"Róbert Csordás"
] | Poster | Models that rely on subword tokenization have significant drawbacks, such as sensitivity to character-level noise like spelling errors and inconsistent compression rates across different languages and scripts. While character- or byte-level models like ByT5 attempt to address these concerns, they have not gained widespread adoption—processing raw byte streams without tokenization results in significantly longer sequence lengths, making training and inference inefficient. This work introduces MrT5 (MergeT5), a more efficient variant of ByT5 that integrates a token deletion mechanism in its encoder to dynamically shorten the input sequence length. After processing through a fixed number of encoder layers, a learned delete gate determines which tokens are to be removed and which are to be retained for subsequent layers. MrT5 effectively "merges" critical information from deleted tokens into a more compact sequence, leveraging contextual information from the remaining tokens. In continued pre-training experiments, we find that MrT5 can achieve significant gains in inference runtime with minimal effect on performance, as measured by bits-per-byte. Additionally, with multilingual training, MrT5 adapts to the orthographic characteristics of each language, learning language-specific compression rates. Furthermore, MrT5 shows comparable accuracy to ByT5 on downstream evaluations such as XNLI, TyDi QA, and character-level tasks while reducing sequence lengths by up to 75%. Our approach presents a solution to the practical limitations of existing byte-level models. | NLP, ByT5, T5, tokenization, byte-level language models, character-level language models | MrT5 improves the efficiency of byte-level models like ByT5 by introducing a token deletion mechanism in its encoder, reducing the sequence length during processing. | 12,835 | 2410.20771 | [
-0.09869027137756348,
-0.06969417631626129,
0.08227576315402985,
0.03127600625157356,
0.044033195823431015,
-0.0516369566321373,
-0.022396905347704887,
0.002932678908109665,
0.04231023043394089,
-0.082163505256176,
-0.03801025450229645,
-0.007702931296080351,
0.004280988126993179,
0.015065664425492287,
-0.01674458384513855,
-0.0025329547934234142,
-0.006570718251168728,
0.11915011703968048,
-0.11407080292701721,
-0.10063000023365021,
0.049236100167036057,
0.0027739887591451406,
0.020518595352768898,
-0.0016119034262374043,
0.08670031279325485,
0.026043659076094627,
-0.01904914900660515,
-0.019902091473340988,
0.0374801866710186,
-0.01876378059387207,
0.07635173201560974,
0.02879014052450657,
-0.010828487575054169,
-0.014420311897993088,
-0.019599659368395805,
0.10152345895767212,
0.0024169168900698423,
-0.05087511986494064,
0.03881367668509483,
-0.06737061589956284,
-0.044462915509939194,
-0.0281191635876894,
-0.025525322183966637,
0.07922248542308807,
0.034541334956884384,
0.024918198585510254,
-0.006879672408103943,
-0.027272021397948265,
-0.06312251091003418,
-0.012256937101483345,
-0.06359352171421051,
-0.017561839893460274,
-0.02045689895749092,
0.06166745722293854,
-0.05370522290468216,
-0.08983121067285538,
-0.03401557356119156,
0.05042353272438049,
-0.07144636660814285,
-0.015523935668170452,
-0.11299669742584229,
-0.04343678429722786,
-0.015370728448033333,
-0.004224833566695452,
0.006791131570935249,
-0.06959349662065506,
0.04605330526828766,
0.0556965246796608,
0.049884650856256485,
0.030809294432401657,
-0.05013985559344292,
0.06263101100921631,
-0.11147090047597885,
0.042176295071840286,
-0.06836714595556259,
0.10135483741760254,
0.033449240028858185,
-0.011762036010622978,
0.019387660548090935,
-0.07517006993293762,
0.0174117274582386,
-0.03943024203181267,
0.06842778623104095,
-0.026517629623413086,
0.07451260834932327,
-0.006519409362226725,
0.00697768060490489,
0.03467278927564621,
-0.021993102505803108,
0.05788056179881096,
-0.05616108328104019,
-0.05804111063480377,
0.1007041335105896,
0.00021026308240834624,
0.0325753316283226,
-0.007407111581414938,
-0.022972119972109795,
0.0035228943452239037,
-0.05342274531722069,
0.09616603702306747,
-0.0005065648001618683,
0.08179280906915665,
-0.02432890422642231,
-0.04024789482355118,
0.02091813273727894,
-0.03240760415792465,
0.08197533339262009,
0.016543244943022728,
0.09248802810907364,
-0.055362481623888016,
-0.0031831758096814156,
0.06520445644855499,
-0.031897030770778656,
-0.05274844542145729,
-0.004297687206417322,
0.016613395884633064,
-0.09092184156179428,
-0.048746854066848755,
0.018030116334557533,
0.08456819504499435,
-0.0368693470954895,
-0.02552969753742218,
-0.10513599961996078,
0.045007675886154175,
-0.004897051490843296,
-0.09365508705377579,
-0.04017206281423569,
3.540169847472265e-33,
-0.01628330908715725,
0.015855034813284874,
-0.04885563999414444,
0.03251255303621292,
-0.018149562180042267,
-0.0004956874763593078,
0.00006772057531634346,
0.06640177220106125,
-0.032255761325359344,
-0.07617803663015366,
-0.013410713523626328,
-0.03543202951550484,
-0.019164742901921272,
0.06670049577951431,
0.021620897576212883,
-0.05956093966960907,
0.005200650542974472,
0.03148697316646576,
-0.07385233044624329,
-0.010913427919149399,
0.05239998176693916,
0.07030591368675232,
-0.028934696689248085,
-0.025036659091711044,
-0.003356165951117873,
0.017741555348038673,
0.06537739187479019,
-0.08817541599273682,
-0.006815864238888025,
0.020849281921982765,
-0.1344676911830902,
-0.04586121067404747,
0.025448553264141083,
0.036925699561834335,
0.04685365781188011,
-0.03711492568254471,
0.005284418351948261,
-0.07606620341539383,
-0.008345702663064003,
0.007342283148318529,
-0.0064480979926884174,
-0.00390133960172534,
-0.00949362013489008,
0.004291468299925327,
-0.059943825006484985,
-0.09729993343353271,
-0.016691116616129875,
0.015728697180747986,
-0.06701412051916122,
-0.01682094857096672,
0.08337688446044922,
0.07294578105211258,
-0.02168116346001625,
0.0020449210423976183,
0.019153010100126266,
0.0014326046220958233,
0.044576823711395264,
-0.001421929569914937,
0.022711584344506264,
0.09813070297241211,
-0.040527164936065674,
0.047425419092178345,
0.03196140378713608,
0.10535544157028198,
0.06942252069711685,
0.02867901138961315,
0.0500069223344326,
0.03457219898700714,
0.008541077375411987,
-0.029521506279706955,
-0.020975273102521896,
0.01877380907535553,
0.0463227815926075,
0.009086424484848976,
-0.028241729363799095,
0.021876895800232887,
0.057851240038871765,
-0.08307161182165146,
-0.026609981432557106,
0.038430530577898026,
0.049710892140865326,
-0.010857982560992241,
0.021123260259628296,
-0.0466981939971447,
-0.009464277885854244,
-0.0218339916318655,
0.05914761871099472,
-0.14914025366306305,
-0.017069173976778984,
-0.032876089215278625,
-0.010779098607599735,
-0.11081789433956146,
0.02496393769979477,
-0.03931931406259537,
0.02271336130797863,
-3.5908685514812455e-33,
0.001138324965722859,
0.00404511159285903,
-0.0368795320391655,
0.08830109238624573,
-0.13657312095165253,
-0.08170409500598907,
-0.0007983134128153324,
0.04751550778746605,
-0.06824670732021332,
-0.055254340171813965,
-0.06769529730081558,
0.0012867082841694355,
0.06730718910694122,
-0.028863495215773582,
0.035623304545879364,
0.00016961098299361765,
0.0004212683707010001,
0.010256660170853138,
0.0593649223446846,
0.09022409468889236,
0.044116925448179245,
0.048602353781461716,
-0.03498842567205429,
0.033816829323768616,
0.01058322936296463,
0.07880234718322754,
-0.06193799152970314,
0.05839046463370323,
0.01842399314045906,
-0.065281942486763,
-0.04736613109707832,
0.03311719745397568,
-0.013266236521303654,
-0.09633965790271759,
0.004632547963410616,
-0.004621641710400581,
0.0781189575791359,
0.046617116779088974,
0.003952250350266695,
0.05236516520380974,
0.10163144022226334,
-0.02611803635954857,
-0.0029543067794293165,
0.047482673078775406,
-0.008332667872309685,
-0.027208851650357246,
-0.1438094675540924,
0.04877844452857971,
0.06766537576913834,
-0.011784753762185574,
0.02638690173625946,
0.00908800307661295,
-0.08025208860635757,
-0.043510615825653076,
0.0014425280969589949,
-0.11769165098667145,
-0.007501138374209404,
-0.033643342554569244,
-0.005020253360271454,
-0.05163644254207611,
-0.020520789548754692,
-0.024072404950857162,
0.06570494174957275,
-0.028548376634716988,
0.029914597049355507,
-0.04387993365526199,
-0.006349888630211353,
-0.03767170384526253,
0.010297287255525589,
-0.03128233551979065,
-0.0066554490476846695,
-0.06478336453437805,
-0.011487667448818684,
0.04349738359451294,
0.012702817097306252,
0.049586221575737,
0.039407193660736084,
-0.09781207144260406,
-0.03190459683537483,
0.011748327873647213,
-0.06160963326692581,
0.02186201512813568,
-0.031106838956475258,
0.005374078173190355,
0.03898937627673149,
0.05850263312458992,
-0.03141014650464058,
0.043593477457761765,
0.041418083012104034,
0.060505013912916183,
0.015821687877178192,
0.10502370446920395,
0.057697705924510956,
0.08319371938705444,
-0.05067459121346474,
-4.277902476701456e-8,
0.035043276846408844,
-0.025517331436276436,
-0.0008010116289369762,
0.05519427731633186,
0.04777739197015762,
-0.05411117896437645,
-0.07466020435094833,
0.07608430087566376,
0.04893907904624939,
-0.016376081854104996,
0.1001635268330574,
-0.05419881269335747,
-0.11497688293457031,
-0.10315147042274475,
-0.004759001079946756,
0.06349614262580872,
0.006393190938979387,
-0.016408374533057213,
-0.015224030241370201,
-0.011902894824743271,
-0.01115721371024847,
0.09262610971927643,
0.005006437189877033,
-0.0027803024277091026,
0.004952577408403158,
-0.03233669698238373,
0.012881443835794926,
0.12488718330860138,
0.045464735478162766,
-0.058575261384248734,
-0.002881651045754552,
0.06405962258577347,
0.0033034200314432383,
-0.057543590664863586,
-0.024769660085439682,
0.09990967810153961,
0.04995734244585037,
0.02647184394299984,
-0.0490814745426178,
0.03582830727100372,
0.08367294073104858,
0.023964935913681984,
-0.027582479640841484,
0.0423881970345974,
-0.006077702157199383,
-0.05725789815187454,
-0.0028246373403817415,
-0.03124523162841797,
0.03804340586066246,
-0.04539884626865387,
0.0048239631578326225,
0.016277512535452843,
-0.018125250935554504,
0.04593058302998543,
0.04849179461598396,
0.04582986980676651,
-0.01911238394677639,
0.018493307754397392,
0.014831899665296078,
0.08517991751432419,
-0.01706407032907009,
-0.010135240852832794,
-0.03466377779841423,
-0.037585947662591934
] | https://github.com/jkallini/mrt5 | 40 | 2 | 0 | 0 |
VVC-Gym: A Fixed-Wing UAV Reinforcement Learning Environment for Multi-Goal Long-Horizon Problems | https://openreview.net/forum?id=5xSRg3eYZz | [
"Xudong Gong",
"Feng Dawei",
"Kele Xu",
"weijia wang",
"Zhangjun Sun",
"Xing Zhou",
"Si Zheng",
"Bo Ding",
"Huaimin Wang"
] | Poster | Multi-goal long-horizon problems are prevalent in real-world applications. The additional goal space introduced by multi-goal problems intensifies the spatial complexity of exploration; meanwhile, the long interaction sequences in long-horizon problems exacerbate the temporal complexity of exploration. Addressing the great exploration challenge posed by multi-goal long-horizon problems depends not only on the design of algorithms but also on the design of environments and the availability of demonstrations to assist in training. To facilitate the above research, we propose a multi-goal long-horizon Reinforcement Learning (RL) environment based on realistic fixed-wing UAV's velocity vector control, named VVC-Gym, and generate multiple demonstration sets of various quality. Through experimentation, we analyze the impact of different environment designs on training, assess the quantity and quality of demonstrations and their influence on training, and assess the effectiveness of various RL algorithms, providing baselines on VVC-Gym and its corresponding demonstrations. The results suggest that VVC-Gym is suitable for studying: (1) the influence of environment designs on addressing multi-goal long-horizon problems with RL. (2) the assistance that demonstrations can provide in overcoming the exploration challenges of multi-goal long-horizon problems. (3) the RL algorithm designs with the least possible impact from environment designs on the efficiency and effectiveness of training. | Reinforcement Learning Environment, Demonstrations, Goal-Conditioned Reinforcement Learning, Fixed-wing UAV Velocity Vector Control | We provide a novel fixed-wing UAV RL environment, demonstrations, and baselines for multi-goal long-horizon problem research. | 12,834 | null | [
0.004329719580709934,
-0.0027537031564861536,
0.02419223077595234,
-0.017628604546189308,
0.02263718657195568,
0.03625888749957085,
-0.0678047388792038,
-0.039716850966215134,
0.004830650519579649,
0.06649236381053925,
-0.044497035443782806,
-0.05797581374645233,
0.003138558706268668,
0.04796275869011879,
0.023058943450450897,
-0.0019555524922907352,
0.03075999952852726,
0.0017974592046812177,
-0.04333198815584183,
0.05134068801999092,
0.013588850386440754,
0.05832290276885033,
0.08774230629205704,
0.053144942969083786,
-0.05597064271569252,
0.033109333366155624,
-0.04131098836660385,
-0.04284500330686569,
0.0019168915459886193,
-0.01516282930970192,
0.06670035421848297,
-0.011800271458923817,
0.04244791716337204,
-0.05854984372854233,
-0.056332629173994064,
0.06531748920679092,
-0.029272209852933884,
-0.021513480693101883,
-0.09087398648262024,
0.018173132091760635,
-0.07501184940338135,
0.010739266872406006,
0.08458133041858673,
-0.0210481658577919,
0.04117507487535477,
0.07747809588909149,
0.018151268362998962,
-0.049308501183986664,
0.0875449851155281,
0.022475186735391617,
-0.018755218014121056,
-0.07889897376298904,
0.02865632250905037,
-0.06521578878164291,
0.04757959023118019,
0.027883972972631454,
-0.00465449970215559,
0.02489583007991314,
-0.0066977678798139095,
-0.048718906939029694,
0.005012573674321175,
-0.03620457276701927,
-0.007561969570815563,
-0.0477442629635334,
-0.08780726045370102,
-0.02212359756231308,
-0.03726320341229439,
0.10979203879833221,
0.09576350450515747,
-0.012679840438067913,
-0.021711019799113274,
-0.00083663419354707,
-0.0044623203575611115,
-0.0002160943258786574,
-0.0015868231421336532,
0.06003781035542488,
-0.013119584880769253,
0.023386547341942787,
-0.02960921823978424,
-0.005925117991864681,
0.08259620517492294,
-0.04494314640760422,
-0.14687800407409668,
0.037795290350914,
0.09030801057815552,
-0.00898725539445877,
-0.008793316781520844,
0.06456326693296432,
0.10258422791957855,
0.03315388411283493,
0.035745568573474884,
-0.010797624476253986,
0.009876081719994545,
-0.026950577273964882,
0.01157129742205143,
-0.043272700160741806,
-0.033450160175561905,
-0.12340506911277771,
-0.026219097897410393,
0.025134893134236336,
-0.024207284674048424,
0.006458083633333445,
0.04953466355800629,
0.029907170683145523,
-0.008110903203487396,
-0.012142043560743332,
0.008957811631262302,
0.027706075459718704,
0.06525829434394836,
-0.09226205945014954,
0.02151031605899334,
-0.03177043795585632,
0.05803908780217171,
0.07360788434743881,
0.002856687642633915,
0.09290825575590134,
-0.09000224620103836,
-0.048214662820100784,
-0.04325399920344353,
0.07713883370161057,
-0.07983063906431198,
-0.097701296210289,
0.10047440230846405,
0.010113170370459557,
0.012523196637630463,
-0.08010589331388474,
-0.027479877695441246,
-2.345012736320834e-34,
-0.04445440694689751,
0.016528576612472534,
-0.004289898555725813,
0.031869497150182724,
0.028067348524928093,
0.0252164788544178,
0.085901640355587,
0.0166454054415226,
0.04990455508232117,
-0.007503936067223549,
0.040701933205127716,
0.010972012765705585,
0.007902160286903381,
0.07288249582052231,
0.11412594467401505,
-0.0027111240196973085,
-0.017567597329616547,
-0.011570921167731285,
-0.03854786604642868,
0.00525604048743844,
0.08660856634378433,
-0.11444851011037827,
-0.08439826965332031,
-0.04043476656079292,
0.08369670808315277,
-0.012670145370066166,
0.018909910693764687,
0.02175886183977127,
-0.03242407366633415,
-0.000374145369278267,
-0.03179285675287247,
0.06343641877174377,
-0.07266324758529663,
0.07433740794658661,
0.0027369242161512375,
0.022129692137241364,
-0.00971113983541727,
-0.036637142300605774,
-0.05629673972725868,
0.06773697584867477,
-0.04962221533060074,
-0.03873506933450699,
0.026708438992500305,
0.029054993763566017,
-0.02815878763794899,
-0.029761632904410362,
0.07670889049768448,
0.0023203955497592688,
-0.11888228356838226,
-0.0372481495141983,
-0.012981774285435677,
0.058767348527908325,
-0.04470151662826538,
-0.14325390756130219,
-0.03819120675325394,
-0.01287779025733471,
0.026134073734283447,
0.0535871759057045,
-0.054996103048324585,
0.03007340058684349,
0.0013857621233910322,
-0.02041204646229744,
-0.02584916166961193,
0.047381691634655,
-0.04930400848388672,
0.08759574592113495,
0.013606095686554909,
0.026973281055688858,
0.11156700551509857,
0.0018670904682949185,
-0.01049787551164627,
0.03861549496650696,
0.020983552560210228,
-0.13452079892158508,
0.11432795226573944,
-0.025157904252409935,
0.023382900282740593,
-0.03785429894924164,
-0.0027942161541432142,
0.010453185997903347,
-0.0006902521126903594,
0.07700572907924652,
-0.002859698375687003,
-0.009151284582912922,
0.034361839294433594,
-0.05965875834226608,
-0.013625471852719784,
-0.049868930131196976,
-0.025183387100696564,
-0.037779636681079865,
-0.12519948184490204,
0.00011367477418389171,
-0.00473327049985528,
0.04999900609254837,
0.03618018329143524,
-4.232081342565114e-34,
0.037691373378038406,
-0.0710662379860878,
-0.017978588119149208,
0.018851032480597496,
0.02875564806163311,
0.02687746472656727,
-0.024680782109498978,
-0.049091558903455734,
-0.016089025884866714,
-0.06221311539411545,
-0.07332546263933182,
0.006363460794091225,
-0.00020692938414867967,
0.030581947416067123,
-0.017795691266655922,
-0.005003388971090317,
-0.0009700412047095597,
-0.026941196992993355,
0.09662394225597382,
-0.028049437329173088,
0.06776551157236099,
0.02314257249236107,
-0.02008136920630932,
-0.029675958678126335,
-0.005580635275691748,
-0.014352946542203426,
0.061029572039842606,
0.023764073848724365,
0.0336725227534771,
0.06110335886478424,
0.00956928264349699,
-0.07368341833353043,
0.0004617743834387511,
0.013395705260336399,
-0.008309398777782917,
0.09796640276908875,
0.09045037627220154,
0.018754659220576286,
-0.05142660811543465,
0.0906149223446846,
0.062306202948093414,
-0.09316978603601456,
-0.0333293192088604,
-0.05083966627717018,
-0.009014475159347057,
0.04118764400482178,
-0.0009167685057036579,
0.04042251035571098,
0.042506154626607895,
0.0008174914983101189,
-0.005481550469994545,
-0.013442454859614372,
-0.04770510643720627,
-0.0343584306538105,
-0.0021573284175246954,
-0.076283760368824,
-0.020689379423856735,
0.051050037145614624,
0.08012119680643082,
-0.04152344912290573,
-0.04067330062389374,
0.015318059362471104,
-0.06021325662732124,
0.08902683854103088,
-0.019125448539853096,
0.05491003021597862,
0.01418613363057375,
0.06925693899393082,
-0.00863733422011137,
0.009339120239019394,
-0.04964332655072212,
0.02755681984126568,
0.03773980215191841,
-0.00764364143833518,
-0.027719151228666306,
-0.04058955982327461,
0.0009865181054919958,
-0.06021469458937645,
-0.006827818229794502,
-0.02518228441476822,
-0.001734577352181077,
-0.06704134494066238,
0.07514764368534088,
0.08286599069833755,
-0.010728701017796993,
0.1363380253314972,
-0.028490185737609863,
0.11668908596038818,
-0.01913832314312458,
-0.015600189566612244,
-0.0005440242821350694,
0.015727508813142776,
0.057832274585962296,
0.0296472255140543,
0.019447630271315575,
-4.227362282449576e-8,
-0.06330356001853943,
0.05201005935668945,
0.06331697851419449,
-0.04193278029561043,
0.045419760048389435,
0.0723535493016243,
-0.029987649992108345,
-0.049942679703235626,
-0.012466786429286003,
-0.025924408808350563,
-0.021343113854527473,
0.005408137571066618,
0.00554745364934206,
0.018438944593071938,
-0.07012464851140976,
0.02602647989988327,
-0.058977264910936356,
-0.06619174033403397,
0.010829007253050804,
0.013830584473907948,
0.02770567312836647,
0.009645584970712662,
0.005664699710905552,
-0.0746280774474144,
-0.07441897690296173,
-0.061585765331983566,
-0.08234632760286331,
-0.00362991401925683,
0.07972193509340286,
-0.036808572709560394,
0.030020710080862045,
-0.002662501996383071,
0.021562179550528526,
-0.013166836462914944,
0.020667366683483124,
-0.013373038731515408,
0.020884037017822266,
-0.059133004397153854,
0.012909889221191406,
-0.00928146205842495,
-0.044686298817396164,
0.08119812607765198,
-0.02572137676179409,
-0.024990908801555634,
0.03647904098033905,
0.01482201274484396,
-0.01120541151612997,
-0.16695986688137054,
-0.02663501538336277,
-0.03300107643008232,
0.03454829379916191,
0.03503241389989853,
-0.05480705946683884,
0.12034258991479874,
0.08838094770908356,
0.10810931771993637,
-0.014313535764813423,
-0.14432251453399658,
0.018019413575530052,
0.08065047860145569,
0.004532041493803263,
0.007218560669571161,
-0.06986341625452042,
0.03109748847782612
] | 0 | 0 | 0 | 0 |
|
Commit0: Library Generation from Scratch | https://openreview.net/forum?id=MMwaQEVsAg | [
"Wenting Zhao",
"Nan Jiang",
"Celine Lee",
"Justin T Chiu",
"Claire Cardie",
"Matthias Gallé",
"Alexander M Rush"
] | Poster | With the goal of benchmarking generative systems beyond expert software development ability, we introduce Commit0, a benchmark that challenges AI agents to write libraries from scratch. Agents are provided with a specification document outlining the library’s API as well as a suite of interactive unit tests, with the goal of producing an implementation of this API accordingly. The implementation is validated through running these unit tests. As a benchmark, Commit0 is designed to move beyond static one-shot code generation towards agents that must process long-form natural language specifications, adapt to multi-stage feedback, and generate code with complex dependencies. Commit0 also offers an interactive environment where models receive static analysis and execution feedback on the code they generate. Our experiments demonstrate that while current agents can pass some unit tests, none can yet fully reproduce full libraries. Results also show that interactive feedback is quite useful for models to generate code that passes more unit tests, validating the benchmarks that facilitate its use. We publicly release the benchmark, the interactive environment, and the leaderboard. | code generation, language model, evaluation, feedback | We introduce Commit0, a novel benchmark designed to challenge AI agents to write libraries from scratch. | 12,832 | 2412.01769 | [
-0.08041838556528091,
-0.05675105005502701,
-0.03004489839076996,
0.09110962599515915,
0.03918726369738579,
-0.06170958653092384,
-0.08326371014118195,
-0.003862538607791066,
-0.026253150776028633,
0.0031879162415862083,
-0.07306814193725586,
-0.1286085993051529,
0.07697214931249619,
-0.011686624959111214,
0.040157854557037354,
-0.014660216867923737,
0.07510437071323395,
0.03230665251612663,
-0.027869034558534622,
-0.047605000436306,
-0.06528162956237793,
0.03441603109240532,
0.0025121497455984354,
0.008454936556518078,
0.046006787568330765,
-0.014641966670751572,
-0.012205860577523708,
-0.0801154300570488,
0.05324280261993408,
-0.002339036902412772,
0.06804472953081131,
-0.028536567464470863,
-0.023584919050335884,
0.08933427929878235,
0.036380086094141006,
0.018772762268781662,
-0.0011771697318181396,
-0.02280646376311779,
0.036891818046569824,
-0.04235862195491791,
-0.05865329131484032,
0.03435561805963516,
0.04231111332774162,
0.004437037277966738,
0.046107858419418335,
-0.04597267881035805,
-0.03775922954082489,
-0.03854725509881973,
-0.017584877088665962,
0.00751410610973835,
-0.09626712650060654,
-0.08452358841896057,
0.027295392006635666,
-0.0281324852257967,
-0.03374490514397621,
0.050334200263023376,
0.07647372782230377,
-0.035566188395023346,
0.0031885227654129267,
0.0033486741594970226,
-0.03745686635375023,
-0.05309097096323967,
0.011300621554255486,
-0.007147867698222399,
0.010062622837722301,
-0.026483850553631783,
0.03632891923189163,
0.03793875873088837,
0.06066242605447769,
-0.04510575905442238,
-0.06587319821119308,
0.042075954377651215,
0.00780480308458209,
-0.001603839104063809,
-0.019925424829125404,
0.05289923772215843,
-0.0752849206328392,
0.0359497144818306,
0.05031220614910126,
-0.10995861142873764,
-0.04881516844034195,
-0.06166776269674301,
-0.006792859639972448,
0.09938040375709534,
0.025179201737046242,
0.05366400629281998,
0.13415765762329102,
0.11030115932226181,
0.009136554785072803,
0.06561757624149323,
-0.018544761463999748,
-0.010494483634829521,
0.09621672332286835,
0.006702460814267397,
0.02418200485408306,
0.08214794099330902,
0.03383910655975342,
-0.13274233043193817,
-0.05929575115442276,
0.08577542006969452,
-0.05706434324383736,
0.1018291786313057,
0.005061279982328415,
-0.08345939964056015,
-0.03865934908390045,
-0.017763769254088402,
0.0946635752916336,
-0.05166955664753914,
-0.0013129357248544693,
-0.06022809445858002,
0.04444682225584984,
0.047957491129636765,
0.030697669833898544,
-0.04023737832903862,
0.03535391762852669,
-0.00983530655503273,
-0.02068714052438736,
-0.006736344192177057,
0.03624742850661278,
0.026660844683647156,
0.09596803039312363,
0.08887242525815964,
-0.02238161675632,
0.047054193913936615,
0.04789987578988075,
0.03091488592326641,
0.0353497639298439,
6.552588465502409e-34,
0.09724287688732147,
0.02966916188597679,
0.04133191704750061,
0.11468951404094696,
0.08000779896974564,
-0.0407952144742012,
0.06261615455150604,
0.0491177998483181,
-0.10472612082958221,
-0.05578294023871422,
0.03371785208582878,
-0.021139072254300117,
-0.04890036582946777,
0.08211366832256317,
0.003245310625061393,
0.01894931122660637,
-0.041534483432769775,
0.04703734442591667,
-0.05535060539841652,
0.061481863260269165,
0.004529634956270456,
-0.015004689805209637,
-0.013233549892902374,
-0.025836888700723648,
0.026503749191761017,
0.049892980605363846,
0.007408703211694956,
-0.06189006567001343,
-0.03152184933423996,
-0.002419770695269108,
-0.13125188648700714,
-0.058396484702825546,
0.004835091065615416,
0.057252489030361176,
-0.03318103402853012,
-0.0042790863662958145,
-0.059929076582193375,
-0.039973825216293335,
0.015120950527489185,
0.05304926633834839,
-0.03965834155678749,
-0.006800525356084108,
-0.04794736206531525,
0.014472301118075848,
-0.013253671117126942,
-0.04640424996614456,
-0.05344842001795769,
0.008456273935735226,
0.020976711064577103,
-0.06059012934565544,
0.029161255806684494,
0.11082343757152557,
0.011581589467823505,
0.005223365500569344,
-0.010722137987613678,
-0.04342775419354439,
0.035347480326890945,
-0.01373776700347662,
0.028523126617074013,
0.057535137981176376,
0.0032729310914874077,
0.046696364879608154,
-0.06929554045200348,
0.0573933906853199,
-0.03874964267015457,
0.032631974667310715,
-0.025151990354061127,
0.0856132060289383,
0.1254854053258896,
0.044975943863391876,
-0.0405927412211895,
0.03646669164299965,
-0.06604645401239395,
-0.01523690763860941,
0.02067144773900509,
-0.005332423839718103,
0.029484517872333527,
-0.09103556722402573,
0.013594957999885082,
-0.07598676532506943,
-0.011602106504142284,
-0.02873818762600422,
0.00035908460267819464,
0.023519249632954597,
0.014723921194672585,
-0.11526482552289963,
-0.02914394438266754,
0.04942499101161957,
-0.057778652757406235,
-0.05311838537454605,
-0.021224435418844223,
-0.02848578430712223,
0.010002742521464825,
-0.014299857430160046,
0.004842032212764025,
-6.291494639795471e-34,
0.016759714111685753,
-0.04994208365678787,
-0.0666542574763298,
0.07946018874645233,
-0.0364210307598114,
-0.0831257775425911,
-0.0830007940530777,
-0.021269457414746284,
-0.02957141399383545,
0.020343033596873283,
-0.05143071338534355,
-0.028131060302257538,
0.054648470133543015,
0.010954567231237888,
0.031997814774513245,
-0.05837763845920563,
-0.03714055195450783,
-0.06224318966269493,
0.00147999229375273,
0.006403743289411068,
0.04562252759933472,
0.1539423018693924,
0.01922263391315937,
-0.02208242565393448,
-0.05123453214764595,
0.0395173542201519,
-0.04049620032310486,
0.04090028256177902,
-0.009756632149219513,
-0.02757512591779232,
0.03247220441699028,
0.051860954612493515,
-0.05021717771887779,
-0.07069887965917587,
0.01106717437505722,
0.03697190433740616,
0.04714895784854889,
0.0008524618460796773,
-0.05942630395293236,
0.09408894926309586,
0.1014188900589943,
-0.06444436311721802,
-0.046750228852033615,
-0.0018881008727476,
0.006747783627361059,
0.08392304182052612,
-0.08713702112436295,
0.011784546077251434,
-0.007021959871053696,
-0.01333880890160799,
-0.059077125042676926,
0.05406009033322334,
-0.03675730526447296,
-0.061587732285261154,
-0.007796093821525574,
-0.08981037139892578,
-0.05861837416887283,
-0.015415629372000694,
0.04835863783955574,
0.026104791089892387,
-0.08314739912748337,
-0.028357628732919693,
0.03221890330314636,
0.005469807889312506,
-0.02254156768321991,
-0.049345482140779495,
-0.057716891169548035,
-0.017626779153943062,
-0.013429826125502586,
0.04955189302563667,
0.0499362051486969,
-0.04283690080046654,
-0.04336661472916603,
0.07115145027637482,
0.035454098135232925,
0.04063190147280693,
0.0027877904940396547,
-0.032495539635419846,
0.044125620275735855,
-0.06618092954158783,
0.01271693967282772,
0.05632591247558594,
0.0802343487739563,
-0.0079667242243886,
0.029934104532003403,
0.024772251024842262,
0.022538719698786736,
0.05765000730752945,
0.0014430857263505459,
0.03831371292471886,
-0.08083689957857132,
0.029091864824295044,
-0.04596986621618271,
0.008141803555190563,
-0.012236514128744602,
-3.7969918764702015e-8,
-0.020010536536574364,
0.06038593500852585,
-0.026410439983010292,
0.06626871973276138,
-0.03412580117583275,
0.004431940149515867,
-0.05441994220018387,
0.04695762321352959,
0.044135913252830505,
-0.03455755487084389,
0.03749982267618179,
-0.04632720723748207,
-0.014708016067743301,
0.007755670230835676,
-0.0007239520200528204,
0.08077555149793625,
-0.053585510700941086,
-0.01445281133055687,
-0.020610086619853973,
-0.07510436326265335,
0.03587501496076584,
0.06278164684772491,
0.037005968391895294,
-0.030350519344210625,
-0.027192886918783188,
-0.0757816806435585,
0.05757126212120056,
0.0976872444152832,
0.04226800799369812,
0.00449289521202445,
0.054672952741384506,
-0.02313324436545372,
0.032992299646139145,
-0.034219175577163696,
0.05648292601108551,
0.07107585668563843,
0.003325976664200425,
0.016207439824938774,
0.07326120883226395,
-0.047716978937387466,
-0.010239556431770325,
0.0375087633728981,
-0.0815107524394989,
-0.040721602737903595,
0.011589417234063148,
-0.09142050892114639,
-0.02907288819551468,
-0.08242008835077286,
0.04158738628029823,
-0.07921379804611206,
-0.05693146213889122,
-0.08122946321964264,
-0.02409680001437664,
0.06994446367025375,
-0.001928477082401514,
0.05605234205722809,
-0.08495649695396423,
-0.014204730279743671,
0.045790817588567734,
0.01418061088770628,
0.025845777243375778,
0.06775243580341339,
0.04241006821393967,
0.0057000936940312386
] | https://github.com/commit-0/commit0 | 144 | 0 | 0 | 0 |
Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control | https://openreview.net/forum?id=1Njl73JKjB | [
"Aleksandar Makelov",
"Georg Lange",
"Neel Nanda"
] | Poster | Disentangling model activations into human-interpretable features is a central
problem in interpretability. Sparse autoencoders (SAEs) have recently attracted
much attention as a scalable unsupervised approach to this problem. However, our
imprecise understanding of ground-truth features in realistic scenarios makes it
difficult to measure the success of SAEs. To address this challenge, we propose
to evaluate SAEs on specific tasks by comparing them to supervised
feature dictionaries computed with knowledge of the concepts relevant to the
task.
Specifically, we suggest that it is possible to (1) compute supervised sparse
feature dictionaries that disentangle model computations for a specific task;
(2) use them to evaluate and contextualize the degree of disentanglement and
control offered by SAE latents on this task. Importantly, we can do this in a
way that is agnostic to whether the SAEs have learned the exact ground-truth
features or a different but similarly useful representation.
As a case study, we apply this framework to the indirect object identification
(IOI) task using GPT-2 Small, with SAEs trained on either the IOI or OpenWebText
datasets. We find that SAEs capture interpretable features for the IOI task, and
that more recent SAE variants such as Gated SAEs and Top-K SAEs are competitive
with supervised features in terms of disentanglement and control over the model.
We also exhibit, through this setup and toy models, some qualitative phenomena
in SAE training illustrating feature splitting and the role of feature
magnitudes in solutions preferred by SAEs. | mechanistic interpretability, sparse autoencoders, evaluations | We compute and validate *supervised* sparse feature dictionaries on the IOI task, and then compare SAEs against them | 12,831 | 2405.08366 | [
-0.08586516976356506,
-0.04358377680182457,
-0.015874633565545082,
0.08122994005680084,
0.08421393483877182,
-0.044661883264780045,
0.025005539879202843,
-0.048759229481220245,
-0.02554558590054512,
-0.06249815970659256,
-0.06014002859592438,
-0.08349364995956421,
0.005259925499558449,
0.087423175573349,
-0.04639950767159462,
0.0017171681392937899,
0.12685583531856537,
0.07902009040117264,
-0.14529739320278168,
-0.007869379594922066,
0.027339469641447067,
0.02117898128926754,
0.016796819865703583,
0.020937098190188408,
0.023034842684864998,
-0.03082650899887085,
0.022590793669223785,
-0.037461522966623306,
0.03294697031378746,
-0.021029407158493996,
0.03306160867214203,
-0.0019777314737439156,
-0.02783183939754963,
0.020768044516444206,
0.011736572720110416,
0.09679640829563141,
0.04480193555355072,
0.08029363304376602,
0.0026695262640714645,
-0.05438603088259697,
0.04339291900396347,
0.019060228019952774,
-0.022715134546160698,
0.0584544874727726,
0.07868392020463943,
0.017452603206038475,
-0.015042594633996487,
-0.01565699465572834,
0.00069730932591483,
-0.04351090267300606,
0.010542026720941067,
0.07511664927005768,
0.012165715917944908,
0.0064698439091444016,
-0.038296815007925034,
-0.016845304518938065,
0.01654324121773243,
0.032922033220529556,
-0.020241504535079002,
0.02848328836262226,
-0.013157029636204243,
-0.10848735272884369,
-0.028027143329381943,
0.00397141557186842,
-0.05145219713449478,
-0.019970010966062546,
0.04644901677966118,
-0.0073287710547447205,
-0.0328650176525116,
-0.008022249676287174,
0.024205295369029045,
0.05828198045492172,
-0.03473944962024689,
0.004909711889922619,
0.05315163731575012,
-0.04090973362326622,
0.03342219814658165,
0.020452480763196945,
0.07228045165538788,
-0.05088822916150093,
-0.02889307029545307,
0.049356553703546524,
0.01654067635536194,
-0.012245108373463154,
0.08104845136404037,
-0.008252960629761219,
-0.004483876749873161,
0.09853740781545639,
-0.015729790553450584,
0.06516615301370621,
-0.08718685805797577,
-0.12816953659057617,
0.0006974493153393269,
-0.050714872777462006,
0.10403487086296082,
0.003140481887385249,
0.024554552510380745,
-0.06593741476535797,
-0.09851862490177155,
0.07083629816770554,
-0.03315567225217819,
0.012240495532751083,
-0.041733596473932266,
-0.0279257632791996,
0.042358506470918655,
0.02249465510249138,
0.015912428498268127,
-0.004104455001652241,
0.022233914583921432,
-0.03331448882818222,
0.03613344952464104,
0.05608752742409706,
0.02736501954495907,
0.009441541507840157,
0.012112442404031754,
-0.031607531011104584,
0.04850197955965996,
-0.021199002861976624,
0.06125165522098541,
0.017038904130458832,
-0.06125261262059212,
-0.031471092253923416,
0.05940530449151993,
0.0473061241209507,
0.07594931870698929,
-0.05164143070578575,
-0.028942150995135307,
3.0364725385258302e-33,
0.02800901234149933,
0.034339722245931625,
0.013542039319872856,
-0.016306914389133453,
-0.0006348152528516948,
-0.023581285029649734,
-0.06034328043460846,
0.07585135847330093,
-0.019703295081853867,
0.11562519520521164,
-0.036158207803964615,
0.06953732669353485,
-0.07757166028022766,
0.02953224442899227,
0.038742758333683014,
0.015842337161302567,
-0.019316360354423523,
0.02231443114578724,
-0.03252999112010002,
-0.053402483463287354,
0.026554124429821968,
-0.008058924227952957,
-0.078470379114151,
-0.015022402629256248,
-0.0138259157538414,
-0.0786387249827385,
0.07426654547452927,
-0.0440150648355484,
-0.032251738011837006,
0.001945576397702098,
-0.10661853849887848,
-0.04492313414812088,
0.06295114755630493,
0.05991833284497261,
-0.020671943202614784,
-0.02695770375430584,
-0.01973714306950569,
-0.06520193815231323,
0.03663339093327522,
-0.00864795595407486,
0.034554675221443176,
0.01652897335588932,
-0.04259651526808739,
-0.08291606605052948,
-0.06927438825368881,
-0.046309944242239,
0.08531691879034042,
-0.01794787123799324,
-0.0836348757147789,
-0.0038901646621525288,
0.029069004580378532,
0.05564204975962639,
-0.014384222216904163,
-0.071769580245018,
-0.03141215816140175,
-0.033470187336206436,
-0.015348275192081928,
0.0022969383280724287,
0.0777430534362793,
0.033401794731616974,
0.07097866386175156,
0.029062125831842422,
-0.019405648112297058,
0.06592752784490585,
0.01196572370827198,
0.0480463020503521,
-0.030889954417943954,
0.054636962711811066,
-0.007172117475420237,
0.00014327683311421424,
-0.09035288542509079,
0.007769260089844465,
0.023278631269931793,
-0.07609544694423676,
0.12912873923778534,
0.004310596734285355,
0.07943565398454666,
-0.07698021084070206,
-0.0053201462142169476,
0.035725921392440796,
0.05409959703683853,
0.008502324111759663,
-0.029187973588705063,
-0.01617489755153656,
-0.09536585956811905,
0.006363169755786657,
0.020455429330468178,
-0.10499419271945953,
-0.08361832797527313,
-0.04526476189494133,
-0.055405497550964355,
-0.046798549592494965,
-0.01109292171895504,
-0.03896934539079666,
0.03230166435241699,
-2.2039235217681602e-33,
-0.05623312294483185,
0.025731774047017097,
0.02209966629743576,
0.06369002163410187,
-0.026478715240955353,
-0.05350199341773987,
-0.0025591079611331224,
0.02498597279191017,
0.03423594310879707,
-0.0395316518843174,
-0.029389336705207825,
-0.08335627615451813,
0.039068471640348434,
-0.00595049886032939,
-0.0032715469133108854,
0.08241792023181915,
-0.0870853066444397,
-0.017785154283046722,
0.01640223152935505,
0.08799777925014496,
-0.023452693596482277,
0.09330844134092331,
-0.08441758900880814,
-0.00047440623166039586,
-0.08237432688474655,
0.052692677825689316,
0.014002989046275616,
0.023003529757261276,
0.03609085455536842,
-0.025436146184802055,
0.0009183585061691701,
0.039575327187776566,
-0.09670490771532059,
-0.06261009722948074,
-0.1133897602558136,
-0.03351794555783272,
0.024293622002005577,
0.004307750146836042,
-0.07369054108858109,
0.051324229687452316,
0.0954873114824295,
-0.028680916875600815,
-0.061489932239055634,
0.023354575037956238,
0.0018913120729848742,
-0.04752055183053017,
-0.058493733406066895,
0.021305853500962257,
0.01695416308939457,
0.0009982737246900797,
-0.017752889543771744,
0.063447505235672,
-0.13222533464431763,
0.08777591586112976,
-0.06253916770219803,
-0.0675211176276207,
0.03740472346544266,
0.005211721174418926,
0.08084290474653244,
0.003429103409871459,
-0.014065379276871681,
-0.024525318294763565,
-0.03464621305465698,
0.06429578363895416,
0.040990639477968216,
0.008696814998984337,
-0.013598572462797165,
0.051396530121564865,
0.017146743834018707,
0.07678946107625961,
0.04467073082923889,
-0.03153005987405777,
0.025337791070342064,
-0.03740157559514046,
0.024482227861881256,
0.03642288222908974,
0.04658889025449753,
-0.07848874479532242,
0.00737830949947238,
-0.005317311733961105,
0.05323951691389084,
-0.04171949625015259,
0.028297586366534233,
0.035630375146865845,
0.012241757474839687,
0.06476574391126633,
0.01210678368806839,
0.043921537697315216,
-0.020587656646966934,
0.05362552031874657,
-0.06945287436246872,
0.03295230492949486,
0.00921545922756195,
0.07239885628223419,
-0.028312256559729576,
-4.665807651349496e-8,
-0.04225531220436096,
0.036012377589941025,
-0.02844647690653801,
-0.054251592606306076,
-0.06322567164897919,
-0.03343094885349274,
-0.008294699713587761,
0.03663147985935211,
-0.04571612924337387,
-0.0036423862911760807,
0.05545786768198013,
-0.02778402529656887,
-0.09797658771276474,
-0.07793769240379333,
0.012411192990839481,
0.15549558401107788,
0.04002237692475319,
-0.008661928586661816,
0.03583665192127228,
0.02127954363822937,
0.13074329495429993,
0.0008511439664289355,
0.0038604438304901123,
-0.02933402918279171,
0.011043988168239594,
-0.053557612001895905,
-0.0274224653840065,
0.15968959033489227,
0.05285242944955826,
0.05365310609340668,
-0.02625306323170662,
0.037530217319726944,
0.018369009718298912,
-0.008713386952877045,
0.00048311761929653585,
0.06275390088558197,
0.0138861620798707,
-0.029024187475442886,
0.006276414729654789,
-0.02890119142830372,
0.014331820420920849,
0.060152798891067505,
-0.035241417586803436,
0.03161662444472313,
-0.034184396266937256,
0.05804523453116417,
0.017029253765940666,
-0.10469421744346619,
0.03223143517971039,
0.010661729611456394,
0.05471619963645935,
-0.017196010798215866,
0.009407267905771732,
0.1269889771938324,
0.05413389950990677,
0.036279190331697464,
0.010580741800367832,
-0.0908576026558876,
0.0183528121560812,
0.058369554579257965,
-0.04470792040228844,
0.08590168505907059,
-0.0636444017291069,
-0.0057943109422922134
] | 0 | 0 | 0 | 0 |
|
CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening | https://openreview.net/forum?id=fv9XU7CyN2 | [
"Gen Zhou",
"Sugitha Janarthanan",
"Yutong Lu",
"Pingzhao Hu"
] | Poster | Due to the rise in antimicrobial resistance, identifying novel compounds with antibiotic potential is crucial for combatting this global health issue. However, traditional drug development methods are costly and inefficient. Recognizing the pressing need for more effective solutions, researchers have turned to machine learning techniques to streamline the prediction and development of novel antibiotic compounds. While foundation models have shown promise in antibiotic discovery, current mainstream efforts still fall short of fully leveraging the potential of multimodal molecular data. Recent studies suggest that contrastive learning frameworks utilizing multimodal data exhibit excellent performance in representation learning across various domains. Building upon this, we introduce CL-MFAP, an unsupervised contrastive learning (CL)-based multimodal foundation (MF) model specifically tailored for discovering small molecules with potential antibiotic properties (AP) using three types of molecular data. This model employs 1.6 million bioactive molecules with drug-like properties from the ChEMBL dataset to jointly pretrain three encoders: (1) a transformer-based encoder with rotary position embedding for processing SMILES strings; (2) another transformer-based encoder, incorporating a novel bi-level routing attention mechanism to handle molecular graph representations; and (3) a Morgan fingerprint encoder using a multilayer perceptron, to achieve the contrastive learning purpose. The CL-MFAP outperforms baseline models in antibiotic property prediction by effectively utilizing different molecular modalities and demonstrates superior domain-specific performance when fine-tuned for antibiotic-related property prediction tasks. | Contrastive Learning, Multimodal Foundation Model, Antibiotic Property Prediction, Bi-level Routing Attention, Transformer | null | 12,828 | null | [
0.033445850014686584,
-0.14832238852977753,
-0.014391890726983547,
0.011011935770511627,
0.10584348440170288,
0.0007998367073014379,
-0.09553483873605728,
0.03756796196103096,
0.0648987889289856,
-0.027729906141757965,
-0.043848250061273575,
-0.09312218427658081,
0.033582620322704315,
0.11849736422300339,
-0.07374740391969681,
0.06190577521920204,
-0.010166537947952747,
0.07075582444667816,
-0.0695459172129631,
0.014132270589470863,
0.005749973934143782,
0.07446905225515366,
0.06953812390565872,
-0.002115316689014435,
0.045273471623659134,
0.05972176417708397,
0.04105186462402344,
0.0015646786196157336,
-0.045121658593416214,
-0.05156299099326134,
0.06676473468542099,
0.02623911201953888,
0.02122107893228531,
-0.0060508958995342255,
-0.004625944420695305,
0.057403564453125,
-0.04811745882034302,
0.03879839926958084,
0.055798228830099106,
-0.03728480264544487,
-0.022974565625190735,
-0.016609912738204002,
0.031846340745687485,
0.022928860038518906,
0.05165907368063927,
-0.037631627172231674,
-0.001365453121252358,
-0.03033997118473053,
0.1535732001066208,
-0.026186129078269005,
-0.02090715616941452,
-0.01153193786740303,
-0.04351865127682686,
0.07555454224348068,
0.03469960764050484,
0.00596057903021574,
-0.04474661499261856,
-0.060025110840797424,
-0.026998061686754227,
-0.06537382304668427,
0.018741166219115257,
-0.08894181996583939,
0.012765388004481792,
-0.002822085050866008,
-0.03672469034790993,
0.01495081651955843,
0.015472187660634518,
0.04333378002047539,
-0.011395803652703762,
-0.007198258303105831,
-0.08025364577770233,
0.027417762205004692,
-0.005319578573107719,
0.058920662850141525,
0.07083165645599365,
0.06552888453006744,
0.11645976454019547,
0.008254663087427616,
0.04613366350531578,
-0.09968176484107971,
0.04673133045434952,
0.02211490273475647,
0.06594929844141006,
0.0322280190885067,
0.14567916095256805,
0.017254171893000603,
0.044846080243587494,
0.045557644218206406,
-0.039790939539670944,
-0.002565340604633093,
0.039055999368429184,
0.012223813682794571,
0.03738650679588318,
-0.014650559052824974,
-0.09034943580627441,
-0.03229789435863495,
0.010261363349854946,
-0.04352428391575813,
-0.05214497447013855,
0.05871967971324921,
-0.07135945558547974,
-0.0037963571958243847,
0.008863292634487152,
-0.014526584185659885,
0.017597440630197525,
-0.09758879989385605,
0.05926918983459473,
-0.010435210540890694,
0.042083900421857834,
-0.046737559139728546,
0.041316304355859756,
0.09189708530902863,
0.01569839008152485,
0.026563512161374092,
0.04106106609106064,
-0.030274327844381332,
0.057003796100616455,
-0.08139583468437195,
0.027361435815691948,
0.041531577706336975,
-0.032686058431863785,
-0.0221248809248209,
-0.031256966292858124,
-0.022583503276109695,
0.019283851608633995,
-0.013632177375257015,
-0.08936715126037598,
1.583533120782296e-34,
0.007544457446783781,
0.01578456163406372,
0.06610631197690964,
0.047263793647289276,
0.043530482798814774,
-0.09401199221611023,
-0.02918146178126335,
0.014071538113057613,
-0.058354537934064865,
0.020208606496453285,
-0.012130764313042164,
0.05333588272333145,
-0.0465802364051342,
0.12500956654548645,
-0.03417104855179787,
-0.03618590533733368,
-0.01653878390789032,
-0.0435062050819397,
0.0023458609357476234,
-0.02275892347097397,
-0.0746157318353653,
-0.020029693841934204,
-0.022924372926354408,
0.011053409427404404,
-0.009879070334136486,
0.015669213607907295,
0.04357781261205673,
0.011265802197158337,
0.033724233508110046,
0.001947508193552494,
-0.16536585986614227,
0.01747812330722809,
-0.04337725415825844,
-0.005581479053944349,
0.0006638295599259436,
-0.07813822478055954,
-0.01940545067191124,
-0.06394540518522263,
0.0511024035513401,
-0.02600240334868431,
-0.06331966817378998,
0.014290450140833855,
-0.03264138102531433,
-0.05278700962662697,
-0.04067371413111687,
-0.009165502153337002,
-0.008058824576437473,
-0.015569004230201244,
-0.060646411031484604,
-0.02024541236460209,
0.03026447631418705,
-0.03681723028421402,
-0.015045360662043095,
-0.03335307538509369,
-0.019746879115700722,
0.007521900348365307,
-0.04454036056995392,
0.055378325283527374,
0.05683216452598572,
0.00856971088796854,
-0.05053243786096573,
0.0565352737903595,
-0.02030145563185215,
0.09473851323127747,
-0.02506042830646038,
0.055584441870450974,
-0.07473736256361008,
0.017070641741156578,
0.07588286697864532,
0.03546122461557388,
0.004875807091593742,
0.06942267715930939,
0.015403330326080322,
-0.030214309692382812,
0.04532181844115257,
0.027006354182958603,
0.06186307594180107,
-0.016443872824311256,
0.06590520590543747,
0.09583256393671036,
-0.0013434397988021374,
0.03183026239275932,
-0.036661867052316666,
0.0024334413465112448,
-0.11034096777439117,
-0.01147108618170023,
0.05537613481283188,
-0.09538202732801437,
-0.04549528285861015,
-0.08344566076993942,
-0.08281511813402176,
-0.049031227827072144,
-0.015572651289403439,
-0.027143919840455055,
0.0017842650413513184,
-1.9570313208580858e-33,
0.0718231350183487,
0.013010459020733833,
0.04093705490231514,
-0.028706343844532967,
-0.0006434418028220534,
0.032004520297050476,
-0.03314721956849098,
0.0004395696159917861,
0.05457591637969017,
-0.044374123215675354,
0.13465087115764618,
-0.009846139699220657,
-0.005015104077756405,
-0.05927121639251709,
0.025622162967920303,
0.015822868794202805,
-0.015576113015413284,
0.02831828035414219,
-0.01953168772161007,
0.014210340566933155,
-0.0397796630859375,
0.07375975698232651,
-0.035628609359264374,
0.06178893521428108,
-0.09049507975578308,
0.02827651984989643,
-0.024151822552084923,
-0.02421201579272747,
-0.06249430030584335,
-0.02399715594947338,
-0.00897468812763691,
-0.022026266902685165,
-0.027000805363059044,
-0.02705688774585724,
-0.08674541115760803,
0.0915747880935669,
-0.02115873619914055,
-0.07759605348110199,
-0.06032712757587433,
0.034245047718286514,
0.057341206818819046,
0.033676523715257645,
-0.12039747834205627,
-0.008214635774493217,
0.00021138535521458834,
-0.01963936537504196,
-0.03137996420264244,
-0.03092358633875847,
0.1078186184167862,
-0.0020279614254832268,
0.04310678690671921,
0.021261775866150856,
-0.026172909885644913,
-0.06849119067192078,
-0.05602290481328964,
-0.06566834449768066,
0.05746563896536827,
-0.04254193231463432,
0.10097305476665497,
0.05127835273742676,
-0.09318795800209045,
-0.08077526837587357,
-0.00857893843203783,
-0.03184438496828079,
-0.09174839407205582,
-0.0255350973457098,
0.018140826374292374,
-0.00895625539124012,
0.05810799077153206,
-0.027803244069218636,
-0.02370244637131691,
0.030401170253753662,
0.05140256509184837,
0.01368134655058384,
0.021129172295331955,
-0.027429580688476562,
-0.02050267532467842,
-0.07283109426498413,
-0.048365868628025055,
0.011975600384175777,
0.07188884913921356,
-0.13479597866535187,
-0.02696477621793747,
0.061935096979141235,
0.0030588675290346146,
0.06332655996084213,
0.050707463175058365,
-0.04132881388068199,
0.08358845114707947,
-0.062002114951610565,
-0.017370065674185753,
0.052261173725128174,
-0.03407479077577591,
0.03243162855505943,
0.026557771489024162,
-3.470904985647394e-8,
0.05295298621058464,
-0.01688236929476261,
0.06998936831951141,
-0.06711982190608978,
0.019802024587988853,
-0.005632313434034586,
-0.07208985090255737,
0.015281478874385357,
0.008694139309227467,
0.047283388674259186,
0.04178429767489433,
0.033549413084983826,
-0.05784868448972702,
-0.048828184604644775,
-0.0371394157409668,
0.06376299262046814,
0.04059759899973869,
-0.03443380817770958,
0.013593875803053379,
0.027290834113955498,
-0.004688269458711147,
0.047066524624824524,
0.012142893858253956,
-0.018374813720583916,
0.1119566559791565,
-0.10959016531705856,
-0.049881238490343094,
-0.06034133955836296,
0.02252061665058136,
0.0215466246008873,
0.03080347366631031,
0.03389224782586098,
0.034101806581020355,
0.012009872123599052,
-0.013106794096529484,
0.020556792616844177,
0.009934420697391033,
-0.05342302471399307,
-0.09953632950782776,
0.06148587167263031,
0.010584634728729725,
0.042958300560712814,
-0.10833752900362015,
0.008118724450469017,
-0.0827292650938034,
0.04298882558941841,
-0.001757809892296791,
0.0006845712196081877,
0.007509693503379822,
-0.08293986320495605,
0.09403727203607559,
0.012547108344733715,
0.054225772619247437,
0.04128831624984741,
0.04436620697379112,
0.024836303666234016,
-0.020046912133693695,
-0.0074929469265043736,
0.035124681890010834,
-0.04533777013421059,
0.01859632320702076,
-0.057244669646024704,
-0.03615986928343773,
0.02173619158565998
] | 0 | 0 | 0 | 0 |
|
SeedLM: Compressing LLM Weights into Seeds of Pseudo-Random Generators | https://openreview.net/forum?id=u3TL0qxLWf | [
"Rasoul Shafipour",
"David Harrison",
"Maxwell Horton",
"JEFFREY MARKER",
"Houman Bedayat",
"Sachin Mehta",
"Mohammad Rastegari",
"Mahyar Najibi",
"Saman Naderiparizi"
] | Poster | Large Language Models (LLMs) have transformed natural language processing, but face significant challenges in widespread deployment due to their high runtime cost. In this paper, we introduce SeedLM, a novel post-training compression method that uses seeds of a pseudo-random generator to encode and compress model weights. Specifically, for each block of weights, we find a seed that is fed into a Linear Feedback Shift Register (LFSR) during inference to efficiently generate a random matrix. This matrix is then linearly combined with compressed coefficients to reconstruct the weight block. SeedLM reduces memory access and leverages idle compute cycles during inference, effectively speeding up memory-bound tasks by trading compute for fewer memory accesses. Unlike state-of-the-art methods that rely on calibration data, our approach is data-free and generalizes well across diverse tasks. Our experiments with Llama3 70B, which is particularly challenging, show zero-shot accuracy retention at 4- and 3-bit compression to be on par with or better than state-of-the-art methods, while maintaining performance comparable to FP16 baselines. Additionally, FPGA-based tests demonstrate that 4-bit SeedLM, as model size increases, approaches a 4x speed-up over an FP16 Llama 2/3 baseline. | Model Compression, Large Language Models, Post-Training Quantization | SeedLM is a novel post-training compression technique that uses pseudo-random generators to compress model weights, achieving significant memory savings and inference speed improvements with minimal accuracy loss in large language models. | 12,825 | 2410.10714 | [
-0.07621403783559799,
-0.015244678594172001,
-0.07339181751012802,
0.05882937088608742,
0.07770479470491409,
0.004872154910117388,
-0.06905721127986908,
-0.07656196504831314,
0.043559279292821884,
-0.013481540605425835,
-0.028101162984967232,
-0.03759533166885376,
-0.0047981045208871365,
0.008951421827077866,
0.014073534868657589,
0.02503088302910328,
0.06117749214172363,
0.10297247022390366,
-0.039198290556669235,
-0.03797439858317375,
-0.01328390371054411,
0.018895287066698074,
0.03505784645676613,
0.04954145848751068,
0.11707916855812073,
-0.06552498787641525,
-0.010735416784882545,
-0.014742516912519932,
0.04333039000630379,
-0.009513067081570625,
0.029145777225494385,
0.040955107659101486,
0.03284033387899399,
0.01706801727414131,
-0.04305943474173546,
0.06330536305904388,
-0.03669100999832153,
-0.010056188330054283,
-0.0024186510127037764,
-0.04092838987708092,
0.009537328034639359,
0.019265927374362946,
0.07022807747125626,
0.0752793550491333,
0.04329811781644821,
0.02477794885635376,
0.023938527330756187,
-0.019458547234535217,
0.012681995518505573,
0.04865565896034241,
-0.0477219820022583,
0.015554534271359444,
-0.010163426399230957,
0.03700512647628784,
-0.15257307887077332,
-0.03364267945289612,
0.07901766896247864,
0.01079635415226221,
-0.049120474606752396,
0.0078108785673975945,
-0.06812699884176254,
-0.09609091281890869,
-0.04216091334819794,
0.018846048042178154,
-0.04212239012122154,
-0.026202095672488213,
0.06066274642944336,
0.057061705738306046,
0.00823878962546587,
-0.03275524079799652,
-0.02271350659430027,
0.04642145335674286,
-0.005073911976069212,
0.1198258325457573,
-0.015345737338066101,
0.05213272199034691,
0.002904654713347554,
-0.025526639074087143,
0.10788900405168533,
-0.05304393917322159,
0.0061974734999239445,
-0.0632954016327858,
0.044232264161109924,
-0.022031322121620178,
0.0324273556470871,
-0.013070652261376381,
0.038362886756658554,
0.08638261258602142,
0.023100120946764946,
0.030983222648501396,
-0.030727123841643333,
-0.022541580721735954,
0.026427313685417175,
0.02617179974913597,
0.0009212175500579178,
0.04010952264070511,
0.01401111576706171,
-0.07238417118787766,
-0.0547255203127861,
0.09631164371967316,
0.0306144617497921,
0.10057244449853897,
0.004257878754287958,
0.005322374403476715,
-0.02576165273785591,
-0.017295844852924347,
0.03561854362487793,
0.05579521879553795,
-0.015862882137298584,
-0.04146571457386017,
0.11034703254699707,
0.09548546373844147,
0.03152306377887726,
0.014934121631085873,
0.050960466265678406,
-0.0003706665011122823,
-0.03808898851275444,
-0.06585909426212311,
0.006929187104105949,
0.06089724227786064,
-0.07461065798997879,
0.022876860573887825,
-0.06506279855966568,
-0.013906695879995823,
0.0090164914727211,
-0.07258985191583633,
-0.056613482534885406,
1.5089832920120205e-33,
0.07395573705434799,
0.060554999858140945,
0.02770143374800682,
0.0429251529276371,
0.05496161803603172,
0.0007035073358565569,
0.08019071817398071,
0.05031107738614082,
-0.05263395980000496,
-0.008190440014004707,
-0.030535096302628517,
-0.02478615567088127,
-0.0222501028329134,
0.06641910970211029,
0.05375190079212189,
-0.03255121409893036,
-0.02327115647494793,
0.05021044984459877,
-0.019583076238632202,
-0.011216964572668076,
0.047645848244428635,
0.018797067925333977,
-0.0031810703221708536,
-0.054229192435741425,
0.04302750527858734,
0.016058996319770813,
0.05788391828536987,
-0.06939371675252914,
-0.06554663181304932,
0.01589229702949524,
-0.09007041156291962,
-0.013787472620606422,
0.0256283450871706,
-0.01578376442193985,
0.04983331263065338,
-0.051939040422439575,
-0.00343116233125329,
-0.052971500903367996,
0.01167940441519022,
-0.02337069995701313,
0.04853425174951553,
0.04262279346585274,
0.04341146722435951,
-0.018209949135780334,
-0.12332180142402649,
-0.07332105189561844,
-0.001150196767412126,
0.04690571501851082,
-0.04817347228527069,
0.008302995003759861,
0.11870259046554565,
0.010491721332073212,
-0.09953812509775162,
0.02980661578476429,
-0.01806538924574852,
-0.036244627088308334,
0.02478604018688202,
-0.014303650707006454,
0.07313261926174164,
0.06632965803146362,
-0.03116181679069996,
0.05902925878763199,
-0.043917570263147354,
0.0819573923945427,
0.030937639996409416,
-0.0012567786034196615,
-0.040229856967926025,
0.021967319771647453,
-0.010087653063237667,
-0.023495929315686226,
0.006893996149301529,
-0.03568625822663307,
0.00028158447821624577,
-0.017374981194734573,
0.03630509227514267,
0.003164480207487941,
0.056353405117988586,
-0.10736376792192459,
-0.054033488035202026,
0.007730886340141296,
0.08404076099395752,
0.028614358976483345,
-0.0014306724769994617,
-0.02710871584713459,
-0.07892822474241257,
-0.08875231444835663,
0.023001648485660553,
-0.09933047741651535,
-0.006624145898967981,
-0.05713953077793121,
-0.034432921558618546,
-0.07353980094194412,
0.045227278023958206,
-0.06191318854689598,
-0.023357192054390907,
-2.522010961485977e-34,
-0.023523125797510147,
0.0014345350209623575,
-0.050648413598537445,
0.12011799216270447,
0.007583634927868843,
-0.11448630690574646,
0.0257958322763443,
-0.0037440257146954536,
-0.02223990298807621,
-0.09456630796194077,
-0.05841108784079552,
-0.02604766935110092,
0.03892989456653595,
0.05135341361165047,
0.07199802994728088,
-0.03766096755862236,
-0.023474251851439476,
0.02215382643043995,
-0.06296219676733017,
0.005456068087369204,
-0.04993497207760811,
0.12685298919677734,
-0.03665686398744583,
0.01965518482029438,
0.008049237541854382,
0.03361260145902634,
-0.029572246596217155,
0.10239938646554947,
-0.010389373637735844,
-0.0228579044342041,
-0.06079685315489769,
0.01696198619902134,
-0.026433760300278664,
-0.06029397249221802,
-0.05755305662751198,
0.05527061969041824,
0.06840360164642334,
0.06546548753976822,
0.00998465996235609,
0.04558851197361946,
0.06585004925727844,
0.0017968323081731796,
-0.07199686765670776,
0.009434280917048454,
0.006681560073047876,
-0.001560287899337709,
-0.09154389053583145,
0.008206713944673538,
0.06903424113988876,
0.022144228219985962,
-0.057922832667827606,
0.07915108650922775,
-0.08023600280284882,
0.05658064782619476,
0.02015325240790844,
-0.13637427985668182,
-0.15957529842853546,
-0.034137994050979614,
-0.006987681146711111,
0.009537034668028355,
-0.06430596858263016,
-0.00893569365143776,
0.01505353581160307,
-0.11097005754709244,
0.0520055815577507,
-0.06646525859832764,
-0.006851355545222759,
0.021217085421085358,
0.004358880687505007,
0.07262083142995834,
0.03282281011343002,
0.0034249224700033665,
0.017503472045063972,
0.07369393110275269,
-0.018201734870672226,
-0.010849600657820702,
0.04401058703660965,
-0.017352068796753883,
-0.01816459745168686,
0.026133807376027107,
0.04587695747613907,
0.013137583620846272,
0.028541168197989464,
0.034536756575107574,
0.05564097315073013,
0.07987985759973526,
0.014936358667910099,
0.0024428723845630884,
0.005882041994482279,
0.012810363434255123,
-0.060999128967523575,
0.059763699769973755,
0.06409779191017151,
0.05263020843267441,
0.006081889383494854,
-4.068273895541097e-8,
-0.04452921077609062,
0.04559079185128212,
-0.014331403188407421,
0.053258828818798065,
0.006772611755877733,
-0.028644395992159843,
-0.011434701271355152,
0.0700177550315857,
0.03848271816968918,
-0.0726398155093193,
0.09398018568754196,
-0.02874187007546425,
-0.06587248295545578,
-0.062394704669713974,
-0.04443061724305153,
0.06684382259845734,
0.0014704434433951974,
-0.014464296400547028,
-0.02108617126941681,
-0.05861764773726463,
-0.024402661249041557,
0.035425975918769836,
0.006713948678225279,
0.007944049313664436,
0.0004609434399753809,
-0.08961858600378036,
-0.009471737779676914,
0.0815213993191719,
0.17886973917484283,
-0.0015068416250869632,
-0.060769855976104736,
-0.0035952564794570208,
0.03503525257110596,
-0.08136473596096039,
0.03894279897212982,
0.05152883380651474,
0.03252340853214264,
0.017846940085291862,
0.0825238972902298,
-0.05426381900906563,
0.05100981891155243,
0.016157569363713264,
-0.06232982501387596,
0.01248862687498331,
-0.04869978502392769,
-0.06527518481016159,
-0.020685764029622078,
-0.06592784076929092,
-0.01851574331521988,
-0.05415507033467293,
0.02164190448820591,
-0.022490356117486954,
-0.03598315641283989,
0.05464361608028412,
0.061215292662382126,
0.014456584118306637,
-0.05057116970419884,
0.023949863389134407,
0.05444813147187233,
0.06270739436149597,
-0.01642213761806488,
-0.029085276648402214,
-0.04820408299565315,
-0.02389172837138176
] | 0 | 0 | 0 | 0 |
|
Dissecting Adversarial Robustness of Multimodal LM Agents | https://openreview.net/forum?id=YauQYh2k1g | [
"Chen Henry Wu",
"Rishi Rajesh Shah",
"Jing Yu Koh",
"Russ Salakhutdinov",
"Daniel Fried",
"Aditi Raghunathan"
] | Poster | As language models (LMs) are used to build autonomous agents in real environments, ensuring their adversarial robustness becomes a critical challenge. Unlike chatbots, agents are compound systems with multiple components taking actions, which existing LMs safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation scripts in a realistic threat model on top of VisualWebArena, a real environment for web agents. To systematically examine the robustness of agents, we propose the Agent Robustness Evaluation (ARE) framework. ARE views the agent as a graph showing the flow of intermediate outputs between components and decomposes robustness as the flow of adversarial information on the graph. We find that we can successfully break latest agents that use black-box frontier LMs, including those that perform reflection and tree search. With imperceptible perturbations to a single image (less than 5% of total web page pixels), an attacker can hijack these agents to execute targeted adversarial goals with success rates up to 67%. We also use ARE to rigorously evaluate how the robustness changes as new components are added. We find that inference-time compute that typically improves benign performance can open up new vulnerabilities and harm robustness. An attacker can compromise the evaluator used by the reflexion agent and the value function of the tree search agent, which increases the attack success relatively by 15% and 20%. Our data and code for attacks, defenses, and evaluation are at https://github.com/ChenWu98/agent-attack | LM agents, multimodal agents, safety, adversarial robustness | null | 12,824 | 2406.12814 | [
-0.043882861733436584,
-0.11896183341741562,
-0.03929384797811508,
0.05566156283020973,
0.1261628270149231,
-0.013768014498054981,
-0.032655563205480576,
-0.06021055951714516,
-0.03381681814789772,
-0.008986511267721653,
-0.009232735261321068,
-0.0878797397017479,
0.10339812934398651,
0.013232680037617683,
-0.015977265313267708,
0.007579404395073652,
0.10204123705625534,
0.06319884955883026,
-0.10043063759803772,
0.014880555681884289,
0.057875856757164,
0.02248169109225273,
-0.01651463285088539,
-0.06758913397789001,
-0.017114704474806786,
-0.033791981637477875,
-0.04298011213541031,
-0.04792485386133194,
0.008023387752473354,
-0.056391216814517975,
0.04123279079794884,
-0.05557616800069809,
-0.049101367592811584,
0.03951999172568321,
-0.07741453498601913,
0.06978362798690796,
-0.061790354549884796,
-0.021462639793753624,
0.07359332591295242,
-0.026243938133120537,
-0.062396395951509476,
0.004233726300299168,
-0.03554503619670868,
-0.06647367030382156,
0.004455627873539925,
-0.05189744383096695,
-0.020434625446796417,
-0.037655141204595566,
-0.0046427748166024685,
-0.03286905214190483,
-0.07395906746387482,
-0.044319722801446915,
0.01596960797905922,
0.05964247137308121,
-0.08860763162374496,
-0.0015414818190038204,
0.008360042236745358,
0.023469215258955956,
0.04569306969642639,
0.03552594780921936,
0.012930191121995449,
-0.009192097932100296,
0.03263673558831215,
0.0017410006839782,
0.032216764986515045,
-0.014698881655931473,
0.0031804582104086876,
-0.023044150322675705,
-0.01110574509948492,
0.10981068015098572,
0.03183997422456741,
0.007519968319684267,
-0.0439133495092392,
-0.02583346702158451,
0.03743208199739456,
0.05361556261777878,
-0.02012190781533718,
-0.04129517823457718,
0.041692525148391724,
-0.025913270190358162,
0.051379263401031494,
-0.01804267056286335,
0.009587271139025688,
0.04688079655170441,
0.11257746070623398,
0.06580682098865509,
-0.06425832957029343,
0.049777381122112274,
0.09742709249258041,
0.03409108892083168,
-0.08562666922807693,
-0.04488040506839752,
0.10341646522283554,
0.09967396408319473,
0.07411214709281921,
-0.022253550589084625,
0.0039597428403794765,
-0.06974313408136368,
-0.039810385555028915,
0.08565811067819595,
-0.03676626458764076,
0.017047494649887085,
0.025914087891578674,
-0.08697006106376648,
0.04743925482034683,
0.010601584799587727,
0.07851086556911469,
0.008679725229740143,
0.038081713020801544,
-0.023602478206157684,
0.023523475974798203,
0.034209515899419785,
0.040262509137392044,
0.025888457894325256,
0.04224075749516487,
-0.07440068572759628,
0.03758527338504791,
-0.028265541419386864,
-0.006974162999540567,
0.042589765042066574,
0.0688062533736229,
0.04827719181776047,
0.08469761162996292,
-0.031448740512132645,
0.08793377131223679,
0.01599895767867565,
-0.007850393652915955,
1.1808542377200525e-33,
0.09853649884462357,
0.040153492242097855,
0.01641850918531418,
0.034035131335258484,
0.08130741864442825,
-0.023171255365014076,
-0.007182836998254061,
0.029578501358628273,
-0.03197953477501869,
-0.009911585599184036,
-0.008769984357059002,
0.007371562533080578,
-0.07216382026672363,
0.1124676764011383,
0.017550060525536537,
0.06893382221460342,
0.021468453109264374,
-0.007238585501909256,
-0.02016345225274563,
0.020956335589289665,
-0.0042386045679450035,
-0.04219939187169075,
-0.027088569477200508,
-0.01076250709593296,
0.03424892947077751,
0.06380877643823624,
0.022425226867198944,
-0.03842608630657196,
-0.007903608493506908,
0.030373968183994293,
-0.1468488872051239,
0.021661194041371346,
0.010002466849982738,
0.04985847696661949,
0.0028895423747599125,
-0.034106720238924026,
-0.1289321631193161,
-0.050158772617578506,
-0.018582910299301147,
0.03408181667327881,
-0.052232902497053146,
0.017093991860747337,
0.021632008254528046,
-0.035375725477933884,
-0.018592553213238716,
-0.00003281266617705114,
-0.06475786119699478,
0.020661303773522377,
-0.09965794533491135,
-0.06158845126628876,
-0.012604817748069763,
0.09293551743030548,
0.030110854655504227,
0.01696586422622204,
-0.03873322531580925,
-0.041644345968961716,
0.0030717048794031143,
0.05297666788101196,
0.0009461734443902969,
0.03641252592206001,
-0.010762807913124561,
0.011892544105648994,
-0.015023039653897285,
0.042604755610227585,
0.011962056159973145,
-0.02750762365758419,
-0.034390293061733246,
0.027387233451008797,
-0.005113157909363508,
-0.004907403606921434,
-0.08798530697822571,
-0.009720221161842346,
0.007801803760230541,
-0.04705742746591568,
-0.023843130096793175,
0.0017188682686537504,
0.09119242429733276,
-0.019174883142113686,
0.051397643983364105,
0.0011295881122350693,
0.008385816588997841,
-0.023383768275380135,
0.055097538977861404,
-0.044527649879455566,
-0.04993109405040741,
-0.07892533391714096,
0.014203243888914585,
-0.06857163459062576,
-0.02686413750052452,
0.011493486352264881,
-0.03681131824851036,
-0.04118024930357933,
-0.060654330998659134,
-0.009802475571632385,
-0.013299418613314629,
-2.480922153383709e-33,
-0.04699781909584999,
0.048297639936208725,
-0.05775950103998184,
0.07882113754749298,
-0.06654469668865204,
-0.060721505433321,
0.0012385336449369788,
0.05454647168517113,
-0.03911931812763214,
-0.013980976305902004,
-0.07493860274553299,
0.008500699885189533,
0.0008666243520565331,
-0.019025851041078568,
0.05280175432562828,
-0.050733502954244614,
0.07098808139562607,
0.02470477484166622,
-0.04624767601490021,
-0.04631076753139496,
0.10959689319133759,
0.10736242681741714,
-0.10862137377262115,
0.017372095957398415,
-0.0027485638856887817,
0.007008731830865145,
-0.03643596172332764,
0.010535725392401218,
0.026402967050671577,
0.01628497987985611,
0.0170906949788332,
-0.002208136720582843,
0.021121205762028694,
-0.021288665011525154,
-0.014156828634440899,
0.08678732812404633,
0.053774233907461166,
-0.032668761909008026,
-0.009567600674927235,
0.09472361207008362,
0.07877903431653976,
-0.03905132785439491,
-0.07710545510053635,
-0.0000783084033173509,
-0.016131965443491936,
-0.011250821873545647,
-0.027040040120482445,
-0.028809616342186928,
-0.016882143914699554,
-0.005712553858757019,
-0.018977336585521698,
-0.008659670129418373,
-0.019839812070131302,
-0.01905704289674759,
-0.043771401047706604,
-0.12466364353895187,
0.024081790819764137,
0.015928054228425026,
0.07390692085027695,
0.005561928730458021,
-0.06749816983938217,
-0.008646444417536259,
-0.0770675390958786,
0.017877094447612762,
-0.01235875766724348,
-0.020530834794044495,
-0.07270639389753342,
0.030304213985800743,
0.036419156938791275,
-0.048197656869888306,
0.13097231090068817,
-0.043484173715114594,
0.007238021120429039,
0.0611676461994648,
0.0422850176692009,
-0.0145634850487113,
-0.0023890738375484943,
-0.03675675392150879,
0.014238642528653145,
-0.016782576218247414,
0.038639090955257416,
-0.024186832830309868,
0.08416051417589188,
0.06301675736904144,
0.05035800114274025,
0.060640182346105576,
0.008248407393693924,
0.01292237639427185,
0.020582402125000954,
0.0159604549407959,
-0.016186313703656197,
-0.01339907106012106,
-0.09713722765445709,
0.05050380155444145,
0.016524111852049828,
-4.161622157994316e-8,
-0.11352358013391495,
0.03895193710923195,
0.026900731027126312,
-0.004967773333191872,
0.018809909000992775,
-0.0124834468588233,
-0.06165871396660805,
0.027549562975764275,
0.04494373872876167,
0.007807351648807526,
0.11560571938753128,
-0.032776519656181335,
0.019839094951748848,
-0.09556087851524353,
-0.05321108549833298,
0.08330804854631424,
-0.08820153027772903,
0.025841660797595978,
-0.019694847986102104,
-0.02113187499344349,
-0.0012774376664310694,
0.0034792020451277494,
-0.04069172590970993,
0.005824604537338018,
0.04558607190847397,
-0.06574716418981552,
-0.027838287875056267,
0.05120464041829109,
-0.02874080091714859,
0.08545868843793869,
-0.08157002925872803,
-0.01122555322945118,
0.014203904196619987,
0.04441504552960396,
-0.039782050997018814,
0.08772961795330048,
-0.07759691774845123,
-0.08961270749568939,
-0.012636122293770313,
-0.04472457244992256,
0.022118471562862396,
0.07790208607912064,
-0.008671550080180168,
-0.019139830023050308,
0.015269332565367222,
0.035485606640577316,
-0.033646807074546814,
-0.13471941649913788,
0.07022678107023239,
-0.08634568005800247,
-0.03931759297847748,
-0.02151726372539997,
0.021877290681004524,
0.15126149356365204,
-0.006367702968418598,
-0.08274421840906143,
0.0025641850661486387,
-0.061860233545303345,
0.07272068411111832,
0.1308051496744156,
0.02266622707247734,
0.045518770813941956,
0.020659727975726128,
-0.03883219510316849
] | https://github.com/chenwu98/agent-attack | 82 | 0 | 0 | 0 |
ACC-Collab: An Actor-Critic Approach to Multi-Agent LLM Collaboration | https://openreview.net/forum?id=nfKfAzkiez | [
"Andrew Estornell",
"Jean-Francois Ton",
"Yuanshun Yao",
"Yang Liu"
] | Poster | Large language models (LLMs) have demonstrated a remarkable ability to serve as general-purpose tools for various language-based tasks.
Recent works have demonstrated that the efficacy of such models can be improved through iterative dialog between multiple models.
While these
paradigms show promise
in
improving model efficacy, most works in this area treat collaboration as an emergent behavior, rather than a learned behavior.
In doing so, current multi-agent frameworks rely on collaborative behaviors to have been sufficiently trained into off-the-shelf models.
To address this limitation, we propose ACC-Collab, an **A**ctor-**C**riti**c** based learning framework to produce a two-agent team (an actor-agent and a critic-agent) specialized in collaboration.
We demonstrate that ACC-Collab outperforms SotA multi-agent techniques on a wide array of benchmarks. | Multi-Agent Collaboration, LLM Agents, Preference Optimization | We train a 2-agent LLM team (one actor-agent and one critic-agent) to collaboratively solve problems. | 12,819 | null | [
-0.03846873715519905,
-0.0912322849035263,
-0.056618593633174896,
0.007712014950811863,
-0.0006070027593523264,
0.04600803554058075,
-0.010664463974535465,
-0.028003064915537834,
0.04441134259104729,
-0.021361397579312325,
-0.04729284718632698,
-0.07918596267700195,
0.04394068568944931,
0.06638926267623901,
0.038015224039554596,
-0.0029699956066906452,
0.06763993203639984,
0.028474116697907448,
-0.0328647680580616,
-0.08555556833744049,
-0.05236552283167839,
0.023399662226438522,
0.0388791523873806,
-0.00041176649392582476,
-0.0010520153446123004,
-0.008996480144560337,
-0.050992898643016815,
-0.07580339163541794,
0.013862238265573978,
-0.014496732503175735,
0.041377898305654526,
-0.0029110999312251806,
-0.026439378038048744,
0.027066387236118317,
-0.056772783398628235,
0.0849912092089653,
-0.10181551426649094,
0.0486886091530323,
0.018658863380551338,
-0.05095873773097992,
-0.044534798711538315,
0.047285594046115875,
0.05545732006430626,
-0.063052698969841,
0.07891279458999634,
-0.04111441597342491,
-0.06738551706075668,
-0.060190677642822266,
-0.003720417618751526,
0.06363599747419357,
-0.0636340007185936,
-0.0785868763923645,
0.06759632378816605,
0.039096564054489136,
-0.007889257743954659,
0.01570696197450161,
0.013122007250785828,
-0.01498547475785017,
-0.017300711944699287,
-0.07946962863206863,
-0.05452851206064224,
-0.11095114797353745,
-0.05922474339604378,
0.040092501789331436,
-0.051803600043058395,
0.01760149374604225,
-0.03889966011047363,
0.08140259236097336,
-0.06731852889060974,
0.02456611581146717,
-0.06326781213283539,
-0.005003532860428095,
0.05318842828273773,
0.0054178801365196705,
0.03347613662481308,
0.01075584813952446,
-0.03223323076963425,
-0.04611145704984665,
0.11344294995069504,
-0.05145462602376938,
-0.012668514624238014,
0.06314895302057266,
0.013197625987231731,
0.019517965614795685,
0.08148222416639328,
-0.01901170052587986,
0.0700356513261795,
0.037666939198970795,
0.008852683939039707,
-0.004122896585613489,
-0.02025148645043373,
0.07011698931455612,
0.08106261491775513,
-0.005070599261671305,
-0.05657251924276352,
0.07149194180965424,
-0.013223557732999325,
-0.08413400501012802,
-0.04176799952983856,
0.09586042165756226,
-0.012733674608170986,
0.08860164880752563,
-0.02179955318570137,
-0.03829612210392952,
-0.054618753492832184,
-0.010510513558983803,
0.04006899148225784,
0.015727290883660316,
0.05731843039393425,
-0.022673655301332474,
-0.004263663198798895,
0.02679641917347908,
-0.02402629517018795,
0.015131860971450806,
0.07603757083415985,
0.03708085045218468,
0.032611213624477386,
-0.01872842013835907,
0.0364573709666729,
0.0012301441747695208,
-0.029473545029759407,
-0.03245535120368004,
0.006712343543767929,
0.03770163282752037,
0.0794045478105545,
0.0437985397875309,
-0.013810737058520317,
-2.025809495324698e-33,
0.11140120774507523,
0.05578551068902016,
0.04620709642767906,
0.10149051994085312,
0.07628361135721207,
-0.0017886669375002384,
0.02388848550617695,
0.01997564174234867,
-0.06657268851995468,
-0.04410576447844505,
0.030601518228650093,
0.03603730350732803,
-0.017758971080183983,
0.07376964390277863,
0.014203320257365704,
-0.014643917791545391,
-0.028852473944425583,
0.0011669084196910262,
-0.045955147594213486,
-0.0740235224366188,
0.032684698700904846,
0.024377524852752686,
0.03328216075897217,
-0.001067819190211594,
0.04325881972908974,
0.025805078446865082,
0.07503117620944977,
0.013477399945259094,
0.028941210359334946,
0.02378729172050953,
-0.0599227212369442,
0.03709730878472328,
0.0008795454050414264,
0.1208784356713295,
0.004910201765596867,
-0.021891966462135315,
-0.08785621076822281,
-0.024909695610404015,
0.05199538543820381,
0.04404795542359352,
-0.012199651449918747,
-0.044219378381967545,
0.03964845463633537,
-0.07932155579328537,
-0.09229356050491333,
-0.07968395948410034,
-0.08000365644693375,
0.0072910236194729805,
0.012682408094406128,
-0.05738973617553711,
0.013262013904750347,
0.020523235201835632,
-0.01958516798913479,
0.03462676703929901,
-0.0014181628357619047,
0.03904447332024574,
-0.040181875228881836,
0.015759536996483803,
-0.00010490072600077838,
-0.02068496309220791,
0.02630491368472576,
0.05490117892622948,
-0.016116300597786903,
0.11315072327852249,
0.01658884808421135,
0.05916992574930191,
-0.022564774379134178,
0.12930764257907867,
0.09909435361623764,
-0.060608815401792526,
-0.05178104341030121,
0.007363177370280027,
-0.06003730744123459,
-0.0571112222969532,
0.061564698815345764,
-0.06335826963186264,
0.0354401133954525,
-0.09207814186811447,
0.03423328697681427,
0.052102748304605484,
-0.008249068632721901,
-0.05484034866094589,
-0.0555947907269001,
-0.03099875897169113,
-0.05188516527414322,
-0.09299281239509583,
-0.017194470390677452,
-0.04881586134433746,
0.04284819960594177,
0.08205904066562653,
-0.09967606514692307,
-0.014031236059963703,
-0.014160802587866783,
0.07891645282506943,
0.0301105547696352,
-3.5449151253119925e-34,
0.02865307405591011,
-0.07125706225633621,
-0.022124307230114937,
0.052081089466810226,
-0.0014948336174711585,
-0.06865695118904114,
-0.01894104667007923,
-0.048887692391872406,
0.0035412870347499847,
0.015201782807707787,
-0.06818071752786636,
0.021353572607040405,
0.02227524295449257,
-0.05919629707932472,
0.09173764288425446,
-0.051207534968853,
0.046259406954050064,
-0.016496187075972557,
0.04440179467201233,
0.007566602434962988,
0.05616322159767151,
0.06656216830015182,
-0.018436653539538383,
0.00035786195076070726,
-0.01777932234108448,
-0.0033356240019202232,
0.024036193266510963,
0.014735250733792782,
0.022594861686229706,
0.05614464730024338,
-0.03601405769586563,
-0.05258126184344292,
0.027522532269358635,
-0.00635424442589283,
-0.09711670875549316,
0.09142479300498962,
0.017573392018675804,
-0.016581794247031212,
-0.09070757776498795,
0.05289768427610397,
0.0518118254840374,
-0.11959440261125565,
-0.0953809842467308,
0.014815757051110268,
-0.009611396118998528,
-0.007244752254337072,
-0.0637439638376236,
0.03861347213387489,
-0.03407231718301773,
-0.02645716443657875,
-0.021937115117907524,
-0.05114082619547844,
-0.03482987359166145,
-0.06252630054950714,
-0.09256434440612793,
-0.05807170644402504,
0.05037660896778107,
-0.061973266303539276,
0.004603092093020678,
-0.031851720064878464,
-0.05404512211680412,
0.017413094639778137,
0.022306498140096664,
-0.009082549251616001,
0.004604417830705643,
-0.019866207614541054,
-0.04534623771905899,
0.03416294604539871,
0.019178710877895355,
0.030081856995821,
0.05420803651213646,
0.02604972943663597,
-0.042014531791210175,
-0.01228113193064928,
0.03742881864309311,
-0.013802479952573776,
-0.08007487654685974,
-0.06997247040271759,
-0.022283487021923065,
-0.0652012825012207,
0.03136615455150604,
0.043941084295511246,
0.08611292392015457,
0.052764587104320526,
0.028317412361502647,
0.08157990127801895,
0.0122600132599473,
0.16891883313655853,
0.005668743513524532,
0.05017722025513649,
-0.021651193499565125,
0.03905104473233223,
0.0676991418004036,
0.03808834031224251,
-0.08670667558908463,
-3.758247402174675e-8,
-0.08236822485923767,
0.04753637686371803,
0.049217477440834045,
0.03694721683859825,
-0.05174243077635765,
-0.06121077015995979,
-0.02153717912733555,
-0.0003225028340239078,
0.03593166545033455,
0.04171153903007507,
0.045427411794662476,
0.05412916839122772,
-0.009576045908033848,
-0.05607405677437782,
-0.018122224137187004,
0.026562673971056938,
-0.0227948147803545,
0.01076703704893589,
-0.002837251638993621,
-0.03025975450873375,
0.016244040802121162,
-0.031112385913729668,
0.008999502286314964,
-0.01839216984808445,
-0.0012531019747257233,
-0.05404706299304962,
-0.03200481832027435,
0.06948763877153397,
0.010841616429388523,
-0.0026536232326179743,
0.0029318956658244133,
-0.04611029103398323,
-0.03798748552799225,
-0.018209196627140045,
0.03845245763659477,
0.06636490672826767,
0.07116087526082993,
-0.10744351893663406,
0.05826709419488907,
-0.08533642441034317,
-0.0018213273724541068,
0.11181821674108505,
-0.04613369330763817,
0.02467290684580803,
-0.016183925792574883,
-0.015401235781610012,
-0.062482014298439026,
-0.15498198568820953,
0.046855174005031586,
-0.08401606976985931,
-0.03865494579076767,
-0.05778688192367554,
-0.05552561208605766,
0.07764114439487457,
0.05065185949206352,
-0.03704593703150749,
-0.006375105120241642,
-0.054755695164203644,
0.09445913881063461,
0.0567285493016243,
0.028348878026008606,
0.006275150924921036,
-0.02099464274942875,
0.013512116856873035
] | 0 | 0 | 0 | 0 |
|
Greener GRASS: Enhancing GNNs with Encoding, Rewiring, and Attention | https://openreview.net/forum?id=rEQqBZIz49 | [
"Tongzhou Liao",
"Barnabas Poczos"
] | Poster | Graph Neural Networks (GNNs) have become important tools for machine learning on graph-structured data. In this paper, we explore the synergistic combination of graph encoding, graph rewiring, and graph attention, by introducing Graph Attention with Stochastic Structures (GRASS), a novel GNN architecture. GRASS utilizes relative random walk probabilities (RRWP) encoding and a novel decomposed variant (D-RRWP) to efficiently capture structural information. It rewires the input graph by superimposing a random regular graph to enhance long-range information propagation. It also employs a novel additive attention mechanism tailored for graph-structured data. Our empirical evaluations demonstrate that GRASS achieves state-of-the-art performance on multiple benchmark datasets, including a 20.3% reduction in mean absolute error on the ZINC dataset. | graph neural networks, graph encoding, graph rewiring, attention mechanism | We propose a novel GNN that achieves SOTA performance with random walk encoding, random rewiring, and a novel additive attention mechanism. | 12,813 | 2407.05649 | [
-0.0318317748606205,
-0.04465366527438164,
0.035560060292482376,
0.02597026154398918,
0.07459214329719543,
0.013230332173407078,
-0.030047228559851646,
-0.06759387254714966,
-0.03386544808745384,
-0.02588444948196411,
0.027685264125466347,
0.00308077409863472,
0.022434771060943604,
0.050191424787044525,
-0.02674201689660549,
0.01409275084733963,
0.11091228574514389,
0.06052413955330849,
-0.06158706545829773,
-0.01732400618493557,
0.008171516470611095,
0.012733053416013718,
0.09681477397680283,
-0.022720318287611008,
0.12420810014009476,
-0.04343824461102486,
-0.04130278527736664,
-0.017288459464907646,
0.04216529056429863,
0.014468682929873466,
0.09787111729383469,
-0.014712032862007618,
0.034315504133701324,
0.027653830125927925,
-0.03434223681688309,
0.0691637322306633,
-0.051738936454057693,
0.04375883936882019,
0.01989634335041046,
-0.031287603080272675,
0.016919733956456184,
0.019353466108441353,
0.014286478981375694,
0.10662061721086502,
0.07772556692361832,
0.008943247608840466,
0.02936537377536297,
-0.055915914475917816,
0.028433330357074738,
0.0028471993282437325,
-0.04976427182555199,
-0.06720606982707977,
-0.05320766568183899,
0.02958088368177414,
0.04157450422644615,
0.058924078941345215,
-0.0418497696518898,
0.030432086437940598,
-0.014501214027404785,
-0.017233435064554214,
0.03920765221118927,
-0.10484239459037781,
-0.011123520322144032,
-0.04431457817554474,
-0.02067023143172264,
-0.004647234454751015,
0.0004794629057869315,
0.028958933427929878,
-0.019118158146739006,
-0.007849878631532192,
0.05088924244046211,
0.0703282505273819,
-0.03768699988722801,
0.04339481517672539,
0.05785135179758072,
0.10831975936889648,
-0.018067533150315285,
-0.01236746832728386,
0.01130300760269165,
-0.0874263346195221,
0.038165271282196045,
-0.03351529315114021,
0.1121998280286789,
-0.010755900293588638,
0.13999472558498383,
0.021288668736815453,
0.019143812358379364,
0.02735486812889576,
-0.0355437695980072,
0.01380237378180027,
-0.11300145834684372,
-0.07289296388626099,
-0.0122402124106884,
-0.032417673617601395,
-0.053526364266872406,
0.039411820471286774,
-0.006555918604135513,
-0.08344107121229172,
-0.05276305601000786,
0.055948276072740555,
0.01040878426283598,
-0.017965078353881836,
0.06644458323717117,
-0.03968819975852966,
0.008545526303350925,
0.01226077415049076,
-0.01283528096973896,
0.1273723691701889,
0.0012808978790417314,
-0.07935841381549835,
0.007384303491562605,
0.12307602167129517,
-0.02711709775030613,
-0.009925092570483685,
0.003393040504306555,
-0.02056765928864479,
-0.08519789576530457,
-0.0057251001708209515,
-0.009756133891642094,
0.07072440534830093,
-0.06305155903100967,
0.06370726972818375,
-0.06006276234984398,
-0.00430962024256587,
-0.048358265310525894,
-0.04223630204796791,
-0.022645922377705574,
1.9398156550773164e-33,
0.0652073547244072,
0.07574749737977982,
0.021522240713238716,
-0.07083956897258759,
0.10585243999958038,
0.028899772092700005,
-0.030214490368962288,
0.022962672635912895,
-0.0016015312867239118,
0.047213099896907806,
-0.10801325738430023,
-0.056082699447870255,
-0.015168244019150734,
0.05425083264708519,
-0.026623234152793884,
-0.0197480209171772,
-0.05959220230579376,
-0.014577120542526245,
-0.038662493228912354,
-0.057631537318229675,
0.03016301803290844,
-0.0359790176153183,
0.06101524829864502,
-0.025205744430422783,
0.07076116651296616,
-0.017610177397727966,
0.00004691589856520295,
-0.00040900066960603,
0.0019493638537824154,
-0.0012005965691059828,
-0.06627078354358673,
0.06270340830087662,
-0.0038420595228672028,
0.05539311468601227,
0.0007545358967036009,
-0.09003652632236481,
-0.02913530170917511,
-0.05831679701805115,
0.047965794801712036,
-0.061039768159389496,
0.005210449919104576,
0.04485662654042244,
0.016342859715223312,
0.012982307933270931,
-0.039197586476802826,
-0.030297568067908287,
-0.0037087714299559593,
0.014920084737241268,
-0.022357909008860588,
-0.038695912808179855,
0.04482058063149452,
-0.00997784361243248,
-0.022591691464185715,
0.0752061977982521,
-0.013463981449604034,
-0.04396158829331398,
0.12770363688468933,
0.11529892683029175,
-0.01758931577205658,
0.060959674417972565,
0.009601801633834839,
0.06860383599996567,
-0.025661898776888847,
0.05270707979798317,
0.027630360797047615,
0.051936306059360504,
-0.08589700609445572,
0.06617482751607895,
0.04740715026855469,
0.007641252130270004,
0.03677307441830635,
0.06139890477061272,
-0.030179589986801147,
-0.004201924428343773,
0.002477548783645034,
0.011583542451262474,
-0.006842340342700481,
-0.11824362725019455,
0.01405048556625843,
0.04613695293664932,
-0.09160305559635162,
-0.012427428737282753,
-0.08151309937238693,
0.004019604995846748,
-0.08805564790964127,
-0.02372262068092823,
0.04082634672522545,
-0.10659579932689667,
0.03233480080962181,
-0.011898349970579147,
-0.027555644512176514,
-0.0596223920583725,
-0.02543853595852852,
-0.0505952462553978,
-0.01638440042734146,
-8.258363930014978e-34,
-0.01909184269607067,
0.15897485613822937,
0.03211377561092377,
0.04209333658218384,
-0.006579290609806776,
-0.033258240669965744,
-0.008127307519316673,
-0.04759905859827995,
-0.038942303508520126,
0.03976673260331154,
-0.02460647188127041,
-0.06268779188394547,
0.05116397887468338,
-0.024288510903716087,
0.05113435164093971,
0.010629119351506233,
-0.050230368971824646,
0.0668264776468277,
-0.044034697115421295,
-0.017096081748604774,
-0.007483765948563814,
0.06221487745642662,
-0.04194959998130798,
0.008223192766308784,
0.06598169356584549,
0.022134704515337944,
-0.013128111138939857,
0.03250398486852646,
0.04194735735654831,
-0.06308380514383316,
-0.08332019299268723,
0.0012236885959282517,
-0.03394821286201477,
-0.05084406957030296,
0.015241679735481739,
0.077593132853508,
0.05279233306646347,
-0.017823193222284317,
-0.022484196349978447,
0.0067674326710402966,
0.06031643599271774,
-0.030714640393853188,
-0.10466958582401276,
0.06465016305446625,
-0.02704279124736786,
0.08706408739089966,
-0.11946791410446167,
0.02583668753504753,
-0.0966142788529396,
0.029341408982872963,
-0.01953565888106823,
0.05117562413215637,
-0.06373864412307739,
-0.012200315482914448,
-0.04929334670305252,
-0.09696686267852783,
-0.007319195196032524,
0.0554216131567955,
-0.011586364358663559,
-0.050894543528556824,
-0.04265802353620529,
-0.03813762217760086,
-0.004061246290802956,
0.008094815537333488,
0.0005462314002215862,
-0.06895311176776886,
0.031980499625205994,
-0.04816942662000656,
0.03975217789411545,
0.01052098348736763,
0.01463265623897314,
-0.012986065819859505,
-0.024772053584456444,
0.006282477173954248,
-0.004446490667760372,
0.006278275512158871,
0.02790459804236889,
-0.01711130142211914,
-0.042918458580970764,
-0.040619004517793655,
0.06167537346482277,
-0.026771632954478264,
0.03222784399986267,
0.05266789719462395,
0.12753847241401672,
0.06945688277482986,
0.008301966823637486,
0.0011498722014948726,
0.09780760854482651,
0.036583833396434784,
-0.020157741382718086,
0.07815694063901901,
-0.03937820717692375,
0.07285289466381073,
0.00984059926122427,
-3.691338079647721e-8,
-0.140928715467453,
0.031049415469169617,
0.010835252702236176,
0.004589450545608997,
0.07316303998231888,
0.017995843663811684,
0.05435718223452568,
0.1476823389530182,
-0.044540684670209885,
0.022573769092559814,
0.0419229157269001,
0.016054755076766014,
0.004188083112239838,
-0.06314940750598907,
-0.02903742529451847,
-0.007266824599355459,
-0.005502942018210888,
-0.012980042956769466,
0.0362662710249424,
0.031179826706647873,
-0.017278233543038368,
-0.02715393900871277,
-0.016539137810468674,
0.06479206681251526,
-0.012846752069890499,
-0.12241502106189728,
-0.017971178516745567,
0.025918064638972282,
0.015463107265532017,
-0.016641788184642792,
-0.004944347310811281,
0.012364711612462997,
0.12601257860660553,
0.011452601291239262,
-0.02803518809378147,
0.09542251378297806,
0.07154764980077744,
0.0615181140601635,
-0.01719498448073864,
0.002577992621809244,
0.0043939449824392796,
0.025173094123601913,
-0.01892063207924366,
0.01574610359966755,
-0.07941219955682755,
-0.05983554571866989,
-0.016957683488726616,
-0.0017058543162420392,
0.07028960436582565,
-0.07645071297883987,
0.032046765089035034,
-0.043758291751146317,
-0.03552921861410141,
-0.0803430899977684,
0.02202213555574417,
0.013288643211126328,
-0.0017799147171899676,
-0.01857512630522251,
0.01532409805804491,
0.03326352313160896,
-0.020041627809405327,
-0.03622835874557495,
-0.07067453116178513,
-0.09262776374816895
] | 0 | 0 | 0 | 0 |
|
NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks in Open Domains | https://openreview.net/forum?id=VoayJihXra | [
"Wonje Choi",
"Jinwoo Park",
"Sanghyun Ahn",
"Daehee Lee",
"Honguk Woo"
] | Poster | We explore neuro-symbolic approaches to generalize actionable knowledge, enabling embodied agents to tackle complex tasks more effectively in open-domain environments. A key challenge for embodied agents is the generalization of knowledge across diverse environments and situations, as limited experiences often confine them to their prior knowledge. To address this issue, we introduce a novel framework, NeSyC, a neuro-symbolic continual learner that emulates the hypothetico-deductive model by continually formulating and validating knowledge from limited experiences through the combined use of Large Language Models (LLMs) and symbolic tools. Specifically, we devise a contrastive generality improvement scheme within NeSyC, which iteratively generates hypotheses using LLMs and conducts contrastive validation via symbolic tools. This scheme reinforces the justification for admissible actions while minimizing the inference of inadmissible ones. Additionally, we incorporate a memory-based monitoring scheme that efficiently detects action errors and triggers the knowledge refinement process across domains. Experiments conducted on diverse embodied task benchmarks—including ALFWorld, VirtualHome, Minecraft, RLBench, and a real-world robotic scenario—demonstrate that NeSyC is highly effective in solving complex embodied tasks across a range of open-domain environments. | Embodied AI, Neuro-symbolic AI | NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks in Open Domains | 12,812 | 2503.00870 | [
0.06935170292854309,
-0.06932367384433746,
0.04374571889638901,
0.04078882932662964,
0.04326761141419411,
0.004439380019903183,
0.024382121860980988,
-0.07450573146343231,
0.01933400146663189,
0.031120093539357185,
-0.05227676033973694,
-0.0590299628674984,
0.014648760668933392,
0.09020271897315979,
0.07595957815647125,
-0.015013476833701134,
0.10139302909374237,
0.06589214503765106,
-0.08900177478790283,
-0.05585474148392677,
0.033709120005369186,
0.0032125262077897787,
0.02625688910484314,
-0.018036123365163803,
-0.04023440182209015,
-0.004578638356178999,
0.0023003132082521915,
-0.050815608352422714,
0.050535961985588074,
-0.03894931450486183,
0.09101857244968414,
0.025115597993135452,
-0.016500424593687057,
-0.03359342738986015,
0.03061819262802601,
0.1593339443206787,
-0.06411361694335938,
0.004403038416057825,
0.018916571512818336,
-0.08552908897399902,
-0.063430555164814,
-0.022648727521300316,
-0.022505776956677437,
0.039186030626297,
0.026179451495409012,
0.023001566529273987,
-0.02737945131957531,
-0.09008296579122543,
0.016021503135561943,
0.006294087506830692,
-0.07240653038024902,
-0.06740371882915497,
0.03588218614459038,
0.0042228419333696365,
0.04244673252105713,
0.07835350930690765,
0.011864691972732544,
0.03128310665488243,
-0.04091056436300278,
-0.07813315093517303,
0.007650637533515692,
-0.1017204001545906,
-0.023871107026934624,
-0.0242381040006876,
-0.03643791005015373,
0.010699423030018806,
-0.00043622616794891655,
0.022572020068764687,
-0.006724832113832235,
0.03900819271802902,
-0.04234744608402252,
-0.006279032677412033,
0.00854763388633728,
0.03511207923293114,
0.01909707859158516,
0.05106605589389801,
-0.03864990919828415,
-0.08760843425989151,
0.06025950238108635,
-0.09252025187015533,
-0.04419002681970596,
0.04757985845208168,
0.0173964723944664,
0.010755864903330803,
0.12997135519981384,
0.08055456727743149,
0.054742481559515,
0.0635487511754036,
0.06955039501190186,
0.05421726405620575,
-0.0021675231400877237,
-0.08134773373603821,
0.07785426080226898,
0.02432866208255291,
0.054466620087623596,
0.0017274424899369478,
-0.014836957678198814,
-0.024273987859487534,
-0.025317825376987457,
0.10717160254716873,
0.010347115807235241,
0.06066258251667023,
-0.0014607824850827456,
-0.003619724651798606,
0.011441361159086227,
-0.028278551995754242,
0.0008827451965771616,
0.01420984324067831,
0.031342148780822754,
-0.04607328027486801,
-0.021424289792776108,
0.05616649612784386,
0.06789178401231766,
0.048734020441770554,
0.027996672317385674,
-0.030186820775270462,
0.044871583580970764,
0.018977222964167595,
0.020984802395105362,
0.02262081764638424,
0.11350307613611221,
0.030378611758351326,
0.056029707193374634,
0.027501149103045464,
0.05152430757880211,
-0.018094176426529884,
-0.03849995508790016,
2.1276397240360517e-34,
0.04134231433272362,
0.02862696349620819,
0.0648537129163742,
0.05211760103702545,
0.04939582198858261,
-0.007251986302435398,
-0.04368937760591507,
0.06055981665849686,
0.028573717921972275,
0.01580963283777237,
0.007005094550549984,
0.060783449560403824,
-0.046527497470378876,
0.09814517945051193,
0.044466737657785416,
-0.03526255488395691,
-0.07308460772037506,
-0.012079296633601189,
-0.03422851487994194,
-0.030410433188080788,
0.041173383593559265,
-0.04662588983774185,
-0.03151446208357811,
-0.022616006433963776,
-0.051898907870054245,
0.013052573427557945,
0.05065849795937538,
0.025442048907279968,
0.01653680019080639,
0.004223083611577749,
-0.09410226345062256,
-0.05088565871119499,
-0.02402420900762081,
0.07087177038192749,
0.005757154896855354,
-0.07994187623262405,
0.016890525817871094,
-0.0748736709356308,
0.03746449947357178,
-0.05606948956847191,
-0.0475006140768528,
0.026040369644761086,
-0.06950034946203232,
-0.009166883304715157,
-0.06474355608224869,
-0.0817168578505516,
0.035706788301467896,
0.0067462860606610775,
-0.03345707803964615,
-0.027073323726654053,
0.02419697679579258,
0.028401924297213554,
0.02721843309700489,
-0.04719275236129761,
-0.06820041686296463,
-0.03321526199579239,
-0.06187934800982475,
0.005272587761282921,
0.005188120063394308,
0.05877293646335602,
0.03340796381235123,
0.04136590659618378,
0.016182493418455124,
0.10474856197834015,
0.05113425850868225,
0.06313978135585785,
-0.05496389418840408,
0.09758298844099045,
0.04429727792739868,
-0.05107052996754646,
-0.021703355014324188,
0.03682190924882889,
0.03378557786345482,
-0.053006675094366074,
0.003779345890507102,
-0.09728234261274338,
0.06675975769758224,
-0.12612789869308472,
-0.05902550742030144,
-0.025388499721884727,
0.03818502277135849,
-0.03317439556121826,
-0.07245995849370956,
-0.004497254732996225,
-0.00481464434415102,
-0.04986187070608139,
0.03083394095301628,
-0.1297617256641388,
-0.07332494854927063,
0.00527593120932579,
-0.08944050222635269,
-0.08012673258781433,
-0.04079240560531616,
-0.002741388976573944,
-0.009094680659472942,
-1.9086527488240407e-33,
-0.0040928334929049015,
-0.03818637132644653,
-0.057099055498838425,
0.03324268013238907,
-0.026728546246886253,
-0.0910382866859436,
-0.020042723044753075,
-0.012817313894629478,
-0.09702911972999573,
-0.14201365411281586,
-0.03194236382842064,
-0.02619941718876362,
0.026231860741972923,
0.05318707600235939,
-0.02124655619263649,
-0.00020163602312095463,
-0.06554917991161346,
0.01242799311876297,
0.012706665322184563,
0.06572452932596207,
-0.006203019060194492,
0.09933187812566757,
-0.09562055766582489,
0.021953295916318893,
-0.023217227309942245,
0.06072130426764488,
-0.006623168010264635,
0.0715096965432167,
-0.030137257650494576,
0.04705316573381424,
-0.00910890381783247,
-0.0031716995872557163,
-0.0076567488722503185,
-0.050099823623895645,
-0.0069434731267392635,
0.02629707008600235,
0.05675690993666649,
-0.0057789236307144165,
-0.08458569645881653,
0.04422284662723541,
0.09710235893726349,
-0.07091642916202545,
-0.034007929265499115,
-0.008372926153242588,
0.007513734977692366,
0.020865513011813164,
-0.15652436017990112,
0.017140250653028488,
0.022957459092140198,
0.011129910126328468,
0.02838185243308544,
-0.02960735745728016,
-0.06984511017799377,
-0.07764286547899246,
-0.07398843765258789,
-0.06813126057386398,
0.01899002678692341,
-0.051403723657131195,
0.014072012156248093,
0.005075149703770876,
-0.0703737661242485,
-0.011632166802883148,
0.00278742634691298,
-0.009435105137526989,
0.014347131364047527,
0.002555390354245901,
-0.025213902816176414,
0.047746188938617706,
0.05044444650411606,
-0.0401364304125309,
0.005506238900125027,
0.056166067719459534,
-0.06691382080316544,
-0.015482867136597633,
0.06000630557537079,
0.016988350078463554,
-0.00929369404911995,
-0.039246901869773865,
-0.021718773990869522,
-0.017673518508672714,
0.02455214411020279,
-0.03399842604994774,
0.04831600934267044,
0.04316059872508049,
0.008992496877908707,
0.09060877561569214,
-0.06461246311664581,
0.12592265009880066,
0.016292259097099304,
0.031204305589199066,
-0.039939265698194504,
0.05923781171441078,
-0.005250032991170883,
0.05067802220582962,
-0.08061438798904419,
-4.272899545298969e-8,
-0.08614249527454376,
-0.027078215032815933,
0.07733377069234848,
-0.01008527260273695,
0.025435492396354675,
-0.0010757442796602845,
-0.05371376499533653,
-0.0416608527302742,
-0.027972500771284103,
-0.030441388487815857,
0.09425151348114014,
0.02054663933813572,
0.012071304023265839,
-0.030783744528889656,
-0.014154749922454357,
0.09591753035783768,
0.04794324189424515,
0.036446426063776016,
-0.03701066970825195,
-0.019621258601546288,
0.07891296595335007,
0.01664568856358528,
-0.005209098570048809,
0.008557443507015705,
0.008189268410205841,
-0.06921476125717163,
-0.04746881127357483,
0.11728370934724808,
0.040729500353336334,
-0.008713348768651485,
0.10875388979911804,
0.012369266711175442,
0.05552609637379646,
-0.02730155922472477,
0.007780689746141434,
0.04411895573139191,
0.004101417493075132,
-0.0820775181055069,
-0.0036730682477355003,
-0.028273867443203926,
-0.007997224107384682,
0.05705936253070831,
-0.039911579340696335,
0.016573842614889145,
-0.0015623485669493675,
-0.008958113379776478,
-0.08815011382102966,
-0.09732146561145782,
0.023249970749020576,
-0.015359412878751755,
-0.021149365231394768,
0.03395075350999832,
-0.047039661556482315,
0.11407365649938583,
0.057780925184488297,
0.001580308540724218,
-0.022692564874887466,
0.03342592343688011,
-0.00009944479825207964,
0.031708985567092896,
0.04437993839383125,
0.057630591094493866,
-0.0739814043045044,
-0.026206254959106445
] | 0 | 0 | 0 | 0 |
|
Conformal Language Model Reasoning with Coherent Factuality | https://openreview.net/forum?id=AJpUZd8Clb | [
"Maxon Rubin-Toles",
"Maya Gambhir",
"Keshav Ramji",
"Aaron Roth",
"Surbhi Goel"
] | Poster | Language models are increasingly being used in important decision pipelines, so ensuring the correctness of their outputs is crucial. Recent work has proposed evaluating the “factuality” of claims decomposed from a language model generation and applying conformal prediction techniques to filter out those claims that are not factual. This can be effective for tasks such as information retrieval, where constituent claims may be evaluated in isolation for factuality, but is not appropriate for reasoning tasks, as steps of a logical argument can be evaluated for correctness only within the context of the claims that precede them. To capture this, we define “coherent factuality” and develop a conformal-prediction-based method to guarantee coherent factuality for language model outputs. Our approach applies split conformal prediction to subgraphs within a "deducibility" graph that represents the steps of a reasoning problem. We evaluate our method on mathematical reasoning problems from the MATH and FELM datasets and find that our algorithm consistently produces correct and substantiated orderings of claims, achieving coherent factuality across target coverage levels. Moreover, we achieve 90\% factuality on our stricter definition while retaining 80\% or more of the original claims, highlighting the utility of our deducibility-graph-guided approach. | language models, reasoning, conformal prediction, factuality, graph representation, coherence | We apply conformal prediction on dependency graphs towards ensuring coherence and factuality in language model reasoning. | 12,804 | null | [
-0.10431739687919617,
-0.021418947726488113,
-0.009908441454172134,
0.027562418952584267,
0.05677505210042,
-0.0444067157804966,
0.006096120458096266,
0.024212758988142014,
-0.007073208224028349,
0.01124218013137579,
-0.05048973485827446,
-0.09980358928442001,
0.07965114712715149,
0.11819693446159363,
0.04880739003419876,
0.009103314951062202,
0.041220806539058685,
0.07075768709182739,
-0.09096535295248032,
-0.07874757796525955,
0.0741930902004242,
0.05228324234485626,
-0.06501895934343338,
0.02646300569176674,
0.0001286874758079648,
-0.01631956361234188,
0.014680287800729275,
-0.048127830028533936,
0.017775539308786392,
-0.0003022827731911093,
-0.037307098507881165,
0.07976667582988739,
-0.05048772320151329,
-0.012585490010678768,
0.002838995773345232,
0.0443948470056057,
-0.047762900590896606,
-0.017402151599526405,
0.04424724727869034,
0.031489282846450806,
-0.020959410816431046,
-0.04034896939992905,
0.0354231595993042,
0.0018736834172159433,
0.04328169301152229,
0.023468591272830963,
-0.005730851553380489,
0.09789861738681793,
-0.02546101063489914,
0.03961281478404999,
-0.10209108889102936,
-0.01569029688835144,
-0.017831813544034958,
-0.036160167306661606,
-0.01703309640288353,
-0.04278993979096413,
0.03545103594660759,
0.0038486008998006582,
-0.03176898881793022,
-0.08531014621257782,
-0.00901787169277668,
-0.07663828879594803,
-0.06391069293022156,
-0.011761205270886421,
-0.03348187357187271,
-0.010149585083127022,
0.008670032024383545,
0.02307935804128647,
-0.05628349259495735,
0.10203589498996735,
-0.039312973618507385,
0.08872756361961365,
-0.06884756684303284,
0.08037309348583221,
0.014134847559034824,
0.10115324705839157,
0.10327164083719254,
-0.02113322913646698,
0.030345765873789787,
-0.09924589097499847,
-0.05928020924329758,
-0.0006371483905240893,
0.0657910704612732,
0.0018809718312695622,
0.10325995087623596,
-0.0629630982875824,
-0.01779608614742756,
0.010499298572540283,
0.03966052457690239,
0.05893959477543831,
-0.042544979602098465,
-0.0636964663863182,
0.09881340712308884,
0.02408750168979168,
0.09030087292194366,
0.09933736175298691,
-0.032885439693927765,
-0.04531588777899742,
0.05303517356514931,
0.06948035955429077,
-0.038862649351358414,
0.16839320957660675,
0.02224772796034813,
-0.09664866328239441,
0.0443381704390049,
-0.060260672122240067,
0.017028966918587685,
0.002870561322197318,
-0.016046738252043724,
-0.12035602331161499,
0.034342557191848755,
0.044886939227581024,
0.0586632639169693,
-0.007808954920619726,
0.07139658182859421,
-0.026255203410983086,
0.03932547941803932,
0.021985448896884918,
-0.03805404156446457,
-0.0022901601623743773,
-0.06221766024827957,
0.11491721123456955,
0.06822333484888077,
0.09019992500543594,
0.0728653147816658,
-0.06709959357976913,
-0.04856949299573898,
2.3477320312798203e-33,
0.014204795472323895,
0.06330908834934235,
0.04658946022391319,
0.028997760266065598,
0.04479939863085747,
-0.014816505834460258,
-0.048274893313646317,
-0.003338653827086091,
-0.006755081936717033,
0.048468951135873795,
-0.026585223153233528,
0.03961695730686188,
-0.023304283618927002,
-0.03266296163201332,
0.05058892071247101,
0.056510601192712784,
-0.014893004670739174,
-0.006877255626022816,
-0.1262078881263733,
-0.013172484934329987,
0.09104866534471512,
-0.013377496972680092,
-0.006361663807183504,
-0.0073763360269367695,
0.029299864545464516,
-0.02662373147904873,
0.02598586492240429,
0.0004641456762328744,
-0.10952044278383255,
0.002083702478557825,
-0.09792623668909073,
-0.06152556091547012,
0.05056828632950783,
0.05514848977327347,
0.013826211914420128,
-0.08004114031791687,
-0.0340229868888855,
-0.06386212259531021,
-0.004953782074153423,
0.00012568854435812682,
-0.04525968059897423,
0.03358914330601692,
-0.026406578719615936,
0.05682150274515152,
-0.007262853905558586,
-0.04122534766793251,
-0.07151749730110168,
0.04404938966035843,
0.01511018630117178,
-0.056178104132413864,
0.10937054455280304,
0.037310000509023666,
-0.04343162849545479,
-0.04909495264291763,
-0.031069118529558182,
-0.005842043086886406,
-0.03091319277882576,
0.03306229040026665,
0.0607941560447216,
0.056847475469112396,
0.004107185639441013,
0.044818926602602005,
-0.01145926397293806,
0.022174878045916557,
-0.06888768076896667,
0.056929368525743484,
-0.10370125621557236,
-0.014849673956632614,
0.09874898940324783,
0.00017958185344468802,
-0.018261177465319633,
-0.01562713086605072,
0.022664552554488182,
-0.01848204806447029,
0.05381482094526291,
-0.031387172639369965,
0.018725082278251648,
-0.054940056055784225,
0.01175650767982006,
0.022467374801635742,
0.0068605332635343075,
0.024179497733712196,
-0.017603952437639236,
-0.002010411350056529,
-0.039466701447963715,
-0.06794131547212601,
0.03185150399804115,
-0.011460413224995136,
-0.07006379216909409,
-0.11956026405096054,
-0.07436569780111313,
-0.020259635522961617,
0.0023198756389319897,
-0.0657869204878807,
0.01242293231189251,
-2.931537076182886e-33,
-0.058770183473825455,
-0.002489242237061262,
0.01903277076780796,
0.0236891508102417,
-0.08742033690214157,
-0.09957726299762726,
-0.035964541137218475,
-0.02209358476102352,
0.016341721639037132,
-0.07622967660427094,
-0.06673016399145126,
-0.04014597833156586,
0.06263953447341919,
-0.007600730285048485,
-0.05839824676513672,
-0.030415916815400124,
0.017990346997976303,
-0.01935434713959694,
-0.07975254207849503,
0.09854326397180557,
0.0360206663608551,
0.05307462811470032,
-0.12198393791913986,
0.0007201913977041841,
0.03409723564982414,
0.037579409778118134,
-0.03268514573574066,
-0.003595330286771059,
0.02179725095629692,
0.06376741826534271,
-0.05259350314736366,
-0.03813569247722626,
-0.01729799248278141,
-0.094403475522995,
-0.04814409464597702,
-0.049003541469573975,
0.026275089010596275,
-0.012893878854811192,
-0.0053941598162055016,
0.034464698284864426,
0.03417595848441124,
-0.06508664786815643,
-0.07537364214658737,
0.010291995480656624,
-0.01273210160434246,
-0.06769023835659027,
0.007331441156566143,
-0.014179110527038574,
-0.03164057806134224,
0.0033135912381112576,
-0.0027019528206437826,
-0.0024353843182325363,
0.014950798824429512,
0.07576444745063782,
-0.03278878331184387,
-0.026833370327949524,
-0.07074475288391113,
-0.0488252267241478,
0.02447592280805111,
0.01295318640768528,
-0.04833173751831055,
0.08993679285049438,
0.04532112181186676,
0.021560583263635635,
0.06344330310821533,
-0.022470517084002495,
-0.05800384283065796,
0.028643500059843063,
0.07200286537408829,
0.01024673692882061,
0.025969529524445534,
-0.03908926621079445,
-0.020031658932566643,
0.006028039380908012,
0.08196359872817993,
-0.010443545877933502,
0.04191000014543533,
-0.0003404458984732628,
-0.03746664524078369,
0.05104527622461319,
0.008760140277445316,
-0.00880858488380909,
0.07540997117757797,
0.052490685135126114,
0.003981925081461668,
-0.09531854093074799,
0.05777868628501892,
0.056807953864336014,
0.06717434525489807,
0.09381117671728134,
-0.0013170554302632809,
-0.004507494159042835,
-0.012386820279061794,
0.09423583000898361,
-0.03575637936592102,
-4.33473985594901e-8,
-0.06117044761776924,
0.04362011328339577,
-0.01263987272977829,
0.08756085485219955,
0.027437618002295494,
-0.04243714362382889,
-0.04417366161942482,
-0.034737035632133484,
-0.03692779690027237,
-0.015368721447885036,
0.04488971084356308,
0.04895760491490364,
-0.10704939812421799,
-0.06264755874872208,
-0.027220385149121284,
0.018205802887678146,
0.055405158549547195,
-0.02046210505068302,
-0.01945204846560955,
-0.025103265419602394,
0.024075208231806755,
0.01217945758253336,
-0.035191651433706284,
0.03318830952048302,
0.04869043081998825,
-0.04557954892516136,
-0.0680353119969368,
0.04116929695010185,
0.034625094383955,
0.04245525225996971,
-0.006261148024350405,
0.015954963862895966,
0.034605029970407486,
0.006633551325649023,
0.029950695112347603,
0.04307043179869652,
0.02153879404067993,
0.004607898183166981,
0.00400341022759676,
-0.023015420883893967,
0.03381036967039108,
0.10883450508117676,
-0.04175420477986336,
0.08735217154026031,
0.07995932549238205,
-0.06346395611763,
-0.024493519216775894,
-0.03634215146303177,
0.05383506789803505,
-0.023440077900886536,
-0.028946716338396072,
0.0024119694717228413,
-0.02276361733675003,
0.054458215832710266,
-0.0047876122407615185,
0.0685792863368988,
-0.06838450580835342,
-0.03563614562153816,
-0.046741992235183716,
-0.013982401229441166,
0.02475648559629917,
0.04531525820493698,
0.043407849967479706,
-0.019545890390872955
] | 0 | 0 | 0 | 0 |
|
Constraint-Conditioned Actor-Critic for Offline Safe Reinforcement Learning | https://openreview.net/forum?id=nrRkAAAufl | [
"Zijian Guo",
"Weichao Zhou",
"Shengao Wang",
"Wenchao Li"
] | Poster | Offline safe reinforcement learning (OSRL) aims to learn policies with high rewards while satisfying safety constraints solely from data collected offline. However, the learned policies often struggle to handle states and actions that are not present or out-of-distribution (OOD) from the offline dataset, which can result in violation of the safety constraints or overly conservative behaviors during their online deployment. Moreover, many existing methods are unable to learn policies that can adapt to varying constraint thresholds. To address these challenges, we propose constraint-conditioned actor-critic (CCAC), a novel OSRL method that models the relationship between state-action distributions and safety constraints, and leverages this relationship to regularize critics and policy learning. CCAC learns policies that can effectively handle OOD data and adapt to varying constraint thresholds. Empirical evaluations on the $\texttt{DSRL}$ benchmarks show that CCAC significantly outperforms existing methods for learning adaptive, safe, and high-reward policies. | Offline Safe Reinforcement Learning, Constraint-conditioned Actor-Critic, Data Generation, Out-of-distribution Detection, Zero-shot Adaptation | null | 12,802 | null | [
-0.09540113061666489,
-0.029919521883130074,
-0.030894601717591286,
0.027778586372733116,
0.06033474951982498,
0.092194102704525,
0.029793227091431618,
-0.022244874387979507,
-0.030704796314239502,
0.05625041201710701,
-0.0011322229402139783,
-0.020399777218699455,
0.05400713533163071,
0.029682254418730736,
0.004985382314771414,
0.017142539843916893,
0.11679476499557495,
-0.0073991031385958195,
-0.1148797869682312,
0.0014521221164613962,
0.04447042569518089,
0.034546129405498505,
0.0593860000371933,
0.012935059145092964,
0.019926488399505615,
-0.01857166364789009,
0.01881609484553337,
-0.0006501770694740117,
-0.01754351146519184,
0.0005771485157310963,
0.030820216983556747,
-0.100287526845932,
-0.038077887147665024,
-0.0673276036977768,
-0.006336072459816933,
0.02218141220510006,
-0.12105786800384521,
-0.02190791815519333,
-0.09231250733137131,
-0.024352991953492165,
-0.05635448545217514,
0.08553889393806458,
-0.003929964732378721,
0.021257998421788216,
0.09310490638017654,
-0.03946118429303169,
0.0010572472820058465,
-0.07661346346139908,
-0.00834613386541605,
0.025155946612358093,
-0.02034681662917137,
0.003407333977520466,
0.040600065141916275,
0.07549922913312912,
0.02965129166841507,
0.006236907560378313,
0.04604843631386757,
0.01961117796599865,
-0.0015634179580956697,
0.01432245597243309,
-0.06763854622840881,
-0.1432729959487915,
-0.0741957500576973,
-0.05544951930642128,
-0.06803788244724274,
0.01174299418926239,
-0.09936325997114182,
0.03466257080435753,
0.01481339056044817,
0.039699628949165344,
-0.11914297193288803,
0.05667560547590256,
0.02937869168817997,
0.006492010783404112,
0.03537219017744064,
0.0031582044903188944,
0.025925016030669212,
-0.08067797869443893,
0.05733892694115639,
-0.022169917821884155,
0.02256682515144348,
-0.026706840842962265,
-0.024767497554421425,
0.07581401616334915,
0.09014779329299927,
-0.035134341567754745,
0.045168496668338776,
0.036560434848070145,
0.1546430140733719,
0.03790431469678879,
0.025843558833003044,
0.05383406952023506,
0.08309537917375565,
0.011746731586754322,
-0.03719836100935936,
-0.012856705114245415,
-0.021081924438476562,
-0.07689166069030762,
-0.087155781686306,
0.10773076862096786,
0.04707005247473717,
0.05677534267306328,
-0.007433874998241663,
-0.001194974989630282,
0.010674906894564629,
-0.022582923993468285,
0.016870662569999695,
0.01678389497101307,
0.03992104157805443,
-0.042241066694259644,
0.054107509553432465,
0.05004988610744476,
0.07294599711894989,
-0.026300987228751183,
-0.04383138194680214,
0.03455188125371933,
-0.03864305093884468,
-0.02385806292295456,
-0.019648004323244095,
0.05060495808720589,
0.0034655772615224123,
-0.08921541273593903,
0.04631396010518074,
0.022457698360085487,
0.014171659015119076,
-0.04497372359037399,
0.002555946819484234,
5.344574247135616e-34,
0.08273220807313919,
-0.0375424288213253,
-0.0021497413981705904,
-0.022256290540099144,
0.02148154005408287,
0.014447246678173542,
0.028752688318490982,
-0.015391327440738678,
-0.09840410947799683,
-0.019206982105970383,
0.01631099358201027,
-0.04578307271003723,
-0.06623173505067825,
0.0555676631629467,
0.08934147655963898,
-0.015829307958483696,
-0.057480309158563614,
0.06597152352333069,
0.042614709585905075,
-0.05720942094922066,
0.05138964205980301,
-0.09879671782255173,
0.0072588250041007996,
-0.027815639972686768,
0.06505171209573746,
-0.013504169881343842,
0.045207679271698,
0.00010669475159374997,
-0.0015743321273475885,
0.02621506154537201,
-0.03132597357034683,
0.040582865476608276,
0.03996901586651802,
0.016530144959688187,
0.06037483736872673,
0.014916014857590199,
-0.173609659075737,
-0.04286380484700203,
0.0065983375534415245,
-0.02028682641685009,
-0.032345034182071686,
-0.041546985507011414,
0.04927970468997955,
0.017564453184604645,
-0.024298250675201416,
-0.1276945173740387,
0.017073748633265495,
-0.041398271918296814,
-0.04830922558903694,
-0.0375494584441185,
-0.04972284287214279,
0.034947481006383896,
0.06222480162978172,
-0.05878914147615433,
-0.08786534518003464,
0.020370054990053177,
0.0021437983959913254,
0.038620345294475555,
-0.0626549944281578,
0.0400649756193161,
0.01214359886944294,
0.02723190002143383,
0.034663744270801544,
0.027220875024795532,
-0.021914366632699966,
0.07057423144578934,
-0.0547078438103199,
0.053047969937324524,
0.028827302157878876,
-0.033442381769418716,
-0.024392500519752502,
0.04830227792263031,
0.0037764294538646936,
-0.06645602732896805,
0.02829684130847454,
0.007396737579256296,
0.05844379588961601,
-0.020386945456266403,
0.002107457024976611,
-0.043638113886117935,
0.03497130423784256,
-0.009726065210998058,
-0.0030463384464383125,
0.04870060458779335,
-0.05197324976325035,
-0.07236726582050323,
-0.015710078179836273,
0.00831715576350689,
-0.03242972865700722,
0.04983600974082947,
-0.07296840101480484,
-0.07057223469018936,
-0.025190364569425583,
0.042596928775310516,
0.0470585823059082,
-2.0696456082122802e-34,
0.012234744615852833,
-0.06142635643482208,
-0.02172739803791046,
-0.039503347128629684,
-0.0655304417014122,
-0.04301050305366516,
-0.07858657836914062,
-0.002234224695712328,
0.038984231650829315,
-0.014995181933045387,
-0.08602990210056305,
0.021858185529708862,
-0.012230304069817066,
-0.03763144835829735,
0.061193108558654785,
-0.05069177597761154,
-0.10166872292757034,
0.07553159445524216,
0.019502559676766396,
0.003972835373133421,
0.0679013580083847,
0.002101854421198368,
-0.10163123905658722,
0.10553966462612152,
-0.009116890840232372,
-0.007556901779025793,
-0.002138565294444561,
0.0971425473690033,
0.0052685821428895,
0.09998364001512527,
-0.023859567940235138,
-0.038496218621730804,
-0.03163094073534012,
0.04548438638448715,
-0.0719558447599411,
0.020068718120455742,
0.01939205266535282,
-0.0029773376882076263,
-0.05113590508699417,
0.09507229924201965,
0.055385854095220566,
-0.024694371968507767,
-0.02679431065917015,
-0.04619339480996132,
-0.01912655495107174,
0.017361264675855637,
-0.014203930273652077,
0.019908029586076736,
-0.007690498139709234,
-0.0009736056090332568,
-0.02345806173980236,
-0.010425377637147903,
-0.03389707952737808,
0.02300719916820526,
-0.04587284475564957,
-0.06571736931800842,
-0.018646975979208946,
0.005827081389725208,
0.057854682207107544,
-0.0214854646474123,
-0.022884447127580643,
0.014773956499993801,
-0.029408933594822884,
0.07732963562011719,
0.013100141659379005,
0.022653289139270782,
0.013273265212774277,
0.037258248776197433,
0.004259706474840641,
-0.0015433861408382654,
0.04841361939907074,
0.05268293246626854,
0.026858031749725342,
-0.024941731244325638,
0.038334351032972336,
-0.0018856726819649339,
-0.007492489647120237,
-0.004381853621453047,
-0.05425450578331947,
0.004403631668537855,
0.07031240314245224,
-0.0013962440425530076,
0.09035222232341766,
0.10325606912374496,
0.05805901065468788,
0.10068988800048828,
-0.007990735583007336,
0.0664261057972908,
0.03576738387346268,
0.009841127321124077,
-0.07205641269683838,
-0.009350660257041454,
-0.0078058126382529736,
-0.004174195230007172,
-0.05309493839740753,
-3.745493515339149e-8,
-0.08728664368391037,
0.018265021964907646,
0.03641018271446228,
0.011299093253910542,
0.053351983428001404,
0.05277983099222183,
0.03882933780550957,
-0.029647784307599068,
0.0006052490789443254,
0.0041414713487029076,
0.08764053136110306,
0.052342720329761505,
0.004910452291369438,
-0.13093100488185883,
-0.09526918828487396,
0.027852410450577736,
0.02549544721841812,
0.03795929625630379,
0.01790538989007473,
0.020881041884422302,
0.03745916113257408,
0.012563342228531837,
0.014916880987584591,
-0.016222866252064705,
0.04140607267618179,
-0.041731514036655426,
0.01100723072886467,
0.13991251587867737,
0.06492091715335846,
0.08112936466932297,
0.04277532547712326,
-0.10689873993396759,
0.040454212576150894,
-0.05501870438456535,
0.024508675560355186,
0.1166996881365776,
0.00947632733732462,
-0.09085185080766678,
0.018256021663546562,
-0.03684290125966072,
0.00509395357221365,
0.050043098628520966,
0.016655975952744484,
0.013320840895175934,
-0.0025750044733285904,
0.030264446511864662,
-0.051352545619010925,
-0.09853924810886383,
0.06788714230060577,
-0.07098423689603806,
0.02891991287469864,
-0.020739326253533363,
-0.013798179104924202,
0.09016797691583633,
0.04510962963104248,
-0.027252735570073128,
0.03112151101231575,
-0.08200222253799438,
0.030999407172203064,
0.12681758403778076,
0.01838499680161476,
-0.014099692925810814,
-0.04164973646402359,
0.01369970478117466
] | 0 | 0 | 0 | 0 |
|
Specialized Foundation Models Struggle to Beat Supervised Baselines | https://openreview.net/forum?id=JYTQ6ELUVO | [
"Zongzhe Xu",
"Ritvik Gupta",
"Wenduo Cheng",
"Alexander Shen",
"Junhong Shen",
"Ameet Talwalkar",
"Mikhail Khodak"
] | Poster | Following its success for vision and text, the "foundation model" (FM) paradigm—pretraining large models on massive data, then fine-tuning on target tasks—has rapidly expanded to domains in the sciences, engineering, healthcare, and beyond. Has this achieved what the original FMs accomplished, i.e. the supplanting of traditional supervised learning in their domains? To answer we look at three modalities—genomics, satellite imaging, and time series—with multiple recent FMs and compare them to a standard supervised learning workflow: model development, hyperparameter tuning, and training, all using only data from the target task. Across these three specialized domains, we find that it is consistently possible to train simple supervised models—no more complicated than a lightly modified wide ResNet or UNet—that match or even outperform the latest foundation models. Our work demonstrates that the benefits of large-scale pretraining have yet to be realized in many specialized areas, reinforces the need to compare new FMs to strong, well-tuned baselines, and introduces two new, easy-to-use, open-source, and automated workflows for doing so. | foundation models, supervised learning, neural architecture search, hyperparameter optimization, convolutional networks, autoregressive models, genomics, satellite imaging, time series | Foundation models in specialized domains have in many cases not yet surpassed much smaller supervised models that are trained on much less data. | 12,799 | 2411.02796 | [
0.0002502964634913951,
-0.1031157597899437,
0.0703023225069046,
0.04707929864525795,
0.14173196256160736,
-0.050505876541137695,
-0.05990320444107056,
0.019001429900527,
-0.059684157371520996,
-0.017435841262340546,
-0.08715809136629105,
-0.024062518030405045,
-0.012078817002475262,
0.08636132627725601,
-0.06179807335138321,
-0.04297281801700592,
0.038282644003629684,
-0.006511158775538206,
-0.08769507706165314,
-0.027575934305787086,
-0.06375255435705185,
0.052596405148506165,
0.022768298164010048,
0.007706781849265099,
-0.021484265103936195,
-0.026827100664377213,
-0.04567943513393402,
-0.03158513829112053,
0.010419703088700771,
-0.042093321681022644,
0.024946294724941254,
0.04269861429929733,
-0.021206064149737358,
0.02314877137541771,
-0.03604283556342125,
0.05393163114786148,
0.014341109432280064,
0.03175805136561394,
-0.01360953040421009,
-0.01744217798113823,
0.010022998787462711,
-0.06674902886152267,
0.007239475380629301,
0.04653496667742729,
0.06057761609554291,
-0.008619409054517746,
0.015226958319544792,
-0.12778161466121674,
0.018704183399677277,
-0.02426101081073284,
-0.0874006524682045,
-0.09970662742853165,
-0.03531809523701668,
0.07348792254924774,
-0.011859080754220486,
0.02017587423324585,
-0.00035844085505232215,
-0.05642008036375046,
-0.030388010665774345,
-0.04810822382569313,
-0.03418773040175438,
-0.08456218242645264,
-0.08942878246307373,
-0.0732714831829071,
0.007524554617702961,
-0.0004076691984664649,
-0.0033083760645240545,
0.01054608915001154,
0.04561173543334007,
0.028593629598617554,
0.007542225532233715,
0.0840430036187172,
-0.02963924966752529,
0.025628939270973206,
0.04680471494793892,
0.07096170634031296,
0.053986500948667526,
0.07276424020528793,
0.09051456302404404,
-0.09423619508743286,
0.020174255594611168,
0.02673996612429619,
0.01054194662719965,
-0.02483821101486683,
0.08640675246715546,
0.02032287046313286,
0.010455816052854061,
0.02981109358370304,
0.015334345400333405,
-0.015614769421517849,
-0.0007111876620911062,
-0.057228125631809235,
0.06789209693670273,
0.00950853805989027,
-0.05105922371149063,
0.02242751605808735,
0.023031774908304214,
-0.10861022770404816,
-0.008777720853686333,
0.11341295391321182,
-0.03383508697152138,
0.0602562315762043,
0.04510929435491562,
0.05012815073132515,
0.020980454981327057,
-0.03873979672789574,
0.060371141880750656,
0.07298225909471512,
0.03150002285838127,
-0.07884006947278976,
0.04075215384364128,
0.07746465504169464,
-0.020939817652106285,
-0.025307519361376762,
-0.006684399675577879,
-0.017949411645531654,
-0.03449702262878418,
-0.08110537379980087,
-0.002897497033700347,
0.052592724561691284,
-0.034639209508895874,
0.01809827983379364,
-0.04871317744255066,
-0.02976507879793644,
0.007938620634377003,
-0.013990853913128376,
-0.13045954704284668,
3.1189450385085296e-33,
0.05344336852431297,
0.05512632802128792,
0.05665403977036476,
0.0396600104868412,
-0.004876415710896254,
-0.03905327245593071,
-0.013547082431614399,
0.020996229723095894,
-0.006164632737636566,
0.08209653198719025,
-0.028354806825518608,
0.06038074940443039,
-0.04292045533657074,
0.13350649178028107,
0.01919323392212391,
-0.08161476999521255,
-0.07115427404642105,
0.05306936800479889,
-0.033161699771881104,
0.03475194424390793,
0.026629166677594185,
-0.01776563934981823,
-0.03688936308026314,
-0.07135548442602158,
0.08040942996740341,
0.0016010408289730549,
0.008797193877398968,
-0.012200256809592247,
-0.039138201624155045,
0.03306744992733002,
-0.10082853585481644,
-0.011490454897284508,
0.0357586108148098,
0.08067338913679123,
0.01781325787305832,
0.0036605584900826216,
0.03992422670125961,
-0.07599876821041107,
0.04205571860074997,
-0.03373872488737106,
0.01244206354022026,
0.05118925869464874,
0.010585417039692402,
-0.062011606991291046,
-0.0237788874655962,
-0.02887887693941593,
-0.008169460110366344,
0.005845409352332354,
-0.05639427900314331,
-0.028703493997454643,
0.024203428998589516,
0.016224144026637077,
-0.043893001973629,
-0.11298536509275436,
-0.040067270398139954,
0.03806393966078758,
-0.01946118101477623,
-0.0002490533224772662,
-0.01639457978308201,
0.03998035565018654,
0.028807373717427254,
-0.03311527892947197,
-0.06374724209308624,
0.11140923947095871,
0.0259613785892725,
0.017162511125206947,
0.01633293740451336,
0.06705054640769958,
0.03216421976685524,
0.04026854410767555,
-0.04590713232755661,
0.04325425997376442,
0.004834618419408798,
-0.10398954898118973,
0.13508473336696625,
-0.008950809948146343,
0.1125696673989296,
-0.0064743501134216785,
-0.02417048066854477,
0.035873930901288986,
0.01728334277868271,
0.04909385368227959,
-0.053811244666576385,
-0.045351557433605194,
-0.005785029847174883,
-0.03889741003513336,
0.024518366903066635,
-0.03771074861288071,
-0.052129097282886505,
-0.04386363551020622,
-0.07108737528324127,
-0.02460167184472084,
0.04063907638192177,
-0.007762637455016375,
-0.04768572747707367,
-2.757841990143554e-33,
0.010934524238109589,
0.005804975517094135,
-0.03967965021729469,
-0.0052734301425516605,
-0.011810197494924068,
-0.0010035092709586024,
0.015191527083516121,
0.03473395109176636,
-0.08565986901521683,
-0.047094669193029404,
0.09262305498123169,
-0.04360336810350418,
0.015112983994185925,
-0.06067108362913132,
-0.03757357597351074,
-0.03413764759898186,
-0.07461320608854294,
-0.023181669414043427,
0.0420512817800045,
0.05675846338272095,
0.010125397704541683,
0.14050191640853882,
-0.08583071082830429,
0.05627482756972313,
-0.039973899722099304,
0.07212333381175995,
-0.05230683833360672,
0.12439693510532379,
-0.05688527598977089,
0.006340802181512117,
-0.09352125227451324,
-0.03411070629954338,
0.016849754378199577,
-0.041590359061956406,
-0.027993975207209587,
0.0818951427936554,
0.06652691215276718,
-0.03507461026310921,
-0.020896367728710175,
0.04865415766835213,
0.004661249928176403,
0.036600109189748764,
-0.06628230214118958,
-0.006661272142082453,
-0.02594524808228016,
-0.02467496134340763,
-0.07639200985431671,
0.057562097907066345,
0.01990494318306446,
-0.06279021501541138,
-0.015918727964162827,
-0.008204027079045773,
-0.07445687800645828,
-0.03425572067499161,
-0.03103136457502842,
-0.06944725662469864,
0.028501270338892937,
-0.02486145682632923,
0.009926930069923401,
0.01484399288892746,
-0.06248347833752632,
-0.006705448031425476,
-0.014676843769848347,
0.004051472060382366,
-0.0066876220516860485,
-0.037121668457984924,
0.0017793712904676795,
-0.035335030406713486,
-0.0448443777859211,
0.06454485654830933,
-0.05136651545763016,
-0.03401602432131767,
0.020221779122948647,
0.007998845539987087,
-0.022115511819720268,
-0.025917714461684227,
0.05968981236219406,
-0.028149234130978584,
-0.03057810291647911,
0.006187666207551956,
0.06315240263938904,
-0.07416083663702011,
0.01838473044335842,
0.08347497135400772,
0.13068324327468872,
0.14492186903953552,
0.024784013628959656,
0.03739576041698456,
0.04081011936068535,
-0.01697893626987934,
-0.09939512610435486,
0.026023302227258682,
-0.014602486975491047,
0.0612419918179512,
-0.011284736916422844,
-4.957023946872141e-8,
-0.014042729511857033,
0.10315767675638199,
-0.046757541596889496,
0.01852254942059517,
-0.03587527573108673,
-0.06944750994443893,
-0.05062137171626091,
0.08262721449136734,
-0.020500831305980682,
0.015744540840387344,
0.032264720648527145,
0.013280411250889301,
-0.04761165753006935,
0.004826310556381941,
0.0451597198843956,
0.043295782059431076,
0.024690357968211174,
0.044672202318906784,
-0.034888237714767456,
0.03819941356778145,
-0.014472411014139652,
0.05745917558670044,
0.07296034693717957,
-0.05539146438241005,
0.11265119910240173,
-0.08509377390146255,
0.018602067604660988,
0.07312700897455215,
0.08408041298389435,
0.04452158510684967,
0.0008470662287436426,
0.037618882954120636,
-0.02364199049770832,
-0.0072705806232988834,
-0.027923867106437683,
-0.01630709506571293,
0.04346553981304169,
-0.02273455820977688,
-0.04560239240527153,
0.01438583992421627,
0.058310169726610184,
0.12797456979751587,
-0.11076301336288452,
0.01169740129262209,
-0.013770568184554577,
-0.02234857715666294,
-0.01336361188441515,
-0.02945629507303238,
0.011072590947151184,
-0.02333986759185791,
0.017833810299634933,
-0.016959920525550842,
0.041708722710609436,
0.0423283576965332,
0.04401054605841637,
0.055132217705249786,
0.0009630136773921549,
-0.0482647642493248,
-0.021347204223275185,
0.0736011192202568,
0.062162552028894424,
-0.04546055570244789,
-0.027361083775758743,
0.030953582376241684
] | https://github.com/ritvikgupta199/DASHA | 2 | 0 | 0 | 0 |
Efficient Biological Data Acquisition through Inference Set Design | https://openreview.net/forum?id=gVkX9QMBO3 | [
"Ihor Neporozhnii",
"Julien Roy",
"Emmanuel Bengio",
"Jason Hartford"
] | Poster | In drug discovery, highly automated high-throughput laboratories are used to screen a large number of compounds in search of effective drugs. These experiments are expensive, so one might hope to reduce their cost by only experimenting on a subset of the compounds, and predicting the outcomes of the remaining experiments. In this work, we model this scenario as a sequential subset selection problem: we aim to select the smallest set of candidates in order to achieve some desired level of accuracy for the system as a whole. Our key observation is that, if there is heterogeneity in the difficulty of the prediction problem across the input space, selectively obtaining the labels for the hardest examples in the acquisition pool will leave only the relatively easy examples to remain in the inference set, leading to better overall system performance. We call this mechanism inference set design, and propose the use of a confidence-based active learning solution to prune out these challenging examples. Our algorithm includes an explicit stopping criterion that interrupts the acquisition loop when it is sufficiently confident that the system has reached the target performance. Our empirical studies on image and molecular datasets, as well as a real-world large-scale biological assay, show that active learning for inference set design leads to significant reduction in experimental cost while retaining high system performance. | Active Learning, Data Acquisition, ML for Drug Discovery | We propose an active learning method for drug discovery, acquiring labels for some samples and training a model to predict the rest. Our models prune the most difficult examples from the target set to achieve high accuracy and reduce costs. | 12,794 | 2410.19631 | [
-0.025956079363822937,
0.0015490929363295436,
-0.0034967572428286076,
0.024931028485298157,
0.09082448482513428,
-0.0489376038312912,
0.06203487887978554,
0.07029645144939423,
0.01889321394264698,
-0.009490610100328922,
-0.058342158794403076,
-0.04463076964020729,
0.016360482200980186,
0.026842864230275154,
-0.025118956342339516,
0.056499890983104706,
0.05579955875873566,
0.023631852120161057,
-0.023558054119348526,
-0.04746834933757782,
-0.0360398069024086,
0.025182759389281273,
0.09686628729104996,
0.03376873955130577,
-0.025607161223888397,
0.023339295759797096,
-0.010744365863502026,
-0.017579514533281326,
-0.05082196369767189,
-0.021396582946181297,
0.01982870325446129,
-0.006006432697176933,
0.015040425583720207,
-0.08503957837820053,
0.013841681182384491,
-0.019439496099948883,
-0.11061162501573563,
0.05301060527563095,
0.041816502809524536,
-0.01230278518050909,
0.014841736294329166,
-0.0022134233731776476,
0.02908431552350521,
0.04963159188628197,
0.019312556833028793,
0.011251416988670826,
-0.038570694625377655,
0.021603425964713097,
0.027560967952013016,
0.005909878760576248,
-0.09976271539926529,
-0.059198349714279175,
-0.03618670627474785,
0.04604976624250412,
-0.040185268968343735,
0.02992614358663559,
0.043378617614507675,
-0.06788704544305801,
-0.01604536361992359,
0.011967874132096767,
-0.017153460532426834,
-0.12041793763637543,
-0.07422806322574615,
-0.06081780418753624,
-0.008322855457663536,
0.05200031027197838,
0.0835581049323082,
0.03788333386182785,
0.08001291751861572,
0.0051229726523160934,
-0.026926163583993912,
0.12756487727165222,
-0.058791808784008026,
0.08103340864181519,
0.05628122761845589,
0.029389405623078346,
0.051337093114852905,
0.0036950823850929737,
0.023551171645522118,
-0.06584218144416809,
0.009678315371274948,
-0.013287866488099098,
0.02932625822722912,
0.02558748796582222,
0.11594963818788528,
-0.03864678367972374,
0.025419792160391808,
0.05183872580528259,
0.059894587844610214,
0.029077807441353798,
-0.043473415076732635,
-0.03454917296767235,
0.014717654325067997,
0.007308488711714745,
0.008398479782044888,
0.005435020662844181,
0.024246208369731903,
-0.05004271864891052,
0.023508697748184204,
0.06395832449197769,
-0.008374708704650402,
0.06766574084758759,
0.028751112520694733,
-0.0277764443308115,
-0.024767860770225525,
-0.05686091259121895,
0.06528390198945999,
0.031191876158118248,
0.09048470109701157,
-0.10230943560600281,
0.016131754964590073,
0.019216090440750122,
0.03207141533493996,
0.05163133516907692,
0.026194898411631584,
0.008161717094480991,
0.019434619694948196,
-0.04999781399965286,
-0.051291994750499725,
0.023877084255218506,
-0.01907109096646309,
-0.014225923456251621,
0.0393080972135067,
-0.03605654835700989,
0.07099936157464981,
-0.03206901252269745,
-0.09506619721651077,
1.714496048433361e-34,
0.01990816928446293,
-0.029246380552649498,
0.09252031147480011,
0.01883223094046116,
0.0021859805565327406,
-0.06429656594991684,
-0.011668295599520206,
0.02496686764061451,
-0.029975809156894684,
0.05180259421467781,
-0.09396492689847946,
-0.05770479887723923,
-0.07947629690170288,
0.06671063601970673,
0.0774623304605484,
-0.015660878270864487,
0.014617844484746456,
0.078680120408535,
-0.018925590440630913,
-0.019036466255784035,
-0.022607022896409035,
-0.12188748270273209,
-0.05352751910686493,
-0.09580845385789871,
-0.018406255170702934,
0.0039889696054160595,
-0.06402675807476044,
0.08237989246845245,
-0.022489167749881744,
-0.008400344289839268,
-0.09016554802656174,
-0.006253235973417759,
0.025789780542254448,
0.11416525393724442,
0.01436704583466053,
-0.06479793041944504,
-0.016391446813941002,
-0.03378031775355339,
0.070816271007061,
0.033582206815481186,
0.02913723699748516,
0.03798041492700577,
0.1400957554578781,
0.07501721382141113,
-0.03362540528178215,
-0.018186330795288086,
-0.1325261890888214,
0.04765039682388306,
-0.06704090535640717,
0.027377810329198837,
0.05223017930984497,
-0.047585438936948776,
0.026140397414565086,
-0.046514369547367096,
-0.07678139954805374,
0.015846487134695053,
-0.07044864445924759,
0.036708783358335495,
0.03248429670929909,
0.08437184244394302,
-0.008938254788517952,
0.03127434849739075,
-0.03004297986626625,
0.10543250292539597,
-0.028646113350987434,
0.0906277447938919,
-0.026813730597496033,
-0.09266242384910583,
0.002815108047798276,
0.011002017185091972,
-0.021658089011907578,
0.020118627697229385,
0.0012538207229226828,
-0.06678745150566101,
0.1254197508096695,
0.051072172820568085,
0.07326149940490723,
0.015038496814668179,
-0.03937734663486481,
0.04086596146225929,
0.024338025599718094,
-0.0006190038984641433,
-0.06532371044158936,
0.008077472448348999,
-0.047663550823926926,
-0.06436929851770401,
0.016613101586699486,
-0.08207781612873077,
-0.066863514482975,
-0.021441703662276268,
-0.11552415043115616,
0.008649352006614208,
-0.08835004270076752,
-0.012161542661488056,
0.048693280667066574,
8.133251841824138e-34,
0.07170934975147247,
-0.011822421103715897,
0.09033157676458359,
0.018810885027050972,
0.03804478794336319,
0.011620762757956982,
-0.0004645017033908516,
-0.0656166598200798,
0.02965821884572506,
-0.07613532990217209,
0.035527125000953674,
0.019631292670965195,
0.025009844452142715,
-0.035357389599084854,
-0.0655786395072937,
0.015999767929315567,
-0.02535947784781456,
-0.027646314352750778,
0.054256148636341095,
0.07058218121528625,
-0.04090128839015961,
0.031502269208431244,
-0.06548545509576797,
-0.013663128949701786,
-0.0543583407998085,
0.04901810735464096,
-0.044608116149902344,
0.06042327731847763,
-0.014439296908676624,
-0.012428072281181812,
-0.09623761475086212,
-0.03533439338207245,
-0.06367964297533035,
-0.0010471446439623833,
-0.0268347579985857,
0.09813031554222107,
0.06760680675506592,
0.0032063706312328577,
-0.01909390278160572,
0.07086190581321716,
0.0601322203874588,
-0.008548831567168236,
-0.15970130264759064,
0.025802167132496834,
0.05532768741250038,
-0.02493797056376934,
-0.006383213214576244,
0.03875886648893356,
0.012481562793254852,
-0.01632120832800865,
-0.006604281719774008,
0.044101424515247345,
-0.04046135023236275,
-0.01255860272794962,
0.0014425920089706779,
-0.021888071671128273,
-0.045951686799526215,
-0.004923737607896328,
0.06897363811731339,
0.0012840344570577145,
-0.08622448891401291,
0.007462289649993181,
-0.02683621272444725,
0.061876747757196426,
-0.005792985204607248,
-0.02199428342282772,
0.01883634552359581,
0.04639079421758652,
0.006245113909244537,
0.023198673501610756,
0.042690783739089966,
0.059683945029973984,
0.06491263210773468,
0.019122803583741188,
-0.007783337030559778,
-0.03245113790035248,
-0.026127658784389496,
-0.10761059820652008,
0.018534084782004356,
0.031306955963373184,
0.036017172038555145,
-0.04206634312868118,
-0.014621193520724773,
0.03491351380944252,
-0.009038346819579601,
0.09855185449123383,
0.026928657665848732,
0.02289988100528717,
0.06274334341287613,
-0.05970810353755951,
-0.015832385048270226,
-0.023027783259749413,
0.04458744823932648,
0.040934719145298004,
-0.030913231894373894,
-4.103848993963766e-8,
0.08709334582090378,
0.014758576638996601,
0.0800105556845665,
-0.010661836713552475,
0.08690611273050308,
0.053893882781267166,
-0.06663249433040619,
0.12930625677108765,
-0.02323700487613678,
-0.057975124567747116,
0.03888332471251488,
0.0439608208835125,
-0.0014060308458283544,
-0.02308478392660618,
-0.033700015395879745,
0.048000164330005646,
0.041552018374204636,
0.021445127204060555,
-0.0423729233443737,
0.01795012317597866,
-0.01910809613764286,
0.011009657755494118,
0.08135318756103516,
-0.05207648500800133,
0.09397993981838226,
-0.1524963676929474,
0.025562070310115814,
-0.010315869934856892,
0.08169456571340561,
0.0242027435451746,
0.01643352024257183,
-0.04738364741206169,
0.04045732691884041,
0.020936908200383186,
0.0711146742105484,
0.008029064163565636,
-0.0735887959599495,
-0.07842808216810226,
-0.047389205545186996,
-0.10457270592451096,
-0.057308707386255264,
-0.0013027895474806428,
-0.09811361879110336,
-0.007255576550960541,
-0.0029489442240446806,
-0.020719831809401512,
0.022229723632335663,
-0.03021182492375374,
0.03668666630983353,
-0.014798054471611977,
0.035236287862062454,
-0.010840410366654396,
0.009197290986776352,
0.010627785697579384,
0.03352202847599983,
0.034546349197626114,
0.007899774238467216,
-0.050419487059116364,
0.0029050782322883606,
-0.035632457584142685,
-0.022880980744957924,
-0.024275779724121094,
-0.039126574993133545,
-0.008428562432527542
] | 0 | 0 | 0 | 0 |
|
Uncertainty Herding: One Active Learning Method for All Label Budgets | https://openreview.net/forum?id=UgPoHhYQ2U | [
"Wonho Bae",
"Danica J. Sutherland",
"Gabriel L. Oliveira"
] | Poster | Most active learning research has focused on methods which perform well when many labels are available, but can be dramatically worse than random selection when label budgets are small.
Other methods have focused on the low-budget regime, but do poorly as label budgets increase.
As the line between "low" and "high" budgets varies by problem,
this is a serious issue in practice.
We propose *uncertainty coverage*,
an objective which generalizes a variety of low- and high-budget objectives,
as well as natural, hyperparameter-light methods to smoothly interpolate between low- and high-budget regimes.
We call greedy optimization of the estimate Uncertainty Herding;
this simple method is computationally fast,
and we prove that it nearly optimizes the distribution-level coverage.
In experimental validation across a variety of active learning tasks,
our proposal matches or beats state-of-the-art performance in essentially all cases;
it is the only method of which we are aware that reliably works well in both low- and high-budget settings. | Active learning | Most active learning methods struggle to perform well across both low- and high-budget settings. We propose Uncertainty Herding, a fast and flexible method that adapts to all regimes, consistently matching or surpassing state-of-the-art performance. | 12,790 | 2412.20644 | [
-0.05340473726391792,
0.018961256369948387,
0.011667616665363312,
0.07386460155248642,
0.06217028945684433,
-0.047018490731716156,
0.01685558445751667,
-0.017023997381329536,
0.004601069260388613,
-0.030296018347144127,
-0.0688152089715004,
-0.023806961253285408,
0.03113136813044548,
-0.028525765985250473,
-0.03260562941431999,
0.0005291479756124318,
0.1158570647239685,
0.0492546372115612,
-0.04840007796883583,
-0.024453604593873024,
-0.0388423427939415,
0.006697302684187889,
0.04307158663868904,
0.02796732448041439,
0.07142031192779541,
-0.051107969135046005,
-0.044503405690193176,
0.0068856095895171165,
-0.012033463455736637,
0.007402314338833094,
0.025791680440306664,
-0.04289871081709862,
-0.014099668711423874,
-0.058866240084171295,
0.03267430514097214,
-0.03681528568267822,
-0.0829523503780365,
0.053252361714839935,
0.08787655830383301,
0.05973786115646362,
0.026256687939167023,
0.0107503617182374,
-0.02451901324093342,
0.026320170611143112,
-0.007112108636647463,
0.027936140075325966,
0.01838378608226776,
-0.0491149015724659,
-0.02434728480875492,
0.05255479738116264,
-0.10255799442529678,
-0.0757773295044899,
-0.006462101358920336,
-0.012913458980619907,
-0.07611773908138275,
0.057854730635881424,
0.035201072692871094,
-0.03122064657509327,
0.016337484121322632,
0.044612038880586624,
-0.05975944921374321,
-0.10649903863668442,
-0.05642859637737274,
0.011304762214422226,
-0.02413853630423546,
0.005315614864230156,
0.04922810196876526,
0.09358715265989304,
-0.018860481679439545,
0.034509655088186264,
0.06565737724304199,
0.16062264144420624,
-0.04672427102923393,
0.06922253221273422,
0.05016527324914932,
0.05478090047836304,
0.07187817245721817,
-0.006242703180760145,
0.07197733968496323,
-0.04690203443169594,
0.022801455110311508,
0.024894990026950836,
-0.017943844199180603,
-0.04196209833025932,
0.16698910295963287,
-0.03278959542512894,
0.04158680886030197,
0.011796281673014164,
-0.007048685569316149,
-0.030625123530626297,
-0.05352930352091789,
-0.011629684828221798,
0.018283331766724586,
0.021926455199718475,
-0.08868405222892761,
0.026532240211963654,
0.004053391050547361,
-0.08259527385234833,
0.03046928346157074,
0.12550126016139984,
-0.06324177980422974,
0.009649868123233318,
0.06444268673658371,
-0.07193177193403244,
-0.02365916408598423,
-0.12078669667243958,
-0.004007373936474323,
0.006014467217028141,
0.017908118665218353,
-0.028084933757781982,
0.03240037336945534,
-0.005150219425559044,
0.026234246790409088,
-0.0069704786874353886,
0.026092713698744774,
0.02571125701069832,
0.010652539320290089,
-0.006129145156592131,
-0.028410309925675392,
0.08503089845180511,
-0.05653853714466095,
0.005857989192008972,
0.03603443503379822,
0.003746584290638566,
0.010005008429288864,
-0.02231813780963421,
-0.0487326979637146,
3.17942532488429e-33,
0.06175728514790535,
-0.06208442524075508,
0.0329277478158474,
-0.0031867888756096363,
0.00014366679533850402,
-0.02923569083213806,
0.020738067105412483,
-0.02054504118859768,
0.0024659214541316032,
0.03611718490719795,
0.029901841655373573,
-0.00413817074149847,
-0.02001436986029148,
0.02088627591729164,
0.03221582993865013,
-0.06265728175640106,
-0.01754792407155037,
0.125339537858963,
-0.05882776901125908,
0.01399043295532465,
-0.036732692271471024,
-0.11692483723163605,
0.015653613954782486,
-0.08172957599163055,
-0.020515279844403267,
-0.028027711436152458,
0.017575165256857872,
0.07180672138929367,
0.020290536805987358,
0.03466147184371948,
-0.05290812999010086,
-0.02637123316526413,
0.054913971573114395,
0.023832999169826508,
0.021528365090489388,
-0.03633492439985275,
-0.06401155143976212,
-0.009367161430418491,
-0.003695910098031163,
0.02059866301715374,
-0.018612602725625038,
0.05611678585410118,
0.13640478253364563,
0.0006566157098859549,
-0.033435795456171036,
0.030983807519078255,
-0.018767615780234337,
-0.03788255900144577,
-0.05423268303275108,
-0.020815778523683548,
-0.034480445086956024,
-0.0798392966389656,
-0.0032495453488081694,
-0.02032529003918171,
-0.0381908193230629,
-0.03655204176902771,
-0.003505587112158537,
-0.008112269453704357,
-0.05038787052035332,
-0.03096838854253292,
-0.032556917518377304,
-0.0027840156108140945,
-0.05735056847333908,
0.042871586978435516,
0.0822729542851448,
0.14095459878444672,
0.07153672724962234,
-0.016720866784453392,
0.0113162100315094,
-0.03324385732412338,
0.0425163172185421,
0.006319351028650999,
-0.04960149526596069,
-0.04254193603992462,
0.09255776554346085,
-0.0012131760595366359,
0.10201938450336456,
0.045282505452632904,
-0.028762703761458397,
-0.056049738079309464,
-0.0056521352380514145,
-0.04512244835495949,
-0.017934491857886314,
0.0017509462777525187,
-0.02988184615969658,
-0.024929465726017952,
-0.0067304763942956924,
0.021282121539115906,
-0.03870851919054985,
-0.010934880934655666,
-0.046934280544519424,
-0.0041945017874240875,
-0.13104599714279175,
-0.009824960492551327,
0.010705776512622833,
-2.3151094916506094e-33,
-0.03568369150161743,
0.06455792486667633,
0.054679568856954575,
0.056907035410404205,
0.010263949632644653,
0.03403249382972717,
-0.01205324474722147,
-0.0029835670720785856,
0.0062588914297521114,
-0.04942813888192177,
-0.06972412765026093,
-0.0008294786093756557,
0.00109676958527416,
0.008308157324790955,
-0.012448173947632313,
-0.01374845951795578,
-0.019582247361540794,
-0.008875696919858456,
0.03705687075853348,
-0.0035887525882571936,
-0.008431759662926197,
0.037130676209926605,
-0.07076739519834518,
0.041637569665908813,
-0.007795990444719791,
0.007526416797190905,
-0.06588588654994965,
0.07397520542144775,
0.019864991307258606,
-0.057545360177755356,
-0.06379427760839462,
-0.0891176238656044,
-0.022587193176150322,
-0.08011835068464279,
-0.05503994971513748,
0.10332337021827698,
0.04716285690665245,
-0.004664267413318157,
-0.03860647603869438,
0.10356679558753967,
0.011957126669585705,
-0.02657790668308735,
-0.06386172026395798,
-0.09885888546705246,
-0.04570586234331131,
-0.01195591501891613,
-0.010392982512712479,
-0.02063647098839283,
-0.0413384810090065,
0.04253886267542839,
-0.041362497955560684,
0.027805479243397713,
-0.07960382103919983,
0.07894238084554672,
-0.05659869313240051,
-0.05742455646395683,
-0.0871252566576004,
-0.0035824724473059177,
0.001585311722010374,
0.0657813549041748,
-0.07426167279481888,
0.08136029541492462,
-0.033815961331129074,
0.03747779130935669,
-0.031989194452762604,
0.008271475322544575,
-0.026023827493190765,
0.006194509565830231,
-0.008551524020731449,
0.016571952030062675,
0.08227727562189102,
0.04985041543841362,
0.009058184921741486,
-0.010746502317488194,
-0.10007016360759735,
0.013838502578437328,
0.07551101595163345,
0.021324995905160904,
0.048607587814331055,
0.07443603128194809,
0.10678612440824509,
0.009861481375992298,
0.0340966135263443,
0.11567293107509613,
0.03888989984989166,
0.10417883098125458,
0.07595452666282654,
0.0006343433633446693,
0.03171024098992348,
-0.037601497024297714,
0.0014179664431139827,
0.02362852543592453,
0.07070588320493698,
0.016392234712839127,
-0.04554624482989311,
-4.650041418585715e-8,
-0.030227720737457275,
0.06690674275159836,
0.025666270405054092,
0.030444150790572166,
0.11991074681282043,
0.00906311720609665,
-0.011241907253861427,
0.06501668691635132,
-0.05736973136663437,
-0.02543220855295658,
0.08002517372369766,
0.01420766580849886,
-0.030331576243042946,
-0.031096870079636574,
-0.04754054173827171,
0.01844804920256138,
-0.002308246912434697,
0.05990750715136528,
-0.04096097871661186,
0.026235805824398994,
-0.013195004314184189,
0.025157799944281578,
0.03820949047803879,
-0.09484566748142242,
0.1117262914776802,
-0.09113598614931107,
0.019995855167508125,
0.0624232292175293,
0.12107552587985992,
0.049735646694898605,
-0.029265329241752625,
-0.032706983387470245,
-0.011528586968779564,
-0.03900280222296715,
0.013720261864364147,
0.10055593401193619,
-0.07407744228839874,
-0.09832055866718292,
0.025247421115636826,
-0.050300631672143936,
-0.03791986033320427,
0.03739413991570473,
-0.040114276111125946,
0.028392065316438675,
0.06924290955066681,
-0.0019733926746994257,
-0.00636389059945941,
-0.08166687190532684,
0.038325000554323196,
-0.09131012856960297,
-0.012628604657948017,
-0.027383480221033096,
0.002507156925275922,
0.03066023252904415,
0.03039608895778656,
-0.025108011439442635,
-0.02650846540927887,
-0.03410196304321289,
-0.024494724348187447,
0.08121905475854874,
0.0779636949300766,
-0.03486741706728935,
-0.08335912227630615,
0.09405963867902756
] | 0 | 0 | 0 | 0 |
|
In-context Time Series Predictor | https://openreview.net/forum?id=dCcY2pyNIO | [
"Jiecheng Lu",
"Yan Sun",
"Shihao Yang"
] | Poster | Recent Transformer-based large language models (LLMs) demonstrate in-context learning ability to perform various functions based solely on the provided context, without updating model parameters. To fully utilize the in-context capabilities in time series forecasting (TSF) problems, unlike previous Transformer-based or LLM-based time series forecasting methods, we reformulate "time series forecasting tasks" as input tokens by constructing a series of (lookback, future) pairs within the tokens. This method aligns more closely with the inherent in-context mechanisms and is more parameter-efficient without the need of using pre-trained LLM parameters. Furthermore, it addresses issues such as overfitting in existing Transformer-based TSF models, consistently achieving better performance across full-data, few-shot, and zero-shot settings compared to previous architectures. | Time Series Forecasting, In-context Learning, Transformer | null | 12,786 | 2405.14982 | [
-0.09741685539484024,
-0.04027597978711128,
0.040061887353658676,
0.06317682564258575,
0.044659506529569626,
-0.006498285569250584,
-0.046319328248500824,
0.054397400468587875,
0.019870469346642494,
-0.03185776248574257,
-0.09551303833723068,
-0.022110017016530037,
-0.003273728070780635,
0.07664946466684341,
0.04107353463768959,
-0.004668375942856073,
-0.008004172705113888,
0.04550257697701454,
-0.05920322239398956,
-0.08990916609764099,
0.009194593876600266,
0.04678262025117874,
-0.01879565231502056,
0.001452959724701941,
0.07821527868509293,
-0.02048920840024948,
-0.018930718302726746,
-0.018870621919631958,
0.009873786941170692,
0.09471452236175537,
-0.015887903049588203,
0.06984292715787888,
-0.08269985765218735,
0.019296834245324135,
-0.07799143344163895,
0.059071123600006104,
-0.05211073160171509,
-0.005113604944199324,
0.015745550394058228,
-0.04455229640007019,
0.01311434619128704,
-0.04005281999707222,
0.02275274507701397,
0.011522914282977581,
0.01544205192476511,
0.007907646708190441,
0.0678129494190216,
-0.0385938435792923,
-0.014693077653646469,
0.0024079878348857164,
-0.08551286160945892,
0.00003165160887874663,
-0.02367568574845791,
0.037128347903490067,
-0.06805089116096497,
0.039548251777887344,
-0.010293398052453995,
0.03870579972863197,
0.03134585916996002,
-0.0455549880862236,
-0.10773555934429169,
-0.09922249615192413,
-0.015945598483085632,
0.04470111429691315,
-0.06685973703861237,
0.02442142926156521,
0.009462110698223114,
0.07706953585147858,
-0.06177506968379021,
-0.037420421838760376,
-0.04356866329908371,
0.0575447641313076,
-0.031559206545352936,
0.05005213990807533,
0.02023657038807869,
0.060377806425094604,
0.07949230074882507,
-0.02967729978263378,
0.06755121052265167,
-0.043592676520347595,
0.06532589346170425,
0.021528787910938263,
0.023191452026367188,
-0.0617930106818676,
0.106733039021492,
0.037967320531606674,
0.014482345432043076,
0.028355980291962624,
0.028428582474589348,
0.021763769909739494,
0.0011136202374473214,
-0.10619714856147766,
0.07436086237430573,
0.016488168388605118,
-0.001608023070730269,
0.04482089728116989,
-0.011024701409041882,
-0.01341234240680933,
-0.06433794647455215,
0.08398917317390442,
-0.020038317888975143,
0.010348409414291382,
0.06861763447523117,
-0.005593922454863787,
-0.023882145062088966,
-0.03382515162229538,
-0.015536107122898102,
0.007506749592721462,
0.02219235897064209,
-0.061704736202955246,
0.025379154831171036,
0.05554765462875366,
0.04961925745010376,
-0.034793876111507416,
0.001571076805703342,
-0.04580842703580856,
0.010581820271909237,
-0.06759507209062576,
0.029330065473914146,
0.11570030450820923,
-0.011731566861271858,
0.03612218797206879,
-0.06886401772499084,
0.04670432582497597,
-0.025361893698573112,
-0.02165522612631321,
-0.013171221129596233,
1.4328577268397922e-33,
-0.01989113539457321,
0.024808477610349655,
0.0566372349858284,
-0.0054817101918160915,
0.003749158000573516,
0.036064986139535904,
0.03745419904589653,
0.032595571130514145,
0.01109873317182064,
-0.03132351115345955,
0.02624109759926796,
0.0717892274260521,
-0.0853043720126152,
0.0371401384472847,
-0.03151305019855499,
-0.02932548150420189,
-0.012248283252120018,
0.054285794496536255,
0.025857985019683838,
0.011342319659888744,
0.022529812529683113,
-0.021583106368780136,
-0.003107623662799597,
-0.08076842129230499,
-0.00004964826803188771,
0.05434829741716385,
0.12996073067188263,
-0.07081618160009384,
-0.01661483198404312,
-0.0006258339853957295,
-0.05259789898991585,
0.0006190019194036722,
-0.04971790686249733,
-0.00014694001583848149,
0.02933703549206257,
-0.06395255029201508,
-0.011597245931625366,
-0.01332945004105568,
-0.005561570171266794,
-0.031294941902160645,
-0.033402178436517715,
0.06852326542139053,
-0.03046133741736412,
-0.011403219774365425,
-0.0987873375415802,
-0.1096106618642807,
0.03262301906943321,
-0.029328400269150734,
-0.0073161376640200615,
0.008542080409824848,
0.0669320821762085,
-0.031161027029156685,
-0.10695070773363113,
-0.06929290294647217,
0.055178288370370865,
0.03624188154935837,
0.052933573722839355,
-0.02084079384803772,
0.013776388019323349,
0.02614463120698929,
-0.05406946316361427,
-0.0036430046893656254,
0.026124129071831703,
0.018935758620500565,
0.012636322528123856,
-0.019876595586538315,
0.03840669244527817,
0.08086231350898743,
0.011582806706428528,
-0.061867475509643555,
0.0028579391073435545,
-0.022090977057814598,
-0.03820780664682388,
0.00853731669485569,
0.058279022574424744,
0.017052140086889267,
0.09250636398792267,
-0.04582979530096054,
-0.0321456603705883,
0.05998993664979935,
0.028362862765789032,
0.014260961674153805,
0.06724368035793304,
-0.013124549761414528,
0.01895148679614067,
-0.001970792654901743,
0.03188607841730118,
-0.07122103124856949,
-0.04367610812187195,
-0.1058318242430687,
-0.10639826208353043,
-0.01071788091212511,
-0.04090731963515282,
0.023814918473362923,
-0.024353493005037308,
-2.432722844238164e-33,
-0.001745323883369565,
-0.005266997031867504,
-0.045465122908353806,
0.032846175134181976,
-0.023877346888184547,
-0.06843060255050659,
0.011425858363509178,
0.037538107484579086,
-0.0027523546013981104,
-0.04198779538273811,
0.005857448559254408,
-0.05496520921587944,
0.0520620197057724,
-0.06350819021463394,
0.04290977120399475,
-0.07112361490726471,
0.009618008509278297,
0.0030390454921871424,
-0.04170948266983032,
-0.012593776918947697,
-0.003360206726938486,
0.015219218097627163,
-0.09503858536481857,
0.07592283934354782,
-0.05547631159424782,
-0.02855336293578148,
-0.07787203043699265,
0.08599486202001572,
-0.01828775182366371,
-0.003445310052484274,
-0.03785311430692673,
-0.0555286705493927,
0.10916204750537872,
-0.025568462908267975,
-0.03681446611881256,
0.039339110255241394,
0.02756454423069954,
-0.04328697547316551,
-0.0037158920895308256,
0.08295958489179611,
0.1380118578672409,
0.02100626565515995,
-0.013496276922523975,
0.04347936436533928,
-0.0995032861828804,
0.011340600438416004,
-0.05805787071585655,
-0.03686392307281494,
0.04626813903450966,
0.015502884052693844,
-0.050277821719646454,
0.03221404552459717,
-0.111565001308918,
-0.03831816092133522,
-0.01722000539302826,
-0.06627333164215088,
0.07626080513000488,
-0.08845006674528122,
0.01454512495547533,
-0.05803052708506584,
0.004999133758246899,
-0.03277021273970604,
0.12909376621246338,
-0.12259265780448914,
0.025944221764802933,
-0.05736794322729111,
-0.021808305755257607,
-0.03823329880833626,
0.037327904254198074,
0.016963064670562744,
0.04949997738003731,
-0.01804947480559349,
-0.0246751569211483,
0.000622789841145277,
-0.008063831366598606,
-0.029635174199938774,
0.03171968087553978,
-0.08190610259771347,
0.03867322579026222,
-0.037274498492479324,
-0.06737347692251205,
0.002478394191712141,
0.014434341341257095,
0.08808384090662003,
0.0700254887342453,
0.09837157279253006,
0.09524349868297577,
0.048251353204250336,
0.07318707555532455,
0.018185744062066078,
-0.0688590481877327,
0.06915362179279327,
0.006952803116291761,
0.12031758576631546,
-0.04138379171490669,
-4.248957807817533e-8,
-0.07858364284038544,
0.04357575252652168,
-0.018193289637565613,
0.03878104314208031,
-0.025614002719521523,
-0.04324758052825928,
0.009855550713837147,
0.06685858219861984,
0.025756947696208954,
-0.0005874496418982744,
0.067162424325943,
-0.07533726841211319,
-0.04508542641997337,
-0.03265153244137764,
0.035043392330408096,
0.05650778114795685,
0.03616369888186455,
0.02464245818555355,
-0.01925872638821602,
0.0029145872686058283,
0.05800158903002739,
0.08063879609107971,
0.004319582134485245,
-0.03783789277076721,
0.11832868307828903,
-0.05771447718143463,
-0.046397723257541656,
0.1205754429101944,
0.056870222091674805,
0.03751456364989281,
-0.037450604140758514,
0.0014186485204845667,
-0.04652022942900658,
-0.07225814461708069,
-0.10190019756555557,
0.034307342022657394,
0.08244893699884415,
-0.02344135381281376,
-0.017748691141605377,
-0.007160150911659002,
0.029698394238948822,
0.03903511166572571,
-0.0921916589140892,
0.06392756849527359,
-0.038328178226947784,
0.006735293194651604,
0.04872961342334747,
-0.10400614142417908,
-0.03240741789340973,
-0.06044474616646767,
0.05158451572060585,
-0.00036640939651988447,
0.0007868282846175134,
0.05652075633406639,
0.10125865042209625,
0.02141718380153179,
0.006696700118482113,
-0.03207315132021904,
0.010408883914351463,
0.0709913969039917,
-0.047173433005809784,
-0.019674871116876602,
-0.07639510929584503,
0.09550664573907852
] | 0 | 0 | 0 | 0 |
|
More Experts Than Galaxies: Conditionally-Overlapping Experts with Biologically-Inspired Fixed Routing | https://openreview.net/forum?id=1qq1QJKM5q | [
"Sagi Shaier",
"Francisco Pereira",
"Katharina von der Wense",
"Lawrence Hunter",
"Matt Jones"
] | Poster | The evolution of biological neural systems has led to both modularity and sparse coding, which enables energy efficiency and robustness across the diversity of tasks in the lifespan. In contrast, standard neural networks rely on dense, non-specialized architectures, where all model parameters are simultaneously updated to learn multiple tasks, leading to interference. Current sparse neural network approaches aim to alleviate this issue but are hindered by limitations such as 1) trainable gating functions that cause representation collapse, 2) disjoint experts that result in redundant computation and slow learning, and 3) reliance on explicit input or task IDs that limit flexibility and scalability.
In this paper we propose Conditionally Overlapping Mixture of ExperTs (COMET), a general deep learning method that addresses these challenges by inducing a modular, sparse architecture with an exponential number of overlapping experts. COMET replaces the trainable gating function used in Sparse Mixture of Experts with a fixed, biologically inspired random projection applied to individual input representations. This design causes the degree of expert overlap to depend on input similarity, so that similar inputs tend to share more parameters. This results in faster learning per update step and improved out-of-sample generalization.
We demonstrate the effectiveness of COMET on a range of tasks, including image classification, language modeling, and regression, using several popular deep learning architectures. | Deep learning, Mixture of Experts, Modularity, Sparsity, Conditional Computation | We propose Conditionally Overlapping Mixture of ExperTs (COMET), a general deep learning method that induces a modular, sparse architecture with an exponential number of overlapping experts | 12,783 | 2410.08003 | [
-0.08561860769987106,
-0.07446039468050003,
0.036976806819438934,
0.09131918847560883,
0.05145138502120972,
0.007199158892035484,
-0.01545137632638216,
-0.03586176782846451,
-0.006585673429071903,
-0.07628141343593597,
-0.03696466237306595,
-0.0478469617664814,
-0.01981167122721672,
0.0318492166697979,
-0.07050065696239471,
-0.028453445062041283,
0.06420066952705383,
0.011855913326144218,
-0.13936610519886017,
-0.06585023552179337,
0.008369523100554943,
0.01912738010287285,
0.01774682104587555,
0.010904932394623756,
0.033993132412433624,
0.03010101243853569,
0.021684948354959488,
-0.014237083494663239,
0.017896978184580803,
-0.0030817214865237474,
0.09864780306816101,
-0.010646644048392773,
-0.02556498721241951,
-0.001466231420636177,
-0.0300181582570076,
0.04718374088406563,
-0.13819502294063568,
0.020198151469230652,
0.046269070357084274,
-0.03434903174638748,
-0.003409789176657796,
0.02456415817141533,
-0.00024771608877927065,
0.041228991001844406,
0.028432417660951614,
0.016200926154851913,
-0.00929123442620039,
-0.037411026656627655,
0.040128760039806366,
0.007145301904529333,
-0.004309182520955801,
-0.04923597350716591,
0.02745787426829338,
0.06744842231273651,
0.07318115234375,
0.025149866938591003,
-0.058188728988170624,
0.012365383096039295,
-0.07943818718194962,
-0.021126599982380867,
-0.025113539770245552,
-0.09716575592756271,
0.020805807784199715,
-0.045618776232004166,
0.03563271462917328,
0.02228485606610775,
0.010876118205487728,
0.06980925053358078,
0.010387878865003586,
0.011434362269937992,
-0.0028279265388846397,
0.09614218026399612,
-0.058892589062452316,
0.035204194486141205,
0.021678714081645012,
0.04256368428468704,
0.021421417593955994,
0.017607439309358597,
0.055772870779037476,
0.01460815966129303,
-0.0017439498333260417,
-0.012378152459859848,
-0.03249868378043175,
-0.013637294061481953,
0.1406019777059555,
-0.03427991643548012,
-0.0030321930535137653,
0.10894668847322464,
0.03178256377577782,
-0.034646667540073395,
-0.059471867978572845,
-0.03166371211409569,
0.09864909201860428,
-0.04625067487359047,
0.0058092279359698296,
0.04471910372376442,
0.02068837732076645,
-0.07647579908370972,
-0.03111879713833332,
0.09648735076189041,
-0.09946972131729126,
-0.04983866959810257,
0.08534666150808334,
0.062436118721961975,
0.028362762182950974,
0.019469287246465683,
-0.007182124070823193,
0.12395612895488739,
0.03991508111357689,
-0.1115003228187561,
0.0014405448455363512,
0.040113676339387894,
0.019327783957123756,
0.044819965958595276,
-0.054791465401649475,
0.044729284942150116,
0.047263335436582565,
-0.030457332730293274,
0.015572655014693737,
0.08157096058130264,
-0.01602722331881523,
-0.03469342365860939,
-0.054741453379392624,
-0.004760930314660072,
0.0368315614759922,
0.006092458497732878,
-0.049143631011247635,
2.5762130519201407e-33,
0.05012109875679016,
0.014067314565181732,
0.02162543684244156,
0.03469012677669525,
0.08176842331886292,
-0.030578970909118652,
-0.028214430436491966,
0.04156552627682686,
0.0006148400134406984,
0.00727337459102273,
-0.0904005765914917,
0.030139468610286713,
-0.014385975897312164,
0.059676643460989,
-0.02611052803695202,
0.03300773352384567,
-0.03619382157921791,
0.021580753847956657,
0.019628029316663742,
-0.10250135511159897,
0.016359630972146988,
-0.03029005229473114,
-0.031402356922626495,
-0.040196485817432404,
-0.0076952469535171986,
-0.000855088816024363,
0.07690536230802536,
-0.0798828974366188,
0.035153526812791824,
-0.020230669528245926,
-0.09525959938764572,
-0.0034938829485327005,
-0.009832738898694515,
0.04845128953456879,
0.03774989768862724,
-0.04308818280696869,
0.0375017449259758,
-0.09678158164024353,
0.022232480347156525,
0.031781744211912155,
0.008505299687385559,
0.04891512170433998,
0.059268586337566376,
-0.03882288187742233,
-0.04947931319475174,
-0.019467126578092575,
0.050009652972221375,
0.028702151030302048,
-0.07116428762674332,
0.047868143767118454,
0.05270225182175636,
-0.010737963020801544,
0.04560939595103264,
-0.05700402334332466,
0.010889986529946327,
-0.03204567730426788,
-0.0017462060786783695,
0.032718975096940994,
0.0742364376783371,
0.032561641186475754,
0.03076382726430893,
0.03110380843281746,
0.008686019107699394,
0.0923333689570427,
0.07823728770017624,
0.06889375299215317,
-0.042007189244031906,
0.11060375720262527,
0.12763352692127228,
-0.009562653489410877,
-0.01131521351635456,
-0.004232298117130995,
0.023380925878882408,
-0.06994760781526566,
0.036294709891080856,
0.027041016146540642,
-0.010990587994456291,
-0.11140255630016327,
-0.09544923901557922,
0.03313329443335533,
-0.055888183414936066,
0.019458729773759842,
-0.011835657060146332,
-0.04759584739804268,
-0.06096833571791649,
-0.001990809803828597,
0.007042707875370979,
-0.03563545271754265,
0.0024601591285318136,
-0.046765197068452835,
-0.0952838808298111,
-0.016092617064714432,
0.01553778350353241,
-0.047774940729141235,
-0.025081096217036247,
-4.479799787434122e-34,
0.01851518824696541,
0.015515938401222229,
-0.01982235722243786,
0.007619970943778753,
0.07034154236316681,
-0.04822022095322609,
0.003190753748640418,
-0.058799486607313156,
-0.033806271851062775,
-0.06411591917276382,
-0.06815357506275177,
0.02411767654120922,
0.11225206404924393,
-0.05940210074186325,
0.046799153089523315,
0.05564383417367935,
-0.09149443358182907,
0.06195990741252899,
0.02544204704463482,
0.014050227589905262,
-0.03258494287729263,
-0.0020467236172407866,
-0.07022243738174438,
-0.012541244737803936,
-0.03801170736551285,
0.05931108444929123,
-0.0565798282623291,
0.10117799043655396,
0.03725223243236542,
0.023974698036909103,
-0.08244124054908752,
-0.032912880182266235,
-0.04849503934383392,
-0.12435746192932129,
0.005602552089840174,
0.14218957722187042,
0.04307215288281441,
0.05228735879063606,
-0.016778217628598213,
-0.00872750859707594,
0.031689926981925964,
-0.0234052911400795,
-0.08927404880523682,
0.0642658919095993,
0.030777230858802795,
-0.0020404341630637646,
-0.020902806892991066,
-0.03067689761519432,
-0.02575083076953888,
0.027441754937171936,
0.009709942154586315,
0.04498615860939026,
-0.11254489421844482,
-0.016610827296972275,
-0.0403330996632576,
-0.061670660972595215,
0.016777541488409042,
0.016840573400259018,
0.08664552122354507,
-0.043954551219940186,
0.051417361944913864,
-0.12158385664224625,
-0.00019500998314470053,
-0.002536344574764371,
-0.008590802550315857,
-0.03360952436923981,
0.010702546685934067,
0.05626918375492096,
0.016003388911485672,
0.09356484562158585,
-0.01120991911739111,
-0.04649536684155464,
0.059961043298244476,
0.02678954415023327,
0.005658962298184633,
0.07778719067573547,
0.08432547748088837,
-0.09076205641031265,
0.01847069524228573,
-0.007058370858430862,
-0.00769324554130435,
-0.04762912541627884,
-0.02529902011156082,
-0.00046906209900043905,
0.08672540634870529,
0.1129629835486412,
-0.012082848697900772,
0.07513851672410965,
0.04106767475605011,
0.0037341287825256586,
-0.042162179946899414,
0.00912631954997778,
-0.01097110565751791,
0.035935867577791214,
-0.052301134914159775,
-4.555105803660808e-8,
0.041459158062934875,
0.08923133462667465,
-0.06563732028007507,
-0.021303493529558182,
0.11182308942079544,
-0.05240577831864357,
-0.013892889022827148,
0.03592703491449356,
0.02746683917939663,
-0.054347142577171326,
0.11292268335819244,
-0.010579629801213741,
0.027050890028476715,
-0.021704738959670067,
0.05770770087838173,
0.07776087522506714,
-0.07504981011152267,
0.02363281324505806,
-0.06852611899375916,
0.051825664937496185,
0.016881752759218216,
-0.0038120197132229805,
-0.041757624596357346,
-0.02980552799999714,
0.09224285930395126,
-0.09580826759338379,
-0.032060880213975906,
0.026786746457219124,
0.03166348114609718,
0.008086973801255226,
-0.09094144403934479,
0.013408023864030838,
0.06213757395744324,
-0.0020275856368243694,
-0.0017404670361429453,
0.07117446511983871,
0.03049153834581375,
0.012787017039954662,
-0.055218469351530075,
-0.054200176149606705,
-0.0024078842252492905,
-0.021632075309753418,
0.02445376105606556,
0.01725771464407444,
0.04267064109444618,
-0.023118572309613228,
0.04943052679300308,
-0.11361315846443176,
0.022777944803237915,
0.06391238421201706,
-0.0029562530107796192,
-0.015060832723975182,
0.008098501712083817,
0.027068039402365685,
0.04786786809563637,
-0.031039131805300713,
0.007783637847751379,
-0.033412784337997437,
0.003574221394956112,
0.02389589697122574,
0.001320605631917715,
-0.02225516363978386,
-0.06121635064482689,
-0.014929741621017456
] | 0 | 0 | 0 | 0 |
|
Can Watermarks be Used to Detect LLM IP Infringement For Free? | https://openreview.net/forum?id=KRMSH1GxUK | [
"Zhengyue Zhao",
"Xiaogeng Liu",
"Somesh Jha",
"Patrick McDaniel",
"Bo Li",
"Chaowei Xiao"
] | Poster | The powerful capabilities of LLMs stem from their rich training data and high-quality labeled datasets, making the training of strong LLMs a resource-intensive process, which elevates the importance of IP protection for such LLMs. Compared to gathering high-quality labeled data, directly sampling outputs from these fully trained LLMs as training data presents a more cost-effective approach. This practice—where a suspect model is fine-tuned using high-quality data derived from these LLMs, thereby gaining capabilities similar to the target model—can be seen as a form of IP infringement against the original LLM. In recent years, LLM watermarks have been proposed and used to detect whether a text is AI-generated. Intuitively, if data sampled from a watermarked LLM is used for training, the resulting model would also be influenced by this watermark. This raises the question: can we directly use such watermarks to detect IP infringement of LLMs? In this paper, we explore the potential of LLM watermarks for detecting model infringement. We find that there are two issues with direct detection: (1) The queries used to sample output from the suspect LLM have a significant impact on detectability. (2) The watermark that is easily learned by LLMs exhibits instability regarding the watermark's hash key during detection. To address these issues, we propose LIDet, a detection method that leverages available anchor LLMs to select suitable queries for sampling from the suspect LLM. Additionally, it adapts the detection threshold to mitigate detection failures caused by different hash keys. To demonstrate the effectiveness of this approach, we construct a challenging model set containing multiple suspect LLMs on which direct detection methods struggle to yield effective results. Our method achieves over 90\% accuracy in distinguishing between infringing and clean models, demonstrating the feasibility of using LLM watermarks to detect LLM IP infringement. | large language models, watermark, model copyright, model infringement detection | null | 12,782 | null | [
-0.09930349886417389,
-0.04542392119765282,
0.03854425996541977,
-0.018555274233222008,
0.16576935350894928,
-0.013200974091887474,
0.05479228124022484,
-0.05522448942065239,
0.04292413592338562,
-0.03075917437672615,
-0.02462528645992279,
0.0061061084270477295,
0.09781408309936523,
-0.04886291176080704,
-0.002000304637476802,
-0.002130372915416956,
0.15553614497184753,
0.07695215195417404,
-0.059653256088495255,
-0.07338037341833115,
0.057898301631212234,
-0.04445837065577507,
0.01668625697493553,
-0.003526954213157296,
-0.05495362728834152,
0.007899017073214054,
0.03848015144467354,
0.03359195962548256,
0.006404277868568897,
-0.037689171731472015,
0.006593178026378155,
0.01767123118042946,
-0.008393775671720505,
0.05429430305957794,
-0.042243268340826035,
-0.07869140803813934,
-0.03957318887114525,
-0.0259746965020895,
0.0625348910689354,
-0.07319588214159012,
0.05527849495410919,
-0.05145499110221863,
-0.029540890827775,
0.03208426386117935,
0.03055671416223049,
0.03702792152762413,
0.06373696029186249,
0.06738529354333878,
-0.07145099341869354,
0.036117054522037506,
-0.05041685700416565,
-0.0493510402739048,
-0.03869275748729706,
-0.02566462755203247,
-0.021224988624453545,
-0.09152332693338394,
-0.0023494160268455744,
0.009193739853799343,
-0.06290347129106522,
0.04192851483821869,
-0.060048673301935196,
-0.07406143844127655,
-0.012463361956179142,
0.004769572056829929,
0.037116941064596176,
-0.04333879053592682,
-0.006185578648000956,
0.14072299003601074,
0.03321269527077675,
0.032962195575237274,
0.05234674736857414,
0.07336458563804626,
-0.023239023983478546,
0.04378326237201691,
-0.043821290135383606,
0.08434180915355682,
-0.050088413059711456,
0.037092175334692,
0.03960631787776947,
-0.11480465531349182,
-0.01472828071564436,
-0.07595934718847275,
0.08075009286403656,
0.039268266409635544,
0.04540049657225609,
0.022019386291503906,
0.036753978580236435,
0.018358943983912468,
0.07460931688547134,
-0.016234343871474266,
-0.041250210255384445,
-0.04703659936785698,
0.0906682163476944,
0.013645722530782223,
0.04808827117085457,
0.07612649351358414,
-0.04775437340140343,
-0.052847567945718765,
0.022190844640135765,
0.06558163464069366,
-0.023184074088931084,
0.04338937997817993,
-0.02704121358692646,
-0.02484203688800335,
0.07446040213108063,
0.027896175161004066,
0.02398872748017311,
0.010883517563343048,
0.08378484845161438,
-0.008894036523997784,
0.004216511733829975,
0.06705295294523239,
0.024063749238848686,
-0.02860230579972267,
0.07885891199111938,
-0.04734659567475319,
0.019511884078383446,
-0.03692907467484474,
-0.05090654641389847,
-0.018417922779917717,
-0.036509934812784195,
0.005301770754158497,
-0.0032322159968316555,
0.0006603380898013711,
-0.01769568771123886,
-0.01456625945866108,
-0.08085284382104874,
2.4186457483740903e-33,
0.054859187453985214,
0.05728406086564064,
-0.02259073406457901,
-0.1126079186797142,
-0.019189294427633286,
0.004601805470883846,
0.03760243579745293,
0.09976005554199219,
-0.024246159940958023,
0.04527292028069496,
-0.06962930411100388,
-0.049650583416223526,
0.001886831596493721,
0.07240671664476395,
0.04740242287516594,
0.03425496444106102,
0.01565629430115223,
0.012372731231153011,
-0.0022365269251167774,
0.005932192783802748,
0.04731766879558563,
-0.07231542468070984,
0.027813062071800232,
-0.05321035906672478,
0.02616158127784729,
-0.0069791716523468494,
-0.019483262673020363,
-0.027809161692857742,
0.08133447915315628,
0.02485438995063305,
-0.040853578597307205,
-0.0262992475181818,
0.07171671092510223,
0.05213656648993492,
0.059906456619501114,
0.03160557523369789,
-0.05704225227236748,
0.014117086306214333,
-0.012896952219307423,
-0.004717729520052671,
-0.0008030497701838613,
-0.012427755631506443,
0.006742737255990505,
-0.042468927800655365,
-0.10827860981225967,
-0.05021144822239876,
-0.11686348915100098,
-0.0811813548207283,
-0.016019996255636215,
-0.08827440440654755,
0.0723162367939949,
0.009078597649931908,
-0.06313397735357285,
-0.04383938014507294,
-0.0773191899061203,
-0.01781187579035759,
-0.012998279184103012,
0.018414387479424477,
0.04221409186720848,
0.076888807117939,
-0.07959485799074173,
0.0035478032659739256,
-0.0188657958060503,
0.11293895542621613,
-0.008448031730949879,
0.031234364956617355,
0.03530890867114067,
0.018541067838668823,
0.007566207088530064,
-0.01293662004172802,
0.056048400700092316,
-0.04907017946243286,
-0.03630775958299637,
0.003647303907200694,
-0.038952670991420746,
-0.07933855801820755,
0.029151543974876404,
0.0005677819135598838,
0.011487823911011219,
0.010036169551312923,
0.03333974629640579,
-0.033890172839164734,
0.04980723187327385,
-0.09409697353839874,
-0.09538910537958145,
-0.06305725127458572,
0.0021724547259509563,
-0.0733242854475975,
0.021990247070789337,
0.05519028753042221,
-0.03998783230781555,
-0.041341736912727356,
-0.048819225281476974,
-0.019200261682271957,
-0.021981891244649887,
-3.310926408542841e-33,
-0.07651776820421219,
0.00624182540923357,
-0.035005562007427216,
0.05823959410190582,
-0.06981075555086136,
-0.06566695123910904,
-0.013274307362735271,
0.06179706007242203,
-0.03293309360742569,
-0.015278339385986328,
-0.06280194967985153,
-0.019064484164118767,
0.004772963467985392,
-0.04996713995933533,
-0.030078697949647903,
-0.03147391602396965,
-0.05743946507573128,
0.05429970100522041,
-0.027335038408637047,
0.013277102261781693,
0.024694137275218964,
0.03096725605428219,
-0.05208480730652809,
0.04923834651708603,
-0.0005961569841019809,
0.023891473188996315,
0.017915962263941765,
0.028108054772019386,
0.015516722574830055,
0.034881364554166794,
-0.03336600959300995,
0.014850486069917679,
-0.05613299459218979,
-0.10256796330213547,
-0.05792688578367233,
0.024695878848433495,
0.09905984252691269,
0.09088192880153656,
-0.021030789241194725,
0.07473032176494598,
0.04371502250432968,
-0.003975473809987307,
-0.07587019354104996,
0.025518184527754784,
-0.03148356452584267,
-0.07656122744083405,
-0.06667807698249817,
0.009147289209067822,
0.08177027106285095,
-0.060907695442438126,
0.02512872777879238,
-0.036824923008680344,
0.060156382620334625,
0.017895828932523727,
-0.019546139985322952,
-0.0469149574637413,
-0.01632659323513508,
0.019749674946069717,
0.02084842324256897,
0.04376249015331268,
-0.0013317144475877285,
-0.0316208079457283,
-0.08656350523233414,
-0.04861627519130707,
0.06226648390293121,
-0.041535958647727966,
-0.01123277097940445,
0.03926821053028107,
0.013611183501780033,
-0.04181020334362984,
0.07843096554279327,
-0.02707270346581936,
-0.044729508459568024,
0.05011929199099541,
0.0006392449722625315,
-0.026600545272231102,
-0.005708232056349516,
0.015170732513070107,
-0.06613600254058838,
0.030510904267430305,
0.05796133354306221,
-0.009630169719457626,
-0.0074293664656579494,
0.06096111610531807,
0.08868487924337387,
0.022722331807017326,
0.010515369474887848,
-0.02195035107433796,
-0.034164655953645706,
0.01583181507885456,
-0.030044803395867348,
0.01309956144541502,
-0.09599293768405914,
0.06478356570005417,
-0.03989062085747719,
-4.1882291412775885e-8,
-0.08372776210308075,
-0.02770041488111019,
0.013962111435830593,
0.08034354448318481,
0.00037934622378088534,
0.06148534268140793,
-0.0515906997025013,
0.0976753756403923,
-0.0069845872931182384,
-0.07862231135368347,
0.06324232369661331,
-0.06146252155303955,
-0.11442193388938904,
-0.05310404300689697,
0.02604244090616703,
0.02082064002752304,
0.010355007834732533,
0.0022750715725123882,
-0.05202037841081619,
-0.007674288470298052,
0.0199800506234169,
0.05268942192196846,
0.006431007757782936,
-0.038073427975177765,
0.028024960309267044,
-0.038052696734666824,
0.048101164400577545,
0.02974487654864788,
-0.007627506274729967,
0.11726630479097366,
-0.039545316249132156,
-0.03607931733131409,
0.030970238149166107,
-0.005014583934098482,
0.08763144910335541,
0.08228766918182373,
-0.045789241790771484,
-0.06610368192195892,
-0.01930224522948265,
0.0060151671059429646,
-0.01837853714823723,
0.06508957594633102,
-0.010409845039248466,
-0.0020753820426762104,
0.12856552004814148,
-0.05510683357715607,
-0.0029747397638857365,
-0.1015632301568985,
0.04462338611483574,
-0.029248230159282684,
0.037970710545778275,
-0.038876645267009735,
0.030890269204974174,
0.07823000848293304,
0.026156149804592133,
0.007250474765896797,
0.011773123405873775,
-0.002100440440699458,
0.03224910423159599,
0.085519939661026,
0.14484000205993652,
-0.058128684759140015,
0.046592384576797485,
0.016673481091856956
] | 0 | 0 | 0 | 0 |
|
Discovering Influential Neuron Path in Vision Transformers | https://openreview.net/forum?id=WQQyJbr5Lh | [
"Yifan Wang",
"Yifei Liu",
"Yingdong Shi",
"Changming Li",
"Anqi Pang",
"Sibei Yang",
"Jingyi Yu",
"Kan Ren"
] | Poster | Vision Transformer models exhibit immense power yet remain opaque to human understanding, posing challenges and risks for practical applications.
While prior research has attempted to demystify these models through input attribution and neuron role analysis,
there's been a notable gap in considering layer-level information and the holistic path of information flow across layers.
In this paper, we investigate the significance of influential neuron paths within vision Transformers, which is a path of neurons from the model input to output that impacts the model inference most significantly.
We first propose a joint influence measure to assess the contribution of a set of neurons to the model outcome.
And we further provide a
layer-progressive neuron locating
approach that efficiently selects the most influential neuron at each layer trying to discover the crucial neuron path from input to output within the target model.
Our experiments demonstrate the superiority of our method finding the most influential neuron path along which the information flows, over the existing baseline solutions.
Additionally, the neuron paths have illustrated that vision Transformers exhibit some specific inner working mechanism for processing the visual information within the same image category.
We further analyze the key effects of these neurons on the image classification task, showcasing that the found neuron paths have already preserved the model capability on downstream tasks, which may also shed some lights on real-world applications like model pruning.
The project website including implementation code is available at https://foundation-model-research.github.io/NeuronPath/. | Explainability, Vision Transformer, Neuron | We propose a method to identify and trace neuron path, the most influential neurons across layers in vision Transformers, revealing their internal mechanisms and enabling applications like model pruning. | 12,777 | 2503.09046 | [
-0.057392798364162445,
-0.04663418233394623,
0.039899732917547226,
0.025354167446494102,
0.11116650700569153,
-0.02308332547545433,
0.016616493463516235,
0.006510477047413588,
0.08465614169836044,
-0.03503343462944031,
-0.06635379046201706,
-0.0016151615418493748,
0.023949123919010162,
0.04904085025191307,
-0.10024267435073853,
0.02720695547759533,
0.06289294362068176,
0.11807667464017868,
-0.15210340917110443,
-0.03881597891449928,
0.08206423372030258,
-0.06486115604639053,
0.041008755564689636,
-0.03985247015953064,
0.032247696071863174,
0.028933191671967506,
0.008389058522880077,
-0.03552335128188133,
0.006778434384614229,
-0.07152789831161499,
-0.017643634229898453,
-0.06391088664531708,
-0.08904868364334106,
0.058621712028980255,
-0.0742756575345993,
0.05634630471467972,
0.005136190447956324,
0.0024072478991001844,
0.04417287930846214,
-0.042009659111499786,
0.03732994571328163,
0.05654098093509674,
0.0563078336417675,
0.016795240342617035,
0.05520779639482498,
0.027200110256671906,
0.08167385309934616,
-0.08110226690769196,
-0.08183611929416656,
-0.013071813620626926,
-0.03242287412285805,
-0.05319582298398018,
-0.023845234885811806,
0.03928648307919502,
-0.017823653295636177,
0.11055033653974533,
0.045040275901556015,
-0.03477934002876282,
-0.03697206452488899,
0.057320673018693924,
0.007006978150457144,
-0.0019755482207983732,
-0.023127242922782898,
-0.010277350433170795,
-0.051041025668382645,
-0.01771911047399044,
0.09270437806844711,
-0.005530303809791803,
-0.04487422853708267,
-0.01718280278146267,
0.06315857172012329,
0.09053236246109009,
-0.006524972151964903,
-0.04394955933094025,
0.05567857623100281,
0.00030097021954134107,
0.09553629159927368,
-0.02958037704229355,
0.021190041676163673,
-0.12982968986034393,
0.07379171997308731,
0.048706840723752975,
0.016279902309179306,
0.05867680534720421,
0.16668684780597687,
0.025159500539302826,
-0.03038669377565384,
0.00015571055701002479,
0.005224600434303284,
0.01860145665705204,
-0.08772812783718109,
-0.10898850113153458,
-0.024019325152039528,
-0.08222415298223495,
0.06680865585803986,
-0.044498443603515625,
-0.03744170442223549,
-0.06948544830083847,
-0.007268480956554413,
-0.002054181881248951,
-0.09561216086149216,
-0.0001668436307227239,
0.03728487342596054,
-0.05633983016014099,
0.0772935003042221,
0.0009414450032636523,
0.04044070839881897,
0.0865083858370781,
0.03876392915844917,
0.021791525185108185,
0.03405553847551346,
-0.01666361466050148,
-0.0058291698805987835,
0.040276650339365005,
0.08548133820295334,
-0.00028320521232672036,
-0.01828395389020443,
0.057766299694776535,
-0.035676561295986176,
0.033077336847782135,
-0.009996166452765465,
0.07324275374412537,
-0.020557062700390816,
-0.010059228166937828,
-0.005338133778423071,
-0.006124940235167742,
-0.056032609194517136,
5.954445971103662e-34,
-0.0382535345852375,
-0.03440698981285095,
0.08390921354293823,
0.06139032542705536,
0.07449359446763992,
0.0013404671335592866,
-0.006846928969025612,
0.035275574773550034,
0.0099928705021739,
0.034319210797548294,
-0.06672952324151993,
-0.02271220088005066,
-0.03941876441240311,
0.1128244400024414,
0.011620713397860527,
-0.03880002722144127,
-0.04678109288215637,
0.02510887011885643,
-0.07518412172794342,
-0.009802171029150486,
0.02733183652162552,
-0.05078379064798355,
-0.023234140127897263,
-0.10736437886953354,
-0.004315690603107214,
-0.07909876853227615,
0.0035410146228969097,
0.008423805236816406,
-0.021012378856539726,
-0.005717119667679071,
-0.053790293633937836,
0.03830043971538544,
0.06388238817453384,
-0.06305814534425735,
0.04270650073885918,
-0.0120597705245018,
0.012192770838737488,
0.025153040885925293,
0.06038465350866318,
0.03860294073820114,
-0.02831823192536831,
0.0483466275036335,
-0.07347626239061356,
-0.01849621906876564,
-0.0617760568857193,
0.021617194637656212,
-0.04548779875040054,
0.08778946846723557,
-0.056579988449811935,
-0.045320603996515274,
0.03709150850772858,
-0.006482552271336317,
0.02026529051363468,
-0.048093970865011215,
-0.0357498936355114,
-0.056091781705617905,
-0.009028078056871891,
-0.007220334839075804,
0.026441557332873344,
-0.0021190918050706387,
0.01171752531081438,
0.04555557668209076,
-0.07123500853776932,
0.11001204699277878,
0.06806283444166183,
0.005582032725214958,
0.0633389800786972,
0.017712941393256187,
0.014423730783164501,
0.045212797820568085,
-0.07939840853214264,
0.06590765714645386,
-0.04420772194862366,
-0.12171708047389984,
0.014703282155096531,
0.0022749004419893026,
-0.037746936082839966,
-0.033207543194293976,
-0.020268401131033897,
0.03516390174627304,
-0.04337688162922859,
-0.07956212759017944,
0.0006450550863519311,
0.0003376878739800304,
0.020940981805324554,
0.05193238705396652,
0.03864569962024689,
-0.05718480050563812,
-0.055977340787649155,
0.0026724664494395256,
0.026990680024027824,
-0.03533243015408516,
-0.032577190548181534,
-0.04403591901063919,
-0.01621571183204651,
-1.2409057441642514e-33,
-0.028034323826432228,
0.05179684981703758,
0.05740226432681084,
-0.000418360170442611,
-0.01191666629165411,
-0.04999828711152077,
-0.0356987901031971,
-0.03176087141036987,
0.0010199014795944095,
-0.012778460048139095,
-0.04304100573062897,
0.0009957622969523072,
-0.0298309288918972,
-0.005448073614388704,
0.056469839066267014,
-0.027145467698574066,
-0.03006780706346035,
0.010618466883897781,
0.018951961770653725,
-0.03512977808713913,
0.006525429431349039,
0.11122890561819077,
-0.1016172543168068,
-0.023125454783439636,
-0.057319123297929764,
0.0793686956167221,
-0.0390910804271698,
0.09044480323791504,
0.027667490765452385,
-0.026200907304883003,
-0.04947105050086975,
-0.023880958557128906,
-0.029149789363145828,
-0.08479666709899902,
-0.06885124742984772,
0.08728370070457458,
-0.02795563079416752,
-0.04150869697332382,
-0.028163723647594452,
0.07653085142374039,
0.02644256316125393,
-0.0696892961859703,
-0.026853561401367188,
0.01796313188970089,
0.0017529993783682585,
-0.06371793895959854,
0.03047974407672882,
0.02209598757326603,
-0.044654376804828644,
0.032756246626377106,
-0.020996585488319397,
0.0505794957280159,
-0.11926091462373734,
0.034771524369716644,
-0.03303410857915878,
0.051767878234386444,
-0.046418845653533936,
0.011595183052122593,
0.1503729671239853,
-0.027439983561635017,
-0.03544872626662254,
0.028400994837284088,
0.012402849271893501,
-0.04492798447608948,
-0.023614700883626938,
0.011252246797084808,
-0.037192389369010925,
-0.006694633048027754,
0.05693233758211136,
0.02403075620532036,
0.003567482577636838,
-0.000307335052639246,
0.013017358258366585,
-0.008538144640624523,
0.003042008262127638,
-0.05241129919886589,
0.0057219346053898335,
0.005398509558290243,
0.026526981964707375,
0.009818720631301403,
-0.022114325314760208,
-0.055410388857126236,
0.010000918060541153,
0.0034540907945483923,
0.0925026684999466,
0.09934214502573013,
0.014094406738877296,
-0.006813523825258017,
0.09523556381464005,
-0.017698273062705994,
-0.05280529707670212,
-0.007357825990766287,
-0.0761546790599823,
-0.0069773090071976185,
-0.07734086364507675,
-4.355989347004652e-8,
-0.07113966345787048,
0.06681989133358002,
0.020642897114157677,
0.008883848786354065,
0.03404155746102333,
-0.04038054868578911,
0.05252731963992119,
0.14521217346191406,
-0.04865783080458641,
0.012260986492037773,
0.07647141069173813,
0.007939359173178673,
-0.015471747145056725,
0.0007694217492826283,
0.03755791485309601,
0.04516930133104324,
-0.0037513498682528734,
0.012728506699204445,
-0.02184927836060524,
-0.04292766749858856,
-0.016788501292467117,
-0.026942601427435875,
-0.010875503532588482,
0.0034205266274511814,
0.0011456453939899802,
-0.08080564439296722,
-0.09540316462516785,
0.057917520403862,
0.030208855867385864,
0.015994103625416756,
0.05937217175960541,
0.05426469445228577,
0.021098896861076355,
-0.010927993804216385,
0.04199223592877388,
0.06592652201652527,
-0.012942441739141941,
-0.007860131561756134,
-0.0050047412514686584,
-0.0077873216941952705,
0.004571013618260622,
0.005545144435018301,
0.051383666694164276,
0.11003891378641129,
-0.009561228565871716,
-0.003988161217421293,
0.0729438066482544,
-0.09842962771654129,
-0.021166617050766945,
0.0027307935524731874,
-0.056294821202754974,
-0.02156480774283409,
-0.021677548065781593,
0.1231352686882019,
0.014922748319804668,
-0.05599842965602875,
-0.020324308425188065,
-0.10859908908605576,
-0.04373723641037941,
0.0900547057390213,
0.01346601266413927,
0.05695139244198799,
-0.020076539367437363,
0.005724604707211256
] | 0 | 0 | 0 | 0 |
|
When narrower is better: the narrow width limit of Bayesian parallel branching neural networks | https://openreview.net/forum?id=CkUHtnyhpY | [
"Zechen Zhang",
"Haim Sompolinsky"
] | Poster | The infinite width limit of random neural networks is known to result in Neural Networks as Gaussian Process (NNGP) (Lee et al. (2018)), characterized by task-independent kernels. It is widely accepted that larger network widths contribute to improved generalization (Park et al. (2019)). However, this work challenges this notion by investigating the narrow width limit of the Bayesian Parallel Branching
Neural Network (BPB-NN), an architecture that resembles neural networks with residual blocks. We demonstrate that when the width of a BPB-NN is significantly smaller compared to the number of training examples, each branch exhibits more robust learning due to a symmetry breaking of branches in kernel renormalization. Surprisingly, the performance of a BPB-NN in the narrow width limit is generally superior to or comparable to that achieved in the wide width limit in bias-limited scenarios. Furthermore, the readout norms of each branch in the narrow width limit are mostly independent of the architectural hyperparameters but generally reflective of the nature of the data. We demonstrate such phenomenon primarily in the branching graph neural networks, where each branch represents a different order of convolutions of the graph; we also extend the results to other more general architectures such as the residual-MLP and demonstrate that the narrow width effect is a general feature of the branching networks. Our results characterize a newly defined narrow-width regime for parallel branching networks in general. | Bayesian Networks, Gaussian Process, Kernel Renormalization, Graph Neural Networks, Residual Network, Theory of Generalization | We use statistical learning theory to understand parallel graph neural networks, and find a narrow width limit where networks generalize better with small width. | 12,773 | 2407.18807 | [
-0.050411444157361984,
-0.12044142931699753,
0.04240997135639191,
-0.0009813208598643541,
0.018326671794056892,
0.006854096427559853,
-0.06819336861371994,
0.016857923939824104,
0.00032833029399625957,
-0.07773367315530777,
-0.035231124609708786,
0.05543511360883713,
-0.03625573217868805,
-0.02453211322426796,
-0.051487069576978683,
-0.05053579434752464,
0.019264506176114082,
0.08757936209440231,
-0.1121150478720665,
-0.027032705023884773,
0.03354958817362785,
0.026671061292290688,
-0.015008518472313881,
-0.00592005392536521,
0.009216583333909512,
-0.019299067556858063,
0.0076246862299740314,
-0.06987590342760086,
0.0015189722180366516,
-0.046914469450712204,
0.08280463516712189,
-0.04068497568368912,
-0.08899983763694763,
-0.0070246863178908825,
0.001406235620379448,
-0.02281365729868412,
-0.010630580596625805,
-0.02335553802549839,
0.05188347399234772,
-0.0025999906938523054,
0.047870200127363205,
0.0456160269677639,
0.05539305880665779,
0.09486037492752075,
0.10067477077245712,
0.03301974758505821,
-0.006229973863810301,
-0.12662221491336823,
0.056807685643434525,
-0.030708536505699158,
-0.04302932694554329,
-0.06284714490175247,
-0.015296967700123787,
0.08918122202157974,
0.012395858764648438,
-0.02566971443593502,
-0.03468769043684006,
0.01418424304574728,
0.02282506600022316,
0.10951091349124908,
-0.025138821452856064,
-0.10156014561653137,
0.00008654133853269741,
-0.05827568843960762,
0.0043892646208405495,
0.05859322473406792,
0.01259771827608347,
0.019507698714733124,
0.025915324687957764,
0.01147266011685133,
0.042163360863924026,
0.08523746579885483,
-0.12127839028835297,
0.04946448653936386,
0.02826802060008049,
0.015765661373734474,
0.03901924192905426,
0.09320631623268127,
0.002759520197287202,
-0.10851367563009262,
-0.07071579247713089,
0.0031335260719060898,
0.01656251959502697,
-0.039599839597940445,
0.03819511830806732,
-0.03071669489145279,
-0.06462061405181885,
0.08274032175540924,
-0.005941972602158785,
0.040389254689216614,
-0.02115040458738804,
0.016858041286468506,
-0.001554084476083517,
-0.0406423956155777,
-0.008081099949777126,
-0.027746668085455894,
-0.06379912048578262,
-0.040924590080976486,
-0.041412174701690674,
0.07202183455228806,
0.042643290013074875,
0.04337190091609955,
0.14369161427021027,
-0.015025715343654156,
0.05223916098475456,
0.0002798142668325454,
0.05968186631798744,
0.02606739103794098,
0.08857543766498566,
-0.14371560513973236,
-0.015474260784685612,
0.08675070852041245,
0.04228704050183296,
0.023389769718050957,
-0.030910249799489975,
-0.03652138262987137,
-0.02356847934424877,
-0.06757114827632904,
0.004241099581122398,
0.12882550060749054,
-0.069056436419487,
0.05578240007162094,
-0.08014599233865738,
0.006896096281707287,
0.006586104165762663,
0.02671053446829319,
-0.09553634375333786,
3.463220150557515e-33,
0.04369749873876572,
0.034688398241996765,
0.07400462031364441,
-0.018468476831912994,
0.09008372575044632,
0.06790373474359512,
-0.024470055475831032,
-0.06310147047042847,
-0.02476395107805729,
0.0083968136459589,
-0.04810260608792305,
-0.016638264060020447,
-0.08115414530038834,
0.035982050001621246,
0.054786741733551025,
-0.00680822366848588,
-0.05723842605948448,
0.06281960010528564,
0.022334780544042587,
0.004361727274954319,
0.047444213181734085,
-0.07507417351007462,
-0.0014983690343797207,
0.019043060019612312,
0.0018170548137277365,
-0.06483184546232224,
0.013317561708390713,
0.014328747056424618,
-0.09327681362628937,
0.02731984667479992,
-0.08506269752979279,
0.012982211075723171,
-0.000624666572548449,
0.051098838448524475,
0.031423695385456085,
-0.07596281915903091,
0.040859024971723557,
-0.04402970150113106,
0.01254135463386774,
-0.07290194928646088,
-0.04365427792072296,
-0.013988060876727104,
0.09357070922851562,
0.008954587392508984,
-0.07317283749580383,
-0.05500999465584755,
0.01911802589893341,
-0.05723767355084419,
-0.004357750993221998,
-0.06173902750015259,
0.02804988995194435,
0.02603044919669628,
0.007387327961623669,
-0.022896604612469673,
0.08680544793605804,
0.03057088516652584,
0.03556986153125763,
0.10302633792161942,
0.022810259833931923,
0.08632712811231613,
0.0225976649671793,
0.06283462047576904,
-0.06442966312170029,
0.045060597360134125,
-0.033975403755903244,
0.06688883900642395,
-0.02850872278213501,
0.05315549671649933,
0.057024531066417694,
-0.07226209342479706,
-0.05477381870150566,
-0.03291730955243111,
0.0035737722646445036,
-0.017759021371603012,
0.020481785759329796,
0.019704466685652733,
-0.012978271581232548,
-0.0850682333111763,
-0.00037918088492006063,
0.022422365844249725,
-0.13348127901554108,
0.05449596419930458,
-0.025127818807959557,
-0.04814135655760765,
-0.03027387149631977,
-0.02977285161614418,
0.045190729200839996,
-0.01725955680012703,
-0.044516339898109436,
-0.016686666756868362,
-0.05739087238907814,
-0.0387415885925293,
0.01718957908451557,
-0.011042466387152672,
-0.0553290955722332,
-3.2353700397758e-33,
-0.05915188789367676,
0.007278612814843655,
-0.06319954991340637,
0.05446594953536987,
-0.03531072288751602,
-0.04271411895751953,
-0.0160182174295187,
-0.036990903317928314,
-0.01667814329266548,
-0.06610628217458725,
0.019127463921904564,
0.024231910705566406,
0.023854224011301994,
-0.0017897351644933224,
0.022439932450652122,
-0.02947954088449478,
0.013843361288309097,
0.0808766633272171,
0.082903653383255,
-0.09592865407466888,
0.021780412644147873,
0.05855090171098709,
-0.021623555570840836,
0.04472607746720314,
0.014134558849036694,
0.029317723587155342,
-0.06751459091901779,
0.10328359156847,
0.05516940355300903,
-0.011850508861243725,
-0.06423906981945038,
-0.00409832876175642,
-0.08980299532413483,
0.0018772183684632182,
-0.004426003899425268,
0.0014803385129198432,
0.05263319984078407,
0.03131021186709404,
0.038248684257268906,
0.0420205220580101,
0.023773914203047752,
-0.04285261034965515,
-0.05701896920800209,
-0.019253583624958992,
-0.0260174423456192,
0.02716965600848198,
0.03761561959981918,
-0.044653017073869705,
-0.09037839621305466,
0.029872028157114983,
-0.017239227890968323,
0.11010821163654327,
-0.045713700354099274,
0.07211361825466156,
-0.03210638463497162,
-0.047517355531454086,
0.02782602049410343,
0.02255842834711075,
0.07181820273399353,
-0.001114496379159391,
-0.047616146504879,
-0.05851922184228897,
0.007975608110427856,
-0.051636990159749985,
-0.02472566068172455,
0.03227828070521355,
0.005462341010570526,
0.044308144599199295,
0.08309691399335861,
0.038887251168489456,
-0.01336726639419794,
0.028939904645085335,
0.0008453888585790992,
0.11294218897819519,
-0.046970006078481674,
0.040417686104774475,
0.02474895678460598,
-0.0761760026216507,
-0.029263073578476906,
-0.009241806343197823,
-0.0338267982006073,
0.009521566331386566,
-0.024429844692349434,
0.05727587640285492,
0.10719141364097595,
0.030630778521299362,
0.03991997241973877,
0.03969081863760948,
0.061668455600738525,
-0.0634632408618927,
0.0026785670779645443,
0.027971118688583374,
-0.048256766051054,
0.05107404291629791,
-0.09650634974241257,
-4.5551672656074516e-8,
-0.053148556500673294,
0.08721807599067688,
-0.006614543031901121,
0.03982169181108475,
0.1399487406015396,
-0.08619965612888336,
0.016699565574526787,
0.07974430173635483,
-0.05689089000225067,
0.013984369114041328,
0.06206826865673065,
-0.014026588760316372,
-0.03793568164110184,
-0.05351237580180168,
0.05336301401257515,
0.04817057028412819,
-0.004927546717226505,
0.00924922525882721,
0.04370063543319702,
-0.0004221092676743865,
0.012116851285099983,
0.00432165851816535,
-0.0035332830157130957,
0.012345151044428349,
-0.009699842892587185,
-0.11981398612260818,
-0.03410749509930611,
0.07376270741224289,
0.040787871927022934,
0.01478993147611618,
0.027622444555163383,
0.046688299626111984,
-0.03594925254583359,
0.0006267457501962781,
0.008203866891562939,
0.03784811124205589,
0.02083956077694893,
0.04268581047654152,
0.03555149957537651,
0.017609350383281708,
-0.03025166504085064,
-0.044195756316185,
0.02659277804195881,
0.019148696213960648,
-0.007196637801826,
-0.009085084311664104,
0.06878636032342911,
0.024127770215272903,
0.041119709610939026,
-0.028129156678915024,
0.044636547565460205,
0.020465511828660965,
0.01421989407390356,
0.014950522221624851,
-0.019264843314886093,
0.030932217836380005,
0.02534835785627365,
-0.09903725981712341,
-0.05449661239981651,
0.0745316669344902,
-0.017396362498402596,
-0.03409887105226517,
-0.09538830071687698,
0.015556965954601765
] | 0 | 0 | 0 | 0 |
|
Directional Gradient Projection for Robust Fine-Tuning of Foundation Models | https://openreview.net/forum?id=goBaGHLAdP | [
"Chengyue Huang",
"Junjiao Tian",
"Brisa Maneechotesuwan",
"Shivang Chopra",
"Zsolt Kira"
] | Poster | Robust fine-tuning aims to adapt large foundation models to downstream tasks while preserving their robustness to distribution shifts. Existing methods primarily focus on constraining and projecting current model towards the pre-trained initialization based on the magnitudes between fine-tuned and pre-trained weights, which often require extensive hyper-parameter tuning and can sometimes result in underfitting. In this work, we propose $\textbf{Di}$rectional $\textbf{Gra}$dient $\textbf{P}$rojection (DiGraP), a novel layer-wise trainable method that incorporates directional information from gradients to bridge regularization and multi-objective optimization. Besides demonstrating our method on image classification, as another contribution we generalize this area to the multi-modal evaluation settings for robust fine-tuning. Specifically, we first bridge the uni-modal and multi-modal gap by performing analysis on Image Classification reformulated Visual Question Answering (VQA) benchmarks and further categorize ten out-of-distribution (OOD) VQA datasets by distribution shift types and degree (i.e. near versus far OOD). Experimental results show that DiGraP consistently outperforms existing baselines across Image Classfication and VQA tasks with discriminative and generative backbones, improving both in-distribution (ID) generalization and OOD robustness. | Fine-tuning, transfer learning, foundation models, robustness, visual question answering | We present a novel regularization technique for robust fine-tuning of large foundation models. | 12,769 | 2502.15895 | [
-0.04263325780630112,
-0.11436345428228378,
0.008543248288333416,
0.026321038603782654,
0.07656031847000122,
0.06892500072717667,
0.02973243035376072,
-0.02685513161122799,
-0.06463801860809326,
-0.04108557850122452,
-0.044588733464479446,
-0.04934808984398842,
-0.0018577580340206623,
0.05739769712090492,
-0.046977076679468155,
-0.0031950261909514666,
0.1102452278137207,
0.09398399293422699,
-0.10045217722654343,
0.010509363375604153,
0.0729474425315857,
0.054246049374341965,
0.04154583066701889,
-0.028105638921260834,
0.03221673518419266,
-0.040006428956985474,
0.007841033861041069,
-0.0206459891051054,
0.044188447296619415,
-0.06588560342788696,
-0.0006858436390757561,
-0.015381879173219204,
-0.1199292317032814,
0.030194656923413277,
-0.07892052084207535,
0.11330398917198181,
-0.019410133361816406,
0.08542734384536743,
-0.0001262590376427397,
0.040475089102983475,
-0.0013617771910503507,
0.012596228159964085,
-0.006916503421962261,
-0.00820072926580906,
0.07267194986343384,
-0.030735790729522705,
0.02464030124247074,
-0.04617768153548241,
0.026160074397921562,
0.003864325350150466,
0.007512201089411974,
-0.031194768846035004,
-0.04716143012046814,
0.05851834639906883,
0.0018344064010307193,
0.01125439815223217,
0.07760791480541229,
-0.022383635863661766,
0.020105287432670593,
-0.0162312351167202,
-0.008598717860877514,
-0.08699551969766617,
-0.01503794640302658,
-0.024322401732206345,
0.015186462551355362,
-0.06504612416028976,
0.004470388870686293,
-0.023687126114964485,
0.02315591275691986,
-0.051538918167352676,
-0.0708245262503624,
0.06528068333864212,
-0.01906665414571762,
-0.0060709696263074875,
0.023880554363131523,
0.08543535321950912,
0.06851521879434586,
0.03748713433742523,
0.05809004604816437,
-0.07218991965055466,
0.08119580149650574,
0.05812351778149605,
0.03542736917734146,
0.0023137242533266544,
0.15234743058681488,
0.029826810583472252,
-0.03393455222249031,
0.0544915534555912,
0.04926171898841858,
0.01825757510960102,
-0.05746132507920265,
-0.012618239969015121,
-0.00363453128375113,
0.056865084916353226,
0.016804510727524757,
-0.02699616551399231,
0.03394808992743492,
-0.16489453613758087,
-0.04978121444582939,
0.1430739164352417,
0.024073639884591103,
0.062275201082229614,
-0.02688497118651867,
0.034618083387613297,
0.0385122187435627,
-0.012634159065783024,
0.04643562436103821,
0.07701856642961502,
-0.026592964306473732,
-0.06417027860879898,
0.041417285799980164,
0.021169345825910568,
0.053138617426157,
-0.027718782424926758,
0.007689889054745436,
-0.04218747094273567,
-0.017607547342777252,
0.001444146502763033,
0.011972525157034397,
-0.003564200596883893,
-0.05346621572971344,
0.005124650429934263,
-0.03958308696746826,
-0.009448212571442127,
0.04027082771062851,
-0.004603030160069466,
-0.06308045983314514,
2.8222003058510742e-33,
0.05510735139250755,
0.06677374988794327,
0.047255344688892365,
0.01785425655543804,
0.020276008173823357,
-0.030337419360876083,
-0.007792341988533735,
0.01948295533657074,
-0.025447512045502663,
0.06037713587284088,
-0.02603127993643284,
0.028690079227089882,
-0.09347426146268845,
0.0989152193069458,
0.022603461518883705,
-0.042846888303756714,
-0.0333823636174202,
-0.006336341146379709,
-0.035979975014925,
0.014685316942632198,
-0.004887610673904419,
-0.08493983000516891,
0.006993274670094252,
-0.06565588712692261,
0.05783670395612717,
-0.051823776215314865,
0.10890419781208038,
-0.05755743756890297,
-0.12175768613815308,
0.011645462363958359,
-0.11841081082820892,
-0.01860867254436016,
0.04404909536242485,
0.005185924004763365,
-0.02350231632590294,
-0.06937254965305328,
-0.02380959689617157,
-0.022592000663280487,
-0.007310977205634117,
-0.01135566271841526,
-0.049106158316135406,
0.06889782845973969,
0.03529210016131401,
-0.0007628448656760156,
-0.03917960822582245,
-0.027785630896687508,
-0.013974983245134354,
0.02433406561613083,
-0.0022409092634916306,
-0.08202111721038818,
-0.0216023251414299,
-0.001469407812692225,
-0.04526834189891815,
-0.03466491028666496,
-0.08313729614019394,
-0.028509723022580147,
0.04069499298930168,
0.037873320281505585,
0.03756948560476303,
-0.0022047257516533136,
0.004681434947997332,
0.07220879942178726,
-0.021380143240094185,
0.10692479461431503,
-0.026859816163778305,
0.009082486853003502,
-0.02767210453748703,
0.023839203640818596,
-0.010369569063186646,
0.05329661816358566,
-0.07531926035881042,
0.03444333001971245,
0.0192530807107687,
-0.06900336593389511,
0.0512775257229805,
0.05211109295487404,
0.05809982120990753,
-0.02340128645300865,
0.05226173624396324,
0.0421869270503521,
0.015067098662257195,
0.12012390047311783,
0.020301971584558487,
-0.06033400446176529,
-0.05940646678209305,
-0.055882249027490616,
0.016828741878271103,
-0.02999163046479225,
-0.014314324595034122,
-0.04338698834180832,
-0.061670586466789246,
0.05388098210096359,
0.0018354739295318723,
-0.01554032787680626,
-0.011530349962413311,
-4.562918258296797e-34,
0.015374981798231602,
0.029272327199578285,
-0.059717871248722076,
0.002811537589877844,
0.02334313467144966,
-0.017696239054203033,
-0.005622477270662785,
0.07210768759250641,
0.01114903949201107,
-0.033787697553634644,
-0.023527046665549278,
-0.0047736456617712975,
-0.03088008612394333,
-0.000573282188270241,
0.02549837902188301,
-0.015354027040302753,
-0.07043661177158356,
0.049072910100221634,
-0.025863833725452423,
0.0074994731694459915,
0.05763997510075569,
0.0771358385682106,
-0.11918506026268005,
0.037126824259757996,
-0.03490182384848595,
0.06848111003637314,
-0.08885686099529266,
0.07124295085668564,
-0.050700169056653976,
0.007100458722561598,
-0.018816715106368065,
-0.009770249016582966,
0.019564440473914146,
-0.011539275757968426,
-0.05683394521474838,
0.06486771255731583,
-0.04496217891573906,
-0.08383256942033768,
-0.008830329403281212,
0.13477040827274323,
0.0001714673708193004,
0.05588803067803383,
-0.08133558928966522,
0.03865164518356323,
0.00819449033588171,
-0.04087013006210327,
-0.01964263990521431,
-0.08716196566820145,
0.03286077827215195,
0.05163855850696564,
-0.020642701536417007,
0.008656919002532959,
-0.0971783995628357,
0.04761140048503876,
-0.0017156947869807482,
-0.07068239152431488,
-0.07478435337543488,
-0.01713869720697403,
0.11474058032035828,
0.032403018325567245,
-0.019351543858647346,
0.010532697662711143,
-0.010599072091281414,
-0.04915820062160492,
-0.01879000850021839,
-0.06374520063400269,
-0.03133796527981758,
-0.011807321570813656,
0.03800288215279579,
0.08378024399280548,
-0.014607221819460392,
0.024408865720033646,
0.05845967307686806,
-0.008496131747961044,
-0.01940041035413742,
-0.019632499665021896,
0.06664299964904785,
0.016051165759563446,
0.03536954149603844,
-0.002643285086378455,
0.0754365473985672,
-0.04239962249994278,
-0.007666998077183962,
0.10431277006864548,
0.09394379705190659,
0.1057858020067215,
0.05525908246636391,
0.04226716235280037,
0.12582752108573914,
-0.031088532879948616,
-0.0013995693298056722,
-0.020682523027062416,
0.030468041077256203,
0.0017194697866216302,
-0.01383264735341072,
-4.8563652654820544e-8,
-0.06561944633722305,
0.07824204862117767,
-0.04347701370716095,
0.0009268914582207799,
-0.026304159313440323,
-0.08283741027116776,
0.042244669049978256,
0.05393388867378235,
0.007909979671239853,
-0.09612669795751572,
0.06823226809501648,
0.02477075345814228,
-0.034846313297748566,
-0.037142638117074966,
-0.05584639310836792,
0.08912790566682816,
-0.019870810210704803,
0.12462436407804489,
0.00040166813414543867,
-0.001770928967744112,
0.003337954171001911,
0.05006905645132065,
0.03274969756603241,
0.002004562644287944,
0.05874637886881828,
-0.08545676618814468,
-0.044596824795007706,
0.011370801366865635,
0.08095722645521164,
0.013913360424339771,
-0.009892675094306469,
0.006086205597966909,
0.018959319218993187,
-0.046898726373910904,
-0.035919297486543655,
0.15388698875904083,
0.004569462966173887,
0.006352890748530626,
-0.018376251682639122,
-0.010493509471416473,
0.0040000127628445625,
0.062350668013095856,
-0.03607582300901413,
0.008387990295886993,
-0.009179714135825634,
0.024529600515961647,
-0.03425003960728645,
-0.05018354579806328,
-0.03902524709701538,
-0.08637608587741852,
0.04446672275662422,
0.016815394163131714,
0.024883044883608818,
0.06574780493974686,
0.04426956549286842,
-0.01806972362101078,
0.0005437296349555254,
-0.10220302641391754,
-0.008701147511601448,
0.033164627850055695,
-0.0018693897873163223,
-0.022918231785297394,
-0.03580142557621002,
-0.009168003685772419
] | 0 | 0 | 0 | 0 |
|
Reward Learning from Multiple Feedback Types | https://openreview.net/forum?id=9Ieq8jQNAl | [
"Yannick Metz",
"Andras Geiszl",
"Raphaël Baur",
"Mennatallah El-Assady"
] | Poster | Learning rewards from preference feedback has become an important tool in the alignment of agentic models. Preference-based feedback, often implemented as a binary comparison between multiple completions, is an established method to acquire large-scale human feedback. However, human feedback in other contexts is often much more diverse. Such diverse feedback can better support the goals of a human annotator, and the simultaneous use of multiple sources might be mutually informative for the learning process or carry type-dependent biases for the reward learning process.
Despite these potential benefits, learning from different feedback types has yet to be explored extensively.
In this paper, we bridge this gap by enabling experimentation and evaluating multi-type feedback in a wide set of environments. We present a process to generate high-quality simulated feedback of six different types. Then, we implement reward models and downstream RL training for all six feedback types.
Based on the simulated feedback, we investigate the use of types of feedback across ten RL environments and compare them to pure preference-based baselines. We show empirically that diverse types of feedback can be utilized and lead to strong reward modeling performance. This work is the first strong indicator of the potential of multi-type feedback for RLHF. | Reinforcement Learning, RLHF, Machine Learning, Multi-Type Feedback | We present a simulation framework and reward model implementations for diverse feedback of multiple types, and show their effectiveness | 12,768 | 2502.21038 | [
-0.03134053945541382,
-0.11993323266506195,
-0.02414754591882229,
-0.017625005915760994,
0.056298717856407166,
0.06159563735127449,
0.010764176957309246,
-0.030169829726219177,
0.03480086103081703,
-0.030177518725395203,
-0.1033487617969513,
-0.09124315530061722,
0.047518689185380936,
0.09271829575300217,
0.06646735221147537,
0.01388043537735939,
0.10950041562318802,
0.03528532385826111,
-0.04033007472753525,
-0.008982471190392971,
0.038057971745729446,
-0.003990331664681435,
0.0646047368645668,
-0.023221390321850777,
-0.08037527650594711,
-0.049601756036281586,
-0.022623851895332336,
-0.02899070829153061,
-0.007394063286483288,
-0.015072209760546684,
0.040213100612163544,
0.016749687492847443,
-0.007394061423838139,
-0.06932442635297775,
-0.030971333384513855,
0.045293889939785004,
-0.14688733220100403,
-0.000670968322083354,
-0.004786182660609484,
0.01780105009675026,
-0.04595215246081352,
0.006544339004904032,
0.0584266260266304,
0.0024941444862633944,
0.031400296837091446,
-0.03825383633375168,
-0.024421226233243942,
-0.06153399869799614,
0.021949520334601402,
-0.001994046615436673,
-0.03932751715183258,
-0.03387119993567467,
0.009642156772315502,
0.03636058419942856,
-0.026762830093503,
0.0014192919479683042,
-0.01773274503648281,
0.033474188297986984,
-0.027882283553481102,
-0.08406081050634384,
-0.032854899764060974,
-0.07859178632497787,
-0.07907728105783463,
-0.02146788127720356,
-0.028147611767053604,
-0.03713337704539299,
-0.07034882158041,
0.04179680347442627,
-0.02607281506061554,
-0.052570369094610214,
-0.0318392813205719,
0.01602729596197605,
0.01062727626413107,
-0.06782849133014679,
0.08393725007772446,
0.06572854518890381,
-0.010451151058077812,
-0.01490186806768179,
0.0006423209561035037,
-0.027061298489570618,
-0.049893539398908615,
0.009898349642753601,
0.00351765681989491,
-0.00844774954020977,
0.082117460668087,
-0.055646318942308426,
0.06861162185668945,
0.039607610553503036,
0.06537676602602005,
0.031753744930028915,
-0.03377320244908333,
0.06679154187440872,
0.05815340578556061,
-0.003440329572185874,
-0.014682866632938385,
0.11057203263044357,
0.010927107185125351,
-0.08948890119791031,
-0.017797859385609627,
0.1209363117814064,
0.03303888812661171,
0.08381559699773788,
-0.02819494903087616,
0.022908346727490425,
0.05828037112951279,
-0.022471150383353233,
-0.0004024504160042852,
0.0911746546626091,
0.05770735815167427,
-0.17509587109088898,
-0.046667419373989105,
0.03872167319059372,
-0.04053759574890137,
0.10098633170127869,
-0.012636319734156132,
-0.009160846471786499,
0.02129025012254715,
-0.024702101945877075,
0.0047618318349123,
0.06528466939926147,
0.04034064710140228,
-0.0951453447341919,
0.07927203178405762,
-0.024005435407161713,
0.06011714041233063,
0.030308343470096588,
-0.03386126086115837,
3.7794476000538956e-33,
0.05172499641776085,
0.08093922585248947,
0.0775633230805397,
0.060404494404792786,
0.048875611275434494,
0.016680801287293434,
0.003342050826177001,
0.021450011059641838,
-0.07937045395374298,
-0.053336501121520996,
0.053797245025634766,
0.035747379064559937,
-0.10622810572385788,
0.12137890607118607,
0.024007601663470268,
-0.00197879271581769,
-0.04756010323762894,
0.044881004840135574,
-0.03607306629419327,
0.0008862853283062577,
0.077479787170887,
0.05688805878162384,
0.014326618053019047,
-0.0039862277917563915,
0.03927319124341011,
0.01509147696197033,
0.022134846076369286,
-0.004923861473798752,
-0.08871070295572281,
-0.037254128605127335,
-0.003133595222607255,
-0.026515966281294823,
0.06196557357907295,
0.012194601818919182,
-0.08065026998519897,
0.06801937520503998,
-0.03639203682541847,
-0.020517904311418533,
0.023775184527039528,
-0.03926720470190048,
0.04715556278824806,
0.004184854682534933,
0.010275913402438164,
0.027086948975920677,
-0.0676034539937973,
-0.03715356066823006,
0.011417522095143795,
-0.007504348177462816,
-0.07361981272697449,
-0.002516885520890355,
0.04735207185149193,
0.016385294497013092,
0.08408850431442261,
-0.014109100215137005,
-0.018109261989593506,
0.007508448325097561,
-0.016583416610956192,
0.05513970926403999,
0.017384426668286324,
-0.017367932945489883,
-0.06199568137526512,
0.029092149809002876,
0.02887096256017685,
-0.04474256932735443,
-0.010812449268996716,
0.059235334396362305,
0.03387514129281044,
0.018411174416542053,
0.09099216759204865,
0.008433257229626179,
-0.03753577172756195,
0.08564566820859909,
-0.03573803976178169,
-0.06356266885995865,
0.058060601353645325,
-0.016904212534427643,
-0.01955595053732395,
-0.12085109949111938,
0.07767289876937866,
-0.02717851661145687,
-0.003161419415846467,
-0.013144112192094326,
-0.006402615923434496,
-0.004140916280448437,
-0.021840805187821388,
-0.08489762991666794,
-0.031384821981191635,
-0.11783183366060257,
0.058687202632427216,
0.010106271132826805,
-0.08121824264526367,
0.02532823570072651,
-0.04183105379343033,
0.018765807151794434,
0.027747107669711113,
-2.8101896923215475e-33,
-0.03693140670657158,
-0.064567431807518,
-0.0421321876347065,
0.03209412470459938,
-0.01824995130300522,
-0.045529987663030624,
-0.03425782173871994,
-0.050283174961805344,
-0.05114986374974251,
-0.02491205558180809,
-0.004781622905284166,
-0.01576288789510727,
0.027060629799962044,
0.007763614412397146,
-0.02428031526505947,
-0.052593912929296494,
-0.028140518814325333,
0.032781731337308884,
0.08723004907369614,
0.01264115422964096,
0.04258701205253601,
0.09602712839841843,
-0.0569949671626091,
0.07524310797452927,
0.0028385219629853964,
0.08917628973722458,
0.022774310782551765,
0.04297439381480217,
0.04224639758467674,
0.012110651470720768,
-0.03846045583486557,
-0.04154833406209946,
-0.013536744751036167,
-0.010448592714965343,
0.02175929956138134,
0.05741452798247337,
0.008184653706848621,
-0.003600975265726447,
-0.06812464445829391,
0.13131387531757355,
0.003351053688675165,
-0.10782008618116379,
-0.07696644961833954,
-0.024865515530109406,
-0.017643561586737633,
-0.022811954841017723,
-0.054074984043836594,
0.05310645326972008,
0.02390296570956707,
-0.0043946849182248116,
-0.058878883719444275,
0.026161033660173416,
-0.10847523808479309,
0.0011754182633012533,
-0.055590976029634476,
-0.07610706984996796,
0.004296566359698772,
-0.03455231711268425,
0.10699819028377533,
-0.03941600024700165,
-0.10550327599048615,
0.02459150366485119,
0.024951759725809097,
-0.006317939143627882,
0.014758395031094551,
-0.0415814109146595,
0.03765156492590904,
-0.03443198651075363,
-0.022049834951758385,
0.029224414378404617,
0.021143091842532158,
0.031580060720443726,
0.02578207291662693,
-0.038011692464351654,
0.09184370934963226,
-0.021024996414780617,
0.013429488055408001,
-0.02900850772857666,
0.02053573541343212,
-0.06062223017215729,
-0.02980944886803627,
0.06349342316389084,
0.06011567264795303,
0.028776196762919426,
0.05382325127720833,
0.09902177006006241,
0.008426873944699764,
0.12136805802583694,
-0.018465621396899223,
-0.03459818661212921,
0.038926348090171814,
0.0659279078245163,
0.05014147609472275,
0.05344421789050102,
-0.06156819686293602,
-4.833703215467722e-8,
-0.049308232963085175,
0.032442156225442886,
0.032989006489515305,
0.06864623725414276,
-0.05125312879681587,
-0.023581955581903458,
-0.007854106836020947,
0.0193164199590683,
-0.009602879174053669,
0.00037806635373272,
0.0349309965968132,
0.07824241369962692,
0.011379982344806194,
-0.06062319502234459,
0.056574221700429916,
0.053017619997262955,
-0.006127405911684036,
-0.006253180559724569,
-0.017551105469465256,
-0.010529587045311928,
0.06025513634085655,
-0.0009207880357280374,
0.028036393225193024,
-0.09528084099292755,
0.04840487241744995,
-0.05939813703298569,
-0.019872436299920082,
0.13052058219909668,
0.014385148882865906,
0.026310661807656288,
0.03984381631016731,
-0.01568857580423355,
0.026213334873318672,
-0.0029213137459009886,
0.08847378194332123,
0.06620583683252335,
0.06651867181062698,
-0.06465309858322144,
0.027542144060134888,
-0.05498874932527542,
-0.014907262288033962,
0.05094999447464943,
-0.04969239607453346,
-0.026831435039639473,
0.028603749349713326,
0.011699583381414413,
-0.06746865808963776,
-0.07198260724544525,
0.03178771957755089,
-0.046497464179992676,
-0.08183491975069046,
-0.07557139545679092,
-0.05570096895098686,
0.048714157193899155,
0.052088070660829544,
0.0018130888929590583,
-0.03213615342974663,
-0.08567006886005402,
0.057440295815467834,
0.04428984224796295,
0.04806544631719589,
-0.027118919417262077,
-0.041061773896217346,
0.024480503052473068
] | https://github.com/ymetz/multi-type-feedback | 3 | 0 | 0 | 0 |
Token-Supervised Value Models for Enhancing Mathematical Problem-Solving Capabilities of Large Language Models | https://openreview.net/forum?id=6HcnC3pPkp | [
"Jung Hyun Lee",
"June Yong Yang",
"Byeongho Heo",
"Dongyoon Han",
"Kyungsu Kim",
"Eunho Yang",
"Kang Min Yoo"
] | Poster | With the rapid advancement of test-time compute search strategies to improve the mathematical problem-solving capabilities of large language models (LLMs), the need for building robust verifiers has become increasingly important. However, all these inference strategies rely on existing verifiers originally designed for Best-of-N search, which makes them sub-optimal for tree search techniques at test time. During tree search, existing verifiers can only offer indirect and implicit assessments of partial solutions or under-value prospective intermediate steps, thus resulting in the premature pruning of promising intermediate steps. To overcome these limitations, we propose token-supervised value models (TVMs) -- a new class of verifiers that assign each token a probability that reflects the likelihood of reaching the correct final answer. This new token-level supervision enables TVMs to directly and explicitly evaluate partial solutions, effectively distinguishing between promising and incorrect intermediate steps during tree search at test time. Experimental results demonstrate that combining tree-search-based inference strategies with TVMs significantly improves the accuracy of LLMs in mathematical problem-solving tasks, surpassing the performance of existing verifiers. | Large Language Models, Mathematical Problem-Solving, Verifiers | null | 12,758 | 2407.12863 | [
-0.03485949710011482,
0.008126236498355865,
0.08142414689064026,
0.03524908050894737,
0.04427625238895416,
-0.011645011603832245,
-0.056715406477451324,
0.05863913148641586,
0.03280063718557358,
-0.031045645475387573,
-0.06554313004016876,
-0.08339893072843552,
0.0456291139125824,
0.06436385959386826,
0.060014642775058746,
0.020486725494265556,
-0.004838971886783838,
0.09648875892162323,
-0.04325110465288162,
-0.12031584233045578,
0.03111628256738186,
0.026185274124145508,
0.00939869973808527,
-0.019823294132947922,
0.03411070629954338,
-0.001216612639836967,
0.007471717428416014,
-0.06540476530790329,
0.013265158981084824,
0.005643879063427448,
0.016231264919042587,
-0.00001215186966874171,
0.0023826947435736656,
0.011134189553558826,
-0.07712259888648987,
0.06473646312952042,
-0.11511608958244324,
0.03240450844168663,
0.04494883492588997,
-0.05746746063232422,
-0.025754358619451523,
-0.005787683185189962,
0.0316045805811882,
-0.00234226044267416,
0.09884516149759293,
-0.006838646251708269,
0.04669097810983658,
-0.05717623233795166,
-0.06344860792160034,
0.001933468971401453,
-0.08932610601186752,
-0.038401808589696884,
-0.08149776607751846,
0.014452227391302586,
-0.039220839738845825,
-0.030186811462044716,
0.04042087495326996,
-0.00995089765638113,
0.05637684091925621,
-0.04510848969221115,
-0.049583036452531815,
-0.13287237286567688,
-0.024768637493252754,
-0.061637457460165024,
0.006685354746878147,
-0.07784745842218399,
0.00011852937313960865,
0.02895730547606945,
-0.057085826992988586,
0.10531681030988693,
0.03425688296556473,
0.10775650292634964,
-0.04929719120264053,
0.09435759484767914,
0.034215547144412994,
0.08926960825920105,
0.06391117721796036,
0.022756272926926613,
0.05767573416233063,
-0.013435926288366318,
-0.031144287437200546,
-0.0016093485755845904,
0.01213426049798727,
0.035185806453228,
0.16259312629699707,
-0.010749287903308868,
0.027182865887880325,
0.08613251894712448,
0.043169111013412476,
0.04472687840461731,
-0.035658545792102814,
-0.12988178431987762,
-0.009117074310779572,
0.0010690585477277637,
0.008897211402654648,
-0.013248411007225513,
-0.0424056313931942,
-0.06666457653045654,
-0.012643848545849323,
0.054154086858034134,
-0.014728458598256111,
0.14034578204154968,
0.08196356147527695,
-0.1375742405653,
-0.0022587694693356752,
0.01173337735235691,
0.010029485449194908,
0.036755822598934174,
0.07208900153636932,
-0.09726135432720184,
0.002192160813137889,
0.03201321139931679,
0.04967863857746124,
0.09407129883766174,
-0.002059805439785123,
-0.02458859607577324,
0.004889166448265314,
-0.09726931154727936,
-0.0015691604930907488,
0.04141228273510933,
0.022911226376891136,
0.015175297856330872,
0.028401287272572517,
-0.002933857962489128,
-0.02867550402879715,
0.010243047960102558,
-0.037739869207143784,
3.0278857359640657e-33,
0.023313172161579132,
0.0583442822098732,
0.0404508076608181,
0.05237576365470886,
0.055599458515644073,
-0.011186960153281689,
-0.03216258063912392,
0.10943996906280518,
-0.00018358505622018129,
0.03535304591059685,
-0.008803675882518291,
-0.03654415160417557,
-0.04707741364836693,
0.008227854035794735,
0.03826454281806946,
-0.011049005202949047,
-0.06030626967549324,
0.022310549393296242,
-0.1101229265332222,
-0.04786556586623192,
0.04326348751783371,
-0.08592173457145691,
0.014812907204031944,
-0.07536676526069641,
0.11283311992883682,
-0.0071301949210464954,
0.09152542054653168,
-0.06759438663721085,
-0.012914265505969524,
-0.005922126583755016,
-0.058302201330661774,
-0.058770690113306046,
-0.04088488593697548,
0.04307735338807106,
0.02240927889943123,
0.0024491194635629654,
0.031146349385380745,
-0.08545023947954178,
-0.006699478719383478,
-0.02408340387046337,
-0.09642280638217926,
0.03783567249774933,
0.018949095159769058,
0.0667579248547554,
-0.052820704877376556,
-0.09676762670278549,
-0.04979465529322624,
-0.02492440491914749,
0.0013250034535303712,
-0.022732682526111603,
0.018228469416499138,
0.051360517740249634,
-0.03775029629468918,
-0.07947107404470444,
-0.0024939754512161016,
0.011688901111483574,
-0.022802911698818207,
0.025572897866368294,
-0.06291020661592484,
0.057209089398384094,
0.058527376502752304,
0.025024371221661568,
0.011456163600087166,
0.052852459251880646,
-0.04784531518816948,
0.09466631710529327,
-0.03508729860186577,
0.04673111438751221,
0.12596049904823303,
-0.016401443630456924,
0.0163059551268816,
0.010217648930847645,
0.026246769353747368,
-0.09058933705091476,
0.021505653858184814,
-0.04049728438258171,
0.12818971276283264,
-0.08516784757375717,
0.025683199986815453,
-0.015544203110039234,
0.031222712248563766,
-0.01321526151150465,
0.0045582749880850315,
-0.009606322273612022,
0.06299283355474472,
-0.08220572024583817,
0.05450593680143356,
-0.058433759957551956,
-0.08331348747015,
-0.10649620741605759,
0.014892392791807652,
-0.0002026567090069875,
-0.04351183399558067,
0.04798879101872444,
0.03825272247195244,
-2.4256059606279045e-33,
-0.03805186599493027,
-0.03853166475892067,
-0.0014262760523706675,
0.03772236779332161,
-0.0418543815612793,
-0.08183843642473221,
-0.027162062004208565,
-0.04075556620955467,
0.011230457574129105,
-0.07081255316734314,
-0.048820991069078445,
-0.04261254519224167,
0.002002222230657935,
0.051453445106744766,
-0.007830771617591381,
-0.0034613690804690123,
-0.05269841477274895,
0.000473665539175272,
0.06795507669448853,
0.09170500934123993,
-0.05556967481970787,
0.1381717175245285,
-0.022656673565506935,
0.044635746628046036,
-0.06744396686553955,
0.019964205101132393,
0.00804957840591669,
0.025359079241752625,
-0.023004494607448578,
0.024246295914053917,
0.0008852557512000203,
-0.0672411397099495,
0.007504864130169153,
-0.021866681054234505,
-0.06611615419387817,
-0.02026381902396679,
0.0754915103316307,
-0.03770751133561134,
0.004897094331681728,
0.11854896694421768,
0.08417633920907974,
-0.04493997246026993,
-0.055959563702344894,
0.029286831617355347,
-0.027698922902345657,
0.015246755443513393,
-0.0026897345669567585,
-0.006177137605845928,
0.028322303667664528,
-0.01347462646663189,
0.024954697117209435,
0.006949905306100845,
-0.0688740462064743,
0.027970995754003525,
-0.05109325796365738,
-0.037999145686626434,
-0.05626322701573372,
-0.005835446994751692,
0.011114362627267838,
-0.0028069631662219763,
-0.04447541385889053,
0.04399418085813522,
0.12484915554523468,
-0.020362885668873787,
0.018855703994631767,
-0.02346179448068142,
-0.0490427166223526,
0.07221181690692902,
-0.03162148594856262,
-0.053412534296512604,
-0.01851203851401806,
-0.023332076147198677,
0.018283706158399582,
0.037563126534223557,
0.017596617341041565,
-0.00920778512954712,
-0.01916159689426422,
-0.08070596307516098,
-0.016540614888072014,
-0.10811106860637665,
0.02434489317238331,
-0.023903820663690567,
-0.0002911704359576106,
0.03953703120350838,
0.0024048443883657455,
0.039299674332141876,
0.09141059219837189,
0.1220359280705452,
0.04783692583441734,
0.02194586768746376,
-0.059480734169483185,
0.050919368863105774,
0.01197841763496399,
0.05049259960651398,
-0.03527066111564636,
-4.055875280073451e-8,
-0.06755141168832779,
0.02177068591117859,
-0.012578487396240234,
0.05621524155139923,
0.04394438862800598,
0.022168677300214767,
-0.06901861727237701,
0.058329518884420395,
-0.011738447472453117,
-0.04770282283425331,
0.01859140582382679,
0.03856253996491432,
-0.07075141370296478,
-0.04586271941661835,
-0.02087331749498844,
0.03351670503616333,
0.03410928696393967,
0.013460998423397541,
-0.026341121643781662,
0.03363054245710373,
0.005908473394811153,
0.0541154220700264,
0.008632661774754524,
0.013136721216142178,
-0.008444251492619514,
-0.030266251415014267,
0.050020597875118256,
0.0871003195643425,
0.05981831252574921,
-0.025735890492796898,
-0.02197045274078846,
0.000043092604755656794,
-0.03627912700176239,
0.026947831735014915,
0.0592467337846756,
0.0695560872554779,
0.01625191606581211,
-0.033937033265829086,
-0.028551515191793442,
-0.009073392488062382,
0.04638995602726936,
0.058464184403419495,
-0.05630115792155266,
0.036846715956926346,
-0.003228114452213049,
0.006683730520308018,
-0.03422186151146889,
-0.03254413604736328,
0.021012887358665466,
-0.07526354491710663,
0.014380918815732002,
-0.006795186549425125,
-0.0322866253554821,
0.012896278873085976,
-0.0021740184165537357,
0.020322540774941444,
-0.03070392832159996,
0.012230235151946545,
0.009755515493452549,
0.04757065325975418,
0.049965158104896545,
-0.036242205649614334,
-0.030786553397774696,
0.029685460031032562
] | 0 | 0 | 0 | 0 |
|
ImProver: Agent-Based Automated Proof Optimization | https://openreview.net/forum?id=dWsdJAXjQD | [
"Riyaz Ahuja",
"Jeremy Avigad",
"Prasad Tetali",
"Sean Welleck"
] | Poster | Large language models (LLMs) have been used to generate formal proofs of mathematical theorems in proofs assistants such as Lean.
However, we often want to optimize a formal proof with respect to various criteria, depending on its downstream use.
For example, we may want a proof to adhere to a certain style, be declaratively structured, or concise. Having suitably optimized proofs is also important for learning tasks, especially since human-written proofs may not optimal for that purpose.
To this end, we study a new problem of automated proof optimization: rewriting a proof so that it is correct and optimizes for an arbitrary criterion, such as length or declarativity.
As a first method for automated proof optimization, we present ImProver, a large-language-model agent that rewrites proofs to optimize arbitrary user-defined metrics in Lean.
We find that naively applying LLMs to proof optimization falls short, and we incorporate various improvements into ImProver, such as the use of symbolic Lean context in a novel Chain-of-States technique, as well as error-correction and retrieval. We test ImProver on rewriting real-world undergraduate, competition, and research-level mathematics theorems, finding that ImProver is capable of rewriting proofs so that they are substantially shorter and more declarative in structure. | Automated Proof Optimization, Neural Theorem Proving, Formal Mathematics, Lean Theorem Prover, Proof Generation, Large Language Models, Symbolic Reasoning, Interactive Theorem Proving | ImProver optimizes formal mathematical proofs for arbitrary metrics using LLM agents, automatically improving proof readability and length in Lean as well as generalizing NTP. | 12,756 | 2410.04753 | [
-0.07464245706796646,
0.08296968787908554,
-0.00028959722840227187,
0.0026189633645117283,
0.004394663497805595,
-0.03613097593188286,
-0.013258487917482853,
0.05556409806013107,
0.01527670118957758,
0.04816083237528801,
-0.051953233778476715,
-0.0752958282828331,
0.06706565618515015,
0.07015649974346161,
0.05445776507258415,
-0.0008017655345611274,
0.06434965878725052,
0.13675729930400848,
-0.049020472913980484,
-0.06936390697956085,
0.024514352902770042,
-0.03488129749894142,
-0.021774310618638992,
-0.02508818916976452,
-0.011491128243505955,
-0.05578126013278961,
-0.011687184683978558,
-0.03872136026620865,
-0.012772144749760628,
-0.0029529540333896875,
-0.02651113271713257,
0.05439373478293419,
0.02823336236178875,
-0.005545740947127342,
0.026119444519281387,
0.15323126316070557,
-0.04490674287080765,
0.03567615896463394,
0.019187696278095245,
0.006943723186850548,
-0.06005547568202019,
-0.06099235638976097,
-0.012849334627389908,
-0.0037807049229741096,
-0.010335257276892662,
0.003988359589129686,
0.01496120449155569,
0.024978963658213615,
-0.04037434607744217,
0.022101107984781265,
-0.09855809062719345,
-0.02069576270878315,
-0.0537874698638916,
-0.040349941700696945,
-0.0037255643401294947,
-0.004623600747436285,
0.10944599658250809,
0.037234093993902206,
-0.030832337215542793,
-0.12410221248865128,
0.008482251316308975,
-0.05617249384522438,
-0.05060823634266853,
0.02485748380422592,
0.01554933562874794,
0.012561552226543427,
0.03447679430246353,
0.0382077656686306,
-0.02965032123029232,
0.11959876865148544,
-0.04951022192835808,
0.018109647557139397,
-0.05721863731741905,
0.03989031910896301,
-0.026921285316348076,
0.03798558562994003,
0.005963073577731848,
-0.008191881701350212,
0.006447008345276117,
-0.033696092665195465,
-0.06850907951593399,
0.007208846975117922,
-0.01736832782626152,
0.010783350095152855,
0.07228410243988037,
-0.022615842521190643,
0.03724974766373634,
-0.060487426817417145,
0.1212892085313797,
0.022046418860554695,
-0.0034164548851549625,
0.0018148539820685983,
0.04260161519050598,
-0.005119027569890022,
0.06824158877134323,
0.015308610163629055,
-0.02470717392861843,
-0.07366041094064713,
0.010050208307802677,
0.085868239402771,
-0.007242455612868071,
0.10842315107584,
0.04776710271835327,
-0.15615375339984894,
0.02304314449429512,
-0.0035175953526049852,
0.03476349264383316,
0.038344670087099075,
0.028204521164298058,
-0.1250876784324646,
0.08120466768741608,
0.05427165329456329,
0.030230164527893066,
0.03617379069328308,
0.08433583378791809,
0.02927420474588871,
0.01967175491154194,
0.012092913500964642,
-0.01459689624607563,
0.006314731668680906,
0.038251373916864395,
0.04173225909471512,
0.031007980927824974,
0.005628112703561783,
0.06438157707452774,
0.014412108808755875,
-0.029565980657935143,
3.8011474094410674e-34,
0.03559251129627228,
0.10846027731895447,
0.03560655191540718,
0.03708260878920555,
0.0517490915954113,
0.018975645303726196,
0.001960275461897254,
0.01717708259820938,
-0.04728199169039726,
-0.013452653773128986,
0.05986527353525162,
0.035285092890262604,
-0.033877894282341,
0.016113342717289925,
0.04538293555378914,
-0.020108608528971672,
-0.005493209231644869,
0.029797565191984177,
-0.05519561097025871,
0.006852908059954643,
0.037235092371702194,
-0.1302417516708374,
-0.013022679835557938,
-0.10117834806442261,
0.027308566495776176,
-0.011835338547825813,
0.04150524362921715,
0.0008341682842001319,
-0.028337718918919563,
-0.04474135860800743,
-0.040644947439432144,
-0.08507119119167328,
0.03328920155763626,
0.11140622198581696,
-0.0027118728030472994,
-0.03880937397480011,
0.016755254939198494,
-0.07348253577947617,
0.027373459190130234,
0.01386682502925396,
0.03277570381760597,
0.02003302052617073,
0.022577719762921333,
0.02744869515299797,
-0.011677354574203491,
0.013956868089735508,
-0.05557083711028099,
0.009598806500434875,
0.009086515754461288,
-0.1044723391532898,
0.08417099714279175,
0.04139717295765877,
0.0011587993940338492,
-0.034670908004045486,
0.0350763164460659,
0.010123470798134804,
-0.0295181293040514,
0.03830700367689133,
-0.011618156917393208,
0.07531797885894775,
0.01489983033388853,
0.05466066673398018,
0.0028805809561163187,
0.056612513959407806,
-0.08783433586359024,
0.013102313503623009,
-0.051804766058921814,
0.07637988030910492,
0.09526491165161133,
0.01580200344324112,
-0.0335148423910141,
-0.007132738828659058,
-0.009180696681141853,
0.02750573121011257,
0.07571781426668167,
-0.05935387313365936,
0.02586665377020836,
-0.11520162224769592,
0.020871145650744438,
-0.052983492612838745,
0.023910146206617355,
0.10295224189758301,
0.015112061984837055,
0.0014223704347386956,
0.01795218139886856,
-0.12409385293722153,
-0.05135278403759003,
-0.015602967701852322,
-0.012462436221539974,
-0.05845998600125313,
-0.11125878244638443,
-0.06455443799495697,
-0.056777097284793854,
0.004976797848939896,
0.012675534933805466,
-1.747886631275815e-33,
-0.018193600699305534,
-0.10295212268829346,
0.00013069433043710887,
0.11051560193300247,
-0.05526667460799217,
-0.08537711948156357,
-0.011402243748307228,
-0.014078999869525433,
-0.004406938329339027,
-0.08513493835926056,
-0.025737054646015167,
-0.04948345571756363,
0.015162135474383831,
-0.012076114304363728,
-0.039294254034757614,
-0.005032897461205721,
-0.024251878261566162,
-0.06752898544073105,
0.05383196845650673,
0.03132762759923935,
0.08298047631978989,
0.06431981176137924,
-0.06877759844064713,
0.05147785320878029,
0.032005734741687775,
-0.0062691387720406055,
-0.03682200238108635,
0.06511656194925308,
0.00580947520211339,
0.058467309921979904,
-0.0041338796727359295,
-0.03659845516085625,
-0.10861356556415558,
-0.0915675237774849,
-0.047225892543792725,
-0.0285929124802351,
0.07305479049682617,
0.01933564990758896,
-0.052217114716768265,
0.1140870749950409,
0.08270525187253952,
-0.12499041110277176,
-0.013753894716501236,
-0.04797867685556412,
0.009294599294662476,
-0.04516446962952614,
-0.05013205483555794,
-0.031058164313435555,
-0.013791690580546856,
0.0020608692429959774,
-0.020896906033158302,
-0.07400711625814438,
-0.039312828332185745,
0.006249685771763325,
-0.08150483667850494,
-0.044957876205444336,
-0.0014538014074787498,
-0.027882305905222893,
0.027840295806527138,
-0.03979984298348427,
-0.10254298150539398,
0.0003162724315188825,
0.0030599762685596943,
0.029484571889042854,
0.013795522972941399,
-0.07287628948688507,
0.013746337033808231,
0.015332168899476528,
0.005663633346557617,
0.01118580624461174,
0.04736100137233734,
-0.04971037060022354,
-0.08583630621433258,
0.06303343176841736,
0.04484063386917114,
0.011648197658360004,
0.04527123272418976,
-0.013530685566365719,
0.047573499381542206,
-0.037112653255462646,
0.03258621692657471,
-0.00007162009569583461,
0.06067782640457153,
0.034040480852127075,
-0.0330081507563591,
0.018656909465789795,
0.028433604165911674,
0.10036918520927429,
0.0033255533780902624,
0.10362652689218521,
-0.0461166612803936,
-0.020719876512885094,
-0.008837622590363026,
0.04653267189860344,
-0.11670812219381332,
-4.0426762382139714e-8,
-0.06187096983194351,
0.04406087100505829,
0.08461207151412964,
-0.010788093321025372,
0.04765082523226738,
-0.0003161410568282008,
-0.028341243043541908,
0.03488032892346382,
0.023746080696582794,
-0.11422833800315857,
0.014451433904469013,
0.051054567098617554,
-0.021695146337151527,
-0.04892934113740921,
-0.06496331840753555,
0.06428518146276474,
-0.06865282356739044,
-0.038787614554166794,
-0.040174514055252075,
-0.028524311259388924,
0.004194348584860563,
0.04423598572611809,
-0.07897745072841644,
0.01269067358225584,
0.03520088270306587,
-0.04796001315116882,
0.02997962385416031,
0.11124713718891144,
0.048031285405159,
-0.04907042905688286,
-0.02373497188091278,
0.010019253939390182,
0.09740903973579407,
-0.012191562913358212,
0.056927718222141266,
0.04597633704543114,
0.0395040400326252,
0.023906061425805092,
0.04454569146037102,
-0.03275725618004799,
-0.026141123846173286,
0.036852773278951645,
0.026712821796536446,
-0.015762757509946823,
0.05099402740597725,
-0.08972810953855515,
-0.03476792946457863,
-0.03254362568259239,
0.0006407429464161396,
-0.04788709804415703,
0.008887000381946564,
-0.062408097088336945,
-0.0244117584079504,
0.011760027147829533,
-0.02031484991312027,
0.03387337923049927,
-0.020083120092749596,
0.0010573890758678317,
0.028376078233122826,
0.0162240881472826,
-0.026655083522200584,
0.06496106833219528,
0.06346068531274796,
-0.03616407513618469
] | https://github.com/riyazahuja/ImProver | 28 | 0 | 0 | 0 |
Learning-Guided Rolling Horizon Optimization for Long-Horizon Flexible Job-Shop Scheduling | https://openreview.net/forum?id=Aly68Y5Es0 | [
"Sirui Li",
"Wenbin Ouyang",
"Yining Ma",
"Cathy Wu"
] | Poster | Long-horizon combinatorial optimization problems (COPs), such as the Flexible Job-Shop Scheduling Problem (FJSP), often involve complex, interdependent decisions over extended time frames, posing significant challenges for existing solvers. While Rolling Horizon Optimization (RHO) addresses this by decomposing problems into overlapping shorter-horizon subproblems, such overlap often involves redundant computations. In this paper, we present L-RHO, the first learning-guided RHO framework for COPs. L-RHO employs a neural network to intelligently fix variables that in hindsight did not need to be re-optimized, resulting in smaller and thus easier-to-solve subproblems. For FJSP, this means identifying operations with unchanged machine assignments between consecutive subproblems. Applied to FJSP, L-RHO accelerates RHO by up to 54\% while significantly improving solution quality, outperforming other heuristic and learning-based baselines. We also provide in-depth discussions and verify the desirable adaptability and generalization of L-RHO across numerous FJSP variates, distributions, online scenarios and benchmark instances. Moreover, we provide a theoretical analysis to elucidate the conditions under which learning is beneficial. | Learning-Guided Optimization, Rolling Horizon Optimization, Flexible Job Shop Scheduling | null | 12,752 | 2502.15791 | [
-0.07537418603897095,
-0.04639622941613197,
0.007444120477885008,
0.07532905042171478,
0.04113558307290077,
0.0029530925676226616,
-0.0420185811817646,
-0.04644652083516121,
-0.1197008416056633,
-0.009398028254508972,
-0.03750224411487579,
0.03083379752933979,
-0.03331650048494339,
0.07508646696805954,
-0.04971875995397568,
0.011998224072158337,
0.046437907963991165,
-0.03857351839542389,
-0.055935680866241455,
-0.08910489827394485,
-0.030906841158866882,
-0.005261904560029507,
-0.03622793033719063,
-0.04530542343854904,
-0.032958243042230606,
0.03464377671480179,
-0.03043469414114952,
-0.024979984387755394,
-0.01673474721610546,
-0.006673880387097597,
0.048310209065675735,
0.011495571583509445,
-0.0049784863367676735,
-0.02011488378047943,
0.016947384923696518,
0.10811544954776764,
-0.09133131802082062,
0.006935932673513889,
0.03951988369226456,
-0.022732863202691078,
0.04522082954645157,
0.0052780951373279095,
0.006976129952818155,
-0.021788666024804115,
0.08668234944343567,
-0.004183067008852959,
-0.033692240715026855,
-0.012832165695726871,
-0.014290209859609604,
0.01019303034991026,
0.0031441631726920605,
-0.0369776152074337,
-0.06133795157074928,
0.10255326330661774,
0.019714366644620895,
0.051462430506944656,
0.04293777793645859,
0.01670139655470848,
-0.07628951221704483,
0.02253277786076069,
-0.06378961354494095,
-0.09231745451688766,
-0.03435952216386795,
-0.015130467712879181,
-0.10020092874765396,
-0.028177127242088318,
-0.0359419584274292,
0.029727110639214516,
-0.05710463598370552,
0.06111770495772362,
0.027443908154964447,
0.05131298303604126,
-0.05995934456586838,
0.015818864107131958,
0.05819552391767502,
0.06403075158596039,
0.13892705738544464,
-0.008275918662548065,
-0.004174902103841305,
-0.09551924467086792,
-0.020403075963258743,
0.032811909914016724,
-0.068437859416008,
0.053554192185401917,
0.09484409540891647,
-0.04657058045268059,
0.008418554440140724,
0.07528606802225113,
0.06335870921611786,
-0.011505553498864174,
0.07638636231422424,
-0.03638693317770958,
-0.010453496128320694,
-0.07957877963781357,
-0.0704408660531044,
-0.008402026258409023,
0.014011438004672527,
0.02415444329380989,
-0.04350382462143898,
0.04914560914039612,
-0.08162302523851395,
0.0367664135992527,
0.08480960130691528,
-0.07315436005592346,
-0.06149285286664963,
-0.0042921132408082485,
-0.03326990455389023,
0.032773036509752274,
-0.05337603762745857,
-0.1217980682849884,
0.03723577782511711,
0.053411297500133514,
0.05467264726758003,
-0.02808791771531105,
0.004393800627440214,
0.0645337849855423,
-0.09520917385816574,
-0.023836294189095497,
-0.05316365510225296,
0.06519082188606262,
-0.09488988667726517,
0.00844465009868145,
0.018337808549404144,
0.0831955149769783,
-0.04229049012064934,
-0.009195413440465927,
-0.08775409311056137,
3.507377963188139e-33,
-0.002866946393623948,
-0.03238200768828392,
0.007968241348862648,
-0.034210748970508575,
0.13301041722297668,
-0.03797025606036186,
0.07917670160531998,
0.07008831948041916,
0.016244830563664436,
0.05159781873226166,
0.0009249567519873381,
0.0036882106214761734,
-0.06115931272506714,
0.007603166159242392,
-0.005862715188413858,
-0.015321733430027962,
-0.01609126105904579,
0.06387487053871155,
-0.05732673034071922,
-0.05990679934620857,
0.028802115470170975,
-0.08442986756563187,
-0.06526979804039001,
-0.011279046535491943,
0.0641503557562828,
-0.007683036848902702,
0.05073565989732742,
-0.03380240872502327,
0.050230324268341064,
0.030725358054041862,
-0.050035253167152405,
0.11509265750646591,
0.013047503307461739,
0.1601092368364334,
-0.011445071548223495,
-0.031445570290088654,
-0.03148059919476509,
-0.06670859456062317,
-0.0012028049677610397,
0.03302081301808357,
-0.07243780046701431,
0.003643587464466691,
0.012161686085164547,
-0.012979299761354923,
-0.06002834066748619,
-0.08426625281572342,
0.01869153045117855,
-0.0645720511674881,
-0.02366407960653305,
0.05278326943516731,
-0.027169421315193176,
0.01679949462413788,
0.06751552224159241,
-0.055749159306287766,
-0.06233053281903267,
-0.016796858981251717,
0.058321513235569,
0.00512291956692934,
0.008779141120612621,
0.08390624076128006,
0.04557725042104721,
0.049387216567993164,
-0.05601560324430466,
0.04831469804048538,
0.024289343506097794,
0.030951417982578278,
0.05192624405026436,
0.06024680659174919,
0.11634178459644318,
-0.01771223358809948,
-0.0024135794956237078,
0.010790213942527771,
0.04829959198832512,
-0.02807961404323578,
0.09842993319034576,
0.011125408113002777,
0.10540929436683655,
0.054926518350839615,
-0.0036785381380468607,
-0.04892200976610184,
-0.026509862393140793,
0.10184790939092636,
-0.0042285555973649025,
-0.046425189822912216,
0.02606653980910778,
-0.038163669407367706,
0.025241998955607414,
0.00015436226385645568,
-0.030030133202672005,
-0.049218110740184784,
-0.06441287696361542,
-0.00756609532982111,
0.021901119500398636,
-0.0016192044131457806,
-0.004331617616117001,
-4.3319042310014465e-33,
0.027134092524647713,
-0.04641835018992424,
-0.0506972037255764,
-0.013919760473072529,
0.0014496918302029371,
0.006764255464076996,
-0.057312555611133575,
-0.09353529661893845,
0.03036509081721306,
-0.05273423343896866,
-0.0930105596780777,
-0.018958399072289467,
0.01764059253036976,
-0.015801923349499702,
0.011831458657979965,
-0.022598225623369217,
-0.05190856382250786,
-0.029925335198640823,
0.05692498758435249,
0.051763180643320084,
0.008111443370580673,
0.03576897084712982,
-0.08219422399997711,
0.019047658890485764,
-0.027646953240036964,
-0.023955555632710457,
0.005573329981416464,
0.06590697914361954,
0.03169272094964981,
0.04268496856093407,
-0.054462045431137085,
-0.09344623982906342,
-0.08633605390787125,
-0.01719485968351364,
0.023054173216223717,
0.06039660796523094,
-0.02391628734767437,
0.04515797272324562,
-0.01221511047333479,
0.09602341055870056,
0.08315905183553696,
-0.08100724220275879,
-0.057458166033029556,
-0.005512760952115059,
-0.006706639658659697,
0.016690470278263092,
-0.057583462446928024,
-0.04034906625747681,
0.02149047888815403,
-0.0065690926276147366,
-0.09357240796089172,
0.07694937288761139,
-0.09823457896709442,
-0.013346042484045029,
-0.08854185789823532,
0.004907662980258465,
-0.011419025249779224,
-0.008374053053557873,
0.06970298290252686,
0.027776068076491356,
0.008137297816574574,
-0.03694083169102669,
0.04521108418703079,
0.01416589505970478,
-0.026115674525499344,
-0.0007410645484924316,
0.03919108957052231,
-0.053064484149217606,
-0.057637203484773636,
0.00611307704821229,
-0.030307047069072723,
0.027674322947859764,
0.09007955342531204,
0.014986326918005943,
-0.07802154868841171,
-0.03104078583419323,
0.002856828970834613,
-0.05431690067052841,
-0.017987508326768875,
0.02367645502090454,
0.05565381050109863,
-0.03749890998005867,
0.030800076201558113,
0.11273571103811264,
-0.04201246425509453,
0.06391043961048126,
0.055687159299850464,
0.07720761746168137,
0.07489001005887985,
-0.050859108567237854,
-0.030679721385240555,
0.008532712236046791,
-0.018142741173505783,
0.06488269567489624,
-0.018010670319199562,
-4.693876931582963e-8,
0.018862765282392502,
0.034578338265419006,
0.015032070688903332,
0.02534509263932705,
0.12760119140148163,
-0.04008593037724495,
0.01764988899230957,
0.04490666463971138,
0.0019184951670467854,
0.027945898473262787,
0.05192810669541359,
0.010351183824241161,
0.04366990551352501,
-0.02329818718135357,
-0.08869099617004395,
0.04236399382352829,
-0.0072895679622888565,
-0.020491331815719604,
0.012010525912046432,
0.03230094164609909,
0.04351617023348808,
0.0221127737313509,
0.03871212527155876,
-0.040738433599472046,
-0.0030706438701599836,
-0.07608193159103394,
-0.09626469761133194,
0.015849094837903976,
0.08580133318901062,
0.009014333598315716,
-0.016136571764945984,
0.03614840283989906,
0.04892496019601822,
0.010805482044816017,
0.07294577360153198,
-0.007049293722957373,
0.024279115721583366,
-0.052180275321006775,
-0.06470742076635361,
0.0008034476195462048,
0.04857213795185089,
0.05506838113069534,
-0.0492115318775177,
0.004464774392545223,
0.05212866887450218,
-0.026567749679088593,
-0.00208615162409842,
-0.0640544444322586,
0.02566980943083763,
-0.021952714771032333,
0.02767469361424446,
0.022243522107601166,
0.017828769981861115,
0.013450117781758308,
0.08665836602449417,
-0.025120934471488,
0.02151321806013584,
-0.10977069288492203,
0.028355736285448074,
0.11392693966627121,
-0.08452706784009933,
-0.01673150062561035,
0.007017659023404121,
0.019662626087665558
] | https://github.com/mit-wu-lab/l-rho | 9 | 0 | 0 | 0 |
Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors | https://openreview.net/forum?id=PIpGN5Ko3v | [
"Tianchun Wang",
"Yuanzhou Chen",
"Zichuan Liu",
"Zhanwen Chen",
"Haifeng Chen",
"Xiang Zhang",
"Wei Cheng"
] | Poster | The advent of large language models (LLMs) has revolutionized the field of text generation, producing outputs that closely mimic human-like writing. Although academic and industrial institutions have developed detectors to prevent the malicious usage of LLM-generated texts, other research has doubt about the robustness of these systems. To stress test these detectors, we introduce a humanized proxy-attack (HUMPA) strategy that effortlessly compromises LLMs, causing them to produce outputs that align with human-written text and mislead detection systems. Our method attacks the source model by leveraging a reinforcement learning (RL) fine-tuned humanized small language model (SLM) in the decoding phase. Through an in-depth analysis, we demonstrate that our attack strategy is capable of generating responses that are indistinguishable to detectors, preventing them from differentiating between machine-generated and human-written text. We conduct systematic evaluations on extensive datasets using proxy-attacked open-source models, including Llama2-13B, Llama3-70B, and Mixtral-8x7B in both white- and black-box settings. Our findings show that the proxy-attack strategy effectively deceives the leading detectors, resulting in an average AUROC drop of 70.4% across multiple datasets, with a maximum drop of 95.0% on a single dataset. Furthermore, in cross-discipline scenarios, our strategy also bypasses these detectors, leading to a significant relative decrease of up to 90.9%, while in cross-language scenario, the drop reaches 91.3%. Despite our proxy-attack strategy successfully bypassing the detectors with such significant relative drops, we find that the generation quality of the attacked models remains preserved, even within a modest utility budget, when compared to the text produced by the original, unattacked source model. | machine-generted text detection; evade detection; fine-tuning | null | 12,740 | 2410.19230 | [
-0.11152137815952301,
-0.0007265941821970046,
0.0334923155605793,
0.07121029496192932,
0.013007516041398048,
-0.02133464440703392,
-0.010416262783110142,
-0.09254224598407745,
0.07930625230073929,
-0.021537158638238907,
-0.035579077899456024,
-0.03903456777334213,
0.09831336885690689,
-0.05256997048854828,
-0.0034865306224673986,
-0.023002076894044876,
0.07171943783760071,
0.06173020973801613,
-0.03387940675020218,
-0.06460252404212952,
0.015184486284852028,
0.007741859648376703,
0.004988834727555513,
-0.011908805929124355,
0.003117689862847328,
0.002193464431911707,
-0.037667904049158096,
-0.014957141131162643,
-0.001996758859604597,
-0.03746143728494644,
0.05316852405667305,
-0.010336761362850666,
-0.006785023957490921,
0.010177649557590485,
-0.039824847131967545,
0.024459686130285263,
0.0017457283101975918,
0.004943485371768475,
0.05091448128223419,
-0.06849422305822372,
-0.008879007771611214,
-0.03155951201915741,
0.048648592084646225,
0.0540136881172657,
0.017721589654684067,
-0.0721692442893982,
0.01262727938592434,
-0.0456690676510334,
-0.029821092262864113,
-0.03981664031744003,
-0.03944820165634155,
-0.005726935341954231,
0.07014770060777664,
0.047573380172252655,
-0.0674828290939331,
-0.11023862659931183,
0.025838511064648628,
0.08672641962766647,
-0.002035800600424409,
-0.010408777743577957,
-0.0886375904083252,
-0.07929238677024841,
-0.04717617109417915,
0.02911687083542347,
-0.03085106424987316,
-0.00823904387652874,
0.018131732940673828,
0.11310919374227524,
0.004589085467159748,
0.04140086844563484,
0.015711579471826553,
-0.007994710467755795,
-0.019773952662944794,
0.06791133433580399,
0.02026514895260334,
0.04445837810635567,
0.013256526552140713,
-0.05953756719827652,
0.09805773198604584,
-0.0695691779255867,
0.04525614529848099,
-0.05761884152889252,
0.06073645129799843,
0.04225427284836769,
0.061489038169384,
0.06785313040018082,
-0.005245743785053492,
0.0551212839782238,
0.08128009736537933,
0.059026677161455154,
-0.035167817026376724,
-0.04982592537999153,
0.1298343688249588,
0.03476930409669876,
0.03212792053818703,
0.006319228559732437,
-0.029063230380415916,
-0.047165848314762115,
-0.03103446774184704,
0.12177916616201401,
-0.011243291199207306,
0.06301891058683395,
0.049426477402448654,
-0.05010882019996643,
0.017634734511375427,
-0.0362769179046154,
0.05048675462603569,
0.0022535547614097595,
0.05090980976819992,
-0.018826480954885483,
0.02455912157893181,
0.05722172185778618,
-0.00885629840195179,
-0.009265369735658169,
0.06355085968971252,
-0.02858622372150421,
0.05822914466261864,
-0.04958993196487427,
-0.0723867118358612,
0.05294928327202797,
0.0070709646679461,
-0.005231147166341543,
-0.02609235607087612,
0.009044640697538853,
0.0555804967880249,
-0.07464874535799026,
-0.04067022353410721,
-1.7254352401880864e-33,
0.0904495045542717,
0.03095261938869953,
-0.04620900750160217,
0.08846070617437363,
0.033371374011039734,
-0.027257971465587616,
0.04199077561497688,
0.07706111669540405,
-0.04095275700092316,
0.017482422292232513,
-0.030369389802217484,
-0.02005869708955288,
-0.012994332239031792,
0.12887504696846008,
0.02322256937623024,
0.052465055137872696,
-0.03888723626732826,
0.018491258844733238,
-0.039480336010456085,
-0.006467027124017477,
0.025752410292625427,
-0.04867899417877197,
0.06330907344818115,
-0.08948485553264618,
0.028499070554971695,
0.0027325465343892574,
0.05501709133386612,
-0.03426194190979004,
-0.026742661371827126,
0.027652109041810036,
-0.07425905764102936,
-0.026417268440127373,
0.020323580130934715,
0.05158042162656784,
0.006315145641565323,
-0.015859151259064674,
-0.021674245595932007,
-0.024070674553513527,
0.0014927502488717437,
0.04251503199338913,
-0.002515166299417615,
0.024453110992908478,
0.016390811651945114,
-0.016567019745707512,
-0.022176070138812065,
0.0007768753566779196,
-0.012169111520051956,
-0.01171779166907072,
-0.0038320196326822042,
-0.03774761036038399,
0.05976782366633415,
0.0964210033416748,
-0.015991272404789925,
0.012048580683767796,
-0.0201348178088665,
-0.026630418375134468,
0.025359388440847397,
-0.02317180298268795,
0.04338667914271355,
0.05316544324159622,
0.008751328103244305,
0.08634496480226517,
0.012362755835056305,
0.08990682661533356,
0.08171961456537247,
-0.026638830080628395,
-0.03210776671767235,
0.03270161524415016,
-0.010111137293279171,
0.016356272622942924,
-0.017994141206145287,
-0.06392204016447067,
-0.027641071006655693,
0.002314481884241104,
-0.011799050495028496,
-0.0026038368232548237,
0.05944371223449707,
0.025711707770824432,
0.05858103558421135,
0.026687033474445343,
0.0603678934276104,
-0.06298024207353592,
0.034603629261255264,
-0.07055573910474777,
-0.11283263564109802,
-0.053347229957580566,
-0.020556436851620674,
-0.09267901629209518,
-0.03698645904660225,
0.024387354031205177,
-0.056938041001558304,
-0.07966969162225723,
-0.05389629304409027,
0.027903389185667038,
-0.009738769382238388,
4.296388230394786e-34,
-0.0927625223994255,
-0.04088670015335083,
-0.055145055055618286,
0.04355059191584587,
-0.09584300220012665,
-0.04772987216711044,
0.05221853405237198,
0.06274569034576416,
0.012448309920728207,
0.03671133145689964,
-0.04240443930029869,
0.0005532637005671859,
0.07675737887620926,
-0.0390411801636219,
0.04300367459654808,
-0.0655386745929718,
0.03724907711148262,
0.03183557093143463,
0.0292521920055151,
0.012523298151791096,
0.021841159090399742,
0.07307934015989304,
-0.07633151113986969,
0.0017614689422771335,
0.023212632164359093,
0.005320933647453785,
0.0028131715953350067,
0.04555438831448555,
0.0402444489300251,
0.02299339696764946,
-0.05607127398252487,
0.0696067288517952,
-0.048912156373262405,
-0.018447037786245346,
-0.11475329101085663,
0.07780686020851135,
0.07007240504026413,
0.08598577231168747,
0.005023318808525801,
0.026264354586601257,
0.07650364935398102,
0.0014736921293660998,
-0.1069750264286995,
-0.0016279101837426424,
-0.00839473120868206,
-0.07294043898582458,
-0.13661770522594452,
-0.0276644304394722,
-0.003960704430937767,
-0.028616059571504593,
0.011500374414026737,
-0.007166767492890358,
-0.005983725190162659,
-0.014955520629882812,
-0.08612262457609177,
-0.07509756088256836,
-0.04087742045521736,
-0.033268753439188004,
0.052074678242206573,
0.025569163262844086,
-0.06787291169166565,
-0.04961048439145088,
-0.0480010099709034,
-0.042356882244348526,
-0.027342000976204872,
-0.041804514825344086,
0.03082231990993023,
0.02516268938779831,
-0.0019676233641803265,
-0.06574352085590363,
0.14314351975917816,
-0.07165798544883728,
-0.0066109043546020985,
0.028971603140234947,
0.021517110988497734,
-0.020982205867767334,
-0.007033499423414469,
-0.08318057656288147,
-0.025911737233400345,
0.0015808548778295517,
0.1034676805138588,
-0.026691317558288574,
0.020365159958600998,
0.06975910812616348,
0.07185281068086624,
0.061980798840522766,
0.0383257158100605,
0.09106095135211945,
-0.011429976671934128,
0.0317462757229805,
-0.06425870209932327,
0.03813467174768448,
-0.03547724336385727,
0.05056799575686455,
-0.018906014040112495,
-3.9168750021190135e-8,
-0.05976787209510803,
-0.019203374162316322,
0.0405716598033905,
0.0677562728524208,
0.003285165410488844,
0.029877137392759323,
-0.1209736317396164,
-0.03548228740692139,
0.005519325844943523,
-0.05329769477248192,
0.05557447299361229,
-0.025382040068507195,
-0.016626667231321335,
-0.06510022282600403,
-0.019266827031970024,
0.07577443867921829,
-0.05680043622851372,
-0.019014863297343254,
-0.05247780680656433,
-0.03741226717829704,
0.06190573424100876,
0.05512024462223053,
0.014554359950125217,
-0.013294348493218422,
0.04108581691980362,
-0.052318744361400604,
-0.06115049123764038,
0.06259596347808838,
-0.013874710537493229,
-0.009728865697979927,
-0.05430256575345993,
0.008379068225622177,
-0.04168003797531128,
-0.01612192578613758,
-0.009699963964521885,
0.14645156264305115,
0.01544348243623972,
-0.05806881934404373,
0.025255749002099037,
0.0014279717579483986,
-0.03258424624800682,
-0.021868761628866196,
0.012664461508393288,
-0.0006368441972881556,
-0.055517882108688354,
-0.02797827497124672,
-0.04475974664092064,
-0.14079098403453827,
0.07257034629583359,
-0.06750859320163727,
0.14997518062591553,
-0.08485758304595947,
0.01613079570233822,
0.06673303991556168,
-0.005768977105617523,
-0.012618131004273891,
0.023095685988664627,
-0.02583211660385132,
0.01733211614191532,
0.11804849654436111,
0.09261087328195572,
-0.033902887254953384,
-0.006514979060739279,
-0.05897726118564606
] | 0 | 0 | 0 | 0 |
|
Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing | https://openreview.net/forum?id=WOt1owGfuN | [
"Qi Le",
"Enmao Diao",
"Ziyan Wang",
"Xinran Wang",
"Jie Ding",
"Li Yang",
"Ali Anwar"
] | Poster | We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner. PP leverages the insight that not all samples and tokens contribute equally to the model's output, and probing a small portion of each batch effectively identifies crucial weights, enabling tailored dynamic pruning for different batches. It comprises three main stages: probing, history-informed pruning, and full inference. In the probing stage, PP selects a small yet crucial set of hidden states, based on residual importance, to run a few model layers ahead. During the history-informed pruning stage, PP strategically integrates the probing states with historical states. Subsequently, it structurally prunes weights based on the integrated states and the PP importance score, a metric developed specifically to assess the importance of each weight channel in maintaining performance. In the final stage, full inference is conducted on the remaining weights. A major advantage of PP is its compatibility with existing models, as it operates without requiring additional neural network modules or fine-tuning. Comprehensive evaluations of PP on LLaMA-2/3 and OPT models reveal that even minimal probing—using just 1.5% of FLOPs—can substantially enhance the efficiency of structured pruning of LLMs. For instance, when evaluated on LLaMA-2-7B with WikiText2, PP achieves a 2.56 times lower ratio of performance degradation per unit of latency reduction compared to the state-of-the-art method at a 40\% pruning ratio. | Large Lanuage Model Pruning, Probe Pruning | We introduce Probe Pruning (PP), a novel framework for online, dynamic, structured pruning of Large Language Models (LLMs) applied in a batch-wise manner. | 12,737 | 2502.15618 | [
-0.031225217506289482,
-0.022712934762239456,
-0.018492845818400383,
0.05251016840338707,
0.04038044810295105,
0.024128640070557594,
-0.03336142376065254,
-0.018106535077095032,
0.002570709679275751,
-0.08613850921392441,
-0.02899353951215744,
-0.000948336033616215,
0.00275123817846179,
-0.004809871315956116,
0.00655017327517271,
-0.026274066418409348,
0.017595507204532623,
0.05278833582997322,
-0.1297474503517151,
-0.12203038483858109,
-0.0032724482007324696,
0.008629614487290382,
-0.006172227207571268,
-0.002615171018987894,
0.06644980609416962,
0.03980020433664322,
0.000043594325688900426,
-0.018932130187749863,
0.03316956385970116,
-0.05787466838955879,
0.001586689380928874,
-0.017215505242347717,
0.05441175401210785,
-0.0003512483381200582,
0.011662664823234081,
0.01074522640556097,
-0.0676511898636818,
-0.05461103841662407,
0.041136402636766434,
-0.07420017570257187,
0.02212871052324772,
-0.02380220778286457,
0.019270315766334534,
0.0022516348399221897,
0.06823981553316116,
-0.014757035300135612,
0.020506028085947037,
-0.029459452256560326,
-0.07204589247703552,
0.032356902956962585,
-0.03995755687355995,
0.005455416161566973,
0.01869811862707138,
0.11133155971765518,
-0.07545918226242065,
0.03718344867229462,
0.0244640801101923,
0.03602588176727295,
-0.0712449699640274,
-0.031844981014728546,
-0.08803317695856094,
-0.08357655256986618,
-0.036417700350284576,
0.03980853036046028,
-0.013510528020560741,
-0.02990555576980114,
0.0815192461013794,
0.05151451379060745,
0.07391998171806335,
0.03982841596007347,
-0.009021944366395473,
0.08176536858081818,
0.005150125361979008,
0.08322503417730331,
0.0404273197054863,
0.05425012856721878,
0.0677320584654808,
-0.021446406841278076,
0.0618867427110672,
-0.056277379393577576,
-0.029437879100441933,
-0.029049264267086983,
0.03696022555232048,
-0.006825541146099567,
0.04333503916859627,
-0.0015590654220432043,
0.0019855534192174673,
0.0031404702458530664,
0.04306606948375702,
0.0012551581021398306,
-0.07692336291074753,
-0.041978318244218826,
0.005251607391983271,
0.032581690698862076,
0.015931112691760063,
0.05250174552202225,
-0.055707331746816635,
-0.05337522178888321,
-0.05166175216436386,
0.09801151603460312,
0.009604590944945812,
0.09723766148090363,
0.08075409382581711,
-0.0796060860157013,
-0.05078079178929329,
-0.019881755113601685,
0.006518018897622824,
0.025309696793556213,
0.00926514994353056,
-0.06732262670993805,
0.024891218170523643,
0.09479867666959763,
0.06663808971643448,
0.0612834133207798,
0.058040644973516464,
-0.017405303195118904,
0.057070113718509674,
-0.08754171431064606,
-0.0239067655056715,
0.03884873539209366,
-0.02347436174750328,
-0.0006637835176661611,
-0.035327792167663574,
0.013339509256184101,
0.01528462115675211,
0.02754407562315464,
-0.02144043520092964,
2.1695265366821846e-33,
0.0471591018140316,
0.03389937803149223,
0.015812577679753304,
-0.012105652131140232,
0.015122789889574051,
0.026115376502275467,
0.06051123887300491,
0.04277900978922844,
-0.06703796982765198,
0.01293005608022213,
-0.07767714560031891,
0.011144955642521381,
-0.026144614443182945,
0.058965690433979034,
0.01672300137579441,
-0.0225149504840374,
-0.003919977229088545,
0.11630889028310776,
0.0035569050814956427,
-0.08608055114746094,
0.041207462549209595,
0.040637265890836716,
0.008257139474153519,
-0.08568516373634338,
0.08343402296304703,
0.03833828866481781,
0.08953225612640381,
-0.014854678884148598,
-0.15294161438941956,
0.015550736337900162,
-0.02603451907634735,
-0.0358552485704422,
-0.08121097832918167,
0.06455323100090027,
0.04704401642084122,
-0.08054599910974503,
-0.019836558029055595,
-0.04136953875422478,
0.06278185546398163,
-0.021971436217427254,
-0.02565089240670204,
-0.007656881585717201,
0.00006870483775855973,
-0.06089707091450691,
-0.09641243517398834,
-0.0726483166217804,
-0.020365361124277115,
0.012028228491544724,
0.01557494979351759,
-0.014366663992404938,
0.09307573735713959,
-0.01679130457341671,
0.005359138827770948,
0.0207924023270607,
-0.07020562887191772,
-0.040532518178224564,
0.00387173262424767,
0.026410561054944992,
0.06494557112455368,
0.07320056855678558,
-0.014121897518634796,
0.054473597556352615,
-0.06863538920879364,
0.1271895170211792,
0.07325520366430283,
0.05422214791178703,
-0.04610366001725197,
0.07876995205879211,
0.013300853781402111,
-0.00389430345967412,
-0.029451066628098488,
-0.02982030250132084,
-0.014470627531409264,
-0.06560862064361572,
0.07259178906679153,
-0.05847402289509773,
0.1519879251718521,
-0.02139884978532791,
0.05418751761317253,
0.06553717702627182,
-0.012547241523861885,
0.02624671906232834,
0.00029394548619166017,
-0.0465911366045475,
-0.04640117287635803,
-0.04223109781742096,
0.060228921473026276,
-0.076860211789608,
0.00834735855460167,
0.011949286796152592,
0.010839501395821571,
-0.021330591291189194,
-0.03229866549372673,
-0.015243388712406158,
0.014334902167320251,
-2.2030483294897745e-33,
-0.04895484820008278,
0.021411443129181862,
0.01916627585887909,
0.08971872925758362,
-0.0756894126534462,
-0.11887271702289581,
-0.0479765385389328,
0.030953392386436462,
-0.08711138367652893,
-0.07435886561870575,
-0.005149145144969225,
-0.030897686257958412,
0.034857843071222305,
-0.00942085962742567,
0.05075062811374664,
0.0329994298517704,
-0.08259183913469315,
-0.017789281904697418,
0.06395696848630905,
0.054824698716402054,
0.04191382974386215,
0.05935659259557724,
-0.0997517853975296,
0.002656655851751566,
-0.04717226326465607,
-0.00035381800262257457,
0.008431852795183659,
-0.007217636797577143,
0.014367376454174519,
-0.06274570524692535,
-0.027678681537508965,
-0.03943626210093498,
-0.0814107283949852,
0.042383529245853424,
-0.0723889097571373,
0.024415496736764908,
0.05206689238548279,
0.0027531918603926897,
0.04892527684569359,
0.14272870123386383,
0.047630876302719116,
0.003958709537982941,
-0.10421887785196304,
0.01184235792607069,
-0.03554408252239227,
0.013440038077533245,
-0.060240715742111206,
0.006784165743738413,
-0.0501079186797142,
-0.012058163993060589,
-0.006074489559978247,
0.0886150673031807,
-0.010597849264740944,
0.030252093449234962,
-0.010973171330988407,
-0.09670184552669525,
-0.023928921669721603,
-0.019984088838100433,
-0.03893886134028435,
-0.026752013713121414,
-0.025000175461173058,
-0.03454133868217468,
0.02282647229731083,
-0.05122608318924904,
0.021237391978502274,
-0.04699292033910751,
0.06242404133081436,
0.04883314296603203,
-0.013951171189546585,
-0.013417898677289486,
0.07724105566740036,
-0.02470528334379196,
-0.019008532166481018,
0.04454588145017624,
-0.020705100148916245,
0.0016859386814758182,
-0.023219363763928413,
-0.04128704592585564,
-0.027874529361724854,
-0.015575287863612175,
0.004507820121943951,
-0.010644173249602318,
0.004490389488637447,
-0.042104221880435944,
0.058998432010412216,
0.11928625404834747,
0.01776188053190708,
0.06810040771961212,
0.03838631138205528,
0.05236576497554779,
-0.0385490283370018,
0.05819761008024216,
0.046667009592056274,
0.0559203065931797,
-0.012280412018299103,
-4.3425853135659054e-8,
-0.07072585821151733,
0.05026760697364807,
-0.06447965651750565,
0.11515235155820847,
0.049368686974048615,
-0.024607717990875244,
-0.058902107179164886,
0.09647320955991745,
-0.017304323613643646,
0.003967352211475372,
0.08822654187679291,
-0.04649979621171951,
-0.055928148329257965,
-0.02777278423309326,
0.01564887911081314,
0.03291090577840805,
-0.015208384022116661,
0.009739495813846588,
-0.05094769969582558,
-0.02455732226371765,
-0.0010011515114456415,
0.07328864187002182,
-0.03456582501530647,
0.0733472928404808,
0.045865029096603394,
-0.017842506989836693,
0.017235668376088142,
0.11715894937515259,
0.04274992644786835,
-0.003340168157592416,
-0.003646000986918807,
0.028862424194812775,
-0.0597817525267601,
-0.06480734050273895,
-0.0022824942134320736,
0.0877537652850151,
0.06748836487531662,
-0.01852431707084179,
-0.013135765679180622,
-0.005231392104178667,
-0.019457537680864334,
0.012557053938508034,
-0.05415477231144905,
0.04453650116920471,
-0.005179579835385084,
-0.08209069818258286,
-0.08382211625576019,
-0.07363066077232361,
0.02618023008108139,
-0.10538055747747421,
0.02686687558889389,
-0.03927944228053093,
-0.047785643488168716,
0.05662444606423378,
-0.026804696768522263,
0.0666780173778534,
-0.015968622639775276,
0.004344250541180372,
-0.03314495086669922,
0.0494760237634182,
0.08904804289340973,
-0.05143411085009575,
-0.09945900738239288,
-0.04124419018626213
] | https://github.com/qi-le1/probe_pruning | 5 | 0 | 0 | 0 |
AgentOccam: A Simple Yet Strong Baseline for LLM-Based Web Agents | https://openreview.net/forum?id=oWdzUpOlkX | [
"Ke Yang",
"Yao Liu",
"Sapana Chaudhary",
"Rasool Fakoor",
"Pratik Chaudhari",
"George Karypis",
"Huzefa Rangwala"
] | Poster | Autonomy via agents based on large language models (LLMs) that can carry out personalized yet standardized tasks presents a significant opportunity to drive human efficiency. There is an emerging need and interest in automating web tasks (e.g., booking a hotel for a given date within a budget). Being a practical use case itself, the web agent also serves as an important proof-of-concept example for various agent grounding scenarios, with its success promising advancements in many future applications. Meanwhile, much prior research focuses on handcrafting their web agent strategies (e.g., agent's prompting templates, reflective workflow, role-play and multi-agent systems, search or sampling methods, etc.) and the corresponding in-context examples. However, these custom strategies often struggle with generalizability across all potential real-world applications. On the other hand, there has been limited study on the misalignment between a web agent's observation and action representation, and the data on which the agent's underlying LLM has been pre-trained. This discrepancy is especially notable when LLMs are primarily trained for language completion rather than tasks involving embodied navigation actions and symbolic web elements. In our study, we enhance an LLM-based web agent by simply refining its observation and action space, aligning these more closely with the LLM's capabilities. This approach enables our base agent to significantly outperform previous methods on a wide variety of web tasks. Specifically, on WebArena, a benchmark featuring general-purpose web interaction tasks, our agent AgentOccam surpasses the previous state-of-the-art and concurrent work by 9.8 (+29.4%) and 5.9 (+15.8%) absolute points respectively, and boosts the success rate by 26.6 points (+161%) over similar plain web agents with its observation and action space alignment. Furthermore, on WebVoyager benchmark comprising tasks defined on real-world websites, AgentOccam exceeds the former best agent by 2.4 points (+4.6%) on tasks with deterministic answers. We achieve this without using in-context examples, new agent roles, online feedback or search strategies. AgentOccam's simple design highlights LLMs' impressive zero-shot performance on web tasks, and underlines the critical role of carefully tuning observation and action spaces for LLM-based agents. | LLM, Agent, LLM-based Agent, Web Agent, Web Navigation | Our method, with new SOTA results on WebArena, demonstrates a strong zero-shot performance on web tasks and sheds light on the importance of observation and action space choice for LLM-based agents. | 12,731 | 2410.13825 | [
-0.015458882786333561,
-0.07449989765882492,
-0.013914646580815315,
-0.0042100874707102776,
0.05140698701143265,
-0.029027976095676422,
0.08011653274297714,
-0.0028784372843801975,
0.02336745895445347,
-0.05225605145096779,
0.0014445725828409195,
-0.04419375956058502,
0.03665746748447418,
0.08419139683246613,
0.14118321239948273,
-0.06357060372829437,
0.16024722158908844,
0.003658723784610629,
-0.0650508776307106,
0.024950804188847542,
0.01587524078786373,
-0.03536234796047211,
0.0393073745071888,
-0.039908140897750854,
-0.029567992314696312,
-0.043123263865709305,
-0.00954735092818737,
-0.03636309131979942,
0.03671102970838547,
-0.027257926762104034,
0.04368359223008156,
0.11095715314149857,
0.01936587691307068,
0.057430658489465714,
0.02632889896631241,
0.07160899043083191,
-0.12484293431043625,
0.00662752753123641,
-0.023776259273290634,
0.013369067572057247,
-0.07290804386138916,
-0.03244679793715477,
0.008773069828748703,
-0.0021578879095613956,
0.011374853551387787,
-0.013090156950056553,
-0.015080682933330536,
0.03205536678433418,
-0.08432506024837494,
0.015627197921276093,
-0.12235096842050552,
-0.07178506255149841,
-0.0028010846581310034,
-0.032663147896528244,
-0.06566019356250763,
-0.030462127178907394,
0.02963995933532715,
0.00961926020681858,
-0.017111150547862053,
-0.07136977463960648,
-0.0085627855733037,
-0.0364069901406765,
-0.01311806682497263,
0.023488864302635193,
-0.028461067005991936,
0.02998659946024418,
0.006275311578065157,
-0.00956825539469719,
0.0003993886348325759,
-0.05729515105485916,
-0.04027743637561798,
-0.010160739533603191,
-0.02109096199274063,
0.029338505119085312,
-0.0038833110593259335,
-0.05609662085771561,
0.005321923177689314,
0.030056703835725784,
0.030977308750152588,
-0.06500861048698425,
-0.032812610268592834,
0.002018526429310441,
-0.009143827483057976,
0.08642195910215378,
0.011493983678519726,
0.04770494997501373,
0.00943189486861229,
0.08017530292272568,
0.055728815495967865,
-0.03633951023221016,
-0.025476116687059402,
-0.09846588969230652,
-0.0006047783535905182,
-0.024585841223597527,
0.0333113819360733,
0.06731827557086945,
0.011808397248387337,
-0.019814394414424896,
-0.0243753120303154,
0.08010223507881165,
0.06117486581206322,
0.027802444994449615,
-0.0015660757198929787,
-0.023465115576982498,
0.002731016371399164,
0.01907702349126339,
0.0252654030919075,
0.06349033117294312,
0.051348742097616196,
-0.06559175997972488,
-0.08072110265493393,
0.010930978693068027,
0.042610108852386475,
-0.001557844807393849,
0.03244279325008392,
-0.04492942616343498,
0.03628672659397125,
0.010737648233771324,
0.04283139482140541,
0.05552414804697037,
0.06458910554647446,
0.019579650834202766,
0.08046045154333115,
-0.004444350488483906,
0.07619385421276093,
0.005335014313459396,
0.028001390397548676,
5.336130432441369e-34,
0.08892785757780075,
0.029037633910775185,
-0.024301527068018913,
0.057768747210502625,
0.10481933504343033,
-0.018647611141204834,
0.03032001666724682,
0.04543289542198181,
-0.018463602289557457,
0.0011114721419289708,
0.046151019632816315,
0.061817701905965805,
-0.0216508898884058,
0.06988275796175003,
-0.013104068115353584,
0.052763938903808594,
-0.03937732055783272,
0.04469984024763107,
0.014482330530881882,
-0.0704926997423172,
0.08116284012794495,
-0.004765399266034365,
0.1014833077788353,
0.011950717307627201,
0.026549477130174637,
0.03098207525908947,
-0.02093825489282608,
0.03070032224059105,
0.05186294764280319,
0.0012102583423256874,
-0.0187095757573843,
-0.042041175067424774,
-0.07827716320753098,
0.059738967567682266,
-0.03230367600917816,
0.037688035517930984,
-0.06949713081121445,
-0.08293554931879044,
0.023904407396912575,
-0.02622227743268013,
-0.09128966927528381,
0.039386142045259476,
0.0036666791420429945,
-0.020539795979857445,
-0.11325836181640625,
-0.1247333213686943,
-0.02532297559082508,
0.0077095492742955685,
-0.021907957270741463,
-0.0122542018070817,
0.01769227720797062,
0.04222744330763817,
0.030053943395614624,
0.006071337498724461,
-0.039258070290088654,
0.02328888326883316,
0.051507867872714996,
0.011604001745581627,
-0.022468138486146927,
-0.023693321272730827,
0.011570995673537254,
-0.029150133952498436,
-0.03629882261157036,
0.10186494141817093,
0.045082658529281616,
-0.05081716552376747,
0.02963854745030403,
0.05831228941679001,
0.07601238042116165,
-0.06811904162168503,
-0.00934233795851469,
0.006248297169804573,
0.01000200491398573,
-0.005718113388866186,
-0.014117035083472729,
-0.05008719116449356,
0.04654237627983093,
-0.0880943313241005,
0.02197352983057499,
0.013903114944696426,
0.03440876305103302,
0.04698817804455757,
0.029834380373358727,
-0.009669973514974117,
-0.043994612991809845,
-0.08355490118265152,
-0.012904077768325806,
-0.1426842212677002,
-0.016304872930049896,
0.06871850043535233,
-0.07696517556905746,
0.009944764897227287,
-0.09450193494558334,
0.05796140432357788,
0.052657268941402435,
-2.0045596798687002e-33,
-0.06769002228975296,
-0.06291233003139496,
-0.07408402860164642,
0.05027642473578453,
0.0016781574813649058,
-0.06545884907245636,
0.0066632856614887714,
-0.051188092678785324,
0.019857564941048622,
-0.0588158518075943,
-0.07243531197309494,
-0.007333288434892893,
0.020236534997820854,
0.027044741436839104,
-0.042656850069761276,
-0.05999720096588135,
0.019966796040534973,
-0.06446868181228638,
0.03947216644883156,
0.04566698148846626,
0.027708111330866814,
0.07759124040603638,
-0.10782168805599213,
-0.009825363755226135,
0.031837861984968185,
0.0352923758327961,
-0.0469389371573925,
0.04419747367501259,
-0.0011695903958752751,
-0.013214345090091228,
0.02794366143643856,
-0.04273325577378273,
-0.020821761339902878,
-0.04043879732489586,
-0.052459411323070526,
0.06478264182806015,
-0.044238608330488205,
0.10221990197896957,
-0.031301349401474,
0.12764699757099152,
0.09258702397346497,
-0.09004156291484833,
-0.03512929752469063,
-0.026207681745290756,
-0.03585335239768028,
0.02016032114624977,
-0.12767094373703003,
0.019503267481923103,
0.06942134350538254,
-0.019753817468881607,
-0.049058470875024796,
0.017065774649381638,
-0.07149887830018997,
-0.030872609466314316,
-0.03351375833153725,
0.01354415062814951,
0.04690197482705116,
-0.06361442804336548,
-0.006579405162483454,
-0.037060678005218506,
-0.05274096503853798,
-0.004233851097524166,
0.056581031531095505,
-0.025357892736792564,
-0.030792463570833206,
-0.024071935564279556,
-0.04952501505613327,
0.025582319125533104,
-0.023475687950849533,
-0.0613730289041996,
0.04895608872175217,
-0.03830866515636444,
-0.07145193219184875,
-0.003391593461856246,
0.05664695054292679,
-0.01262197270989418,
0.013652623631060123,
-0.02776055969297886,
0.0413418672978878,
-0.09012101590633392,
-0.028530579060316086,
0.02434271015226841,
-0.001131524215452373,
0.03904100880026817,
-0.02280343696475029,
0.06772474944591522,
-0.036730434745550156,
0.0840236097574234,
0.01351763866841793,
-0.03391338139772415,
0.02690146304666996,
0.02690456621348858,
-0.04353385791182518,
0.07715340703725815,
-0.06830082833766937,
-5.071111175425358e-8,
-0.09813546389341354,
0.05232560634613037,
0.05627933889627457,
0.02563152089715004,
0.021655481308698654,
-0.04793412983417511,
0.030781451612710953,
0.03819151595234871,
0.01385498046875,
-0.020366644486784935,
0.05934663861989975,
-0.057253409177064896,
-0.008973365649580956,
-0.027656812220811844,
0.037260230630636215,
0.044525664299726486,
-0.009946654550731182,
-0.010501415468752384,
-0.04190799966454506,
-0.030579697340726852,
0.07776089757680893,
-0.02504342794418335,
-0.0028761972207576036,
-0.03446110337972641,
0.05031164735555649,
0.00078146334271878,
-0.07936316728591919,
0.11587357521057129,
-0.007227844092994928,
0.048504408448934555,
0.010095219127833843,
0.015997104346752167,
0.006932515185326338,
-0.014872854575514793,
-0.002233649604022503,
0.03678428381681442,
-0.03727905824780464,
-0.10571695864200592,
0.0044531975872814655,
-0.05270479992032051,
0.04501227289438248,
0.04595712944865227,
-0.06505192816257477,
0.029072869569063187,
0.07575942575931549,
0.04133334010839462,
-0.05483143404126167,
-0.12237637490034103,
0.06616304069757462,
-0.05489083006978035,
-0.07224465161561966,
-0.045756399631500244,
-0.0077421157620847225,
0.11552543938159943,
0.09351740032434464,
0.04834512621164322,
0.10284367948770523,
-0.05597420781850815,
0.11260123550891876,
0.09614036977291107,
-0.007239527069032192,
0.03249724581837654,
-0.055067937821149826,
0.009962962940335274
] | 0 | 0 | 0 | 0 |
|
Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF | https://openreview.net/forum?id=SQnitDuow6 | [
"Shicong Cen",
"Jincheng Mei",
"Katayoon Goshvadi",
"Hanjun Dai",
"Tong Yang",
"Sherry Yang",
"Dale Schuurmans",
"Yuejie Chi",
"Bo Dai"
] | Poster | Reinforcement learning from human feedback (RLHF) has demonstrated great promise in aligning large language models (LLMs) with human preference. Depending on the availability of preference data, both online and offline RLHF are active areas of investigation. A key bottleneck is understanding how to incorporate uncertainty estimation in the reward function learned from the preference data for RLHF, regardless of how the preference data is collected. While the principles of optimism or pessimism under uncertainty are well-established in standard reinforcement learning (RL), a practically-implementable and theoretically-grounded form amenable to large language models is not yet available, as standard techniques for constructing confidence intervals become intractable under arbitrary policy parameterizations.
In this paper, we introduce a unified approach to online and offline RLHF --- value-incentivized preference optimization (VPO) --- which regularizes the maximum-likelihood estimate of the reward function with the corresponding value function, modulated by a sign to indicate whether the optimism or pessimism is chosen. VPO also directly optimizes the policy with implicit reward modeling, and therefore shares a simpler RLHF pipeline similar to direct preference optimization. Theoretical guarantees of VPO are provided for both online and offline settings, matching the rates of their standard RL counterparts. Moreover, experiments on text summarization, dialogue, and standard benchmarks verify the practicality and effectiveness of VPO. | preference optimization, the principle of optimism/pessimism, RLHF theory | Principled and practical exploration for preference optimization (e.g., RLHF) can be achieved without requiring explicit uncertainty modelling. | 12,730 | 2405.19320 | [
-0.07789848744869232,
-0.04623465985059738,
-0.00749646220356226,
0.0284125916659832,
-0.0003042219905182719,
0.0746007040143013,
0.009011315181851387,
0.09289006143808365,
0.052454378455877304,
-0.009666940197348595,
-0.03408540412783623,
-0.05955076962709427,
0.09806397557258606,
0.01563059166073799,
0.06279357522726059,
0.025759737938642502,
0.08119390904903412,
0.0008754122536629438,
-0.09919018298387527,
-0.05573827028274536,
-0.004276654217392206,
0.027210796251893044,
-0.01757865771651268,
-0.0192337017506361,
0.04862580448389053,
-0.0545034185051918,
-0.010761311277747154,
0.03493355214595795,
0.0019032404525205493,
0.059804681688547134,
0.013267633505165577,
-0.04123508185148239,
0.05816648155450821,
-0.10327332466840744,
-0.045792628079652786,
0.022936252877116203,
-0.15145157277584076,
-0.0645400807261467,
-0.028174545615911484,
-0.029740888625383377,
-0.062417082488536835,
-0.038184478878974915,
0.026226703077554703,
0.012526025995612144,
0.04022656008601189,
0.02389996312558651,
0.020038796588778496,
-0.02290881983935833,
-0.033990319818258286,
0.026648886501789093,
0.009139073081314564,
-0.042514994740486145,
0.010082239285111427,
0.016991080716252327,
-0.014092907309532166,
0.023158421739935875,
-0.00029327618540264666,
0.01780642196536064,
0.0030015239026397467,
-0.045817647129297256,
-0.0833229124546051,
-0.1253819316625595,
0.007731495425105095,
0.02229285053908825,
-0.05834940820932388,
-0.031685344874858856,
0.002410757588222623,
0.05570027977228165,
-0.03625155985355377,
0.05654005706310272,
-0.00754511309787631,
0.07181771099567413,
0.04030971601605415,
-0.009456897154450417,
0.00980080384761095,
-0.00473514199256897,
0.04478083550930023,
-0.03962619975209236,
0.02774382382631302,
0.03126293793320656,
-0.05642574280500412,
0.05327162519097328,
0.009904231876134872,
-0.023208661004900932,
0.05462390556931496,
0.029792917892336845,
0.037857431918382645,
0.08746127784252167,
0.10639242082834244,
-0.008597156964242458,
-0.005645992234349251,
-0.011489380151033401,
-0.034063421189785004,
0.0022460261825472116,
-0.060270726680755615,
0.051356781274080276,
-0.09515086561441422,
-0.09520047903060913,
-0.0856567993760109,
0.07646948844194412,
0.04551738128066063,
0.056365687400102615,
-0.02293366938829422,
-0.06844782084226608,
-0.04732964187860489,
-0.04589783400297165,
-0.05408013239502907,
0.0818033441901207,
0.04244771972298622,
-0.08531387895345688,
-0.047273240983486176,
0.018750684335827827,
0.03485264256596565,
0.06452923268079758,
-0.009478691965341568,
0.01489800401031971,
0.06678294390439987,
-0.048643093556165695,
0.029382992535829544,
0.01618833653628826,
0.014023547060787678,
-0.05546083673834801,
0.10858559608459473,
0.014127789065241814,
-0.012265377677977085,
0.01405999157577753,
0.02112974226474762,
2.265580955533674e-33,
0.05200105533003807,
0.043901003897190094,
0.07515470683574677,
-0.011257102712988853,
0.06060251593589783,
0.022465066984295845,
0.02991306595504284,
0.018378971144557,
-0.08412806689739227,
-0.0575486496090889,
0.030339550226926804,
0.06458069384098053,
-0.028104182332754135,
0.05662894621491432,
0.038504622876644135,
0.00023761419288348407,
0.008441795594990253,
0.02481825463473797,
0.012180780060589314,
-0.048798490315675735,
0.09341850876808167,
-0.004895162768661976,
0.05449927970767021,
-0.04543408378958702,
0.051439810544252396,
0.02271394617855549,
0.06101928651332855,
-0.03180266171693802,
-0.017783498391509056,
-0.027215832844376564,
-0.0220298171043396,
-0.026224719360470772,
0.02607821486890316,
-0.02465985156595707,
-0.031540125608444214,
-0.003969475626945496,
-0.1072017177939415,
-0.0277925543487072,
-0.009873315691947937,
-0.04342249780893326,
0.06665734946727753,
0.03734555467963219,
0.043058354407548904,
0.06049293652176857,
-0.05017509311437607,
-0.08629774302244186,
0.04461169242858887,
-0.015816636383533478,
-0.07145298272371292,
-0.041756849735975266,
0.02818736620247364,
0.048109788447618484,
-0.017426518723368645,
0.00645030802115798,
-0.04501630738377571,
-0.02594550885260105,
-0.03621026501059532,
0.09690801799297333,
-0.03952410817146301,
-0.05597050487995148,
-0.008730718865990639,
0.028965502977371216,
0.05867135152220726,
-0.03882881999015808,
-0.008545669727027416,
0.11354322731494904,
-0.06757666170597076,
-0.02024810016155243,
0.01592390611767769,
0.00491733942180872,
0.031126050278544426,
0.03715472295880318,
0.0033779500517994165,
-0.0652734637260437,
0.01843826286494732,
0.007156480103731155,
0.02192836068570614,
-0.03595777228474617,
0.09001792967319489,
-0.036036211997270584,
-0.010600720532238483,
-0.04372260347008705,
-0.019804030656814575,
-0.04652326926589012,
0.027105776593089104,
-0.10769573599100113,
0.035312261432409286,
-0.039711154997348785,
-0.012622208334505558,
-0.005378073081374168,
-0.03204193338751793,
-0.005847324151545763,
0.00045254401629790664,
0.04540078714489937,
0.04873737320303917,
-3.180507514371016e-33,
-0.035799071192741394,
-0.07480952888727188,
0.027992039918899536,
0.051283206790685654,
-0.003998501691967249,
0.0011877270881086588,
-0.0011509835021570325,
-0.009339849464595318,
-0.01344283390790224,
-0.04927506670355797,
-0.08856532722711563,
0.0463712252676487,
0.07815514504909515,
0.02113146148622036,
0.013008153066039085,
-0.06875155866146088,
0.016826560720801353,
0.019831154495477676,
0.07973062992095947,
0.01989457756280899,
0.00019163929391652346,
0.06003136560320854,
-0.09888289123773575,
0.12729567289352417,
-0.014755665324628353,
0.01566055603325367,
0.018579479306936264,
0.01834123395383358,
-0.07100872695446014,
0.004799975082278252,
-0.08783115446567535,
-0.05848272144794464,
-0.08131366223096848,
-0.053038373589515686,
-0.011341582983732224,
0.0484410859644413,
0.03259958326816559,
0.02072843350470066,
-0.09741014242172241,
0.13915051519870758,
0.032594356685876846,
-0.05530232563614845,
-0.05753196403384209,
-0.06725643575191498,
-0.040527842938899994,
-0.0004309286014176905,
-0.072413869202137,
-0.03776884078979492,
-0.0014761011116206646,
-0.027952220290899277,
-0.0056958794593811035,
0.02698979526758194,
-0.04851020500063896,
0.06824660301208496,
-0.053039778023958206,
-0.05931574106216431,
-0.07884424924850464,
-0.07090217620134354,
0.08157816529273987,
0.016163524240255356,
-0.10166383534669876,
0.03726978227496147,
-0.01407388225197792,
0.028891408815979958,
-0.04650198295712471,
-0.000048873778723645955,
0.00793809350579977,
0.027334172278642654,
-0.028911592438817024,
-0.01578420028090477,
0.029777228832244873,
-0.008720660582184792,
0.06328576058149338,
-0.04231691360473633,
0.004353316500782967,
-0.0010848419042304158,
0.026521172374486923,
0.026476910337805748,
0.023845940828323364,
-0.008989894762635231,
0.08519665896892548,
0.07578730583190918,
0.08965320885181427,
0.04704437032341957,
0.049066945910453796,
0.05870334059000015,
0.03902887925505638,
0.15064001083374023,
0.0041038235649466515,
-0.02845834009349346,
-0.008437703363597393,
-0.005289315711706877,
0.016415594145655632,
0.03755458816885948,
0.004328318405896425,
-4.4334097282217044e-8,
-0.06379861384630203,
0.0025718414690345526,
0.10686243325471878,
0.05028378590941429,
0.029813962057232857,
-0.03852483630180359,
0.032367952167987823,
-0.05825437232851982,
-0.021526077762246132,
-0.02140599675476551,
0.025378582999110222,
0.07966400682926178,
0.0057222116738557816,
-0.15561212599277496,
-0.06782695651054382,
0.11074449867010117,
0.012086262926459312,
-0.010593394748866558,
0.044017575681209564,
-0.007340625859797001,
0.09336233884096146,
-0.023834457620978355,
-0.014322462491691113,
-0.06950999796390533,
0.06841576099395752,
-0.04218432307243347,
-0.017802022397518158,
0.12096616625785828,
0.01016130018979311,
0.005361681804060936,
0.007006523665040731,
-0.0952114388346672,
0.004195020068436861,
0.015063448809087276,
0.09369762241840363,
0.056156739592552185,
0.03954295441508293,
-0.05487300083041191,
0.014093715697526932,
-0.06167501583695412,
0.03796817362308502,
0.07031428068876266,
-0.03669395297765732,
-0.0451919250190258,
0.017832500860095024,
0.017262868583202362,
-0.062335528433322906,
-0.11500155925750732,
0.0013716332614421844,
-0.0696922168135643,
0.03548521175980568,
-0.015112612396478653,
-0.061078209429979324,
0.08068989217281342,
0.023553946986794472,
-0.05012648552656174,
-0.017444374039769173,
-0.060500480234622955,
-0.030392104759812355,
0.06129520758986473,
0.07434267550706863,
-0.044909343123435974,
-0.0687478631734848,
0.012517369352281094
] | 0 | 0 | 0 | 0 |
|
Beyond-Expert Performance with Limited Demonstrations: Efficient Imitation Learning with Double Exploration | https://openreview.net/forum?id=FviefuxmeW | [
"Heyang Zhao",
"Xingrui Yu",
"David Mark Bossens",
"Ivor Tsang",
"Quanquan Gu"
] | Poster | Imitation learning is a central problem in reinforcement learning where the goal is to learn a policy that mimics the expert's behavior. In practice, it is often challenging to learn the expert policy from a limited number of demonstrations accurately due to the complexity of the state space. Moreover, it is essential to explore the environment and collect data to achieve beyond-expert performance. To overcome these challenges, we propose a novel imitation learning algorithm called Imitation Learning with Double Exploration (ILDE), which implements exploration in two aspects: (1) optimistic policy optimization via an exploration bonus that rewards state-action pairs with high uncertainty to potentially improve the convergence to the expert policy, and (2) curiosity-driven exploration of the states that deviate from the demonstration trajectories to potentially yield beyond-expert performance. Empirically, we demonstrate that ILDE outperforms the state-of-the-art imitation learning algorithms in terms of sample efficiency and achieves beyond-expert performance on Atari and MuJoCo tasks with fewer demonstrations than in previous work. We also provide a theoretical justification of ILDE as an uncertainty-regularized policy optimization method with optimistic exploration, leading to a regret growing sublinearly in the number of episodes. | Optimistic exploration, Curiosity, Imitation learning, Inverse reinforcement learning, Reinforcement learning | null | 12,729 | null | [
-0.03962469473481178,
-0.055944278836250305,
0.09856110066175461,
0.013971356675028801,
-0.023100562393665314,
-0.03522805869579315,
0.03179718554019928,
-0.006493457593023777,
0.006947822403162718,
0.04021334648132324,
-0.04503215476870537,
-0.012816318310797215,
0.06671345233917236,
0.011997895315289497,
-0.026949824765324593,
0.04962639510631561,
0.06316325068473816,
0.04260234162211418,
-0.027665020897984505,
-0.008096235804259777,
0.029315032064914703,
-0.04632851108908653,
0.06775885820388794,
-0.03158904239535332,
0.027561718598008156,
0.007755571510642767,
-0.005953066051006317,
-0.004105092491954565,
0.00537507189437747,
0.01227460615336895,
0.01815730705857277,
-0.05222330614924431,
0.03600584343075752,
-0.10035552084445953,
-0.06949010491371155,
0.049401067197322845,
-0.12007804214954376,
0.011405441910028458,
0.03436993807554245,
-0.011465621180832386,
-0.00012416686513461173,
0.024774089455604553,
0.0016345193143934011,
0.015875080600380898,
0.038803163915872574,
0.027899907901883125,
0.04120898246765137,
-0.09834936261177063,
0.10889264941215515,
0.030898546800017357,
-0.011744402348995209,
-0.0429823212325573,
0.018574852496385574,
-0.08462872356176376,
0.03623908758163452,
0.07967592775821686,
0.004251935984939337,
0.03477600961923599,
0.013439936563372612,
-0.023850884288549423,
-0.0023071470204740763,
-0.09968648850917816,
0.01029008999466896,
-0.04236040636897087,
0.007433570921421051,
-0.048981692641973495,
-0.01921471394598484,
0.0596865639090538,
0.008518599905073643,
0.060232751071453094,
-0.03976799175143242,
0.04872400313615799,
0.019450051710009575,
0.022130433470010757,
0.01885305717587471,
-0.048175692558288574,
0.016376065090298653,
-0.045705560594797134,
-0.039642054587602615,
-0.055621422827243805,
0.004311443772166967,
-0.04105183854699135,
-0.020413821563124657,
0.02660559117794037,
0.03348074480891228,
-0.0073369150049984455,
-0.0028893002308905125,
0.030981840565800667,
0.1634529083967209,
-0.0267946720123291,
-0.018834879621863365,
-0.03620418533682823,
-0.04128841310739517,
-0.022857697680592537,
0.03945879265666008,
0.024675238877534866,
-0.029222631826996803,
-0.032607439905405045,
-0.050518907606601715,
0.07390745729207993,
0.04582171514630318,
0.05522201582789421,
0.00941962469369173,
-0.014800317585468292,
-0.04270365089178085,
-0.09827499836683273,
0.01594618149101734,
0.0035866890102624893,
0.04835936427116394,
-0.023138847202062607,
0.046404656022787094,
-0.024239489808678627,
0.09082998335361481,
0.06961460411548615,
0.0013186956057325006,
0.03250923380255699,
-0.016569601371884346,
0.004381822422146797,
-0.028187913820147514,
0.08866381645202637,
-0.00625515216961503,
-0.04885590821504593,
0.09651322662830353,
0.04566138982772827,
0.015191270038485527,
-0.019891276955604553,
0.00865523423999548,
-1.589233097481208e-33,
0.05851643905043602,
0.015388375148177147,
0.08387918025255203,
-0.0030772704631090164,
0.01918899454176426,
0.02914826013147831,
0.0657043531537056,
-0.027354605495929718,
-0.09906366467475891,
-0.0035415291786193848,
0.05550025776028633,
-0.01315446849912405,
-0.04119483381509781,
0.07515702396631241,
0.03277108818292618,
-0.023278187960386276,
-0.05727672949433327,
0.010331560857594013,
-0.007729053497314453,
-0.02672659605741501,
0.063008151948452,
-0.040156640112400055,
-0.030543537810444832,
-0.05570521205663681,
0.028307458385825157,
-0.039697278290987015,
0.025256182998418808,
0.0067129177041351795,
0.008743601851165295,
-0.006080884952098131,
-0.11261892318725586,
0.010196804068982601,
-0.060533471405506134,
0.057655394077301025,
0.04944176599383354,
0.040976688265800476,
0.03679361566901207,
-0.03113744594156742,
-0.045699428766965866,
0.002074474934488535,
-0.016357380896806717,
-0.01763211563229561,
0.07073023915290833,
-0.00690508633852005,
-0.10912378132343292,
-0.12182416021823883,
0.04099239408969879,
-0.016971515491604805,
-0.03364000469446182,
-0.03716824948787689,
-0.00535714952275157,
0.010403184220194817,
-0.030897248536348343,
-0.10479085147380829,
-0.01579117402434349,
-0.0299930851906538,
0.027407409623265266,
0.02534204162657261,
-0.029503248631954193,
0.019894106313586235,
-0.02554928883910179,
0.018916117027401924,
-0.015069594606757164,
0.06897444278001785,
-0.06043248251080513,
0.1637704074382782,
-0.0507478341460228,
0.05827333778142929,
0.06470894068479538,
0.02661203034222126,
-0.0128905875608325,
-0.019951483234763145,
0.0011775677558034658,
-0.13597357273101807,
0.11278627812862396,
0.0076457420364022255,
0.025357453152537346,
-0.057897694408893585,
0.032227788120508194,
-0.02817116118967533,
-0.05554291605949402,
-0.047211457043886185,
-0.09495126456022263,
-0.033740997314453125,
-0.00902298092842102,
-0.05891317129135132,
0.017936598509550095,
-0.04560428485274315,
0.001613290631212294,
0.018747229129076004,
-0.13513177633285522,
-0.03327377513051033,
-0.038552992045879364,
0.04082648083567619,
0.01762620359659195,
1.0512167029084137e-34,
0.059719592332839966,
-0.05999871715903282,
0.05981593579053879,
0.04070655629038811,
-0.017912840470671654,
-0.028646821156144142,
-0.012052989564836025,
-0.05761408433318138,
-0.061082422733306885,
-0.09668556600809097,
-0.07518704980611801,
0.05426096171140671,
0.09567739814519882,
0.012840662151575089,
-0.018972450867295265,
0.007186632603406906,
-0.00595448911190033,
0.08836450427770615,
0.03369716554880142,
-0.003290470689535141,
0.012353284284472466,
0.05177510157227516,
-0.05564184486865997,
0.07596658170223236,
-0.034225303679704666,
0.060650747269392014,
-0.005111086647957563,
0.12393444031476974,
0.0257200188934803,
0.003973189741373062,
-0.06015933305025101,
-0.05444817245006561,
0.01879943534731865,
0.018872464075684547,
-0.06353465467691422,
0.1337696760892868,
0.0795707106590271,
0.07331494241952896,
-0.1080731451511383,
0.046355072408914566,
-0.0038142164703458548,
-0.05492793396115303,
0.015372555702924728,
0.011760789901018143,
0.0016313696978613734,
-0.02666846290230751,
-0.009523600339889526,
-0.01892703026533127,
-0.02260597236454487,
-0.004199325572699308,
-0.02748536318540573,
0.0333881638944149,
-0.09111116081476212,
0.009138260036706924,
-0.047872256487607956,
-0.05004517361521721,
-0.04720674827694893,
0.04024621844291687,
0.055594224482774734,
0.010683692060410976,
-0.013031847774982452,
0.013621206395328045,
-0.050647348165512085,
0.055803973227739334,
-0.03153696283698082,
0.0573127418756485,
-0.0492168627679348,
0.0706525593996048,
0.017022697255015373,
-0.011104388162493706,
-0.026050416752696037,
0.014653244987130165,
-0.008686492219567299,
-0.05656207352876663,
0.055206298828125,
-0.03676455095410347,
0.04020432010293007,
0.02951882965862751,
-0.00791048165410757,
0.01405105646699667,
0.05660521239042282,
0.015256932936608791,
0.07516313344240189,
0.03592343255877495,
0.014541996642947197,
0.12620380520820618,
-0.064923495054245,
0.144452303647995,
-0.016998425126075745,
-0.012120288796722889,
-0.016202423721551895,
-0.013437834568321705,
0.015255953185260296,
-0.016082335263490677,
-0.0029665932524949312,
-3.882863097715017e-8,
-0.13757385313510895,
0.025863520801067352,
0.1345706731081009,
0.005588583182543516,
0.08251707255840302,
0.002328030299395323,
-0.047119542956352234,
-0.056186478585004807,
-0.05593287572264671,
-0.0899684876203537,
0.056321997195482254,
0.0312616191804409,
0.01386625412851572,
-0.07923191040754318,
-0.006465525832027197,
0.02715849131345749,
0.01736474595963955,
-0.03409372270107269,
-0.003300985088571906,
0.04295492172241211,
0.008702573366463184,
-0.02841537818312645,
0.01985440030694008,
-0.05312979221343994,
-0.001811634749174118,
-0.038382638245821,
-0.04627024754881859,
0.030463891103863716,
0.08682603389024734,
0.0030547683127224445,
-0.030177516862750053,
-0.05516182258725166,
0.06946586817502975,
-0.0022336854599416256,
0.016267728060483932,
-0.004207481164485216,
0.06269844621419907,
-0.051627956330776215,
0.007189867552369833,
-0.09312839061021805,
-0.018777716904878616,
0.028755949810147285,
0.027842586860060692,
-0.0018120553577318788,
-0.018728965893387794,
0.004645076114684343,
-0.05252623185515404,
-0.10018902271986008,
0.0638621523976326,
-0.057799916714429855,
0.05191510170698166,
-0.027748169377446175,
-0.012800519354641438,
0.1236179992556572,
0.10251117497682571,
0.028330715373158455,
-0.06859590858221054,
-0.010671090334653854,
-0.02462160587310791,
0.11679666489362717,
0.08222656697034836,
-0.017559893429279327,
-0.09040196985006332,
-0.027103673666715622
] | 0 | 0 | 0 | 0 |
|
ORSO: Accelerating Reward Design via Online Reward Selection and Policy Optimization | https://openreview.net/forum?id=0uRc3CfJIQ | [
"Chen Bo Calvin Zhang",
"Zhang-Wei Hong",
"Aldo Pacchiano",
"Pulkit Agrawal"
] | Poster | Reward shaping is critical in reinforcement learning (RL), particularly for complex tasks where sparse rewards can hinder learning. However, choosing effective shaping rewards from a set of reward functions in a computationally efficient manner remains an open challenge. We propose Online Reward Selection and Policy Optimization (ORSO), a novel approach that frames the selection of shaping reward function as an online model selection problem. ORSO automatically identifies performant shaping reward functions without human intervention with provable regret guarantees. We demonstrate ORSO's effectiveness across various continuous control tasks. Compared to prior approaches, ORSO significantly reduces the amount of data required to evaluate a shaping reward function, resulting in superior data efficiency and a significant reduction in computational time (up to 8×). ORSO consistently identifies high-quality reward functions outperforming prior methods by more than 50% and on average identifies policies as performant as the ones learned using manually engineered reward functions by domain experts. | Reinforcement Learning, Reward Design, Reward Selection | We present ORSO, a method for efficiently selecting shaping rewards in reinforcement learning. | 12,721 | 2410.13837 | [
-0.11117365956306458,
-0.023345494642853737,
0.02127961441874504,
0.04694706201553345,
0.005886164028197527,
0.019261112436652184,
-0.026294250041246414,
0.019629403948783875,
0.028992628678679466,
0.0440039336681366,
-0.0387689433991909,
-0.07082907110452652,
0.061419472098350525,
0.042979896068573,
0.020463978871703148,
0.052669983357191086,
0.08344237506389618,
0.007886392064392567,
-0.020723354071378708,
-0.05412356182932854,
0.05503496900200844,
0.02104017324745655,
0.09747571498155594,
-0.017121898010373116,
0.015310538932681084,
-0.016634879633784294,
0.01567731611430645,
0.0285983607172966,
0.03476961329579353,
-0.03886030986905098,
-0.006485790014266968,
-0.09681548923254013,
0.03068752959370613,
-0.11722620576620102,
-0.11058084666728973,
-0.0067133973352611065,
-0.10243348032236099,
-0.05783965438604355,
-0.049254536628723145,
-0.034928400069475174,
-0.0572015680372715,
-0.03789832070469856,
0.002506694756448269,
0.07294216006994247,
0.025679675862193108,
-0.010073620826005936,
0.032983873039484024,
-0.02845541574060917,
0.05055428668856621,
-0.01585899293422699,
0.015313449315726757,
-0.05210903286933899,
0.03729945048689842,
-0.0029131202027201653,
-0.0279660876840353,
0.07783716171979904,
0.003541478654369712,
0.017192600294947624,
0.003386238357052207,
-0.03266929090023041,
-0.07685244083404541,
-0.16782915592193604,
-0.003784365253522992,
0.03830758482217789,
-0.06749158352613449,
0.03807210549712181,
0.00415677297860384,
0.043801575899124146,
-0.0015251908916980028,
0.01655026525259018,
0.01728793978691101,
0.09056386351585388,
0.018212798982858658,
0.022846460342407227,
0.1001284047961235,
-0.005943438038229942,
0.0027631393168121576,
-0.065821573138237,
0.024344509467482567,
-0.009172049351036549,
-0.045033104717731476,
-0.0017546048620715737,
-0.026717042550444603,
0.02902781218290329,
0.07603949308395386,
-0.03202688321471214,
0.044341083616018295,
0.01091625913977623,
0.15314339101314545,
0.00937446765601635,
0.0033546495251357555,
0.0022109171841293573,
0.047154296189546585,
-0.06879398971796036,
-0.12764772772789001,
0.04058409482240677,
-0.056787267327308655,
-0.08533315360546112,
-0.10109689831733704,
0.044599831104278564,
-0.006838192697614431,
0.08530160784721375,
-0.061642542481422424,
0.01161203347146511,
0.02844892628490925,
-0.048600826412439346,
-0.0054395184852182865,
0.09990351647138596,
0.04822809249162674,
-0.07029518485069275,
-0.005860623437911272,
0.005783827975392342,
0.043595608323812485,
0.06361841410398483,
-0.03213934972882271,
0.06271745264530182,
0.006949425674974918,
-0.036477431654930115,
-0.02301458828151226,
0.09601056575775146,
0.03558202087879181,
-0.04924002289772034,
0.03900456801056862,
0.015982570126652718,
-0.02165142074227333,
-0.03272389620542526,
-0.0785161480307579,
4.03128771382597e-33,
0.061152953654527664,
0.007269109599292278,
0.07219517976045609,
-0.07276685535907745,
0.048391472548246384,
0.014858975075185299,
0.0639517679810524,
0.06201651692390442,
-0.09929361194372177,
-0.028511453419923782,
-0.0058846608735620975,
-0.02562195435166359,
-0.06369014084339142,
0.1551108956336975,
0.10267613083124161,
-0.06667499989271164,
-0.0460876040160656,
0.07959099113941193,
0.020568901672959328,
-0.047671638429164886,
0.056594863533973694,
-0.017493119463324547,
-0.008198700845241547,
-0.07358727604150772,
0.007774769328534603,
0.048777759075164795,
0.05287306010723114,
0.005586635787039995,
0.00777994142845273,
-0.009157787077128887,
-0.02177157625555992,
-0.027148203924298286,
0.014870339073240757,
-0.006578646134585142,
0.02988634817302227,
0.04578690603375435,
-0.016826743260025978,
0.016316570341587067,
0.005146547220647335,
-0.006369440816342831,
0.025232132524251938,
-0.004889969248324633,
-0.021319223567843437,
0.04457695782184601,
-0.09237444400787354,
-0.0965094119310379,
0.032340504229068756,
0.024074897170066833,
-0.08414200693368912,
-0.005120549350976944,
0.06174436956644058,
0.051515404134988785,
0.019366499036550522,
0.0103911142796278,
-0.058577556163072586,
-0.029412604868412018,
-0.01580890826880932,
0.036487072706222534,
-0.04424530267715454,
-0.014408515766263008,
0.08997167646884918,
-0.04750562086701393,
-0.0016047816025093198,
0.03549555689096451,
-0.038935575634241104,
0.04691473767161369,
0.009017353877425194,
0.0006928204675205052,
0.024218743667006493,
0.014110651798546314,
0.021496960893273354,
0.029944298788905144,
0.016140667721629143,
-0.11737428605556488,
0.07462966442108154,
0.01607440412044525,
0.08467647433280945,
-0.06950562447309494,
0.020756516605615616,
-0.029352251440286636,
-0.04139032959938049,
0.00015493104001507163,
-0.030176818370819092,
-0.07329325377941132,
0.02905689738690853,
-0.07631048560142517,
0.02221767045557499,
-0.055652014911174774,
-0.03175484761595726,
-0.012066174298524857,
-0.0550355389714241,
-0.026510441675782204,
-0.005074246786534786,
0.008680828846991062,
0.014851793646812439,
-3.4341703790708344e-33,
-0.019444383680820465,
-0.06358899921178818,
0.026473071426153183,
0.005018256139010191,
-0.0018109699012711644,
-0.04741404950618744,
-0.0575123131275177,
-0.06162770837545395,
-0.000343161984346807,
-0.07095813006162643,
0.010082296095788479,
0.02926185540854931,
0.050303276628255844,
-0.005869084969162941,
0.020988861098885536,
0.010291964747011662,
-0.04907207190990448,
0.013947875238955021,
-0.01877681352198124,
-0.011810509487986565,
0.007565604988485575,
0.07429429888725281,
-0.11018962413072586,
0.07347001135349274,
-0.029108954593539238,
0.0503217875957489,
-0.002240200759842992,
0.0608530268073082,
0.01650848425924778,
0.06054758280515671,
-0.08349768817424774,
-0.02778358943760395,
-0.06907325237989426,
-0.03002513013780117,
-0.02947109006345272,
0.06197456642985344,
0.0753042921423912,
0.0500868521630764,
-0.0669628158211708,
0.08935152739286423,
0.0017293996643275023,
-0.030700217932462692,
-0.0667685717344284,
-0.01730620488524437,
-0.009065967053174973,
-0.015792686492204666,
-0.06581271439790726,
0.03591655194759369,
-0.03010396473109722,
0.009586532600224018,
-0.03078720159828663,
0.03647415712475777,
-0.10149121284484863,
-0.011099401861429214,
-0.06868723034858704,
-0.058462318032979965,
-0.07248743623495102,
-0.06786687672138214,
0.13897587358951569,
0.00043582989019341767,
-0.06424333155155182,
0.05821105092763901,
-0.03898230195045471,
0.04109355807304382,
-0.011411958374083042,
0.0053588105365633965,
0.0334591381251812,
0.012336882762610912,
-0.011522571556270123,
0.035972896963357925,
-0.01584705337882042,
0.039539504796266556,
0.06312299519777298,
-0.052723683416843414,
0.015588114969432354,
-0.021104581654071808,
-0.023741571232676506,
0.006158270873129368,
-0.01941351778805256,
0.020901260897517204,
0.04577294737100601,
0.05266628414392471,
0.07075690478086472,
-0.015717055648565292,
-0.013396654278039932,
0.09888430684804916,
-0.001892670406959951,
0.14427688717842102,
-0.002349134301766753,
0.020036207512021065,
-0.011171472258865833,
-0.021319061517715454,
0.041239842772483826,
0.06829480826854706,
-0.039349157363176346,
-4.076712301070984e-8,
-0.030673226341605186,
0.010179680772125721,
0.07862116396427155,
0.052492689341306686,
0.048361338675022125,
0.012728994712233543,
-0.027119319885969162,
-0.015887700021266937,
-0.046210043132305145,
-0.0957200899720192,
0.08928041905164719,
0.08134873956441879,
0.051594581454992294,
-0.1099957600235939,
-0.004666369408369064,
0.008786626160144806,
0.049094412475824356,
0.04602481797337532,
-0.014082391746342182,
-0.06348196417093277,
0.04453691840171814,
-0.012956964783370495,
-0.009530967101454735,
-0.08840328454971313,
0.03243190422654152,
-0.0891244113445282,
-0.01721741259098053,
0.08702472597360611,
0.1274418979883194,
0.03067842498421669,
-0.010381991043686867,
-0.030134661123156548,
0.07652397453784943,
0.014585026539862156,
0.09276119619607925,
0.025039374828338623,
0.06589698791503906,
-0.05324440449476242,
-0.03310302272439003,
-0.012038923799991608,
0.03641270473599434,
0.030031422153115273,
-0.022186024114489555,
-0.028298785910010338,
-0.029209937900304794,
0.024980273097753525,
-0.017493316903710365,
-0.05918469652533531,
0.08318889886140823,
-0.038341835141181946,
0.0009674078901298344,
-0.006273684091866016,
-0.037521470338106155,
0.08994312584400177,
0.06414634734392166,
-0.019236644729971886,
-0.018864454701542854,
-0.04338544234633446,
-0.0104175740852952,
0.06432792544364929,
0.01812957040965557,
-0.0302699226886034,
-0.016176018863916397,
0.0006821989081799984
] | https://github.com/calvincbzhang/orso | 2 | 0 | 0 | 0 |
Risk-Sensitive Variational Actor-Critic: A Model-Based Approach | https://openreview.net/forum?id=irrtPRFksw | [
"Alonso Granados",
"Reza Ebrahimi",
"Jason Pacheco"
] | Poster | Risk-sensitive reinforcement learning (RL) with an entropic risk measure typically requires knowledge of the transition kernel or performs unstable updates w.r.t. exponential Bellman equations. As a consequence, algorithms that optimize this objective have been restricted to tabular or low-dimensional continuous environments. In this work we leverage the connection between the entropic risk measure and the RL-as-inference framework to develop a risk-sensitive variational actor-critic algorithm (rsVAC). Our work extends the variational framework to incorporate stochastic rewards and proposes a variational model-based actor-critic approach that modulates policy risk via a risk parameter. We consider, both, the risk-seeking and risk-averse regimes and present rsVAC learning variants for each setting. Our experiments demonstrate that this approach produces risk-sensitive policies and yields improvements in both tabular and risk-aware variants of complex continuous control tasks in MuJoCo. | reinforcement learning, variational inference, risk sensitive RL | null | 12,719 | null | [
-0.04130494222044945,
-0.03252946212887764,
-0.014977517537772655,
0.02537493035197258,
0.03127771615982056,
0.11557688564062119,
0.014272407628595829,
0.00216027838177979,
-0.0033901319839060307,
0.07649590075016022,
-0.02831033617258072,
-0.041473619639873505,
0.0792371854186058,
0.08124738186597824,
-0.026044640690088272,
0.02545475773513317,
0.0379277840256691,
0.02441103383898735,
-0.024568533524870872,
0.04255112260580063,
0.040639154613018036,
0.0913996547460556,
0.05306670814752579,
-0.0199319776147604,
0.014759141020476818,
-0.023375144228339195,
0.052261389791965485,
0.04134056344628334,
-0.019476475194096565,
-0.033849284052848816,
0.03756356239318848,
-0.05206816643476486,
-0.024653756991028786,
-0.09363722801208496,
-0.10269805788993835,
0.05473178252577782,
-0.15548019111156464,
-0.0023773019202053547,
-0.028056439012289047,
-0.013934397138655186,
-0.06261224299669266,
0.015106474980711937,
-0.014132906682789326,
0.029958391562104225,
0.04330306500196457,
-0.04920292645692825,
0.00226817955262959,
-0.08535462617874146,
0.050102561712265015,
-0.026604648679494858,
0.013088633306324482,
-0.012033152393996716,
0.039604686200618744,
-0.05901610851287842,
0.04414500296115875,
-0.01754947565495968,
0.019337216392159462,
-0.026309190317988396,
-0.0257081501185894,
-0.040474824607372284,
0.002179780974984169,
-0.1743914633989334,
-0.03790544345974922,
-0.04809381067752838,
-0.08552150428295135,
-0.014437281526625156,
-0.045229721814394,
0.04477578401565552,
0.01713624969124794,
0.004001433961093426,
-0.051498640328645706,
0.04534080624580383,
-0.03684234246611595,
-0.0022217605728656054,
-0.028762180358171463,
-0.019289851188659668,
0.04716210439801216,
-0.04475012794137001,
0.021801432594656944,
-0.01360283698886633,
0.009713827632367611,
-0.03374553844332695,
0.02424624375998974,
0.0412897914648056,
0.06379655003547668,
0.0015179408947005868,
0.03722728416323662,
0.02835671789944172,
0.0993230938911438,
0.004361466038972139,
0.07036174833774567,
0.016013817861676216,
0.013424371369183064,
0.00735284760594368,
-0.003500819206237793,
0.013939283788204193,
-0.06358525902032852,
-0.0754682868719101,
-0.04726984724402428,
0.08870722353458405,
-0.017356539145112038,
0.018610360100865364,
0.01719866506755352,
0.05727279186248779,
0.0257500521838665,
-0.1106482744216919,
-0.004820540081709623,
-0.027819858863949776,
0.025191286578774452,
-0.08523070812225342,
0.05136025324463844,
-0.01527724415063858,
0.05972175672650337,
0.01567087695002556,
-0.05269495025277138,
0.053905852138996124,
-0.026033498346805573,
-0.05765882879495621,
0.00008877512300387025,
0.10298099368810654,
0.011985273100435734,
-0.07897220551967621,
-0.04810911417007446,
0.07823538035154343,
-0.020257839933037758,
0.003479383885860443,
0.007750993128865957,
-3.473143618765859e-34,
0.06812570989131927,
-0.07341565936803818,
0.04052252322435379,
0.0019224664429202676,
0.03756658732891083,
0.005561861675232649,
0.03327744081616402,
-0.01989387534558773,
-0.044816724956035614,
-0.052369870245456696,
0.005565188825130463,
0.0049794078804552555,
-0.08251115679740906,
0.05082130804657936,
0.0676204040646553,
-0.004135750234127045,
-0.10870365053415298,
0.0007719292188994586,
0.01931283250451088,
-0.08242714405059814,
0.016959605738520622,
-0.057095445692539215,
-0.023154528811573982,
-0.00694723566994071,
-0.02572803758084774,
-0.01739243045449257,
0.07209377735853195,
-0.015136945992708206,
0.03396656736731529,
0.014733687974512577,
-0.026292340829968452,
0.07025063782930374,
-0.06381047517061234,
-0.0027215841691941023,
0.04071211814880371,
0.06026493385434151,
-0.08888877183198929,
0.03563721850514412,
0.017851077020168304,
-0.02374187856912613,
0.04262223094701767,
0.0059744794853031635,
0.021393030881881714,
0.0659872367978096,
-0.07297170162200928,
-0.10836800932884216,
0.03428944572806358,
0.008265216834843159,
-0.059787850826978683,
-0.046672921627759933,
-0.024345379322767258,
0.03713109716773033,
0.06632275134325027,
-0.04033106938004494,
-0.04025755822658539,
0.01409424003213644,
0.0015908981440588832,
0.07163209468126297,
-0.030667496845126152,
0.02803119458258152,
0.005405571311712265,
0.01722325198352337,
0.055151790380477905,
0.03407217189669609,
-0.020573506131768227,
0.12599599361419678,
-0.07147209346294403,
0.04106377810239792,
0.0075383675284683704,
-0.007225676905363798,
0.0005655941204167902,
0.021458599716424942,
0.029991673305630684,
-0.14154596626758575,
0.027824625372886658,
-0.009060605429112911,
-0.004615874495357275,
-0.04398464038968086,
0.01888713426887989,
-0.0254705511033535,
-0.02769101969897747,
0.03600546345114708,
-0.0800010934472084,
0.04412045329809189,
-0.006001169793307781,
-0.05882861092686653,
-0.032321516424417496,
-0.07812001556158066,
-0.008528879843652248,
0.027154972776770592,
-0.07857369631528854,
-0.03971868008375168,
0.08282912522554398,
0.1044798344373703,
0.010196552611887455,
-5.0984849639684505e-34,
-0.03661226108670235,
-0.04943748936057091,
-0.010294410400092602,
0.012674400582909584,
0.0014477658551186323,
0.017673667520284653,
-0.029281560331583023,
-0.032238781452178955,
0.025348950177431107,
-0.011149227619171143,
-0.10809147357940674,
0.07037948071956635,
0.030635055154561996,
-0.03494314104318619,
0.02545991726219654,
-0.041694384068250656,
-0.017376437783241272,
0.04500245302915573,
-0.037737887352705,
-0.0397445410490036,
0.05585136637091637,
0.0338507778942585,
-0.07776378095149994,
0.047952987253665924,
0.01270349696278572,
0.03730010613799095,
0.008564840070903301,
0.1055556982755661,
-0.0008460739045403898,
0.03539475426077843,
-0.10064631700515747,
-0.04287511482834816,
0.02816017158329487,
0.051155343651771545,
-0.115730419754982,
0.06166212260723114,
0.08975428342819214,
-0.006882549729198217,
-0.036278195679187775,
0.13001573085784912,
-0.0017383747035637498,
-0.026269886642694473,
0.024774722754955292,
-0.017003081738948822,
0.0621454231441021,
-0.02509372867643833,
0.07673463225364685,
-0.06547701358795166,
-0.04997863620519638,
-0.04220297932624817,
-0.02065887860953808,
0.042052049189805984,
-0.09873265027999878,
0.0009870959911495447,
-0.07753491401672363,
-0.05708320811390877,
-0.04861260950565338,
-0.022856755182147026,
0.0528724268078804,
0.06767960637807846,
-0.033880941569805145,
0.045859381556510925,
0.005818669218569994,
0.08208853006362915,
-0.05776246637105942,
0.02528722584247589,
-0.026958253234624863,
-0.04407009482383728,
0.07003314793109894,
-0.04361310228705406,
0.00847743358463049,
0.011519798077642918,
0.05449691042304039,
-0.052740275859832764,
0.09572917222976685,
-0.05848180875182152,
-0.02589232102036476,
0.011597191914916039,
-0.038855113089084625,
-0.056697797030210495,
0.07052037864923477,
0.01660907454788685,
0.10171818733215332,
0.013747751712799072,
0.02994643896818161,
0.09061993658542633,
-0.03212227299809456,
0.06285128742456436,
0.006520147435367107,
0.03996890410780907,
-0.05889487266540527,
-0.0004980888916179538,
0.05344119295477867,
0.019344089552760124,
-0.041394177824258804,
-4.121307739524127e-8,
-0.04457354545593262,
0.06090109422802925,
0.0798201784491539,
0.00009062582103069872,
0.09824547171592712,
-0.02247059904038906,
0.019180385395884514,
-0.03382858633995056,
-0.009874479845166206,
-0.0017154037486761808,
0.1105315238237381,
0.05951067805290222,
0.08005436509847641,
-0.12631256878376007,
-0.06700405478477478,
0.07755231112241745,
0.014003342017531395,
0.02525511384010315,
0.03248602896928787,
0.05868550390005112,
0.09361664950847626,
-0.0019443000201135874,
0.003675948129966855,
-0.03345863148570061,
0.039286043494939804,
-0.053702179342508316,
-0.030746828764677048,
0.10217907279729843,
0.09348543733358383,
0.006706856656819582,
0.00008428120781900361,
-0.0023635628167539835,
0.020579570904374123,
0.03725937381386757,
-0.03202877938747406,
0.040292538702487946,
0.035323284566402435,
-0.07739036530256271,
-0.017833758145570755,
-0.0017737533198669553,
0.03780112788081169,
0.022542905062437057,
-0.009159084409475327,
0.010857452638447285,
-0.03301068767905235,
0.057738736271858215,
-0.03552336245775223,
-0.10248133540153503,
0.09572151303291321,
-0.024181753396987915,
0.05932997912168503,
-0.010376575402915478,
-0.023586126044392586,
0.10029131919145584,
0.04275452345609665,
-0.03688308596611023,
0.014984654262661934,
-0.07137898355722427,
0.03844141215085983,
0.040631189942359924,
-0.0031361060682684183,
-0.02488952875137329,
-0.07398880273103714,
-0.002291847951710224
] | 0 | 0 | 0 | 0 |
|
L3Ms — Lagrange Large Language Models | https://openreview.net/forum?id=ULGbw2URE3 | [
"Guneet S. Dhillon",
"Xingjian Shi",
"Yee Whye Teh",
"Alex Smola"
] | Poster | Supervised fine-tuning (SFT) and alignment of large language models (LLMs) are key steps in providing a good user experience. However, the concept of an appropriate alignment is inherently application-dependent, and current methods often rely on heuristic choices to drive optimization. In this work, we formulate SFT and alignment as a constrained optimization problem: the LLM is fine-tuned on a task while being required to meet application-specific requirements, without resorting to heuristics. To solve this, we propose Lagrange Large Language Models (L3Ms), which employ logarithmic barriers to enforce the constraints. This approach allows for the customization of L3Ms across diverse applications while avoiding heuristic-driven processes. We experimentally demonstrate the versatility and efficacy of L3Ms in achieving tailored alignments for various applications. | LLM, alignment | null | 12,708 | 2410.21533 | [
0.0031764330342411995,
-0.10931564122438431,
-0.012572159059345722,
-0.032212238758802414,
0.04832812771201134,
0.011643823236227036,
-0.05500275641679764,
-0.005736024118959904,
0.028542570769786835,
-0.02068786695599556,
0.004576141946017742,
-0.08174807578325272,
0.016421206295490265,
0.05018319562077522,
0.034225963056087494,
-0.038460176438093185,
0.05495448410511017,
0.033598642796278,
-0.11102961003780365,
-0.07763427495956421,
0.0577692911028862,
0.04037054255604744,
0.03207482770085335,
0.02987208217382431,
0.001712276367470622,
-0.002734607318416238,
0.006649672985076904,
-0.016853636130690575,
0.038725901395082474,
-0.07022393494844437,
0.026155900210142136,
0.010914508253335953,
-0.0035386853851377964,
0.052280936390161514,
-0.004560946021229029,
0.018011141568422318,
-0.16367147862911224,
-0.08969462662935257,
0.013061506673693657,
-0.03723333775997162,
-0.05475199222564697,
0.022583425045013428,
0.0326106958091259,
-0.019075138494372368,
0.05274142697453499,
-0.010464197024703026,
-0.03997763618826866,
-0.012673591263592243,
-0.013581403531134129,
0.03184932842850685,
-0.11714956164360046,
-0.0527498833835125,
0.003989117685705423,
0.09369503706693649,
-0.09994320571422577,
0.006759957876056433,
0.034227773547172546,
0.03147774562239647,
0.014931779354810715,
-0.04823383688926697,
-0.03561883047223091,
-0.06191439554095268,
-0.032852157950401306,
0.02273254282772541,
0.007405649404972792,
-0.04752632603049278,
0.011077430099248886,
0.01646537147462368,
-0.03877149522304535,
0.1225108653306961,
-0.0382390134036541,
0.06475075334310532,
-0.003893799614161253,
0.048967938870191574,
0.013374737463891506,
0.08798856288194656,
0.02952142059803009,
0.015495417639613152,
0.029363106936216354,
-0.03778357058763504,
0.013620637357234955,
0.05183156579732895,
0.04444589093327522,
0.06435302644968033,
0.08416002988815308,
-0.04227709397673607,
-0.04216574504971504,
0.018292458727955818,
0.08887111395597458,
0.05362148955464363,
-0.06425638496875763,
-0.08328301459550858,
0.01837967522442341,
0.014997707679867744,
-0.04622088000178337,
0.0825459212064743,
0.0929146483540535,
-0.006980155128985643,
-0.06992731243371964,
0.07073372602462769,
0.01329787913709879,
0.15375348925590515,
0.011795148253440857,
-0.05944160744547844,
-0.0029427106492221355,
-0.028439778834581375,
-0.03919491171836853,
0.0035533455666154623,
-0.04096829146146774,
-0.12026142328977585,
0.05335491523146629,
0.06454230099916458,
0.022154681384563446,
0.02657083421945572,
-0.006902857217937708,
-0.032329682260751724,
0.04318932816386223,
-0.11751031130552292,
-0.026366500183939934,
0.02952137403190136,
-0.028775664046406746,
0.03260931000113487,
-0.013714064843952656,
-0.03693680837750435,
-0.012871747836470604,
0.010047433897852898,
-0.05893661081790924,
1.425706818178206e-34,
0.06581910699605942,
0.07594633847475052,
-0.00181670137681067,
0.08162132650613785,
0.02830876223742962,
0.018931981176137924,
-0.018752126023173332,
0.0579831525683403,
-0.05833692476153374,
-0.024522745981812477,
-0.01314346119761467,
-0.003007300430908799,
0.009713515639305115,
-0.03665388375520706,
-0.0033571100793778896,
-0.02596767619252205,
-0.05872591212391853,
0.09118986874818802,
-0.09562336653470993,
0.0061373282223939896,
0.0349891297519207,
-0.01842178776860237,
-0.008525628596544266,
-0.032434284687042236,
0.02047516405582428,
-0.0011737202294170856,
0.03717871755361557,
-0.06639767438173294,
-0.04866869002580643,
0.010206441394984722,
-0.07821284234523773,
-0.06278355419635773,
0.044982101768255234,
-0.0030750513542443514,
0.09527071565389633,
-0.06004950776696205,
-0.054279327392578125,
-0.10464054346084595,
0.03303341940045357,
-0.005231826566159725,
-0.010380476713180542,
0.04941324517130852,
0.04944039136171341,
0.007088346406817436,
-0.07476469874382019,
-0.017405224964022636,
0.006983851548284292,
0.012681864202022552,
0.02495015226304531,
-0.07212073355913162,
0.05409107729792595,
0.08808737993240356,
-0.03315692022442818,
0.004300558939576149,
0.015007839538156986,
0.002737932838499546,
0.01068045012652874,
0.05072958394885063,
-0.013491981662809849,
0.034620001912117004,
-0.006195770110934973,
0.0024372348561882973,
0.03648550063371658,
0.07281976193189621,
0.10086537897586823,
-0.02411988377571106,
0.02350340597331524,
0.08169645816087723,
0.07579212635755539,
-0.026796288788318634,
-0.028769532218575478,
-0.10076230019330978,
0.04400421306490898,
0.03434644639492035,
0.007251079194247723,
0.01924220286309719,
0.07525946944952011,
-0.05155809596180916,
-0.003126169089227915,
-0.033616501837968826,
-0.0006296741776168346,
0.01435704343020916,
0.034828923642635345,
-0.05005035176873207,
-0.06975367665290833,
-0.1734132170677185,
0.06756894290447235,
-0.058320414274930954,
-0.05145637318491936,
-0.019037755206227303,
-0.03810970485210419,
0.036023203283548355,
0.05570964515209198,
-0.017187492921948433,
0.02915988862514496,
-1.2683637309960254e-33,
-0.02458241954445839,
-0.07779068499803543,
0.05214010551571846,
0.028177132830023766,
0.0053055305033922195,
-0.08626572787761688,
-0.022227972745895386,
0.03444332629442215,
-0.02790534310042858,
0.010520003736019135,
-0.007063933648169041,
-0.08279336988925934,
0.04494280740618706,
-0.05699644610285759,
-0.003972430247813463,
0.05430730804800987,
0.03801487013697624,
0.046566981822252274,
0.02787490375339985,
0.07121971249580383,
0.019025208428502083,
0.10759060084819794,
-0.1070423275232315,
0.07494911551475525,
-0.020856186747550964,
0.003172073047608137,
-0.03222864121198654,
-0.015786409378051758,
0.002167412545531988,
0.024679964408278465,
-0.07668519020080566,
0.00658823037520051,
-0.11412974447011948,
-0.05755182355642319,
-0.07513213157653809,
0.035244252532720566,
-0.021557144820690155,
0.01461088564246893,
0.0682569295167923,
0.07858578115701675,
0.046552903950214386,
-0.01602286659181118,
-0.025429023429751396,
-0.04816826060414314,
-0.0063839079812169075,
0.013227582909166813,
-0.08514367043972015,
-0.05728381872177124,
0.011961269192397594,
-0.012364404276013374,
-0.06257057934999466,
0.026454538106918335,
0.00882139801979065,
0.0008902279078029096,
0.00577583909034729,
-0.0619661808013916,
-0.07783441245555878,
-0.04558467119932175,
0.028904905542731285,
-0.019616300240159035,
-0.027295926585793495,
-0.008150819689035416,
0.10099651664495468,
-0.04888557270169258,
0.007626419421285391,
-0.05790551006793976,
-0.0035243667662143707,
0.0016887584934011102,
0.0028263169806450605,
-0.0651816800236702,
0.0712086409330368,
-0.06755983084440231,
-0.021939687430858612,
0.0995030328631401,
-0.07568122446537018,
-0.05524348467588425,
0.04170297086238861,
-0.04132208228111267,
0.018532313406467438,
-0.10322356969118118,
0.0938107967376709,
0.0666062980890274,
0.019641727209091187,
0.06271573156118393,
0.020017335191369057,
0.006652697920799255,
0.06727416068315506,
0.1087409108877182,
0.03557036817073822,
0.014222285710275173,
-0.004343489650636911,
0.03566037863492966,
0.05529317259788513,
0.03902440145611763,
-0.055527448654174805,
-3.752172972326662e-8,
-0.05753627419471741,
0.008777067996561527,
-0.03206688538193703,
0.09795699268579483,
-0.05587887391448021,
-0.029948530718684196,
-0.039871301501989365,
0.0967884287238121,
0.035651691257953644,
-0.007051405496895313,
-0.0008848137804307044,
0.01740231364965439,
-0.03566167503595352,
-0.058341387659311295,
-0.027854010462760925,
0.034570805728435516,
-0.0932554379105568,
0.07802693545818329,
-0.03623082488775253,
0.05766436457633972,
-0.05640896409749985,
0.09245871752500534,
-0.021079914644360542,
0.04383016750216484,
0.09146654605865479,
-0.015100792981684208,
0.004625586327165365,
0.05953492224216461,
0.07173830270767212,
-0.00967018585652113,
-0.057611726224422455,
0.0029811616986989975,
-0.002898560604080558,
-0.005563241429626942,
-0.02452397719025612,
0.03811202943325043,
-0.017615118995308876,
-0.009578476659953594,
-0.017693474888801575,
-0.03674706071615219,
0.030341118574142456,
0.06793253868818283,
-0.05987081304192543,
0.05407589301466942,
0.009252488613128662,
-0.022717567160725594,
-0.04782619699835777,
-0.07476320117712021,
0.011399378068745136,
-0.06120320037007332,
-0.015092716552317142,
-0.029897890985012054,
0.03709021583199501,
-0.0217630285769701,
0.0158874299377203,
0.04912184923887253,
-0.016596060246229172,
0.006254199426621199,
0.10818745195865631,
-0.02612101100385189,
-0.0010362031171098351,
0.034314293414354324,
0.0011524651199579239,
0.027972903102636337
] | 0 | 0 | 0 | 0 |
|
Locality Alignment Improves Vision-Language Models | https://openreview.net/forum?id=qssVptHTPN | [
"Ian Connick Covert",
"Tony Sun",
"James Zou",
"Tatsunori Hashimoto"
] | Poster | Vision language models (VLMs) have seen growing adoption in recent years, but many still struggle with basic spatial reasoning errors. We hypothesize that this is due to VLMs adopting pre-trained vision backbones, specifically vision transformers (ViTs) trained with image-level supervision and minimal inductive biases. Such models may fail to encode the class contents at each position in the image, and our goal is to resolve this with a vision backbone that effectively captures both local and global image semantics. Our main insight is that we do not require new supervision to learn this capability – pre-trained models contain significant knowledge of local semantics that we can extract and use for scalable self-supervision. We propose a new efficient post-training stage for ViTs called locality alignment and a novel fine-tuning procedure called MaskEmbed that uses a masked reconstruction loss to learn semantic contributions for each image patch. We first evaluate locality alignment with a vision-only benchmark, finding that it improves a model’s performance at patch-level semantic segmentation, especially for strong backbones trained with image-caption pairs (e.g., CLIP and SigLIP). We then train a series of VLMs with and without locality alignment, and show that locality-aligned backbones improve performance across a range of benchmarks, particularly ones that involve spatial understanding (e.g., RefCOCO, OCID-Ref, TallyQA, VSR, AI2D). Overall, we demonstrate that we can efficiently learn local semantic extraction via a locality alignment stage, and that this procedure benefits VLM training recipes that use off-the-shelf vision backbones. | Multimodal language models, vision-language models, locality, alignment, vision transformers | locality alignment post-training helps ViTs recover localized features, improves VLMs | 12,706 | 2410.11087 | [
0.025193748995661736,
-0.06491049379110336,
0.03798892721533775,
-0.022296210750937462,
0.0901838168501854,
0.027023926377296448,
-0.019828811287879944,
-0.008898925967514515,
0.021769670769572258,
-0.06646687537431717,
0.000576453865505755,
-0.061824847012758255,
-0.008909683674573898,
0.13495363295078278,
-0.03637827932834625,
-0.028816847130656242,
0.11408277601003647,
0.08909883350133896,
-0.10060770064592361,
-0.0664220005273819,
0.01959901489317417,
0.0139334537088871,
0.06286840885877609,
0.043192002922296524,
0.05136658251285553,
0.008717935532331467,
0.02168145217001438,
-0.08847547322511673,
0.05303005129098892,
-0.04968629032373428,
0.015363086014986038,
-0.014384834095835686,
-0.03305354341864586,
0.046624958515167236,
0.002363287378102541,
0.09753796458244324,
0.0057123457081615925,
0.007679140195250511,
0.007785629015415907,
-0.04354386776685715,
-0.032661572098731995,
0.025795821100473404,
0.008330556564033031,
-0.012627042829990387,
0.05216148868203163,
0.02508537471294403,
0.0363386832177639,
-0.010028333403170109,
-0.032631658017635345,
-0.023948732763528824,
-0.06837776303291321,
-0.034155696630477905,
-0.06178748607635498,
-0.0062431157566607,
-0.023439256474375725,
0.03311929851770401,
0.09923382103443146,
-0.02077653631567955,
0.037006925791502,
0.0049439044669270515,
0.05359213799238205,
-0.053825199604034424,
-0.008170383982360363,
0.034237440675497055,
-0.057137325406074524,
-0.06647270917892456,
0.05239531770348549,
-0.002604034496471286,
-0.02260698191821575,
-0.02057502791285515,
-0.03643202409148216,
0.0409567691385746,
-0.0008458765223622322,
0.0602705292403698,
0.009821159765124321,
0.054380130022764206,
0.060907963663339615,
-0.05475713312625885,
0.08035508543252945,
-0.09793558716773987,
0.03396174684166908,
0.0037736352533102036,
0.056344885379076004,
0.011361095122992992,
0.13672275841236115,
-0.012102718465030193,
0.019056985154747963,
0.035848356783390045,
0.0339229516685009,
-0.0008495390065945685,
-0.0671657994389534,
-0.10194692760705948,
-0.0106067368760705,
-0.008250308223068714,
0.06035938113927841,
-0.01775873266160488,
-0.021251700818538666,
-0.11503802239894867,
-0.013774536550045013,
0.07074545323848724,
-0.01955164223909378,
0.04871194437146187,
0.00022629099839832634,
-0.006474571768194437,
0.04061289504170418,
0.019666526466608047,
-0.0008253300911746919,
0.039897095412015915,
0.017317749559879303,
-0.06899064779281616,
0.0676334872841835,
0.03383062779903412,
-0.020276455208659172,
-0.032105639576911926,
0.02331933006644249,
0.04224763065576553,
-0.0032592841889709234,
-0.013863643631339073,
0.03623466566205025,
-0.04202916473150253,
-0.03894152119755745,
0.021528644487261772,
0.010194533504545689,
-0.0026890698354691267,
0.05834130942821503,
-0.07035019248723984,
-0.061853837221860886,
-4.463556384051177e-35,
0.08868162333965302,
0.05057249590754509,
0.06840787082910538,
0.013352463021874428,
0.051615629345178604,
-0.014397512190043926,
0.03857862204313278,
0.05867611616849899,
-0.06534507125616074,
-0.017480557784438133,
0.005117562133818865,
-0.0740395113825798,
-0.004240279085934162,
0.08542608469724655,
-0.04197349771857262,
0.012876271270215511,
-0.03473507612943649,
-0.0004025878442917019,
-0.06884127855300903,
0.05628314986824989,
0.029756242409348488,
0.02227170579135418,
-0.027530476450920105,
0.0006851550424471498,
0.02659030258655548,
-0.037020955234766006,
0.04575624316930771,
-0.07241218537092209,
-0.023151174187660217,
0.00878980290144682,
-0.12258562445640564,
0.026940492913126945,
0.03611505776643753,
0.0656857118010521,
0.06720618903636932,
-0.05445026606321335,
0.000625181826762855,
-0.08928757160902023,
0.03950339928269386,
-0.018619712442159653,
0.028934277594089508,
0.08307461440563202,
-0.0656747817993164,
-0.08710306137800217,
-0.03428738936781883,
-0.03189104422926903,
-0.01392505131661892,
0.05758676677942276,
-0.09307783097028732,
-0.03180772811174393,
0.07845178246498108,
0.059408482164144516,
-0.08758149296045303,
-0.06483536213636398,
-0.06605295836925507,
0.0007906905375421047,
0.0564463846385479,
0.0490693673491478,
0.08065258711576462,
0.011053310707211494,
0.02284284494817257,
-0.007110114209353924,
-0.05416826903820038,
0.10291149467229843,
0.028797568753361702,
-0.06275254487991333,
0.09795039147138596,
0.054862651973962784,
0.033370643854141235,
-0.0075515215285122395,
-0.04968806728720665,
0.040154073387384415,
-0.03126210719347,
-0.05440199375152588,
0.05829653888940811,
-0.011685943230986595,
0.032856810837984085,
-0.05856615677475929,
-0.008349008858203888,
0.0462644137442112,
-0.027501439675688744,
-0.020682431757450104,
-0.030700640752911568,
-0.08958803117275238,
-0.06277530640363693,
-0.045476991683244705,
0.09240228682756424,
-0.07560582458972931,
0.04801540449261665,
-0.00008341093052877113,
-0.017135512083768845,
-0.09285415709018707,
0.03179429471492767,
-0.04029564559459686,
-0.011186333373188972,
-9.247566418255304e-35,
0.06051903963088989,
0.005453767254948616,
-0.011841757223010063,
0.02886389009654522,
-0.06812871992588043,
-0.09476035088300705,
0.05893053859472275,
0.07415786385536194,
-0.05156853795051575,
-0.06576912850141525,
-0.07666099816560745,
-0.02216586470603943,
-0.02923355996608734,
-0.0006558863678947091,
-0.010180185548961163,
-0.0033604467753320932,
-0.016556067392230034,
-0.03156834840774536,
-0.01617237739264965,
0.05150201544165611,
-0.010610340163111687,
0.08850708603858948,
-0.06237447261810303,
0.05419822782278061,
-0.08562427759170532,
0.10275232791900635,
0.0001100515146390535,
0.034088216722011566,
-0.041990287601947784,
0.01757439598441124,
0.01979866810142994,
0.008100579492747784,
0.007948033511638641,
-0.06503936648368835,
-0.09611441940069199,
0.09074107557535172,
0.010397396050393581,
-0.0587807260453701,
-0.05041319504380226,
0.06340686231851578,
0.0029792299028486013,
-0.061377037316560745,
-0.07947223633527756,
0.021345417946577072,
-0.09498333930969238,
-0.07865015417337418,
-0.01671990007162094,
-0.0016928097466006875,
-0.027203332632780075,
0.0510881282389164,
-0.061680395156145096,
0.05170118808746338,
-0.10338841378688812,
0.050180818885564804,
-0.03908739238977432,
-0.0573105588555336,
-0.04040477052330971,
0.008335561491549015,
0.07788597047328949,
-0.03312864154577255,
-0.016836220398545265,
-0.051683586090803146,
0.0035139513202011585,
-0.01319202221930027,
0.009522399865090847,
-0.024604693055152893,
-0.09058376401662827,
-0.010685494169592857,
0.06924717873334885,
0.03545720875263214,
0.031263548880815506,
-0.015088626183569431,
-0.00981127843260765,
-0.021441489458084106,
-0.00656850403174758,
-0.03583799675107002,
0.04353298619389534,
0.02189188078045845,
0.06671422719955444,
-0.03460417687892914,
0.02512473054230213,
-0.04940390959382057,
0.035569556057453156,
0.057148147374391556,
0.07783759385347366,
0.053629469126462936,
0.047435976564884186,
0.04041769355535507,
0.002239908091723919,
0.01980544626712799,
-0.10082750022411346,
-0.0036318213678896427,
-0.019653797149658203,
-0.00645793741568923,
-0.010442142374813557,
-3.901270417827618e-8,
-0.1182350441813469,
0.04161302372813225,
-0.041948579251766205,
0.04545331746339798,
-0.05739818513393402,
-0.06389643251895905,
-0.01865246519446373,
0.12391483783721924,
-0.022918948903679848,
0.03821374103426933,
0.05219706520438194,
0.0028282098937779665,
-0.008811529725790024,
-0.05405709519982338,
-0.006264031399041414,
0.060801610350608826,
0.042811982333660126,
0.06848230957984924,
-0.011866816319525242,
0.06853517889976501,
-0.05576162040233612,
0.057501837611198425,
0.03952431306242943,
-0.020447738468647003,
0.03376244008541107,
-0.036913879215717316,
0.003183252178132534,
0.13772746920585632,
0.12110377848148346,
-0.003499578218907118,
-0.034631796181201935,
-0.005670812446624041,
0.06766670197248459,
-0.04140797257423401,
0.05005819350481033,
0.02753179520368576,
-0.02563582919538021,
-0.030538184568285942,
0.02726743556559086,
-0.06439065933227539,
0.059273336082696915,
0.022844506427645683,
-0.018465092405676842,
0.031238839030265808,
-0.021317409351468086,
0.01878361403942108,
0.011149794794619083,
-0.08372385054826736,
-0.0221130158752203,
-0.010072825476527214,
-0.0314655639231205,
-0.02514740452170372,
-0.05282941460609436,
0.12883822619915009,
0.002488186117261648,
-0.002257314743474126,
-0.017158275470137596,
-0.056326497346162796,
0.05063144117593765,
0.1206107959151268,
-0.05074039474129677,
0.025205086916685104,
0.026955464854836464,
-0.03601889684796333
] | 0 | 0 | 0 | 0 |
|
Benign Overfitting in Out-of-Distribution Generalization of Linear Models | https://openreview.net/forum?id=6jxUsDAdAu | [
"Shange Tang",
"Jiayun Wu",
"Jianqing Fan",
"Chi Jin"
] | Poster | Benign overfitting refers to the phenomenon where an over-parameterized model fits the training data perfectly, including noise in the data, but still generalizes well to the unseen test data. While prior work provides some theoretical understanding of this phenomenon under the in-distribution setup, modern machine learning often operates in a more challenging Out-of-Distribution (OOD) regime, where the target (test) distribution can be rather different from the source (training) distribution. In this work, we take an initial step towards understanding benign overfitting in the OOD regime by focusing on the basic setup of over-parameterized linear models under covariate shift. We provide non-asymptotic guarantees proving that benign overfitting occurs in standard ridge regression, even under the OOD regime when the target covariance satisfies certain structural conditions. We identify several vital quantities relating to source and target covariance, which govern the performance of OOD generalization. Our result is sharp, which provably recovers prior in-distribution benign overfitting guarantee (Tsigler & Bartlett, 2023), as well as under-parameterized OOD guarantee (Ge et al., 2024) when specializing to each setup. Moreover, we also present theoretical results for a more general family of target covariance matrix, where standard ridge regression only achieves a slow statistical rate of $\mathcal{O}(1/\sqrt{n})$ for the excess risk, while Principal Component Regression (PCR) is guaranteed to achieve the fast rate $\mathcal{O}(1/n)$, where $n$ is the number of samples. | Over-parameterization, benign overfitting, OOD generalization, principal component regression, minimum norm interpolation, ridge regression | We provide non-asymptotic guarantees of over-parameterized ridge regression under general covariate shift. | 12,693 | 2412.14474 | [
-0.09274450689554214,
-0.0895279049873352,
0.05280507728457451,
0.09033748507499695,
0.1813923418521881,
-0.03753313422203064,
0.03674459457397461,
-0.021821020171046257,
0.015347525477409363,
0.0017186058685183525,
0.0530373752117157,
0.07166887819766998,
-0.01663784869015217,
-0.06415864080190659,
-0.049400631338357925,
-0.06294858455657959,
0.07034910470247269,
0.02585863694548607,
-0.06455747783184052,
0.05157727375626564,
0.03076826222240925,
0.08596917241811752,
-0.02486223727464676,
0.07447242736816406,
0.044665832072496414,
-0.10490218549966812,
-0.028448620811104774,
-0.05179540067911148,
-0.04762466251850128,
-0.018962085247039795,
0.014021608047187328,
0.023862794041633606,
-0.0635920986533165,
0.005783713888376951,
0.018880119547247887,
0.0019458658061921597,
0.055011212825775146,
0.042096469551324844,
0.05533583089709282,
0.0482790507376194,
0.06107845902442932,
-0.013703588396310806,
0.0040192874148488045,
0.021931659430265427,
0.03139759972691536,
-0.06015003100037575,
0.0505233071744442,
-0.134455144405365,
0.01834453083574772,
0.04494427144527435,
-0.05733870714902878,
-0.013394021429121494,
-0.09044352918863297,
0.06729702651500702,
-0.051414355635643005,
-0.0944778174161911,
0.0030733710154891014,
-0.005784894339740276,
0.03491455689072609,
0.04440543055534363,
-0.06997986882925034,
-0.07953549921512604,
-0.015232726000249386,
0.044027119874954224,
-0.03130398318171501,
-0.03449571132659912,
0.03481314331293106,
0.038374725729227066,
-0.05507783219218254,
0.05414556711912155,
0.006428527645766735,
0.06633622944355011,
-0.06200332194566727,
0.03699833154678345,
0.08007564395666122,
0.010450542904436588,
-0.0018772681942209601,
0.02259737253189087,
0.03170428425073624,
-0.05028447508811951,
0.03894035145640373,
0.08251653611660004,
0.029529642313718796,
-0.10527971386909485,
0.09057731181383133,
0.025350406765937805,
-0.03156231716275215,
0.02245326153934002,
0.07073426991701126,
0.007416455540806055,
-0.037157025188207626,
-0.0475255623459816,
-0.040781740099191666,
0.03897984325885773,
0.056851617991924286,
-0.03471595048904419,
0.05589427798986435,
-0.08340317755937576,
0.04840497672557831,
0.09218999743461609,
-0.042422425001859665,
0.08038942515850067,
0.01705063134431839,
0.03555203974246979,
0.011848942376673222,
-0.08657725155353546,
0.05373344197869301,
0.005750410258769989,
0.05572443827986717,
-0.05809302628040314,
0.029854655265808105,
-0.02271571196615696,
0.07420773804187775,
-0.025526396930217743,
-0.04264611005783081,
-0.018597211688756943,
-0.00982060469686985,
0.041589826345443726,
-0.061658747494220734,
0.0010151165770366788,
-0.05280096456408501,
0.03022122196853161,
-0.023403534665703773,
0.06581638008356094,
-0.0619761124253273,
0.06161130964756012,
-0.10800963640213013,
4.131312364847362e-33,
0.05882374942302704,
0.057362183928489685,
0.028566841036081314,
-0.0013049391563981771,
-0.05337899550795555,
-0.053151749074459076,
0.01719886250793934,
-0.004156431648880243,
0.018775388598442078,
0.04108639806509018,
-0.004064289387315512,
-0.058326464146375656,
-0.06621282547712326,
0.0008789379498921335,
0.030262984335422516,
0.04865267127752304,
-0.05365639552474022,
0.08347895741462708,
0.031349021941423416,
0.022112315520644188,
0.07876572012901306,
-0.09403043985366821,
-0.08079040050506592,
-0.016829855740070343,
-0.024664126336574554,
-0.005581018514931202,
-0.004614464472979307,
-0.012174554169178009,
-0.10334821045398712,
0.04655761271715164,
-0.011283441446721554,
0.0602167546749115,
0.052074287086725235,
0.06119503825902939,
0.0308961421251297,
-0.09492859989404678,
-0.07892134040594101,
-0.05511791631579399,
0.03306357190012932,
-0.0018818986136466265,
-0.037696875631809235,
0.010401537641882896,
-0.010753568261861801,
0.028694303706288338,
-0.0040875948034226894,
-0.022495221346616745,
0.05644748732447624,
-0.013816144317388535,
-0.049146391451358795,
-0.09355858713388443,
-0.048056453466415405,
0.07226751744747162,
-0.04924025014042854,
-0.000623655563686043,
-0.1307419389486313,
0.086284339427948,
-0.003948035649955273,
-0.061294108629226685,
-0.007324154023081064,
0.028971923515200615,
-0.004508463200181723,
0.003869813634082675,
-0.02465737797319889,
-0.023220019415020943,
-0.008505161851644516,
-0.049610089510679245,
-0.013380619697272778,
-0.04334304854273796,
0.03203267231583595,
0.01444388460367918,
0.02957642264664173,
-0.07263586670160294,
-0.13131879270076752,
0.00952858105301857,
0.07755771279335022,
-0.005654113367199898,
0.05799174681305885,
0.00352118699811399,
0.03075963631272316,
-0.031401000916957855,
0.007264398504048586,
-0.015318555757403374,
0.023669734597206116,
-0.05460906773805618,
-0.05069650337100029,
-0.0300485547631979,
0.0203933697193861,
-0.04589272662997246,
-0.06272415816783905,
-0.012118814513087273,
-0.052602384239435196,
-0.0451456755399704,
-0.10691031068563461,
-0.05228380486369133,
0.018101422116160393,
-3.874196118150018e-33,
-0.09376224130392075,
0.02159704454243183,
0.0010410749819129705,
0.025367651134729385,
-0.04560083523392677,
-0.057172808796167374,
-0.00872332975268364,
0.022998111322522163,
-0.019691232591867447,
-0.0548454150557518,
-0.015578056685626507,
-0.0468624085187912,
0.008930113166570663,
-0.014526722952723503,
-0.022377386689186096,
0.014122474007308483,
0.003910258878022432,
0.11202947050333023,
-0.014878829009830952,
0.00028790082433260977,
0.02755696140229702,
-0.026953665539622307,
-0.00938128586858511,
0.026332322508096695,
0.0024886210449039936,
0.07147852331399918,
-0.11023098975419998,
0.11950766295194626,
0.012434731237590313,
-0.0018490981310606003,
-0.06182299926877022,
0.06585316359996796,
0.0012448076158761978,
-0.003746657632291317,
0.0007273775991052389,
0.08735909312963486,
-0.02520950511097908,
-0.027502400800585747,
-0.011021503247320652,
0.1128871962428093,
0.08369365334510803,
0.01439546700567007,
-0.05440094321966171,
-0.026737533509731293,
0.060211412608623505,
-0.026445506140589714,
0.0382133387029171,
0.009387366473674774,
0.06389489024877548,
-0.014611857011914253,
-0.029811732470989227,
0.09734561294317245,
0.02231811173260212,
0.10662410408258438,
-0.045613616704940796,
-0.023844541981816292,
-0.0022337150294333696,
0.0016418829327449203,
-0.023891784250736237,
-0.01695128343999386,
-0.00894449558109045,
0.06397613883018494,
-0.002022131811827421,
-0.013895820826292038,
0.0030881401617079973,
-0.043810341507196426,
0.04569082707166672,
-0.00451615359634161,
0.06078063324093819,
0.03493623435497284,
-0.06172643601894379,
-0.09446151554584503,
-0.001719735562801361,
-0.03732442855834961,
-0.017135415226221085,
0.01686016656458378,
-0.017613345757126808,
-0.07991909235715866,
-0.05486968904733658,
0.01971741020679474,
-0.02119746059179306,
-0.03831952437758446,
0.047500524669885635,
0.08901825547218323,
0.07555149495601654,
0.10038787871599197,
0.09306718409061432,
0.0019258302636444569,
-0.021904971450567245,
-0.034317705780267715,
-0.03009212203323841,
-0.008205682039260864,
-0.027703793719410896,
0.025440320372581482,
-0.05993375927209854,
-4.3678607397623637e-8,
-0.06082215905189514,
0.054119501262903214,
-0.013141625560820103,
-0.05232320353388786,
0.117984339594841,
0.028187040239572525,
-0.02611401304602623,
0.06076429784297943,
-0.049366071820259094,
-0.01104911882430315,
0.0006521043251268566,
-0.012553403154015541,
0.030750928446650505,
-0.04534856602549553,
-0.04269842430949211,
0.05570029839873314,
-0.0500924251973629,
0.0830657035112381,
-0.025977106764912605,
-0.023658640682697296,
-0.019428428262472153,
0.016446391120553017,
-0.0039544375613331795,
-0.01800382323563099,
0.04731539636850357,
-0.04828114062547684,
0.003498593345284462,
0.08994398266077042,
0.10729385912418365,
0.02464510127902031,
0.011355084367096424,
0.018714342266321182,
0.014169573783874512,
-0.0233810693025589,
-0.06366897374391556,
0.08448923379182816,
0.06005416437983513,
-0.04363144934177399,
0.016285689547657967,
0.04227804020047188,
-0.025325749069452286,
0.06053101643919945,
-0.03227242827415466,
0.036364078521728516,
0.02304505743086338,
0.028966687619686127,
0.053767960518598557,
-0.026896046474575996,
-0.07331347465515137,
-0.0661339983344078,
0.0800047218799591,
0.006499490235000849,
0.054404571652412415,
0.08033298701047897,
0.003999519161880016,
-0.0462796725332737,
-0.05037977918982506,
-0.016285762190818787,
-0.022669097408652306,
0.05680084973573685,
-0.01715853624045849,
-0.0447002649307251,
-0.010852932929992676,
0.04461871460080147
] | 0 | 0 | 0 | 0 |
|
ClimaQA: An Automated Evaluation Framework for Climate Question Answering Models | https://openreview.net/forum?id=goFpCuJalN | [
"Veeramakali Vignesh Manivannan",
"Yasaman Jafari",
"Srikar Eranky",
"Spencer Ho",
"Rose Yu",
"Duncan Watson-Parris",
"Yian Ma",
"Leon Bergen",
"Taylor Berg-Kirkpatrick"
] | Poster | The use of Large Language Models (LLMs) in climate science has recently gained significant attention. However, a critical issue remains: the lack of a comprehensive evaluation framework capable of assessing the quality and scientific validity of model outputs. To address this issue, we develop *ClimaGen* (Climate QA Generator), an adaptive learning framework that generates question-answer pairs from graduate textbooks with climate scientists in the loop. As a result, we present *ClimaQA-Gold*, an expert-annotated benchmark dataset alongside *ClimaQA-Silver*, a large-scale, comprehensive synthetic QA dataset for climate science. Finally, we develop evaluation strategies and compare different LLMs on our benchmarks. Our results offer novel insights into various approaches used to enhance knowledge of climate LLMs. ClimaQA’s source code is publicly available at https://github.com/Rose-STL-Lab/genie-climaqa | Climate Benchmark, Scientific Foundation Models, Scientific Question Answering, Large Language Models, Automated QA generation | We develop ClimaGen, an automated framework for generating climate science QA datasets, and introduce ClimaQA-Gold and ClimaQA-Silver as benchmark datasets to evaluate and improve the performance of foundation models in climate science. | 12,692 | 2410.16701 | [
-0.06712598353624344,
0.03081495687365532,
0.027936400845646858,
0.08589951694011688,
0.04000686854124069,
0.0004643453285098076,
-0.032092541456222534,
0.003058734117075801,
0.027504319325089455,
0.01482152845710516,
-0.06161059811711311,
-0.14469029009342194,
0.05380064994096756,
0.0443294420838356,
-0.0016352299135178328,
0.06626860797405243,
-0.02896149456501007,
-0.03701665624976158,
-0.05213166028261185,
-0.05664124712347984,
-0.009417173452675343,
0.107559934258461,
0.04937371984124184,
0.050373680889606476,
0.060871269553899765,
-0.004122294485569,
-0.031370628625154495,
0.040362630039453506,
-0.04613649100065231,
0.09084975719451904,
-0.01667219027876854,
0.09211376309394836,
-0.0006870661745779216,
0.043416596949100494,
0.05185607820749283,
0.07576547563076019,
-0.03880505636334419,
-0.015072312206029892,
-0.046274106949567795,
0.054574113339185715,
-0.0608707033097744,
-0.03833305835723877,
0.05410882458090782,
-0.04714483395218849,
0.10938464105129242,
-0.030531078577041626,
-0.04793550819158554,
-0.030991025269031525,
-0.0003296006179880351,
-0.0050794463604688644,
-0.04576076939702034,
-0.053880639374256134,
-0.05107012763619423,
0.006582905538380146,
0.003165231551975012,
-0.007028679829090834,
-0.019176367670297623,
-0.011396055109798908,
-0.020274970680475235,
-0.07444463670253754,
-0.024509631097316742,
-0.10804164409637451,
-0.12535133957862854,
0.07017572969198227,
0.06678011268377304,
-0.040859900414943695,
-0.03650759905576706,
0.09731000661849976,
-0.05891377106308937,
-0.03292710334062576,
-0.11358806490898132,
0.11794496327638626,
-0.024156678467988968,
0.02547473832964897,
0.030765362083911896,
0.04656727612018585,
0.008929313160479069,
0.017689049243927002,
0.11865999549627304,
-0.0527350977063179,
0.03952912613749504,
-0.04256302863359451,
0.04605592042207718,
0.04365994408726692,
0.09200701862573624,
0.014567513018846512,
0.09816890209913254,
0.05962542071938515,
-0.009070921689271927,
0.005424905568361282,
-0.03632631152868271,
-0.06303475052118301,
0.11783739179372787,
0.0009773351484909654,
-0.016100365668535233,
0.12799496948719025,
0.051164209842681885,
-0.1289495974779129,
-0.023534514009952545,
0.06160453334450722,
0.0711069107055664,
0.049404747784137726,
-0.04643436521291733,
-0.09177820384502411,
-0.03702951595187187,
0.002455078298225999,
-0.005289118736982346,
-0.011891208589076996,
0.02213956043124199,
-0.07696186751127243,
-0.04996686428785324,
0.02665722742676735,
0.025492075830698013,
-0.016755979508161545,
0.039677221328020096,
-0.01671222783625126,
0.05872254818677902,
0.04224061220884323,
-0.0028993713203817606,
-0.029903864488005638,
-0.09485691785812378,
0.020065603777766228,
0.01252197939902544,
0.03405497595667839,
0.06262614578008652,
-0.015080157667398453,
-0.0060000671073794365,
1.3788583632633973e-33,
0.12694300711154938,
0.023043492808938026,
0.07292217761278152,
0.02821326069533825,
0.05529247596859932,
0.019915513694286346,
-0.011741235852241516,
0.0038994471542537212,
-0.06858553737401962,
-0.01891453005373478,
-0.01067644078284502,
0.07697867602109909,
-0.07453567534685135,
0.02880381979048252,
0.05749639496207237,
-0.010271739214658737,
-0.11143598705530167,
0.020357608795166016,
-0.06892305612564087,
0.012573186308145523,
0.013672673143446445,
-0.0115031898021698,
0.09148465842008591,
-0.03348918631672859,
0.06145710498094559,
-0.01747855171561241,
0.13562974333763123,
-0.026156999170780182,
-0.060738708823919296,
-0.0008618963183835149,
-0.03894753009080887,
-0.04460582137107849,
-0.057361602783203125,
0.0018605554942041636,
0.026009485125541687,
-0.04528174176812172,
-0.010599512606859207,
-0.005904073361307383,
-0.006495429668575525,
0.022546425461769104,
0.014955569058656693,
0.005660035647451878,
0.064188651740551,
0.008043431676924229,
-0.021114414557814598,
-0.04803458973765373,
0.026306355372071266,
-0.016644824296236038,
0.043204404413700104,
-0.016888301819562912,
-0.03914074972271919,
-0.03272027522325516,
-0.03363800048828125,
-0.04673619940876961,
0.018249018117785454,
0.0421801395714283,
0.028801329433918,
0.005164837930351496,
-0.05867611989378929,
-0.04484071955084801,
-0.02508927322924137,
0.07076992094516754,
-0.015561272390186787,
0.009536508470773697,
0.051358871161937714,
0.05331125855445862,
-0.06974118202924728,
0.06950466334819794,
0.05358634516596794,
0.02034999616444111,
0.05434742197394371,
-0.01364625059068203,
-0.04038475453853607,
0.003578566713258624,
0.03478282690048218,
-0.027464641258120537,
0.02959493175148964,
-0.06563135981559753,
0.016748890280723572,
0.04938884824514389,
0.03813977167010307,
-0.03337010741233826,
0.034290608018636703,
-0.1150728389620781,
-0.10869912058115005,
-0.10669011622667313,
0.06482904404401779,
-0.04234655946493149,
0.030390070751309395,
-0.040973227471113205,
-0.01865243725478649,
-0.06909172981977463,
0.04347856715321541,
-0.018169818446040154,
-0.025025010108947754,
-1.5628024743449535e-33,
-0.016078371554613113,
-0.03753211721777916,
-0.031201735138893127,
0.07489214837551117,
-0.04030536860227585,
-0.062413446605205536,
-0.08102302253246307,
0.02131389081478119,
0.007909312844276428,
-0.04874251410365105,
0.00583728076890111,
0.023308580741286278,
0.024797169491648674,
0.0061801946721971035,
0.05212707072496414,
-0.04591788724064827,
-0.033068012446165085,
-0.018847262486815453,
-0.08040612936019897,
0.03355923667550087,
-0.07723183929920197,
0.04915682226419449,
-0.04557429626584053,
0.05688848719000816,
-0.0691971480846405,
0.005060361698269844,
0.059901848435401917,
0.03384234011173248,
-0.04907558485865593,
-0.005389383528381586,
-0.06295280158519745,
-0.036425963044166565,
-0.05977194383740425,
0.02219311334192753,
-0.10362459719181061,
0.05156589299440384,
0.10313341021537781,
-0.10595977306365967,
-0.025635333731770515,
0.07175730913877487,
0.02465124800801277,
-0.04467250406742096,
-0.05683470517396927,
-0.03185857832431793,
-0.03223785385489464,
-0.02170700766146183,
-0.07677014172077179,
-0.006717934738844633,
0.0017995212692767382,
0.03496783971786499,
0.024236859753727913,
-0.026753779500722885,
-0.07942278683185577,
0.02992922253906727,
0.016657115891575813,
-0.0807393491268158,
-0.08592630177736282,
-0.025609605014324188,
-0.0005468211020343006,
-0.0008402368985116482,
-0.03450695797801018,
0.03489193692803383,
0.03133661299943924,
0.018261758610606194,
0.020557750016450882,
-0.07962027937173843,
-0.08558498322963715,
0.037145741283893585,
0.009590468369424343,
-0.018837228417396545,
0.04279586300253868,
-0.0019150524167343974,
-0.025884514674544334,
-0.03354145213961601,
0.049675166606903076,
0.01760909892618656,
-0.022750750184059143,
-0.03698023036122322,
-0.0005536144017241895,
-0.02682127058506012,
0.02896774373948574,
0.0010362775065004826,
0.035953838378190994,
0.11124758422374725,
0.03703511506319046,
0.007971152663230896,
0.10327970236539841,
0.00034885830245912075,
0.021540682762861252,
0.11553299427032471,
-0.01851584203541279,
0.00033635905128903687,
-0.03275173902511597,
0.08404001593589783,
-0.03891552612185478,
-3.894693278994055e-8,
-0.008013173006474972,
0.024983305484056473,
-0.01871214061975479,
0.06917957216501236,
-0.014679051004350185,
-0.015138941816985607,
-0.04257284849882126,
0.01300322450697422,
-0.03440135344862938,
0.014127636328339577,
0.02312159165740013,
0.0758964866399765,
0.015631144866347313,
0.024261241778731346,
-0.002289413008838892,
0.10770116001367569,
0.0713868960738182,
0.04442387446761131,
-0.036448270082473755,
-0.045275647193193436,
0.07985115051269531,
0.06993318349123001,
-0.011276730336248875,
0.05288996919989586,
0.08768410235643387,
0.006020193453878164,
-0.04783257842063904,
0.1185942217707634,
0.04824867472052574,
-0.00490419939160347,
-0.010387982241809368,
-0.04891292005777359,
-0.035489521920681,
-0.06841222941875458,
0.06770875304937363,
0.05240175500512123,
0.0016370599623769522,
-0.02667655050754547,
0.014338687993586063,
0.009205765090882778,
-0.020914532244205475,
0.06118495389819145,
-0.08042614907026291,
0.02985297702252865,
0.003990525845438242,
-0.023956172168254852,
-0.06993366032838821,
-0.03360408917069435,
-0.01823018118739128,
-0.020195582881569862,
-0.0028981699142605066,
-0.08771112561225891,
-0.007241320796310902,
0.005209467839449644,
0.025029947981238365,
0.009729591198265553,
-0.03230631351470947,
0.0004699233977589756,
0.007518470752984285,
0.04773043841123581,
0.02978370152413845,
-0.024498460814356804,
-0.030460916459560394,
0.04349162057042122
] | 0 | 0 | 1 | 0 |
|
Lossy Compression with Pretrained Diffusion Models | https://openreview.net/forum?id=raUnLe0Z04 | [
"Jeremy Vonderfecht",
"Feng Liu"
] | Poster | We apply Theis et al. (2022)'s DiffC algorithm to Stable Diffusion 1.5, 2.1, XL, and and Flux-dev, and demonstrate that these pretrained models are remarkably capable lossy image compressors. A principled algorithm for compression using pretrained diffusion models has been understood since at least 2020 (Ho et al.), but challenges in reverse-channel coding have prevented such algorithms from ever being fully implemented. We introduce simple workarounds that lead to the first complete implementation of DiffC, which is capable of compressing and decompressing images using Stable Diffusion in under 10 seconds. Despite requiring no additional training, our method is competitive with other state-of-the-art generative compression methods at low ultra-low bitrates. | Compression, Stable Diffusion, Flux, DDPM, DiffC | We implement a principled algorithm for compressing images using pretrained diffusion models (e.g. Stable Diffusion and Flux) | 12,671 | 2501.09815 | [
-0.061863020062446594,
-0.018906017765402794,
0.04458670690655708,
0.05150400474667549,
0.0333397202193737,
-0.09107807278633118,
-0.0841379463672638,
-0.11321056634187698,
-0.006513978820294142,
0.012753166258335114,
-0.015679067000746727,
0.018127795308828354,
-0.013899560086429119,
0.02895345352590084,
-0.06572304666042328,
-0.05529935657978058,
0.020986638963222504,
0.09637190401554108,
-0.08617912977933884,
-0.05570317059755325,
-0.040691230446100235,
-0.05471903458237648,
-0.02536858431994915,
-0.02801200933754444,
0.1487005054950714,
0.022243674844503403,
0.05157000571489334,
-0.06159668788313866,
0.09109752625226974,
-0.06177462264895439,
0.054556697607040405,
-0.048113349825143814,
0.015310925431549549,
0.04524069279432297,
-0.07070007920265198,
0.04788116365671158,
0.04831332713365555,
-0.025597596541047096,
0.03151506930589676,
0.07536812126636505,
-0.03808121755719185,
0.057163920253515244,
0.020559821277856827,
0.01641748659312725,
-0.02437060885131359,
-0.03932350128889084,
-0.01472847256809473,
-0.008106118999421597,
0.016958454623818398,
-0.018458057194948196,
0.0586216077208519,
0.06257953494787216,
-0.06659123301506042,
0.0873512551188469,
-0.04777515307068825,
0.02736724726855755,
0.033862385898828506,
0.0037044405471533537,
-0.058153118938207626,
0.008173448033630848,
-0.025802908465266228,
-0.017179245129227638,
-0.021520862355828285,
0.011668913997709751,
-0.025886116549372673,
0.0005675024003721774,
0.0939466655254364,
0.04517880082130432,
0.012369421310722828,
-0.015591492876410484,
-0.07043042033910751,
-0.014843826182186604,
-0.025938179343938828,
0.049008533358573914,
0.014187400229275227,
-0.00998677033931017,
0.1286383420228958,
0.05588578060269356,
0.004923964384943247,
-0.0355280265212059,
0.06013103574514389,
-0.05080419033765793,
0.0545736700296402,
-0.04495029151439667,
-0.018780073150992393,
-0.011109625920653343,
-0.01986195333302021,
0.05653878301382065,
-0.01671810820698738,
-0.02383206970989704,
-0.06949316710233688,
0.07228709012269974,
0.0033982396125793457,
-0.00947558507323265,
-0.00030291927396319807,
0.017482075840234756,
-0.029485316947102547,
0.0424891896545887,
-0.041607074439525604,
0.07307550311088562,
0.017138691619038582,
-0.06354080140590668,
-0.06590239703655243,
0.01563873514533043,
0.02979094907641411,
-0.11087782680988312,
0.11417169123888016,
0.07369224727153778,
-0.015440554358065128,
0.060581840574741364,
0.04105638340115547,
0.059985797852277756,
0.029820077121257782,
-0.042280592024326324,
0.0768531784415245,
-0.021496986970305443,
-0.051711514592170715,
0.01535705290734768,
-0.0014552620705217123,
-0.029989834874868393,
-0.022881433367729187,
-0.02067507430911064,
-0.12555977702140808,
-0.0325273796916008,
0.02679532952606678,
-0.09203850477933884,
-0.05779269337654114,
1.0397561610419993e-33,
0.0035761657636612654,
-0.0051139299757778645,
-0.01986049860715866,
0.026539763435721397,
-0.00842902809381485,
0.07201191037893295,
-0.024251887574791908,
-0.03447271138429642,
0.014426608569920063,
-0.005947288125753403,
-0.03269175440073013,
-0.018490921705961227,
-0.10482564568519592,
0.11573143303394318,
-0.03982316702604294,
-0.06031667813658714,
-0.023941004648804665,
0.03659882768988609,
0.028399845585227013,
0.0014101165579631925,
0.04389798268675804,
-0.04160487651824951,
-0.04193439334630966,
-0.036903221160173416,
-0.025324519723653793,
0.012689925730228424,
0.05753025785088539,
-0.1065531075000763,
0.09868670254945755,
0.03641066700220108,
-0.12701964378356934,
-0.06385583430528641,
0.020522497594356537,
-0.0857553780078888,
0.02924586459994316,
0.02761504240334034,
-0.047824159264564514,
0.021343780681490898,
-0.020990664139389992,
-0.019274897873401642,
0.02534985914826393,
0.08495766669511795,
-0.04048208147287369,
-0.09013152867555618,
-0.10980914533138275,
-0.055744778364896774,
0.059103239327669144,
0.017877809703350067,
-0.07910298556089401,
-0.03687260299921036,
0.07587578892707825,
-0.05433635413646698,
-0.06369059532880783,
-0.03365316614508629,
0.01769094169139862,
-0.009097152389585972,
0.09966694563627243,
-0.017671532928943634,
0.05782797187566757,
0.09499666094779968,
0.07376876473426819,
0.05024512857198715,
-0.05974828079342842,
0.06963778287172318,
0.0011227919021621346,
-0.07191267609596252,
-0.005489376373589039,
0.04981847479939461,
-0.09127236902713776,
0.014284908771514893,
-0.07612794637680054,
0.00976449716836214,
0.012377727776765823,
-0.09008824825286865,
0.0745493620634079,
-0.022150028496980667,
0.04854689538478851,
-0.01605118438601494,
-0.018517712131142616,
-0.07209118455648422,
-0.10578516870737076,
0.006367415655404329,
-0.022269675508141518,
-0.05385030061006546,
0.006294259335845709,
-0.0033347494900226593,
-0.03465111926198006,
-0.0007654291111975908,
-0.04208168759942055,
-0.09013223648071289,
-0.033276982605457306,
-0.04580320417881012,
0.064537912607193,
0.0023611627984791994,
0.053320739418268204,
-1.4716815501825677e-33,
0.006958975922316313,
0.10261006653308868,
-0.039773114025592804,
0.07712318748235703,
-0.06825244426727295,
-0.006687976885586977,
0.06723842024803162,
0.06382583826780319,
0.08152037113904953,
-0.03170951455831528,
0.014448017813265324,
-0.012324669398367405,
-0.06710143387317657,
0.0029161320999264717,
0.03971507400274277,
-0.07420092821121216,
0.013734221458435059,
-0.05218824744224548,
-0.015692593529820442,
-0.009871195070445538,
-0.016774022951722145,
0.10464154928922653,
0.011983506381511688,
-0.025728920474648476,
-0.002224136143922806,
0.030754810199141502,
-0.034759119153022766,
0.12970702350139618,
0.0004614111385308206,
-0.026131263002753258,
-0.040180277079343796,
0.014555390924215317,
0.02099626697599888,
-0.05164376646280289,
-0.09616681188344955,
0.026747969910502434,
-0.020959530025720596,
0.04812297970056534,
-0.03505483642220497,
0.003459138795733452,
0.05119168013334274,
0.0009823006112128496,
-0.04689369350671768,
0.052521005272865295,
0.03810366243124008,
-0.014406074769794941,
-0.08687800914049149,
-0.06448791921138763,
0.09305451065301895,
0.06509771943092346,
0.022140005603432655,
0.03739764541387558,
-0.07070077210664749,
0.08411309123039246,
0.05685912445187569,
-0.03864078223705292,
-0.0926261842250824,
-0.004999611992388964,
0.08310838788747787,
0.012815524823963642,
-0.023711731657385826,
-0.059854019433259964,
0.05122756585478783,
-0.06955894827842712,
-0.033607132732868195,
-0.012814316898584366,
-0.052130457013845444,
0.04719061404466629,
0.03279832378029823,
0.11389197409152985,
0.016287043690681458,
0.07412310689687729,
-0.012092391960322857,
0.025705445557832718,
0.004032936412841082,
-0.05376361683011055,
0.031414154917001724,
0.012313524261116982,
0.005735752638429403,
0.01399792917072773,
0.007608259096741676,
0.039530813694000244,
0.0007808901136741042,
0.015284117311239243,
0.07084362208843231,
0.028316393494606018,
0.016473274677991867,
0.0010935647878795862,
0.023331884294748306,
0.04373776540160179,
-0.03514289855957031,
0.014508839696645737,
0.10835257917642593,
0.017468515783548355,
-0.004693120718002319,
-4.008240495068094e-8,
-0.1251102089881897,
0.0271272212266922,
0.04133974388241768,
0.005428524687886238,
-0.016277281567454338,
-0.08919229358434677,
0.028575999662280083,
0.0662652850151062,
0.05064855515956879,
-0.06748327612876892,
0.09904734790325165,
0.008177150972187519,
-0.03874373435974121,
-0.0028904881328344345,
0.01273451279848814,
0.09871463477611542,
0.012693880125880241,
-0.002457674592733383,
0.0470893494784832,
-0.012325477786362171,
-0.04726133123040199,
0.008200244978070259,
0.04423817619681358,
0.0004213764623273164,
0.034044958651065826,
-0.04725629463791847,
0.03975474461913109,
-0.012961691245436668,
0.04456111416220665,
-0.03287120908498764,
-0.06185610592365265,
0.012006191536784172,
0.01474547665566206,
-0.06561592221260071,
-0.027275968343019485,
0.03007243201136589,
0.027475425973534584,
0.028136305510997772,
0.016971111297607422,
-0.05568661913275719,
-0.030986038967967033,
-0.0014450384769588709,
-0.02564120478928089,
-0.03880727291107178,
-0.014309059828519821,
-0.005975283216685057,
0.1007644534111023,
0.04578414559364319,
-0.060873132199048996,
0.008796615526080132,
0.03774331882596016,
-0.006810895167291164,
-0.023146428167819977,
-0.024275939911603928,
0.06675714999437332,
-0.047895073890686035,
-0.004956177435815334,
-0.013745090924203396,
0.02434011735022068,
0.0544445663690567,
-0.017564838752150536,
0.043500300496816635,
0.013275709934532642,
0.038926564157009125
] | https://github.com/jeremyiv/diffc | 15 | 0 | 0 | 0 |
Human-inspired Episodic Memory for Infinite Context LLMs | https://openreview.net/forum?id=BI2int5SAC | [
"Zafeirios Fountas",
"Martin Benfeghoul",
"Adnan Oomerjee",
"Fenia Christopoulou",
"Gerasimos Lampouras",
"Haitham Bou Ammar",
"Jun Wang"
] | Poster | Large language models (LLMs) have shown remarkable capabilities, but still struggle with processing extensive contexts, limiting their ability to maintain coherence and accuracy over long sequences. In contrast, the human brain excels at organising and retrieving episodic experiences across vast temporal scales, spanning a lifetime. In this work, we introduce EM-LLM, a novel approach that integrates key aspects of human episodic memory and event cognition into LLMs with no fine-tuning, enabling them to handle practically infinite context lengths while maintaining computational efficiency. EM-LLM organises sequences of tokens into coherent episodic events using a combination of Bayesian surprise and graph-theoretic boundary refinement in an online fashion. When needed, these events are retrieved through a two-stage memory process, combining similarity-based and temporally contiguous retrieval for efficient, human-inspired access to relevant information. Experiments on the LongBench and $\infty$-Bench benchmarks demonstrate EM-LLM's superior performance, consistently outperforming the state-of-the-art retrieval model InfLLM across various baseline LLMs. In addition, EM-LLM outperforms its popular counterpart, RAG, in a wide range of tasks, while requiring similar resources. Notably, EM-LLM's performance even surpasses full-context models in most tasks, while successfully performing retrieval across 10 million tokens -- a scale computationally infeasible for such models. Finally, our analysis reveals strong correlations between EM-LLM's event segmentation and human-perceived events, suggesting parallels between this artificial system and its biological counterpart, thereby offering a novel computational framework for exploring human memory mechanisms. | large language models, long context, retrieval, episodic memory, event cognition, training-free | null | 12,669 | null | [
-0.027093717828392982,
-0.054996244609355927,
0.018771419301629066,
-0.00020504659914877266,
0.07175002992153168,
0.016680607572197914,
-0.029504265636205673,
-0.0071736350655555725,
0.10887593030929565,
-0.06490885466337204,
0.0003519053861964494,
-0.00399049324914813,
0.04764418303966522,
0.008632487617433071,
0.056199926882982254,
-0.006921460386365652,
0.06432788819074631,
0.032671861350536346,
-0.04870573803782463,
-0.05089511349797249,
0.07765830308198929,
0.0111816031858325,
0.05926961451768875,
-0.04503306373953819,
0.049559395760297775,
0.03305685892701149,
0.01181480661034584,
-0.02250928059220314,
0.09086033701896667,
-0.01591605506837368,
0.15317319333553314,
0.039495404809713364,
-0.0005963938892818987,
0.009761841036379337,
-0.01939374953508377,
0.0380902998149395,
-0.10273681581020355,
0.00651784660294652,
-0.01647695153951645,
-0.039358269423246384,
-0.0032144279684871435,
0.03630336746573448,
0.0215277262032032,
0.11442271620035172,
0.07233404368162155,
-0.019285740330815315,
-0.04019367694854736,
-0.014172552153468132,
0.010731587186455727,
-0.0030853068456053734,
-0.058451954275369644,
0.01852314919233322,
-0.049101363867521286,
0.05267756059765816,
-0.03281642124056816,
0.05971694365143776,
0.023838970810174942,
0.04152105748653412,
-0.03652705252170563,
-0.08518485724925995,
-0.04618380218744278,
-0.15295806527137756,
-0.008319872431457043,
-0.014552155509591103,
-0.04669835418462753,
0.008038760162889957,
0.013142630457878113,
0.08345600217580795,
-0.00828672107309103,
0.053875479847192764,
-0.009236239828169346,
0.028935326263308525,
-0.04697253927588463,
0.11700870096683502,
0.0010474073933437467,
0.061366595327854156,
0.009680546820163727,
-0.025612154975533485,
0.016678599640727043,
-0.01706118881702423,
-0.008455895818769932,
-0.02189077064394951,
0.09028762578964233,
-0.030951473861932755,
0.1295829564332962,
0.02831900492310524,
-0.018925856798887253,
0.03592831268906593,
-0.003815531963482499,
0.07755997776985168,
-0.04583562910556793,
-0.1482659876346588,
0.042415060102939606,
-0.014047159813344479,
0.05400296673178673,
0.026093292981386185,
0.019030997529625893,
0.002284964080899954,
0.0073129478842020035,
0.0717666745185852,
0.02155553176999092,
0.053749676793813705,
-0.06797143071889877,
-0.026488784700632095,
-0.01733432151377201,
-0.014489562250673771,
-0.03310133516788483,
0.06606072187423706,
0.007951168343424797,
-0.0913296565413475,
0.007336072623729706,
0.023661304265260696,
0.0636836364865303,
0.05335649102926254,
0.02913675270974636,
-0.09001508355140686,
0.06339724361896515,
-0.04317223280668259,
0.06062476709485054,
0.06818660348653793,
0.019088249653577805,
0.044629231095314026,
-0.030850807204842567,
0.01956869103014469,
-0.05812229961156845,
-0.05206722766160965,
0.018490422517061234,
7.182253953134868e-34,
0.034742116928100586,
-0.025206631049513817,
-0.02687365747988224,
0.0758269652724266,
0.08508056402206421,
0.07348879426717758,
-0.04260387271642685,
0.06405382603406906,
-0.005635202396661043,
0.07030994445085526,
-0.04312055557966232,
0.06978519260883331,
-0.01898828335106373,
0.028369812294840813,
0.021240001544356346,
0.02194209210574627,
-0.07672660797834396,
0.06950382888317108,
-0.032384250313043594,
-0.0637015774846077,
-0.0010967523558065295,
-0.01867169886827469,
0.046679649502038956,
-0.09596145898103714,
0.041584111750125885,
0.0028833062388002872,
0.03439824655652046,
-0.037156153470277786,
0.019770773127675056,
0.010184815153479576,
-0.12733237445354462,
0.027837570756673813,
-0.03955571725964546,
0.017905766144394875,
-0.03476989269256592,
-0.033765148371458054,
-0.032299257814884186,
-0.07574687153100967,
0.001796314725652337,
-0.04963969066739082,
-0.023342086002230644,
0.03273520618677139,
0.0007947656558826566,
-0.03683953359723091,
-0.11303253471851349,
-0.13244612514972687,
0.02289680764079094,
-0.00776521023362875,
-0.020650114864110947,
-0.07655078172683716,
0.08768592029809952,
0.01717868074774742,
-0.05678747966885567,
0.006502019707113504,
-0.024533629417419434,
-0.0031954289879649878,
-0.005911278538405895,
0.03447575122117996,
-0.0020966301672160625,
0.0973130464553833,
0.07081132382154465,
0.06252864003181458,
0.0310538187623024,
0.08232250809669495,
0.07448693364858627,
0.045942068099975586,
-0.10039231181144714,
0.09896516799926758,
0.03498458489775658,
0.025543637573719025,
0.039510104805231094,
0.0008113447693176568,
0.04082352668046951,
-0.08513050526380539,
-0.008013269864022732,
-0.036555834114551544,
0.04464438930153847,
-0.060753945261240005,
-0.07679010927677155,
0.027354054152965546,
0.035610079765319824,
-0.08446605503559113,
-0.012606717646121979,
0.005536263808608055,
-0.033076319843530655,
-0.04420502111315727,
0.0469859354197979,
-0.06264746189117432,
-0.052486151456832886,
-0.05111332982778549,
0.008257684297859669,
-0.03511431813240051,
0.06346140056848526,
0.012587346136569977,
-0.03712774068117142,
-1.7467023207173617e-33,
-0.02593362145125866,
-0.06183076649904251,
0.014000910334289074,
0.006641838699579239,
0.022270342335104942,
-0.06998268514871597,
-0.008859667927026749,
0.04085453599691391,
-0.13175718486309052,
-0.04765082523226738,
-0.06980190426111221,
-0.06358689814805984,
0.05910887196660042,
-0.0019347447669133544,
0.05853163078427315,
-0.055638089776039124,
0.017296450212597847,
0.016858210787177086,
-0.006470929365605116,
0.0018697109771892428,
0.006432854570448399,
-0.03385103866457939,
-0.08360709249973297,
-0.020957976579666138,
0.005126845557242632,
0.0531756654381752,
-0.0025012067053467035,
-0.05070578306913376,
-0.05500228703022003,
-0.007376206573098898,
-0.10208169370889664,
0.005374423693865538,
-0.04382876306772232,
-0.04220116510987282,
-0.10586222261190414,
0.06472449749708176,
0.041233256459236145,
0.013449721038341522,
-0.04338940978050232,
-0.0011210704687982798,
0.05533820763230324,
0.00023776906891725957,
-0.022195320576429367,
-0.02052600122988224,
-0.024584941565990448,
0.01069470215588808,
-0.13469819724559784,
0.021867170929908752,
0.06683506816625595,
0.04836440458893776,
-0.00855126790702343,
0.08325092494487762,
-0.10288970917463303,
-0.03686690330505371,
-0.07179834693670273,
-0.08879418671131134,
-0.0934462770819664,
-0.06945069134235382,
-0.016838649287819862,
-0.039566703140735626,
-0.0487031564116478,
-0.023080110549926758,
-0.03902807459235191,
-0.029820993542671204,
0.04241824522614479,
-0.0026190427597612143,
0.018353596329689026,
-0.004300753120332956,
-0.06819966435432434,
-0.004118435084819794,
0.043140169233083725,
-0.008493323810398579,
-0.1085960641503334,
-0.035130660980939865,
0.06478464603424072,
0.05475683882832527,
-0.0013256684178486466,
-0.04408455267548561,
-0.032236672937870026,
-0.0678429901599884,
-0.05743159353733063,
-0.00683259591460228,
-0.043024785816669464,
0.09349708259105682,
0.014197186566889286,
0.07767906785011292,
0.0015904135070741177,
0.09335459023714066,
0.09175039082765579,
0.0406348817050457,
0.006910170428454876,
0.004857058636844158,
-0.018982961773872375,
0.06107797473669052,
0.03258698806166649,
-4.340957460158279e-8,
-0.09172604978084564,
0.025901436805725098,
0.00028081604978069663,
0.037559933960437775,
0.026307661086320877,
-0.04303812235593796,
0.011358313262462616,
0.08586779236793518,
-0.013699659146368504,
-0.05721079558134079,
0.03259800001978874,
-0.012942900881171227,
-0.01783190853893757,
-0.020041940733790398,
0.014588491059839725,
0.005665898323059082,
0.06264134496450424,
0.01777755282819271,
0.034680284559726715,
-0.012495714239776134,
0.043082013726234436,
0.0473768450319767,
0.0032154868822544813,
0.00014699969324283302,
-0.018487639725208282,
-0.02184976264834404,
-0.023461192846298218,
0.12707725167274475,
0.0324377603828907,
-0.04125150293111801,
-0.046177882701158524,
0.011133961379528046,
-0.056682176887989044,
-0.0036140603478997946,
-0.03601142019033432,
-0.029470983892679214,
0.02230546623468399,
-0.0334312841296196,
-0.0023289171513170004,
-0.04012473672628403,
0.039040230214595795,
0.0012005375465378165,
0.04524737969040871,
0.07384414970874786,
0.02238217182457447,
0.02889898233115673,
0.029697393998503685,
-0.013832678087055683,
0.03343133628368378,
-0.009814385324716568,
-0.012009036727249622,
-0.022211402654647827,
-0.003970189951360226,
0.018538327887654305,
0.09816673398017883,
0.0683131217956543,
0.004908975213766098,
-0.02570071443915367,
0.03401247039437294,
-0.018810467794537544,
0.08966679126024246,
0.007515633013099432,
-0.11725037544965744,
-0.048107363283634186
] | 0 | 0 | 0 | 0 |
|
Global Convergence of Policy Gradient in Average Reward MDPs | https://openreview.net/forum?id=2PRpcmJecX | [
"Navdeep Kumar",
"Yashaswini Murthy",
"Itai Shufaro",
"Kfir Yehuda Levy",
"R. Srikant",
"Shie Mannor"
] | Poster | We present the first comprehensive finite-time global convergence analysis of policy gradient for infinite horizon average reward Markov decision processes (MDPs). Specifically, we focus on ergodic tabular MDPs with finite state and action spaces. Our analysis shows that the policy gradient iterates converge to the optimal policy at a sublinear rate of $O(\frac{1}{T})$, where $T$ represents the number of iterations. Performance bounds for discounted reward MDPs cannot be easily extended to average reward MDPs as the bounds grow proportional to the fifth power of the effective horizon. Recent work on such extensions makes a smoothness assumption that has not been verified. Thus, our primary contribution is in providing the first complete proof that the policy gradient algorithm converges globally for average-reward MDPs, without such an assumption. We also obtain the corresponding finite-time performance guarantees. In contrast to the existing discounted reward performance bounds, our performance bounds have an explicit dependence on constants that capture the complexity of the underlying MDP. Motivated by this observation, we reexamine and improve the existing performance bounds for discounted reward MDPs. We also present simulations that empirically validate the result. | Policy Gradient, Reinforcement Learning, Average Reward MDPs | null | 12,657 | null | [
-0.10477973520755768,
-0.06271377950906754,
-0.005019109696149826,
-0.07137483358383179,
0.038821715861558914,
0.00545540964230895,
0.009490217082202435,
-0.024869229644536972,
0.033495109528303146,
0.07180053740739822,
-0.07182516902685165,
0.020302096381783485,
0.046864159405231476,
0.03457043692469597,
-0.0337848961353302,
0.0528106614947319,
0.016634179279208183,
-0.007423939649015665,
0.022525617852807045,
-0.007824533618986607,
0.001865497324615717,
0.015810811892151833,
0.0634193867444992,
0.004937687423080206,
-0.04166612774133682,
-0.012216722592711449,
-0.0352846197783947,
-0.0215842816978693,
-0.008026400581002235,
0.00682058883830905,
0.009472399950027466,
-0.09616851061582565,
0.06459010392427444,
-0.0825057104229927,
-0.009902121499180794,
0.06795929372310638,
-0.1051449179649353,
-0.020642682909965515,
-0.05548004060983658,
0.014508958905935287,
0.027620982378721237,
0.020983992144465446,
-0.0043828473426401615,
0.002691066125407815,
0.024116789922118187,
-0.038989707827568054,
0.030226411297917366,
-0.048752639442682266,
-0.028399396687746048,
-0.04052469879388809,
0.026364246383309364,
0.03056793101131916,
-0.03209960460662842,
-0.041081417351961136,
-0.06990248709917068,
0.031064528971910477,
0.04176454618573189,
0.0026121260598301888,
-0.04664365202188492,
-0.06947449594736099,
-0.057411812245845795,
-0.13283498585224152,
-0.010922462679445744,
-0.019152885302901268,
-0.05692537873983383,
-0.03871827945113182,
0.04093141853809357,
0.005840839818120003,
-0.007747665047645569,
0.0751292034983635,
-0.08180724084377289,
0.0039128502830863,
-0.045064039528369904,
-0.028050284832715988,
0.0018174732103943825,
-0.002264101756736636,
0.027990683913230896,
-0.042850106954574585,
-0.005607803352177143,
-0.033958759158849716,
-0.0002291798300575465,
0.03260543569922447,
0.026826370507478714,
-0.08061971515417099,
0.013497346080839634,
-0.03183966502547264,
-0.018887747079133987,
0.016304491087794304,
0.19419923424720764,
0.0011487457668408751,
-0.010635415092110634,
-0.004589724820107222,
-0.07288836687803268,
-0.017756322398781776,
0.020596561953425407,
-0.010901707224547863,
-0.013329374603927135,
-0.013693153858184814,
-0.07700882852077484,
0.09230230003595352,
0.002931928960606456,
-0.01623539812862873,
-0.028483234345912933,
-0.008561283349990845,
0.06795401871204376,
-0.05233939737081528,
0.017603708431124687,
-0.009719683788716793,
0.03064296394586563,
-0.07757982611656189,
0.03235585242509842,
0.007130860351026058,
0.06259793043136597,
0.04646071419119835,
-0.03868498280644417,
-0.018018191680312157,
-0.022961314767599106,
-0.07371106743812561,
-0.010485967621207237,
0.054940130561590195,
-0.008268624544143677,
-0.01689896546304226,
0.06414429843425751,
0.07739360630512238,
-0.03236067295074463,
0.030861765146255493,
-0.007068436127156019,
2.2163574969349753e-33,
0.001990648452192545,
-0.009687845595180988,
0.09426257014274597,
-0.08294074237346649,
0.039854418486356735,
0.06115071848034859,
0.09552931785583496,
-0.01287919282913208,
-0.05930931493639946,
0.028923386707901955,
-0.015451679937541485,
0.041850604116916656,
-0.06585035473108292,
0.07302226126194,
0.07484973222017288,
-0.030715767294168472,
-0.010604563169181347,
0.04781295731663704,
0.05518961325287819,
-0.06570740044116974,
0.08218841254711151,
-0.013298097997903824,
-0.025973975658416748,
-0.00863423477858305,
0.05617624893784523,
-0.011457695625722408,
-0.026536010205745697,
0.0012042248854413629,
0.028521709144115448,
0.012281382456421852,
-0.015352505259215832,
0.009262971580028534,
-0.04627910628914833,
0.013592513278126717,
0.012916533276438713,
0.05054039508104324,
-0.008429426699876785,
0.01599358581006527,
-0.009792093187570572,
-0.06127152219414711,
-0.009156442247331142,
0.045431774109601974,
0.04478756710886955,
0.018508337438106537,
-0.13928519189357758,
-0.10464338213205338,
0.03518145531415939,
-0.01602792739868164,
-0.05499360337853432,
-0.018999850377440453,
0.03894125670194626,
-0.012956991791725159,
-0.010804837569594383,
-0.06921093165874481,
-0.03748949617147446,
-0.03959651663899422,
-0.03901579603552818,
0.08197207748889923,
0.009145320393145084,
0.08162355422973633,
0.04714159667491913,
-0.06271465122699738,
-0.01808933913707733,
-0.0038772174157202244,
-0.0842856615781784,
0.0930941179394722,
-0.024576369673013687,
0.031193088740110397,
0.06661716103553772,
-0.03194597363471985,
0.06441184878349304,
0.043808382004499435,
0.04684533551335335,
-0.10716643929481506,
0.08614630252122879,
-0.01736375503242016,
0.08954650163650513,
-0.08006113022565842,
-0.02615564689040184,
0.03554162383079529,
-0.06393706798553467,
-0.056924007833004,
-0.04182276129722595,
-0.04059410095214844,
0.0007274710806086659,
-0.042704787105321884,
0.023903172463178635,
-0.04662344604730606,
0.017789416015148163,
-0.009921125136315823,
-0.09806384891271591,
-0.013050940819084644,
0.04718968644738197,
0.054821524769067764,
0.014225920662283897,
-3.854670054956906e-33,
0.09159749746322632,
-0.11317921429872513,
0.04345739632844925,
0.039420176297426224,
0.02688475139439106,
0.02229023166000843,
-0.07327690720558167,
0.011727575212717056,
-0.03248842805624008,
-0.06484340876340866,
-0.06577033549547195,
0.0839960128068924,
0.050382133573293686,
0.043898481875658035,
0.035204481333494186,
-0.09190495312213898,
-0.0034592594020068645,
0.10195782035589218,
0.004047637805342674,
0.0021042197477072477,
0.07507608085870743,
0.056167762726545334,
-0.0728457048535347,
0.08984583616256714,
0.045654296875,
0.037285950034856796,
-0.032974015921354294,
0.013188090175390244,
-0.03203847259283066,
0.03495138883590698,
-0.09218988567590714,
-0.0596257708966732,
-0.12710298597812653,
0.03257336840033531,
-0.03869931027293205,
0.00011805404938058928,
0.04497135058045387,
0.04611486569046974,
-0.08427052199840546,
0.13260744512081146,
0.01987329125404358,
-0.07428028434515,
0.05061177536845207,
0.019095027819275856,
0.04544302076101303,
-0.06155520677566528,
-0.056716859340667725,
0.030839787796139717,
-0.019270066171884537,
-0.008511902764439583,
-0.03292278200387955,
-0.01840907335281372,
-0.033223673701286316,
0.09016889333724976,
-0.07703762501478195,
0.017989130690693855,
-0.08015633374452591,
0.011369060724973679,
0.03570594638586044,
-0.05472230538725853,
-0.024896884337067604,
0.1048375815153122,
-0.10845521092414856,
0.07635797560214996,
0.055119678378105164,
0.035591818392276764,
-0.012569711543619633,
-0.07375284284353256,
0.065923310816288,
0.041726239025592804,
-0.009575508534908295,
0.051929816603660583,
-0.04687584564089775,
0.03846050798892975,
0.07228277623653412,
0.05033472925424576,
0.07205495983362198,
-0.03796333819627762,
-0.02276363968849182,
-0.04247349128127098,
0.020280374214053154,
0.08262186497449875,
0.055374354124069214,
-0.05717429146170616,
-0.02704959362745285,
-0.01689431630074978,
0.10741692036390305,
0.007321817800402641,
-0.01985062099993229,
-0.010350242257118225,
0.03614312410354614,
-0.062050383538007736,
0.019354907795786858,
0.04495915770530701,
-0.07384999096393585,
-4.401905684403573e-8,
-0.017960602417588234,
0.03513370826840401,
0.02954530529677868,
0.0019976156763732433,
0.05878904089331627,
0.03001968190073967,
0.002173825167119503,
-0.020339127629995346,
0.009640355594456196,
-0.07139728218317032,
0.12214986979961395,
0.07584187388420105,
0.03230541944503784,
-0.060174521058797836,
-0.033066801726818085,
0.0368582047522068,
-0.08421056717634201,
0.005350680556148291,
0.0065199341624975204,
0.02106454409658909,
-0.02481718547642231,
0.06665537506341934,
0.02341596409678459,
-0.07037996500730515,
-0.04850117862224579,
-0.048493392765522,
-0.02820381335914135,
0.05561189725995064,
0.044183433055877686,
0.0010203160345554352,
-0.00007886246748967096,
0.026702597737312317,
0.07651545852422714,
0.03820207715034485,
0.03963641822338104,
-0.020469844341278076,
-0.0001985500566661358,
0.0006834319210611284,
0.02137554995715618,
-0.026441318914294243,
-0.0016822631005197763,
0.039993464946746826,
0.04547952115535736,
-0.06035466864705086,
-0.03771321102976799,
0.026286812499165535,
-0.09402092546224594,
-0.010946236550807953,
0.0574319027364254,
-0.04689464345574379,
0.03582444787025452,
0.07166001945734024,
-0.04744589701294899,
0.05513496696949005,
0.07936900109052658,
-0.0076944706961512566,
0.04884649068117142,
-0.13790559768676758,
0.03144102916121483,
0.041270215064287186,
-0.0004334081895649433,
-0.04888765886425972,
-0.03435476869344711,
-0.0020124686416238546
] | 0 | 0 | 0 | 0 |
|
What Matters in Learning from Large-Scale Datasets for Robot Manipulation | https://openreview.net/forum?id=LqhorpRLIm | [
"Vaibhav Saxena",
"Matthew Bronars",
"Nadun Ranawaka Arachchige",
"Kuancheng Wang",
"Woo Chul Shin",
"Soroush Nasiriany",
"Ajay Mandlekar",
"Danfei Xu"
] | Poster | Imitation learning from large multi-task demonstration datasets has emerged as a promising path for building generally-capable robots. As a result, 1000s of hours have been spent on building such large-scale datasets around the globe. Despite the continuous growth of such efforts, we still lack a systematic understanding of what data should be collected to improve the utility of a robotics dataset and facilitate downstream policy learning. In this work, we conduct a large-scale dataset composition study to answer this question. We develop a data generation framework to procedurally emulate common sources of diversity in existing datasets (such as sensor placements and object types and arrangements), and use it to generate large-scale robot datasets with controlled compositions, enabling a suite of dataset composition studies that would be prohibitively expensive in the real world. We focus on two practical settings: (1) what types of diversity should be emphasized when future researchers collect large-scale datasets for robotics, and (2) how should current practitioners retrieve relevant demonstrations from existing datasets to maximize downstream policy performance on tasks of interest. Our study yields several critical insights -- for example, we find that camera poses and spatial arrangements are crucial dimensions for both diversity in collection and alignment in retrieval. In real-world robot learning settings, we find that not only do our insights from simulation carry over, but our retrieval strategies on existing datasets such as DROID allow us to consistently outperform existing training strategies by up to 70\%. | imitation learning, robotics, dataset composition | null | 12,652 | null | [
-0.030477110296487808,
-0.10147327929735184,
0.08449387550354004,
-0.0035252668894827366,
0.08144887536764145,
-0.0013940357603132725,
-0.032201576977968216,
-0.06895070523023605,
0.008034419268369675,
0.02018846943974495,
-0.0247940830886364,
-0.07284244894981384,
0.05511441081762314,
0.018719976767897606,
0.033084969967603683,
0.0482647530734539,
0.11788563430309296,
0.05449426546692848,
-0.10098502784967422,
-0.011117258109152317,
-0.01623503491282463,
0.011606109328567982,
0.10271210223436356,
0.02030886523425579,
-0.026329122483730316,
0.03606915473937988,
0.052154239267110825,
-0.10835319012403488,
0.00899887178093195,
-0.03234744817018509,
0.04313696548342705,
0.025436336174607277,
-0.056231241673231125,
-0.03361701965332031,
0.002932640491053462,
0.13106432557106018,
-0.0701366513967514,
0.02751711942255497,
-0.009234520606696606,
-0.032898738980293274,
-0.00799871701747179,
0.018446842208504677,
0.05313553661108017,
0.008607841096818447,
0.02725900523364544,
0.06017173454165459,
0.040386050939559937,
-0.11873401701450348,
0.06495260447263718,
0.028550731018185616,
-0.08409386873245239,
-0.02518860064446926,
0.041006576269865036,
0.011476130224764347,
-0.005069821607321501,
-0.031183699145913124,
0.06436408311128616,
-0.009309152141213417,
0.0718640685081482,
0.00038516547647304833,
0.011640931479632854,
-0.103745236992836,
-0.007664005272090435,
-0.051342666149139404,
-0.04199083894491196,
-0.05915869027376175,
-0.030511166900396347,
0.03367771953344345,
-0.015263061970472336,
-0.0020859502255916595,
-0.0174094270914793,
0.05419732630252838,
0.018210453912615776,
0.03989917412400246,
0.020395301282405853,
0.015354678966104984,
0.00700679374858737,
-0.0677194818854332,
-0.02113375812768936,
-0.10230672359466553,
0.02385568991303444,
-0.026749681681394577,
-0.010500884614884853,
0.07989729195833206,
0.07384824752807617,
-0.043876197189092636,
0.029044585302472115,
0.06324080377817154,
-0.00974119734019041,
0.03687429800629616,
-0.07059726119041443,
-0.008971196599304676,
0.08545321226119995,
-0.037691861391067505,
-0.004728778265416622,
0.02845190279185772,
-0.018763260915875435,
-0.1061655580997467,
0.06329279392957687,
0.0541352741420269,
-0.008244466967880726,
0.07807274907827377,
0.03767410293221474,
0.017624661326408386,
-0.03547482565045357,
-0.06674809753894806,
-0.031746286898851395,
0.0066557517275214195,
0.062003858387470245,
-0.016238277778029442,
0.05594184249639511,
-0.008135530166327953,
0.020699668675661087,
0.03798788785934448,
0.011855321004986763,
-0.05072682723402977,
-0.03171459585428238,
-0.007878276519477367,
-0.032991837710142136,
0.1417795568704605,
0.007427703123539686,
-0.04242744669318199,
0.03336689621210098,
0.025876447558403015,
0.07773656398057938,
-0.008945832028985023,
-0.01715191826224327,
8.142004682962956e-34,
0.10327865183353424,
0.004158799070864916,
0.05942663177847862,
0.05097616836428642,
0.07127730548381805,
-0.027749236673116684,
0.05896473675966263,
0.023781653493642807,
-0.05621938407421112,
0.03534203767776489,
-0.00508520845323801,
0.02445010282099247,
-0.065201535820961,
0.12715177237987518,
0.013903128914535046,
0.0029019454959779978,
-0.009821929968893528,
0.059311337769031525,
-0.05376483127474785,
0.050185129046440125,
0.07272839546203613,
-0.006089052185416222,
-0.01891779713332653,
-0.01584489271044731,
0.04861926659941673,
-0.06408601999282837,
0.004578183405101299,
0.02323848381638527,
-0.020232321694493294,
-0.009844510816037655,
-0.09240574389696121,
0.017104538157582283,
-0.051894497126340866,
0.11629092693328857,
0.04423769563436508,
0.03042246215045452,
-0.04033753648400307,
-0.05421606823801994,
0.03145040199160576,
0.016095919534564018,
0.07672271877527237,
-0.011302224360406399,
-0.014293831773102283,
0.001015480374917388,
-0.047860462218523026,
-0.07154586166143417,
0.032086972147226334,
0.029373211786150932,
-0.03788677975535393,
-0.04833700880408287,
0.06559906899929047,
0.04537927731871605,
0.015942752361297607,
-0.09457910060882568,
0.016163550317287445,
-0.056237440556287766,
-0.037274014204740524,
-0.010073845274746418,
0.07207896560430527,
-0.023571966215968132,
-0.025584612041711807,
0.010441984049975872,
-0.04093949496746063,
0.0996139720082283,
-0.0014524515718221664,
0.10016532242298126,
0.006140219513326883,
0.04877430945634842,
0.10110577195882797,
-0.008309929631650448,
-0.017011189833283424,
0.026849418878555298,
-0.06561814993619919,
-0.0811171755194664,
0.07340496778488159,
0.009490292519330978,
0.0004350745875854045,
-0.0819057896733284,
0.026124559342861176,
0.0402325876057148,
-0.027704579755663872,
0.01593894138932228,
-0.03220932185649872,
0.00041888264240697026,
0.0034758553374558687,
-0.01698310859501362,
0.03802771866321564,
-0.057073045521974564,
-0.0315556526184082,
-0.017292918637394905,
-0.12743882834911346,
-0.11292915046215057,
-0.1700170636177063,
-0.007609919179230928,
0.014190015383064747,
-2.7796347370528376e-33,
-0.04140063002705574,
0.003927373327314854,
0.029532497748732567,
0.0300845205783844,
-0.03545001149177551,
-0.05420805141329765,
-0.04673776030540466,
-0.037616338580846786,
-0.01326050516217947,
-0.03598262742161751,
-0.08956485986709595,
-0.0476900115609169,
0.04315047711133957,
-0.07786907255649567,
-0.04678256809711456,
-0.03132238611578941,
0.004857166670262814,
0.04400426894426346,
0.03404746204614639,
-0.026695700362324715,
-0.00019297408289276063,
0.08474387973546982,
-0.0014435130869969726,
0.05098317563533783,
-0.03051682934165001,
0.02764379419386387,
-0.053639478981494904,
0.017814425751566887,
0.019416913390159607,
0.04924551770091057,
-0.07631785422563553,
-0.09171499311923981,
-0.006450238637626171,
-0.015619108453392982,
-0.07944874465465546,
0.02033190056681633,
0.046778708696365356,
0.031255628913640976,
-0.045181937515735626,
0.06252046674489975,
-0.01561194472014904,
-0.08634284883737564,
-0.05299104005098343,
0.022169500589370728,
-0.035332005470991135,
-0.08364176750183105,
-0.058566514402627945,
0.05825940892100334,
-0.007496973499655724,
-0.00254248920828104,
-0.050207629799842834,
0.05568685382604599,
-0.08131574839353561,
-0.0485101044178009,
-0.01589512825012207,
-0.07018500566482544,
0.02929225005209446,
0.002421385608613491,
0.03357110172510147,
-0.02537425048649311,
-0.08480417728424072,
0.011297359131276608,
0.001971662975847721,
0.0906260535120964,
-0.005023594945669174,
-0.015443867072463036,
-0.0231256652623415,
-0.0664350613951683,
0.006303758360445499,
0.005080333445221186,
0.025306297466158867,
-0.04820473864674568,
0.04269582778215408,
-0.003177792299538851,
0.029254816472530365,
-0.007442172151058912,
0.015713922679424286,
-0.012020221911370754,
0.03660863637924194,
0.061944734305143356,
0.04395417496562004,
0.00908985547721386,
0.0898253470659256,
0.08671627193689346,
0.03753777593374252,
0.11565794795751572,
-0.0243669506162405,
0.053264912217855453,
0.0068988134153187275,
0.0152311185374856,
-0.04404091462492943,
-0.02724791131913662,
-0.04961929842829704,
0.047854408621788025,
-0.018999114632606506,
-4.0206874274417714e-8,
-0.022575877606868744,
0.10556595027446747,
0.07626300305128098,
0.030127162113785744,
0.002338853431865573,
0.01975850760936737,
-0.05735133960843086,
0.06786300241947174,
-0.044127944856882095,
-0.00003787664900301024,
0.08718927204608917,
0.03270738199353218,
0.02849874086678028,
-0.04610651358962059,
0.04312106966972351,
0.03946326673030853,
0.057690177112817764,
-0.05347228795289993,
-0.011461500078439713,
0.054694876074790955,
0.027188187465071678,
-0.028062405064702034,
0.05937815457582474,
-0.037918392568826675,
0.013205639086663723,
-0.0609482079744339,
-0.05997458100318909,
0.022860435768961906,
0.0099643524736166,
0.016580309718847275,
-0.04117773473262787,
-0.08198365569114685,
-0.007789739407598972,
-0.028454873710870743,
0.0413995124399662,
-0.036701079457998276,
-0.06581562012434006,
-0.029812827706336975,
-0.00001950527439475991,
-0.10001799464225769,
-0.008164092898368835,
0.043968990445137024,
-0.004342952743172646,
0.015894241631031036,
-0.008733874186873436,
0.004122807178646326,
-0.08437418937683105,
-0.06224985048174858,
-0.008803758770227432,
-0.028613144531846046,
-0.037842340767383575,
-0.05153260752558708,
-0.020549451932311058,
0.11267074942588806,
0.12855246663093567,
0.05640329793095589,
-0.018986359238624573,
-0.09290307760238647,
0.011508563533425331,
0.059116508811712265,
0.013117303140461445,
-0.0006093637202866375,
-0.10218843817710876,
-0.04893767461180687
] | 0 | 0 | 0 | 0 |
|
Rapidly Adapting Policies to the Real-World via Simulation-Guided Fine-Tuning | https://openreview.net/forum?id=XwUrzurG94 | [
"Patrick Yin",
"Tyler Westenbroek",
"Ching-An Cheng",
"Andrey Kolobov",
"Abhishek Gupta"
] | Poster | Robot learning requires a considerable amount of high-quality data to realize the promise of generalization. However, large data sets are costly to collect in the real world. Physics simulators can cheaply generate vast data sets with broad coverage over states, actions, and environments. However, physics engines are fundamentally misspecified approximations to reality. This makes direct zero-shot transfer from simulation to reality challenging, especially in tasks where precise and force-sensitive manipulation is necessary. Thus, fine-tuning these policies with small real-world data sets is an appealing pathway for scaling robot learning. However, current reinforcement learning fine-tuning frameworks leverage general, unstructured exploration strategies which are too inefficient to make real-world adaptation practical. This paper introduces the \emph{Simulation-Guided Fine-tuning} (SGFT) framework, which demonstrates how to extract structural priors from physics simulators to substantially accelerate real-world adaptation. Specifically, our approach uses a value function learned in simulation to guide real-world exploration. We demonstrate this approach across five real-world dexterous manipulation tasks where zero-shot sim-to-real transfer fails. We further demonstrate our framework substantially outperforms baseline fine-tuning methods, requiring up to an order of magnitude fewer real-world samples and succeeding at difficult tasks where prior approaches fail entirely. Last but not least, we provide theoretical justification for this new paradigm which underpins how SGFT can rapidly learn high-performance policies in the face of large sim-to-real dynamics gaps. | Robot Learning, Reinforcement Learning, Fine-Tuning | We use value functions trained in simulation to guide efficient exploration for efficient real-world finetuning, with robot hardware and theoretical results | 12,648 | 2502.02705 | [
-0.005913380533456802,
-0.0669027790427208,
0.05949472635984421,
0.058413974940776825,
0.03215290978550911,
-0.04257136210799217,
-0.027616610750555992,
-0.05973086878657341,
-0.03599826619029045,
0.01889949105679989,
0.031076019629836082,
-0.06415533274412155,
-0.01595846191048622,
0.021469932049512863,
0.001581392833031714,
-0.005143467336893082,
0.09550508111715317,
0.07799989730119705,
-0.12876446545124054,
0.006310055963695049,
0.010455770418047905,
0.009171003475785255,
0.09596272557973862,
-0.019405778497457504,
0.004140579141676426,
-0.0005524575826711953,
-0.006630916614085436,
-0.05617742985486984,
-0.000564988178666681,
-0.06089140847325325,
0.011990421451628208,
-0.0340161956846714,
-0.12265599519014359,
-0.12679234147071838,
-0.016928428784012794,
0.08085936307907104,
-0.0929555594921112,
0.01338875014334917,
-0.06704989075660706,
-0.009286349639296532,
-0.0008290657424367964,
0.04101771488785744,
0.06785956770181656,
0.03720711171627045,
0.08827203512191772,
0.06931612640619278,
0.03772517666220665,
-0.05992559716105461,
0.07471282035112381,
-0.019562896341085434,
-0.09530617296695709,
-0.03409416601061821,
0.03274838998913765,
-0.008271773345768452,
0.06828664988279343,
0.06397995352745056,
0.05628257244825363,
-0.02589198760688305,
0.02649252861738205,
-0.03561326116323471,
0.0033119937870651484,
-0.09463997185230255,
0.0136070866137743,
-0.03557414188981056,
-0.0486026369035244,
-0.026151757687330246,
0.03759928420186043,
0.016674892976880074,
0.0500025637447834,
0.005870366003364325,
0.004529573954641819,
0.0357595831155777,
0.01192929781973362,
0.013596674427390099,
0.07158518582582474,
0.03160365670919418,
-0.07339918613433838,
-0.003783621359616518,
0.07764258980751038,
0.04091409966349602,
0.01931055635213852,
-0.046673960983753204,
-0.04112829640507698,
0.05978148430585861,
0.07476507872343063,
-0.010986974462866783,
0.05091536045074463,
0.044973816722631454,
0.08846549689769745,
0.05595056340098381,
0.03081197291612625,
-0.03796028718352318,
-0.05799006670713425,
-0.039140619337558746,
-0.06022854894399643,
-0.03422865644097328,
-0.004786522127687931,
-0.1148078590631485,
-0.08211692422628403,
0.031623952090740204,
0.03837234526872635,
0.08162204921245575,
0.010396049357950687,
0.07113339751958847,
-0.003076908178627491,
-0.01672891154885292,
0.003982932306826115,
0.039451759308576584,
0.00033729177084751427,
-0.06409181654453278,
0.04967542737722397,
-0.015701696276664734,
0.0705505907535553,
-0.003860070602968335,
0.011676517315208912,
-0.016005393117666245,
-0.002067367546260357,
-0.02121148258447647,
-0.04050368815660477,
0.08340204507112503,
-0.030454710125923157,
-0.06361968070268631,
-0.01379935909062624,
0.08103737980127335,
0.048392824828624725,
0.00365472212433815,
-0.053120508790016174,
8.30531482742575e-34,
0.051145393401384354,
0.04995294287800789,
0.0656215250492096,
0.027926746755838394,
0.04207267239689827,
-0.04677596688270569,
0.05332385376095772,
0.04011489823460579,
-0.08007222414016724,
0.05142708867788315,
0.003196548903360963,
0.051614910364151,
-0.05869479849934578,
0.08083736151456833,
0.046102724969387054,
0.0007423886563628912,
-0.050248123705387115,
0.029987948015332222,
-0.0156786497682333,
0.02108851633965969,
0.04952467978000641,
-0.030959637835621834,
-0.04074324294924736,
-0.13641583919525146,
0.035131558775901794,
0.02976902388036251,
-0.0263845045119524,
0.06218760460615158,
-0.0071557131595909595,
-0.02160252071917057,
-0.069344662129879,
0.05572516471147537,
-0.03132706880569458,
0.060320671647787094,
-0.013658648356795311,
-0.0016656792722642422,
-0.04017849266529083,
-0.07475069910287857,
-0.05801631882786751,
0.002258321736007929,
0.015515131875872612,
0.026746602728962898,
0.007182467728853226,
0.02741059474647045,
-0.06416209042072296,
-0.08042572438716888,
0.03289315477013588,
0.03414119407534599,
-0.07283119112253189,
-0.04555433616042137,
0.01817506179213524,
0.014741523191332817,
0.0038982275873422623,
-0.06356094032526016,
-0.03062756359577179,
-0.03367272764444351,
0.05256591737270355,
0.009080042131245136,
-0.05439165234565735,
0.04148947075009346,
0.03584929183125496,
-0.0199287086725235,
-0.06213178485631943,
0.13136012852191925,
-0.02237594686448574,
0.08686530590057373,
-0.0452602244913578,
0.06375172734260559,
0.06559060513973236,
0.008413142524659634,
-0.028187735006213188,
-0.009425501339137554,
-0.0011964195873588324,
-0.05531534180045128,
0.051764585077762604,
0.02532663755118847,
0.05733180791139603,
-0.06114134564995766,
0.024241004139184952,
0.04776646941900253,
0.0022504145745187998,
0.0673617422580719,
-0.04308760166168213,
0.030858537182211876,
-0.0008266975055448711,
-0.09321525692939758,
-0.002555667422711849,
-0.07886820286512375,
-0.0318271704018116,
-0.07988746464252472,
-0.10861131548881531,
-0.09304408729076385,
-0.04266226664185524,
-0.008866524323821068,
-0.022513898089528084,
-1.1000799597323067e-33,
0.0205784160643816,
-0.054329898208379745,
0.03542245924472809,
0.08286648243665695,
-0.03151002153754234,
-0.013933755457401276,
-0.04013671353459358,
-0.06869763135910034,
0.07529092580080032,
-0.07956922799348831,
-0.11638066917657852,
-0.04114875942468643,
-0.02179330214858055,
-0.030971864238381386,
0.005443055182695389,
0.005969451740384102,
-0.043089400976896286,
-0.03470763936638832,
0.00712075550109148,
-0.01815633475780487,
0.016876276582479477,
0.09171388298273087,
-0.01770278997719288,
0.08647768944501877,
0.0034959258046001196,
0.004241996444761753,
0.014447459951043129,
0.001347930752672255,
-0.002172050066292286,
0.03320331871509552,
-0.051278430968523026,
-0.023902535438537598,
0.012702257372438908,
0.02892959862947464,
-0.06299490481615067,
0.09082774817943573,
0.10535029321908951,
0.04266689717769623,
0.03464633226394653,
0.09212301671504974,
0.007353508844971657,
-0.02339853160083294,
-0.04950220137834549,
0.04774721711874008,
0.05700479820370674,
-0.009100720286369324,
-0.046258922666311264,
-0.01982099749147892,
-0.018428828567266464,
0.0486399345099926,
-0.05976433306932449,
0.024821065366268158,
-0.07277894020080566,
-0.05147038400173187,
-0.06166971102356911,
-0.05906262993812561,
-0.07285971939563751,
0.011992759071290493,
0.110195092856884,
0.015413814224302769,
-0.09117940068244934,
0.005889134481549263,
-0.044831350445747375,
0.040645282715559006,
-0.06008106842637062,
0.028104009106755257,
-0.02633732557296753,
0.023348607122898102,
0.0191832035779953,
-0.04792797937989235,
-0.026704473420977592,
-0.005523995496332645,
0.0879252701997757,
-0.03900908678770065,
0.009064665995538235,
-0.03624872490763664,
0.04433629289269447,
-0.006251335144042969,
0.04961983859539032,
0.011616463772952557,
0.09250090271234512,
0.03255796059966087,
0.06631051003932953,
0.0904226303100586,
0.0251749400049448,
0.1093529760837555,
0.01791643351316452,
0.08655312657356262,
0.013869695365428925,
-0.01322183571755886,
-0.039632346481084824,
0.0031438881997019053,
0.026043176651000977,
0.04524392634630203,
-0.0763126090168953,
-4.018707500108576e-8,
-0.05699224770069122,
0.07159610092639923,
0.053850699216127396,
0.03975353017449379,
0.02148480899631977,
0.057392481714487076,
-0.0454128235578537,
0.018724702298641205,
-0.05608860403299332,
-0.01330859586596489,
0.07275433093309402,
-0.02304939553141594,
0.004711034707725048,
-0.012528744526207447,
-0.018920935690402985,
0.07393960654735565,
0.047013696283102036,
-0.028562212362885475,
-0.014063572511076927,
0.033174820244312286,
0.011225011199712753,
0.030208606272935867,
0.05418790504336357,
-0.07675367593765259,
0.017268583178520203,
-0.09674587100744247,
-0.06016279384493828,
0.037972938269376755,
0.06438332796096802,
-0.02163546159863472,
0.021622952073812485,
-0.08365342766046524,
0.06596723943948746,
0.02975519560277462,
0.015025128610432148,
0.03244196996092796,
0.006065158639103174,
-0.024502018466591835,
0.014062936417758465,
-0.05153702199459076,
-0.023268969729542732,
0.1238163411617279,
-0.004395507741719484,
-0.006981360260397196,
-0.06558893620967865,
0.0160051416605711,
-0.10461777448654175,
-0.07040441036224365,
0.05415453016757965,
0.02169089950621128,
0.07920140773057938,
0.015618361532688141,
-0.052216093987226486,
0.06772646307945251,
0.08991318941116333,
0.0335150808095932,
0.032489825040102005,
-0.059536684304475784,
0.008244511671364307,
0.07274185866117477,
-0.010286186821758747,
-0.014251507818698883,
-0.08068674057722092,
0.0132662458345294
] | 0 | 0 | 0 | 0 |
|
Empowering Users in Digital Privacy Management through Interactive LLM-Based Agents | https://openreview.net/forum?id=FEpAUnS7f7 | [
"BOLUN SUN",
"Yifan Zhou",
"Haiyun Jiang"
] | Poster | This paper presents a novel application of large language models (LLMs) to enhance user comprehension of privacy policies through an interactive dialogue agent. We demonstrate that LLMs significantly outperform traditional models in tasks like Data Practice Identification, Choice Identification, Policy Summarization, and Privacy Question Answering, setting new benchmarks in privacy policy analysis. Building on these findings, we introduce an innovative LLM-based agent that functions as an expert system for processing website privacy policies, guiding users through complex legal language without requiring them to pose specific questions. A user study with 100 participants showed that users assisted by the agent had higher comprehension levels (mean score of 2.6 out of 3 vs. 1.8 in the control group), reduced cognitive load (task difficulty ratings of 3.2 out of 10 vs. 7.8), increased confidence in managing privacy, and completed tasks in less time (5.5 minutes vs. 15.8 minutes). This work highlights the potential of LLM-based agents to transform user interaction with privacy policies, leading to more informed consent and empowering users in the digital services landscape. | LLM, Agent, Usable Privacy Policies, Benchmarking, HCI | null | 12,643 | 2410.11906 | [
-0.009713147766888142,
-0.01938481070101261,
-0.04790634289383888,
0.018755214288830757,
0.0006819118862040341,
0.044837433844804764,
0.06535745412111282,
0.04061158746480942,
0.01685907691717148,
0.006147619336843491,
-0.023169778287410736,
-0.027292082086205482,
0.056046225130558014,
-0.014202598482370377,
0.11428964138031006,
0.03308715671300888,
0.12508884072303772,
-0.028244605287909508,
-0.13823142647743225,
-0.026021229103207588,
0.012583047151565552,
0.01847950369119644,
0.060706157237291336,
-0.026515668258070946,
-0.06585267931222916,
-0.01915793865919113,
0.007690484635531902,
-0.027703706175088882,
-0.02262154594063759,
0.00516035221517086,
0.07320718467235565,
-0.0036835302598774433,
0.016966721042990685,
0.05053005367517471,
0.0039471047930419445,
-0.07402558624744415,
-0.07123537361621857,
-0.006884031463414431,
-0.049905963242053986,
-0.03278740495443344,
-0.08299382030963898,
-0.10326185822486877,
-0.034921638667583466,
0.02815035916864872,
0.004086795262992382,
-0.029078204184770584,
-0.016215404495596886,
0.03442706540226936,
-0.04224473983049393,
0.004214668180793524,
-0.04354163631796837,
-0.036613889038562775,
0.060223158448934555,
-0.017476767301559448,
-0.1007174551486969,
0.004101485013961792,
-0.021321425214409828,
0.0032370935659855604,
-0.06290736049413681,
-0.032477788627147675,
-0.11592458188533783,
-0.13472330570220947,
-0.015937747433781624,
0.053436823189258575,
-0.019226569682359695,
0.060801729559898376,
-0.015033739618957043,
0.007996120490133762,
-0.006134068127721548,
-0.06430850178003311,
-0.09940539300441742,
0.059297796338796616,
0.04805919528007507,
0.040118616074323654,
-0.006721684709191322,
-0.06263955682516098,
-0.038001298904418945,
0.0013752863742411137,
0.07151442766189575,
-0.014669348485767841,
-0.032274823635816574,
-0.013292541727423668,
-0.0020495459903031588,
0.11288853734731674,
-0.01903558149933815,
0.048689547926187515,
-0.03827045112848282,
0.08057014644145966,
0.0002490966289769858,
-0.004023226909339428,
0.06152679771184921,
-0.0019986291881650686,
0.005313722416758537,
0.04374720901250839,
0.06207580864429474,
0.04376772418618202,
-0.011508860625326633,
-0.041938651353120804,
-0.07798746228218079,
0.005076856818050146,
0.000012900341971544549,
0.05777525529265404,
-0.027257978916168213,
-0.06071086600422859,
-0.039202865213155746,
0.08717363327741623,
0.035795435309410095,
0.0541357696056366,
0.010790524072945118,
0.03297494724392891,
0.010234824381768703,
-0.04087640345096588,
0.021074941381812096,
0.013339770957827568,
0.03960065543651581,
-0.08397111296653748,
0.005745031870901585,
0.07457958906888962,
0.09578725695610046,
0.023321060463786125,
0.04482410103082657,
0.051572445780038834,
0.08897820115089417,
0.0024489169009029865,
0.07165167480707169,
0.036611173301935196,
-0.04382383078336716,
7.079645464511323e-35,
0.06852228194475174,
0.03871283680200577,
-0.08729865401983261,
0.033371906727552414,
0.04625272378325462,
0.0194927379488945,
-0.012844624929130077,
0.09804127365350723,
-0.055197808891534805,
0.043585799634456635,
0.0603371299803257,
0.029152587056159973,
-0.011099527589976788,
0.0077062807977199554,
0.018875807523727417,
0.014838374219834805,
-0.010043434798717499,
0.0798715204000473,
-0.021360523998737335,
-0.04890832304954529,
0.12994208931922913,
0.04121795669198036,
0.11987751722335815,
0.01088983565568924,
0.008027167990803719,
0.017207223922014236,
0.05219770595431328,
-0.07803589105606079,
0.08163553476333618,
0.00839941669255495,
-0.07520173490047455,
-0.042131949216127396,
0.019239887595176697,
-0.03517216071486473,
0.07600462436676025,
0.028813425451517105,
-0.02074204385280609,
-0.03410424664616585,
0.09807968139648438,
0.02766047790646553,
-0.0713965967297554,
-0.04518957808613777,
0.08580989390611649,
0.007141607813537121,
-0.13551051914691925,
0.007090930826961994,
-0.037697646766901016,
0.013785682618618011,
0.031840354204177856,
-0.002440165262669325,
-0.00849996693432331,
0.047091271728277206,
-0.02993253618478775,
-0.01802304945886135,
0.02211695909500122,
-0.017992397770285606,
-0.01596747152507305,
0.00999198853969574,
-0.05270792916417122,
-0.022457251325249672,
0.004471834748983383,
0.09092291444540024,
0.006297665182501078,
0.03709422051906586,
0.018092233687639236,
-0.07160136103630066,
-0.057584334164857864,
0.07699164003133774,
0.03922843933105469,
-0.032424356788396835,
-0.007818941958248615,
0.01496867649257183,
0.036434125155210495,
0.02257210575044155,
-0.01352526806294918,
0.012098153121769428,
0.05319516733288765,
-0.031096408143639565,
0.01810716837644577,
-0.06581635773181915,
0.07440701872110367,
0.0008245701901614666,
0.01865703985095024,
-0.08453656733036041,
-0.044537145644426346,
-0.071037158370018,
-0.02059219405055046,
0.03594422712922096,
0.026055727154016495,
-0.01930703967809677,
0.0067513082176446915,
-0.03413286805152893,
-0.06918082386255264,
0.05961961671710014,
-0.032146453857421875,
-8.348205677551379e-34,
-0.10572919249534607,
-0.08504858613014221,
-0.08191300183534622,
0.06174056604504585,
0.043293796479701996,
-0.022285370156168938,
-0.007769053336232901,
0.032072268426418304,
0.08434656262397766,
-0.07722827792167664,
-0.07604667544364929,
0.04346729442477226,
0.06168009340763092,
-0.014544522389769554,
0.036425646394491196,
0.010038251988589764,
-0.025495821610093117,
-0.01659814827144146,
-0.03330216929316521,
0.024845065549016,
-0.07277361303567886,
0.06956448405981064,
-0.07542378455400467,
0.011431818827986717,
-0.014729101210832596,
-0.0095511544495821,
-0.057830810546875,
-0.004341568797826767,
0.08714843541383743,
0.010556364431977272,
0.008336449041962624,
-0.025529490783810616,
-0.03181352838873863,
-0.07752525061368942,
-0.0918964073061943,
0.0016943297814577818,
0.019727934151887894,
0.03548818826675415,
-0.03166015446186066,
0.029489297419786453,
0.006127047818154097,
-0.027555309236049652,
-0.08820004761219025,
-0.03184379264712334,
-0.05185066536068916,
0.05586022138595581,
-0.05765030160546303,
-0.05312780290842056,
0.00823940522968769,
0.019084507599473,
0.011611077934503555,
-0.03284410759806633,
0.03036818467080593,
-0.029040025547146797,
-0.03960985317826271,
-0.06221712753176689,
0.021869661286473274,
-0.05515290051698685,
0.01544167473912239,
-0.022907448932528496,
-0.0316145084798336,
0.09069434553384781,
0.024343131110072136,
-0.02289043366909027,
0.014869064092636108,
-0.02657412551343441,
-0.05975406616926193,
0.01883954368531704,
-0.03950591757893562,
-0.012227253057062626,
0.0015042470768094063,
-0.1405031383037567,
-0.07763062417507172,
-0.03399820998311043,
0.09239750355482101,
0.004668420646339655,
-0.006669209338724613,
-0.07867207378149033,
-0.03188612684607506,
0.028176486492156982,
0.06368152052164078,
0.07310560345649719,
0.022762540727853775,
-0.01252501830458641,
0.06287363171577454,
0.027337197214365005,
0.04287797957658768,
0.04921890050172806,
-0.009806056506931782,
0.02619270421564579,
0.023419534787535667,
-0.01159390714019537,
-0.08296757191419601,
0.08755605667829514,
-0.06446979194879532,
-4.3445751884974015e-8,
0.03418271988630295,
-0.015216629020869732,
0.026410434395074844,
0.0028478505555540323,
0.03275580331683159,
-0.04733293130993843,
-0.06389927119016647,
0.04059050604701042,
-0.02262524515390396,
-0.08108267933130264,
0.10768936574459076,
-0.02341761626303196,
-0.021938711404800415,
-0.05339136719703674,
0.045205939561128616,
0.042065370827913284,
-0.0006273249164223671,
0.017315484583377838,
0.012498253956437111,
-0.038569048047065735,
0.03451891615986824,
-0.013278447091579437,
-0.055055487900972366,
-0.004682639148086309,
0.043335024267435074,
0.08343570679426193,
-0.08042211085557938,
0.04130461812019348,
0.07032917439937592,
0.09454178810119629,
-0.012494651600718498,
-0.04594632238149643,
-0.009714881889522076,
0.023510422557592392,
-0.0387016125023365,
0.025511126965284348,
-0.04672735929489136,
-0.05239636451005936,
0.03876938670873642,
-0.04047249257564545,
0.033720217645168304,
0.033496372401714325,
0.08595253527164459,
0.07757151871919632,
-0.01693923957645893,
-0.015652978792786598,
-0.06135281175374985,
-0.16446451842784882,
0.06651973724365234,
-0.06238287314772606,
-0.08158519119024277,
-0.040708694607019424,
0.07100053131580353,
0.10040617734193802,
0.0814935639500618,
0.012953235767781734,
0.05094800889492035,
0.02630528248846531,
0.0007701203576289117,
0.03537647798657417,
0.07454877346754074,
0.05590597540140152,
-0.031162086874246597,
-0.012931674718856812
] | 0 | 0 | 0 | 0 |
|
Object-Centric Pretraining via Target Encoder Bootstrapping | https://openreview.net/forum?id=7d2JwGbxhA | [
"Nikola Đukić",
"Tim Lebailly",
"Tinne Tuytelaars"
] | Poster | Object-centric representation learning has recently been successfully applied to real-world datasets. This success can be attributed to pretrained non-object-centric foundation models, whose features serve as reconstruction targets for slot attention. However, targets must remain frozen throughout the training, which sets an upper bound on the performance object-centric models can attain. Attempts to update the target encoder by bootstrapping result in large performance drops, which can be attributed to its lack of object-centric inductive biases, causing the object-centric model’s encoder to drift away from representations useful as reconstruction targets. To address these limitations, we propose **O**bject-**CE**ntric Pretraining by Target Encoder **BO**otstrapping, a self-distillation setup for training object-centric models from scratch, on real-world data, for the first time ever. In OCEBO, the target encoder is updated as an exponential moving average of the object-centric model, thus explicitly being enriched with object-centric inductive biases introduced by slot attention while removing the upper bound on performance present in other models. We mitigate the slot collapse caused by random initialization of the target encoder by introducing a novel cross-view patch filtering approach that limits the supervision to sufficiently informative patches. When pretrained on 241k images from COCO, OCEBO achieves unsupervised object discovery performance comparable to that of object-centric models with frozen non-object-centric target encoders pretrained on hundreds of millions of images. The code and pretrained models are publicly available at https://github.com/djukicn/ocebo. | Object-centric learning, bootstrapping, self-supervised pretraining | null | 12,641 | 2503.15141 | [
-0.02876533754169941,
-0.10670863091945648,
-0.0012575172586366534,
0.04082275554537773,
0.06356780976057053,
-0.031171873211860657,
-0.04746032506227493,
-0.04780442640185356,
0.004760230891406536,
-0.05131370201706886,
-0.06999162584543228,
-0.06535191088914871,
-0.015799934044480324,
0.028527703136205673,
0.010920392349362373,
-0.0005999945569783449,
0.13859246671199799,
0.10127204656600952,
-0.10721292346715927,
-0.0016473132418468595,
-0.012122957967221737,
-0.015957295894622803,
0.048545971512794495,
0.0541546568274498,
0.02482660859823227,
-0.04125469923019409,
0.020794613286852837,
-0.0503862090408802,
0.03193967416882515,
-0.09269028902053833,
-0.035796791315078735,
-0.0045964764431118965,
-0.03410548344254494,
0.007056098897010088,
-0.09007314592599869,
0.029167788103222847,
-0.025961052626371384,
-0.01830196939408779,
-0.0093354731798172,
-0.011899562552571297,
0.06310287863016129,
0.0311276838183403,
0.0011088160099461675,
0.015895448625087738,
0.11202181875705719,
-0.008486039936542511,
0.04592479392886162,
-0.020320678129792213,
0.00033919597626663744,
-0.06255055963993073,
-0.044692959636449814,
-0.052100472152233124,
-0.06801494210958481,
0.06586414575576782,
-0.019526321440935135,
0.03489809110760689,
0.042126186192035675,
0.015166519209742546,
-0.012371422722935677,
0.03709385544061661,
-0.0007859003962948918,
-0.015034960582852364,
0.0310573298484087,
0.0016717141261324286,
-0.03256789222359657,
-0.06882639229297638,
0.04145744815468788,
0.04280473291873932,
-0.015116952359676361,
0.0470418743789196,
0.010287073440849781,
0.040505651384592056,
-0.025572942569851875,
-0.004438136704266071,
0.10708464682102203,
0.07162012904882431,
0.05883903056383133,
0.10017382353544235,
0.10211635380983353,
-0.10826606303453445,
0.010527577251195908,
0.0008858173387125134,
0.05415256693959236,
-0.06087006628513336,
0.060555681586265564,
0.006599602289497852,
0.02828902378678322,
0.0679350271821022,
-0.02830534614622593,
-0.007027389016002417,
-0.05185547471046448,
-0.06260393559932709,
0.017142195254564285,
-0.008975976146757603,
-0.026012703776359558,
-0.04180422052741051,
0.01672678254544735,
-0.0033859219402074814,
-0.044792335480451584,
0.04393180087208748,
-0.003926622681319714,
0.055040013045072556,
0.013490264303982258,
-0.006822179537266493,
-0.030877772718667984,
-0.037695229053497314,
0.08852502703666687,
0.04955996945500374,
0.0009938213042914867,
-0.00977865606546402,
0.0489240400493145,
0.03657875582575798,
-0.016474727541208267,
-0.03567516431212425,
-0.016987809911370277,
-0.004444883204996586,
-0.10990976542234421,
-0.01566159911453724,
0.048465218394994736,
0.01603476144373417,
-0.03983710706233978,
-0.043962642550468445,
-0.0075506605207920074,
0.048344239592552185,
0.018406422808766365,
-0.0892924815416336,
-0.08064652234315872,
3.4040321732836895e-33,
0.040328823029994965,
-0.0002648241934366524,
0.001326694735325873,
-0.0566374845802784,
-0.006667631212621927,
-0.05564281344413757,
0.04091383516788483,
0.0798347145318985,
0.008056268095970154,
0.0587141327559948,
-0.004548606462776661,
0.01904972642660141,
-0.019423332065343857,
0.10267715156078339,
0.040743544697761536,
-0.030992107465863228,
-0.07181096822023392,
0.039790742099285126,
-0.0833091288805008,
0.021314995363354683,
0.018324827775359154,
0.017391357570886612,
-0.10011275112628937,
-0.08075922727584839,
0.04018983989953995,
0.07649821043014526,
-0.009286828339099884,
-0.05110786482691765,
-0.06068221107125282,
0.03132525831460953,
-0.009899698197841644,
0.02277487702667713,
0.005842736456543207,
-0.012149105779826641,
0.013108567334711552,
-0.09666239470243454,
0.05009013041853905,
0.027638902887701988,
-0.017360860481858253,
-0.044454846531152725,
0.05107717961072922,
0.09671732038259506,
0.009225457906723022,
-0.046810392290353775,
-0.05542473495006561,
-0.07329230010509491,
0.03537112474441528,
0.04806264862418175,
-0.06735149770975113,
-0.04641040414571762,
0.023763371631503105,
0.004958422388881445,
-0.04792632907629013,
-0.07753755897283554,
-0.07249999791383743,
0.011829402297735214,
0.032027970999479294,
0.050864409655332565,
0.06560231745243073,
0.0020594659727066755,
0.0056794798001646996,
0.030245238915085793,
0.04090747609734535,
0.10397522151470184,
-0.021078290417790413,
0.009264523163437843,
-0.006549432873725891,
-0.009648961946368217,
0.0001344700576737523,
0.052980028092861176,
-0.0431603267788887,
0.0030148166697472334,
-0.05261008441448212,
-0.06720232218503952,
0.10743184387683868,
-0.024783428758382797,
0.06453591585159302,
-0.060169901698827744,
-0.0307475533336401,
0.0010665749432519078,
0.01095776166766882,
0.021763786673545837,
-0.024661481380462646,
-0.08206555247306824,
-0.041834551841020584,
-0.023456349968910217,
0.0336780920624733,
-0.11656414717435837,
-0.009231442585587502,
0.020598990842700005,
0.016740895807743073,
-0.04478604719042778,
-0.03111419267952442,
-0.05400441214442253,
-0.05678107589483261,
-2.2633708424831357e-33,
0.06088407710194588,
0.08007174730300903,
-0.050379350781440735,
0.0939943939447403,
-0.00912520382553339,
-0.041298799216747284,
0.00603728462010622,
0.022726653143763542,
0.0007461606292054057,
-0.049697112292051315,
-0.050679970532655716,
-0.03473632410168648,
0.04952124506235123,
0.03567863628268242,
0.012120758183300495,
-0.021286651492118835,
-0.07587872445583344,
0.008790465071797371,
-0.05406033620238304,
0.0001709691423457116,
0.03132404386997223,
0.06138661503791809,
-0.03318331763148308,
-0.010667871683835983,
-0.021010642871260643,
0.0836719274520874,
-0.03057221695780754,
0.11195464432239532,
-0.0032920653466135263,
-0.058173954486846924,
-0.07512163370847702,
-0.012805881910026073,
-0.029268251731991768,
-0.0656324177980423,
-0.1459583342075348,
0.06635782122612,
0.03977905586361885,
-0.03876759856939316,
-0.04335997253656387,
0.057158712297677994,
0.08296222984790802,
-0.013223697431385517,
-0.09143149852752686,
0.04823112487792969,
-0.04024292901158333,
-0.058782681822776794,
-0.08614211529493332,
0.05957719683647156,
0.07856088876724243,
0.0776437297463417,
-0.03352686017751694,
0.05114813521504402,
-0.024869421496987343,
0.008156922645866871,
-0.04390144720673561,
-0.04470790550112724,
-0.050468992441892624,
-0.0708441510796547,
0.07671250402927399,
0.04504486173391342,
-0.0538751482963562,
-0.02265588566660881,
0.08446045219898224,
-0.0741942897439003,
0.02824321947991848,
-0.021637707948684692,
-0.05100155249238014,
0.04281163960695267,
0.034270476549863815,
-0.016920186579227448,
0.003932217601686716,
0.03817060962319374,
0.004659589380025864,
0.024281661957502365,
-0.0652911439538002,
0.004886194597929716,
0.05168452486395836,
-0.0992884710431099,
0.0016602147370576859,
-0.028195850551128387,
-0.023614604026079178,
-0.0332043394446373,
-0.005396888591349125,
0.06525549292564392,
0.0560234934091568,
0.09185518324375153,
0.06327763944864273,
0.00465488201007247,
-0.008801999501883984,
-0.01686674728989601,
-0.042746543884277344,
0.016576047986745834,
0.03094332106411457,
0.09761573374271393,
-0.03451155871152878,
-4.2818484757845e-8,
-0.10624205321073532,
0.061238255351781845,
0.0034921930637210608,
0.1428019404411316,
0.017053905874490738,
-0.030370788648724556,
-0.009663218632340431,
0.07163038104772568,
-0.06140657514333725,
-0.02084563858807087,
0.07007168978452682,
0.011073342524468899,
-0.02893819473683834,
-0.030261537060141563,
0.07206735014915466,
0.12312137335538864,
0.05372350662946701,
-0.02771288901567459,
-0.01705988496541977,
0.028712322935461998,
0.0749313086271286,
0.02613738365471363,
0.044200725853443146,
-0.04354025423526764,
0.0756819024682045,
-0.07048602402210236,
-0.0136574637144804,
0.12475638091564178,
0.10104912519454956,
0.02425142377614975,
-0.045365817844867706,
0.020296689122915268,
0.08860115706920624,
-0.03027811460196972,
0.031617671251297,
0.041326697915792465,
-0.0011768657714128494,
0.06822647899389267,
-0.03578309342265129,
-0.04045921191573143,
0.05231136828660965,
0.034984104335308075,
-0.0811803787946701,
0.007845640182495117,
-0.07987281680107117,
-0.021492110565304756,
-0.03133152797818184,
-0.017813727259635925,
0.0007263067527674139,
-0.04196591675281525,
0.01827709749341011,
-0.030390700325369835,
-0.021032778546214104,
0.061192046850919724,
0.040799837559461594,
0.005776899866759777,
0.02691124938428402,
-0.03744111955165863,
0.0734231248497963,
0.09388380497694016,
-0.009931102395057678,
0.011051280423998833,
-0.017162254080176353,
0.017986923456192017
] | 0 | 0 | 0 | 0 |
|
Modality-Specialized Synergizers for Interleaved Vision-Language Generalists | https://openreview.net/forum?id=7UgQjFEadn | [
"Zhiyang Xu",
"Minqian Liu",
"Ying Shen",
"Joy Rimchala",
"Jiaxin Zhang",
"Qifan Wang",
"Yu Cheng",
"Lifu Huang"
] | Poster | Recent advancements in Vision-Language Models (VLMs) have led to the emergence of Vision-Language Generalists (VLGs) capable of understanding and generating both text and images. However, seamlessly generating an arbitrary sequence of text and images remains a challenging task for the current VLGs. One primary limitation lies in applying a unified architecture and the same set of parameters to simultaneously model discrete text tokens and continuous image features. Recent works attempt to tackle this fundamental problem by introducing modality-aware expert models. However, they employ identical architectures to process both text and images, disregarding the intrinsic inductive biases in these two modalities. In this work, we introduce Modality-Specialized Synergizers (MoSS), a novel design that efficiently optimizes existing unified architectures of VLGs with modality-specialized adaptation layers, i.e., a Convolutional LoRA for modeling the local priors of image patches and a Linear LoRA for processing sequential text. This design enables more effective modeling of modality-specific features while maintaining the strong cross-modal integration gained from pretraining. In addition, to improve the instruction-following capability on interleaved text-and-image generation, we introduce LeafInstruct, the first open-sourced interleaved instruction tuning dataset comprising 184,982 high-quality instances on more than 10 diverse domains. Extensive experiments show that VLGs integrated with MoSS achieve state-of-the-art performance, significantly surpassing baseline VLGs in complex interleaved generation tasks. Furthermore, our method exhibits strong generalizability on different VLGs. | vision-language generation, interleaved vision-language instruction tuning | null | 12,640 | null | [
-0.04136047139763832,
-0.11755011230707169,
0.05322880670428276,
0.018832940608263016,
0.11564773321151733,
-0.034169603139162064,
0.0010723518207669258,
-0.010012684389948845,
0.03212963789701462,
-0.09868680685758591,
-0.03653620555996895,
-0.054545655846595764,
0.028232285752892494,
0.07387612015008926,
0.013106824830174446,
0.06159840151667595,
0.06415942311286926,
0.10009876638650894,
-0.15548010170459747,
-0.04266650602221489,
0.05351574718952179,
0.0017013660399243236,
0.04888632521033287,
-0.05474471300840378,
0.04355558007955551,
0.08079881966114044,
0.0011014682240784168,
-0.036600399762392044,
0.06421075761318207,
-0.02129773609340191,
-0.021063709631562233,
0.03834052383899689,
-0.06113598495721817,
0.05763426795601845,
-0.022624049335718155,
0.08947192877531052,
-0.03410273417830467,
0.028160499408841133,
0.00026534704375080764,
-0.04318080469965935,
0.014918738976120949,
-0.007276267744600773,
-0.014290153048932552,
0.01743810996413231,
0.06589601188898087,
0.009300678968429565,
0.01176516804844141,
-0.02555202692747116,
0.006966369226574898,
0.037834737449884415,
-0.051740020513534546,
-0.05657380819320679,
-0.020154012367129326,
-0.0072712283581495285,
-0.03686649724841118,
0.02091814950108528,
0.03853779658675194,
-0.0376395508646965,
0.047759395092725754,
0.022172199562191963,
-0.010561182163655758,
-0.03228349983692169,
0.005012890323996544,
-0.0018255237955600023,
-0.0008795080939307809,
-0.06550958752632141,
0.06961143761873245,
0.031776003539562225,
0.004316693637520075,
-0.06827603280544281,
-0.053773630410432816,
0.04578087851405144,
0.0062498897314071655,
0.029172733426094055,
0.011693434789776802,
-0.00046372227370738983,
0.07713517546653748,
-0.019190018996596336,
0.05228221043944359,
-0.10596387833356857,
0.046444546431303024,
0.04132310301065445,
0.12694893777370453,
-0.002855421509593725,
0.12672479450702667,
0.017639880999922752,
-0.02723948284983635,
0.08462978154420853,
-0.023557331413030624,
0.02620488777756691,
-0.07555405050516129,
-0.10700573027133942,
0.05958704650402069,
-0.03730818256735802,
0.051082052290439606,
0.04034799337387085,
-0.01039013359695673,
-0.1320277899503708,
-0.02644774504005909,
0.11074318736791611,
-0.03532126173377037,
-0.043952811509370804,
0.06105682626366615,
-0.006571351550519466,
0.03478647395968437,
-0.021678615361452103,
0.06061519682407379,
0.04188217967748642,
-0.03761931136250496,
-0.0814814418554306,
0.06324556469917297,
0.026718799024820328,
0.055288612842559814,
0.027294525876641273,
0.007086188066750765,
-0.01746036671102047,
-0.012658470310270786,
-0.019180497154593468,
0.07522006332874298,
-0.03212093189358711,
-0.047813285142183304,
-0.005174522288143635,
-0.015160396695137024,
0.017012011259794235,
-0.028592946007847786,
0.03501664474606514,
-0.02857981249690056,
1.1468082708592668e-34,
0.07919874042272568,
0.011192326433956623,
0.049833815544843674,
0.053715549409389496,
0.026716697961091995,
-0.047400347888469696,
-0.006265813484787941,
0.022451117634773254,
-0.09971478581428528,
-0.0008730273111723363,
-0.0430276021361351,
-0.03779367357492447,
-0.053353797644376755,
0.11834505945444107,
0.04131181538105011,
-0.034460049122571945,
-0.06233163923025131,
-0.008263352327048779,
-0.010798083618283272,
0.009775960817933083,
-0.020642131567001343,
-0.012645522132515907,
0.0027834551874548197,
0.023080991581082344,
-0.00867524091154337,
0.02676287479698658,
0.09304608404636383,
-0.10068783909082413,
-0.03880978375673294,
-0.010140414349734783,
-0.10546029359102249,
-0.00036229853867553174,
0.0867956280708313,
0.014250818639993668,
-0.020355643704533577,
-0.05052177608013153,
-0.0025589060969650745,
-0.048095300793647766,
0.0532706156373024,
-0.022527506574988365,
-0.008313464932143688,
0.055061351507902145,
-0.010127337649464607,
-0.06984235346317291,
-0.11054917424917221,
0.0008155283867381513,
0.012419392354786396,
0.11365877836942673,
-0.08101847767829895,
-0.04036656394600868,
0.03757203370332718,
0.007788361050188541,
-0.042699918150901794,
-0.05955854803323746,
-0.020674552768468857,
-0.014136035926640034,
-0.0058605726808309555,
0.049678295850753784,
0.07767623662948608,
-0.0024745857808738947,
0.0335351787507534,
0.0648658350110054,
-0.00006914452387718484,
0.08904086798429489,
0.05998387932777405,
0.0010694254888221622,
-0.02503332868218422,
0.04522816836833954,
0.010407038033008575,
-0.007996242493391037,
-0.04841379448771477,
0.04642340540885925,
-0.04452916979789734,
-0.024882813915610313,
0.01748870126903057,
-0.019789718091487885,
-0.004824692849069834,
-0.08880870044231415,
0.017407424747943878,
0.10520163923501968,
-0.008998746983706951,
0.007841305807232857,
0.01227247528731823,
-0.08967947959899902,
-0.048415470868349075,
-0.04571884125471115,
0.047610022127628326,
-0.09430445730686188,
0.033731263130903244,
-0.07092379778623581,
-0.054640013724565506,
-0.014319181442260742,
0.02722305990755558,
-0.03646551072597504,
0.037748005241155624,
2.884990760050437e-34,
0.02338101528584957,
0.026062525808811188,
-0.05587036535143852,
0.08078419417142868,
0.015908893197774887,
-0.039482854306697845,
0.06668546050786972,
-0.013526678085327148,
-0.018807116895914078,
-0.030835900455713272,
-0.012053269892930984,
0.034064922481775284,
-0.027166131883859634,
-0.05737514793872833,
0.008142000995576382,
-0.025188375264406204,
-0.05052315443754196,
0.026935793459415436,
0.007122456561774015,
0.03455419838428497,
0.04880233854055405,
0.026802781969308853,
-0.08833880722522736,
0.027320021763443947,
-0.046069223433732986,
0.056243646889925,
-0.08472125232219696,
0.034624502062797546,
-0.054486267268657684,
-0.010326310060918331,
0.004661704413592815,
0.033358849585056305,
0.024794992059469223,
-0.05515579506754875,
-0.04281742498278618,
0.036838073283433914,
0.023445317521691322,
-0.03689992427825928,
-0.06722043454647064,
0.10505782812833786,
0.07651437073945999,
-0.061871036887168884,
-0.04771841689944267,
0.03342019394040108,
-0.051296770572662354,
-0.030543947592377663,
-0.06748066842556,
-0.06337594240903854,
0.0509609617292881,
0.03666706755757332,
-0.06023890897631645,
-0.005734290461987257,
-0.05323885753750801,
-0.02200014516711235,
-0.06524847447872162,
-0.06477070599794388,
-0.02018950879573822,
-0.019837429746985435,
0.0845138356089592,
-0.017662297934293747,
-0.025306250900030136,
-0.06290220469236374,
0.03265036642551422,
-0.07176496088504791,
0.056173089891672134,
-0.019100921228528023,
-0.07245148718357086,
-0.011975288391113281,
-0.02648511342704296,
0.0010947291739284992,
0.05253737419843674,
-0.0623314306139946,
-0.0027975321281701326,
0.013628712855279446,
0.03117898479104042,
-0.05753485485911369,
0.017054539173841476,
-0.020585089921951294,
0.0021971731912344694,
-0.08416378498077393,
0.03840074688196182,
-0.02309746854007244,
-0.007770720403641462,
0.10821448266506195,
0.05644669383764267,
0.13079170882701874,
0.013396549969911575,
0.05517103150486946,
0.0653749629855156,
0.005747808609157801,
-0.034540481865406036,
0.016580011695623398,
0.02130398340523243,
0.07257772982120514,
-0.02473183162510395,
-3.931235426080093e-8,
-0.09253811836242676,
0.013570869341492653,
-0.12520259618759155,
0.017543619498610497,
-0.04507770016789436,
-0.0872158408164978,
-0.024803901091217995,
0.008054951205849648,
-0.002517987508326769,
-0.06263382732868195,
0.049045126885175705,
-0.01898377574980259,
-0.008230390027165413,
-0.07775744795799255,
0.02686496265232563,
0.04607135057449341,
0.0581529475748539,
0.056923527270555496,
0.0271939467638731,
-0.01721540465950966,
0.06894205510616302,
0.030548160895705223,
0.0016313132364302874,
0.053534217178821564,
0.027742717415094376,
-0.05893847718834877,
-0.0777045488357544,
0.05799479782581329,
0.0995793342590332,
-0.000512863858602941,
0.0348549485206604,
0.027908790856599808,
0.03193655237555504,
-0.042967893183231354,
0.005154462531208992,
0.06528548151254654,
-0.007942575961351395,
-0.037147488445043564,
0.004006483126431704,
-0.06316032260656357,
0.06226930394768715,
0.06944065541028976,
-0.031527332961559296,
0.005885701160877943,
0.017384832724928856,
-0.015956759452819824,
0.06185692921280861,
-0.13771553337574005,
-0.010441038757562637,
-0.0014830983709543943,
-0.000579548068344593,
-0.021762032061815262,
0.0298597514629364,
0.12360524386167526,
0.03265034779906273,
-0.018696254119277,
0.049453068524599075,
0.015229903161525726,
0.119315005838871,
0.019197864457964897,
-0.015434643253684044,
0.06132771447300911,
-0.005671517923474312,
-0.04297422617673874
] | 0 | 0 | 0 | 0 |
|
Self-Normalized Resets for Plasticity in Continual Learning | https://openreview.net/forum?id=G82uQztzxl | [
"Vivek Farias",
"Adam Daniel Jozefiak"
] | Poster | Plasticity Loss is an increasingly important phenomenon that refers to the empirical observation that as a neural network is continually trained on a sequence of changing tasks, its ability to adapt to a new task diminishes over time. We introduce Self-Normalized Resets (SNR), a simple adaptive algorithm that mitigates plasticity loss by resetting a neuron’s weights when evidence suggests its firing rate has effectively dropped to zero. Across a battery of continual learning problems and network architectures, we demonstrate that SNR consistently attains superior performance compared to its competitor algorithms. We also demonstrate that SNR is robust to its sole hyperparameter, its rejection percentile threshold, while competitor algorithms show significant sensitivity. SNR’s threshold-based reset mechanism is motivated by a simple hypothesis test we derive. Seen through the lens of this hypothesis test, competing reset proposals yield suboptimal error rates in correctly detecting inactive neurons, potentially explaining our experimental observations. We also conduct a theoretical investigation of the optimization landscape for the problem of learning a single ReLU. We show that even when initialized adversarially, an idealized version of SNR learns the target ReLU, while regularization based approaches can fail to learn. | Continual Learning, Plasticity, Lifelong Learning, Neuron Resets | null | 12,620 | 2410.20098 | [
-0.07719918340444565,
-0.08610095083713531,
0.06062759459018707,
0.11977359652519226,
-0.010446173138916492,
-0.0024006348103284836,
0.008204196579754353,
-0.06266897916793823,
0.02936403825879097,
-0.05521450936794281,
-0.04530606418848038,
0.02172224596142769,
-0.03433266654610634,
0.03699807822704315,
-0.08550790697336197,
-0.008878308348357677,
0.07684152573347092,
0.14402033388614655,
-0.06654384732246399,
-0.014715348370373249,
0.022989455610513687,
0.04883863404393196,
0.05110727250576019,
0.065935418009758,
0.07784346491098404,
-0.02977875806391239,
-0.01592888869345188,
-0.04643862694501877,
0.030037730932235718,
-0.06882495433092117,
0.0748443603515625,
-0.03862917795777321,
-0.05134664848446846,
0.020990481600165367,
-0.047370005398988724,
0.008471955545246601,
-0.0639730766415596,
0.044239580631256104,
0.056358978152275085,
0.021389583125710487,
0.021022407338023186,
-0.03453207388520241,
-0.0630086362361908,
-0.010774962604045868,
0.03178722783923149,
0.03570837900042534,
-0.0030968422070145607,
-0.048041123896837234,
0.04402189701795578,
-0.014830101281404495,
0.0604436993598938,
-0.036169230937957764,
-0.03819134831428528,
0.03259670361876488,
0.07032787799835205,
0.033531010150909424,
0.07243985682725906,
0.0621299222111702,
-0.07126065343618393,
0.037979040294885635,
-0.03335478901863098,
-0.08972541987895966,
0.011364392936229706,
-0.027469146996736526,
0.03564917668700218,
0.051332391798496246,
-0.015558538027107716,
-0.005476012360304594,
0.032079461961984634,
0.030707400292158127,
0.06035957112908363,
0.11532896012067795,
-0.038365721702575684,
0.003828867804259062,
0.05832592025399208,
-0.016404855996370316,
-0.026924870908260345,
0.01782768964767456,
0.03261522203683853,
0.0066180177964270115,
0.05278619006276131,
-0.020123600959777832,
-0.02434219978749752,
-0.07143671810626984,
0.1106603667140007,
-0.0054129124619066715,
0.01178399845957756,
0.03329114615917206,
0.04769853875041008,
-0.019910374656319618,
-0.038661908358335495,
0.02263052575290203,
-0.09964121133089066,
-0.020236516371369362,
0.007964296266436577,
-0.010895023122429848,
0.009365632198750973,
-0.07371795922517776,
-0.032058797776699066,
0.07070541381835938,
-0.015402036719024181,
0.00901816412806511,
-0.024553360417485237,
0.005124079994857311,
0.01639898493885994,
-0.005117330234497786,
0.052580155432224274,
0.067601278424263,
0.03342413902282715,
0.0126543790102005,
0.02836027182638645,
-0.04248413071036339,
0.012249177321791649,
0.04741015285253525,
0.014944266527891159,
-0.010061277076601982,
-0.030790768563747406,
0.0383981354534626,
0.00933744478970766,
0.058954738080501556,
-0.02781464159488678,
-0.01955033838748932,
-0.03961663320660591,
-0.048074573278427124,
0.039916157722473145,
0.023850657045841217,
-0.1032574474811554,
4.0673673088725526e-33,
0.017983339726924896,
0.03599840775132179,
-0.03833310306072235,
-0.044922541826963425,
0.007462463807314634,
-0.050824325531721115,
-0.02313072793185711,
0.06387252360582352,
0.025672398507595062,
-0.04542596638202667,
-0.01905161701142788,
-0.004221809096634388,
-0.08593881875276566,
0.028239887207746506,
0.0835091844201088,
-0.024529539048671722,
-0.03790018707513809,
0.021628020331263542,
0.04715273529291153,
-0.15457041561603546,
0.008410257287323475,
-0.0003234493196941912,
-0.012659425847232342,
-0.040193118155002594,
-0.0473456047475338,
-0.08754751086235046,
0.051222436130046844,
0.013652881607413292,
-0.08088277280330658,
0.020494703203439713,
-0.07853395491838455,
0.08389812707901001,
0.019141914322972298,
-0.006353448610752821,
0.016477292403578758,
-0.024646829813718796,
0.0417269803583622,
-0.0031357405241578817,
0.030043384060263634,
-0.014810122549533844,
-0.006482441909611225,
0.10071227699518204,
0.025736626237630844,
-0.038034118711948395,
-0.03507843241095543,
-0.1176808699965477,
-0.050461381673812866,
-0.004324642941355705,
-0.07459071278572083,
-0.06987267732620239,
-0.014194183982908726,
0.04250595346093178,
0.0006411114009097219,
-0.03187469393014908,
-0.04372323676943779,
-0.03704913705587387,
-0.024822844192385674,
0.014884190633893013,
-0.06768430024385452,
0.048800379037857056,
0.08042734861373901,
0.02641795389354229,
-0.0071908654645085335,
0.11727894097566605,
0.0656341016292572,
0.12423007935285568,
0.01659863069653511,
0.05493796244263649,
-0.019360018894076347,
0.004431892652064562,
0.002745712408795953,
0.017391745001077652,
-0.030254140496253967,
-0.10789499431848526,
0.04439008980989456,
-0.054070595651865005,
0.11846556514501572,
-0.0473950020968914,
-0.05388446897268295,
-0.08016851544380188,
0.037938203662633896,
0.011849183589220047,
-0.05422480031847954,
0.005658273585140705,
-0.050751835107803345,
-0.06615383923053741,
0.0568833164870739,
-0.028497949242591858,
-0.03168734163045883,
0.03989076241850853,
0.0048348731361329556,
-0.06008157506585121,
0.04359960928559303,
-0.012185240164399147,
-0.008607610128819942,
-2.5703190497767208e-33,
-0.04415145888924599,
0.04146299511194229,
-0.03662846237421036,
0.0876259133219719,
0.017159784212708473,
0.019267642870545387,
-0.0006719597149640322,
0.07238113880157471,
-0.06672852486371994,
-0.06143314018845558,
-0.0178080964833498,
-0.02109309285879135,
-0.003976346459239721,
0.06037409231066704,
-0.04901117458939552,
-0.0371965728700161,
-0.06871377676725388,
0.04384509474039078,
-0.012510981410741806,
0.007135448977351189,
0.02949478104710579,
0.08777337521314621,
-0.06455502659082413,
0.028964582830667496,
0.02096025086939335,
0.03780753165483475,
-0.12232004851102829,
0.15111632645130157,
0.01867586001753807,
-0.04810969531536102,
-0.04686947911977768,
-0.008500151336193085,
-0.05761260166764259,
-0.03304686397314072,
0.05428589880466461,
0.0587235651910305,
0.0589081272482872,
0.005319648887962103,
0.03634107857942581,
0.058091502636671066,
0.11972782760858536,
0.004746715072542429,
-0.053144581615924835,
0.03556623309850693,
0.07064112275838852,
-0.09534677863121033,
-0.049532774835824966,
-0.016411077231168747,
-0.001765066641382873,
-0.02114281989634037,
-0.01021691132336855,
-0.012415161356329918,
-0.09015762805938721,
-0.01409825962036848,
-0.030354715883731842,
-0.004415349569171667,
0.02713954821228981,
0.008849654346704483,
0.009540146216750145,
0.030701126903295517,
-0.06688667088747025,
-0.02487734518945217,
-0.012490839697420597,
0.04291528835892677,
-0.027714595198631287,
0.0037388124037534,
0.042533308267593384,
0.005900894291698933,
0.03986072167754173,
0.008022312074899673,
0.027042511850595474,
0.10152030736207962,
0.026977255940437317,
-0.007114039734005928,
-0.01884201169013977,
-0.053607113659381866,
-0.0535113699734211,
-0.06265483796596527,
-0.0930764451622963,
0.0041463421657681465,
-0.013800171203911304,
0.00015009709750302136,
-0.0783643126487732,
0.03355793654918671,
0.07054934650659561,
0.107816681265831,
0.019793126732110977,
0.08148592710494995,
0.04522285237908363,
-0.017286626622080803,
0.0027966329362243414,
-0.025225909426808357,
-0.06329748034477234,
-0.021196620538830757,
-0.05271630734205246,
-4.731921166012398e-8,
-0.08708979189395905,
0.022055013105273247,
0.00930316373705864,
0.0834314301609993,
0.1255253702402115,
-0.010121739469468594,
0.016947703436017036,
0.0266774483025074,
-0.07093042880296707,
-0.10176412761211395,
0.11679311841726303,
0.02248058281838894,
-0.01202212180942297,
-0.061226509511470795,
-0.016336841508746147,
0.09025980532169342,
-0.008264445699751377,
-0.005456285551190376,
0.01122917141765356,
-0.040326982736587524,
-0.01419797632843256,
-0.0011203702306374907,
0.04180096089839935,
-0.05331035330891609,
0.0016723277512937784,
-0.06281322985887527,
0.0011585410684347153,
0.0905839055776596,
0.059190258383750916,
-0.03587142378091812,
0.04258905351161957,
-0.05999830365180969,
0.08293867856264114,
0.04112965241074562,
-0.000843001005705446,
0.06232594698667526,
0.03807550668716431,
0.008058849722146988,
-0.013124912977218628,
-0.025755641981959343,
-0.005109704565256834,
0.050008226186037064,
0.023319635540246964,
0.03705195337533951,
-0.0675593912601471,
-0.054817359894514084,
0.02538020722568035,
-0.03970585763454437,
0.025154760107398033,
-0.08267737179994583,
0.02583167888224125,
0.02622906304895878,
-0.02485150657594204,
0.05162831395864487,
0.07931765168905258,
-0.0044309357181191444,
-0.020786551758646965,
-0.02723923698067665,
-0.09800107777118683,
0.09742208570241928,
0.08710042387247086,
0.03587740287184715,
-0.07662875950336456,
0.023760974407196045
] | 0 | 0 | 0 | 0 |
|
Homomorphism Counts as Structural Encodings for Graph Learning | https://openreview.net/forum?id=qFw2RFJS5g | [
"Linus Bao",
"Emily Jin",
"Michael M. Bronstein",
"Ismail Ilkan Ceylan",
"Matthias Lanzinger"
] | Poster | Graph Transformers are popular neural networks that extend the well-known Transformer architecture to the graph domain. These architectures operate by applying self-attention on graph nodes and incorporating graph structure through the use of positional encodings (e.g., Laplacian positional encoding) or structural encodings (e.g., random-walk structural encoding). The quality of such encodings is critical, since they provide the necessary \emph{graph inductive biases} to condition the model on graph structure. In this work, we propose \emph{motif structural encoding} (MoSE) as a flexible and powerful structural encoding framework based on counting graph homomorphisms. Theoretically, we compare the expressive power of MoSE to random-walk structural encoding and relate both encodings to the expressive power of standard message passing neural networks. Empirically, we observe that MoSE outperforms other well-known positional and structural encodings across a range of architectures, and it achieves state-of-the-art performance on a widely studied molecular property prediction dataset. | graph transformers, structural encodings, homomorphism counts, expressivity | null | 12,618 | 2410.18676 | [
-0.09536487609148026,
-0.04868021607398987,
0.05117693543434143,
0.013721609488129616,
0.015974687412381172,
-0.0207144096493721,
-0.025191348046064377,
-0.012795993126928806,
0.03797631338238716,
-0.04286397621035576,
0.03382500633597374,
-0.03757916018366814,
0.008756537921726704,
0.11093344539403915,
0.027476297691464424,
0.055449310690164566,
-0.01814756542444229,
0.09926643967628479,
-0.028843475505709648,
-0.025753341615200043,
0.0718551054596901,
0.038968775421381,
0.039142873138189316,
-0.017914550378918648,
0.040939755737781525,
0.049787357449531555,
-0.014396783895790577,
0.003959338180720806,
0.06758400052785873,
-0.02083418518304825,
0.09693185240030289,
-0.04164093732833862,
-0.019636506214737892,
0.0739547610282898,
-0.022425666451454163,
0.06496302783489227,
-0.036585111171007156,
-0.052782636135816574,
0.00285199168138206,
0.010516171343624592,
0.03133430704474449,
0.0000812895450508222,
0.08115965127944946,
0.08911216259002686,
0.10093920677900314,
0.039433278143405914,
0.04696580767631531,
-0.0010573473991826177,
-0.056762680411338806,
-0.01543639600276947,
-0.022165734320878983,
-0.053313516080379486,
-0.054950520396232605,
0.04794146493077278,
0.012202521786093712,
0.01616635173559189,
-0.027010707184672356,
-0.0015425513265654445,
0.001514913747087121,
0.010781070217490196,
0.009964051656425,
-0.06438986957073212,
0.02058667317032814,
-0.026064587756991386,
-0.04216285049915314,
0.05997951328754425,
0.06921770423650742,
0.04755588620901108,
-0.015538870356976986,
-0.0010597428772598505,
0.007030094042420387,
0.025317559018731117,
-0.05526738613843918,
0.060044482350349426,
0.07596709579229355,
0.128341406583786,
0.0332225039601326,
0.003811329137533903,
-0.007246009074151516,
-0.0955204963684082,
0.009027726948261261,
-0.013838698156177998,
0.07463996112346649,
0.0446319617331028,
0.10277904570102692,
0.0079867048189044,
-0.02767818234860897,
0.02899985760450363,
-0.041856326162815094,
0.026046616956591606,
-0.06622680276632309,
-0.05950434133410454,
0.002604144159704447,
-0.032496217638254166,
-0.030051665380597115,
0.062410324811935425,
0.0391516238451004,
-0.025059498846530914,
-0.049995601177215576,
0.04471154138445854,
-0.02046632394194603,
0.03428208455443382,
0.03812837973237038,
-0.009300929494202137,
-0.028800442814826965,
0.009579027071595192,
-0.021309835836291313,
0.06819596886634827,
-0.015125170350074768,
-0.04237280786037445,
-0.0711367279291153,
0.13399629294872284,
0.04022367671132088,
0.018857698887586594,
0.028835777193307877,
-0.08384348452091217,
-0.016150759533047676,
-0.030641060322523117,
0.054771896451711655,
0.07235390692949295,
-0.05740026384592056,
0.009461299516260624,
-0.07289601862430573,
-0.016520420089364052,
-0.055299852043390274,
-0.01206500455737114,
-0.13153959810733795,
4.475273675005975e-34,
0.06727936118841171,
0.0873575434088707,
0.05631600320339203,
0.042349908500909805,
0.05048108845949173,
-0.04114948958158493,
-0.028480730950832367,
0.01177375391125679,
-0.03898417949676514,
0.029226746410131454,
-0.07264501601457596,
0.028453625738620758,
-0.003469939576461911,
0.06977857649326324,
-0.0033420962281525135,
-0.0058900779113173485,
0.020653029903769493,
-0.040955543518066406,
0.015455860644578934,
-0.09111981093883514,
-0.013501793146133423,
0.00994523148983717,
0.02559705264866352,
0.003837337251752615,
0.014275169931352139,
0.04342425987124443,
-0.053484611213207245,
-0.014633269980549812,
-0.05323515832424164,
-0.010293570347130299,
-0.0934404656291008,
0.049877338111400604,
-0.0017198181012645364,
0.024923227727413177,
0.03178156912326813,
-0.08518186956644058,
0.01480802521109581,
-0.08023998886346817,
0.01531356479972601,
-0.091577909886837,
0.045528851449489594,
0.004780193325132132,
-0.024175560101866722,
0.002992302179336548,
-0.03107619844377041,
0.01679036393761635,
-0.009786153212189674,
-0.05690845102071762,
-0.0025130175054073334,
-0.07468725740909576,
0.024562915787100792,
-0.044071901589632034,
-0.0508127361536026,
0.05722670629620552,
0.06606574356555939,
-0.03757074102759361,
0.021646983921527863,
0.06731332838535309,
0.015751419588923454,
0.06679240614175797,
-0.0711895301938057,
0.07424300909042358,
0.014555326662957668,
0.011357104405760765,
0.016363661736249924,
-0.02068486250936985,
-0.060106027871370316,
0.006249826401472092,
0.041763029992580414,
0.015300056897103786,
0.061768073588609695,
0.09570999443531036,
-0.08818842470645905,
-0.028181323781609535,
-0.0037614929024130106,
-0.014074474573135376,
-0.030671440064907074,
-0.15597662329673767,
-0.006889294367283583,
0.11173320561647415,
-0.02212360128760338,
-0.030587930232286453,
0.006962203420698643,
-0.036798376590013504,
-0.034657616168260574,
-0.04136024788022041,
0.06608731299638748,
-0.07816090434789658,
-0.019934816285967827,
-0.027551760897040367,
-0.03750476986169815,
-0.09209838509559631,
0.02386307343840599,
-0.013513057492673397,
0.01490192860364914,
-1.4971224621767275e-33,
-0.07237648218870163,
0.06620360910892487,
0.05268425494432449,
0.047584451735019684,
-0.02298426628112793,
-0.05782204121351242,
0.010893656872212887,
-0.018677743151783943,
0.0055666109547019005,
0.0706716999411583,
-0.025189809501171112,
-0.04866914451122284,
0.036389052867889404,
-0.07050000131130219,
0.08495911955833435,
0.008326046168804169,
-0.045567695051431656,
-0.03476567566394806,
0.010816700756549835,
-0.017590098083019257,
-0.009919025003910065,
0.12275298684835434,
-0.12357018887996674,
0.08359610289335251,
0.06322088092565536,
0.07065268605947495,
-0.05346117541193962,
0.052533768117427826,
0.05794607475399971,
0.008427207358181477,
-0.05330547317862511,
-0.024987326934933662,
-0.006902888882905245,
-0.025635337457060814,
0.014569618739187717,
0.026657672598958015,
0.038773149251937866,
0.010075429454445839,
-0.025685710832476616,
-0.0016068076947703958,
0.010559940710663795,
-0.030887596309185028,
-0.0812988132238388,
0.0644763633608818,
0.035435330122709274,
0.08852717280387878,
-0.031757619231939316,
0.046440884470939636,
-0.006089834962040186,
-0.00004043782246299088,
0.016033947467803955,
0.04490301012992859,
-0.03772830218076706,
-0.04969518631696701,
-0.042787738144397736,
-0.10694195330142975,
-0.014525099657475948,
0.044289764016866684,
0.014797894284129143,
-0.024734467267990112,
-0.030482498928904533,
-0.06316028535366058,
0.011536531150341034,
-0.017004963010549545,
-0.09422361105680466,
-0.06483908742666245,
-0.026228871196508408,
-0.08434935659170151,
0.019314363598823547,
-0.002793180989101529,
0.003273952519521117,
-0.045523639768362045,
-0.030458761379122734,
-0.009381181560456753,
0.03616964444518089,
-0.0646490752696991,
-0.009060065262019634,
-0.0004666062304750085,
-0.07778715342283249,
-0.04579341039061546,
-0.0316205658018589,
-0.015613223426043987,
0.01054875273257494,
0.035836122930049896,
0.08911646157503128,
0.01759144477546215,
0.03346552699804306,
0.03281605243682861,
0.05366361141204834,
-0.014599899761378765,
0.004920763894915581,
0.09147617220878601,
-0.09356164187192917,
0.10211973637342453,
0.013355030678212643,
-4.2743085515439816e-8,
-0.10394588857889175,
-0.042723607271909714,
0.033026598393917084,
-0.04187621921300888,
0.05512482672929764,
0.0448191873729229,
0.0986807644367218,
0.13430571556091309,
0.013221661560237408,
0.06224728003144264,
0.0972362756729126,
-0.024260595440864563,
-0.015673398971557617,
-0.01894996128976345,
-0.0075758774764835835,
-0.011896414682269096,
0.04400152340531349,
-0.04641532897949219,
0.040717076510190964,
0.0066042630933225155,
0.0071295988745987415,
0.021930478513240814,
-0.011107555590569973,
0.01600828766822815,
0.013979210518300533,
-0.16365692019462585,
-0.006048786919564009,
-0.006358274258673191,
-0.051096584647893906,
0.024580897763371468,
-0.023710452020168304,
-0.0033985069021582603,
0.10585807263851166,
0.01200389489531517,
-0.01425593439489603,
0.09606903046369553,
0.12323876470327377,
-0.050894223153591156,
-0.050170060247182846,
0.002791804261505604,
0.03001217544078827,
0.024716008454561234,
-0.06933372467756271,
0.02493535913527012,
-0.0648646354675293,
-0.01815587840974331,
0.048502638936042786,
0.04553132504224777,
0.014660950750112534,
-0.09745440632104874,
0.031599875539541245,
-0.05090241879224777,
-0.09049788862466812,
-0.05485326051712036,
-0.033616743981838226,
-0.006958927493542433,
-0.017788635566830635,
-0.1003982424736023,
0.04914983734488487,
-0.00033024881849996746,
0.015123924240469933,
-0.04280288890004158,
-0.04836542531847954,
-0.07587052881717682
] | https://github.com/linusbao/MoSE | 4 | 0 | 0 | 0 |
Real2Code: Reconstruct Articulated Objects via Code Generation | https://openreview.net/forum?id=CAssIgPN4I | [
"Zhao Mandi",
"Yijia Weng",
"Dominik Bauer",
"Shuran Song"
] | Poster | We present Real2Code, a novel approach to reconstructing articulated objects via code generation. Given visual observations of an object, we first reconstruct its part geometry using image segmentation and shape completion. We represent these object parts with oriented bounding boxes, from which a fine-tuned large language model (LLM) predicts joint articulation as code. By leveraging pre-trained vision and language models, our approach scales elegantly with the number of articulated parts, and generalizes from synthetic training data to real world objects in unstructured environments. Experimental results demonstrate that Real2Code significantly outperforms the previous state-of-the-art in terms of reconstruction accuracy, and is the first approach to extrapolate beyond objects' structural complexity in the training set, as we show for objects with up to 10 articulated parts. When incorporated with a stereo reconstruction model, Real2Code moreover generalizes to real-world objects, given only a handful of multi-view RGB images and without the need for depth or camera information. | articulated objects, code generation LLMs, foundation models | null | 12,609 | 2406.08474 | [
-0.06291543692350388,
-0.056071937084198,
0.008176968432962894,
0.026655476540327072,
0.06310731917619705,
0.0021851288620382547,
-0.06921064853668213,
-0.059886686503887177,
0.00638413755223155,
-0.07762567698955536,
-0.0477057509124279,
-0.03849001228809357,
-0.02877960167825222,
0.0672648549079895,
0.0009164829389192164,
-0.050178900361061096,
0.09124903380870819,
0.13540537655353546,
-0.08308035135269165,
0.04362437129020691,
0.03223181143403053,
0.023438405245542526,
0.03285188227891922,
-0.028064755722880363,
0.07907533645629883,
0.023429296910762787,
0.042917411774396896,
-0.09721183776855469,
0.018857691437005997,
-0.07020853459835052,
0.038232140243053436,
0.0008476729271933436,
-0.03139045834541321,
0.010955149307847023,
-0.024476291611790657,
0.09481944888830185,
-0.008841518312692642,
-0.022836929187178612,
-0.016181975603103638,
-0.04890580102801323,
-0.06741175055503845,
0.08072514086961746,
0.019315587356686592,
0.042733293026685715,
0.11981884390115738,
0.036931995302438736,
0.032121822237968445,
-0.035416897386312485,
0.021523114293813705,
-0.032191287726163864,
-0.06415057927370071,
-0.02395515888929367,
0.01658690720796585,
0.04335102438926697,
-0.010722692124545574,
0.07812801003456116,
0.04247957468032837,
-0.0602119117975235,
0.021443592384457588,
0.03992599993944168,
0.06258399784564972,
0.017706962302327156,
0.060085177421569824,
-0.02266901731491089,
-0.09873542189598083,
-0.09074212610721588,
0.040055807679891586,
-0.006856509950011969,
0.006722826510667801,
0.05230699107050896,
-0.005526312626898289,
0.0074709621258080006,
0.008229906670749187,
0.02133108861744404,
0.09596149623394012,
-0.004353390075266361,
0.008210615254938602,
-0.03601536899805069,
0.0273443553596735,
-0.06211840361356735,
0.013815825805068016,
-0.03467811644077301,
0.024232305586338043,
0.01080559566617012,
-0.013942506164312363,
0.03755709156394005,
0.008256115019321442,
0.11077562719583511,
-0.06783609837293625,
0.0520363375544548,
-0.11141602694988251,
-0.036740049719810486,
0.022869640961289406,
-0.017524369060993195,
-0.02490827813744545,
-0.0282916147261858,
0.02174570970237255,
-0.03131427615880966,
0.0057505397126078606,
0.041953060775995255,
-0.0030381979886442423,
-0.0019302057335153222,
0.06701552122831345,
-0.04008082300424576,
0.0220172256231308,
-0.04998089373111725,
0.03972611576318741,
-0.005042308010160923,
0.0019693004433065653,
0.020104093477129936,
0.06973657011985779,
-0.010168535634875298,
-0.033084556460380554,
-0.00951233971863985,
0.023780016228556633,
0.04938292130827904,
-0.03619065135717392,
-0.02011403627693653,
-0.03782690316438675,
-0.0005255489377304912,
0.016815686598420143,
0.01746681146323681,
-0.02931678295135498,
0.0445290133357048,
-0.03127489984035492,
-0.09382161498069763,
-0.03561308979988098,
1.5440202786707212e-33,
-0.0008083455031737685,
0.012542304582893848,
0.05007495358586311,
0.017057958990335464,
0.06665664911270142,
-0.03458142280578613,
-0.004681762773543596,
-0.018798300996422768,
0.03850375488400459,
0.0040769269689917564,
-0.010488923639059067,
-0.0531991608440876,
0.0013060298515483737,
0.1271413117647171,
0.013631572015583515,
-0.01623455248773098,
-0.06570055335760117,
0.024258842691779137,
-0.07606709003448486,
0.13571976125240326,
-0.03653646633028984,
0.0069388472475111485,
-0.012915513478219509,
-0.013925171457231045,
0.007452555000782013,
0.05050378665328026,
0.026954809203743935,
-0.004170629661530256,
-0.0019913872238248587,
-0.009502877481281757,
-0.04539883881807327,
-0.03425108641386032,
0.035647839307785034,
0.032313816249370575,
0.027246348559856415,
-0.07059404253959656,
0.02160513401031494,
-0.059070974588394165,
-0.07820937037467957,
0.03949398919939995,
0.09849245846271515,
0.09336371719837189,
-0.043547071516513824,
-0.006895206402987242,
-0.02768165059387684,
-0.07360871136188507,
0.061594635248184204,
0.1284492164850235,
-0.10865749418735504,
0.033443111926317215,
0.07464688271284103,
0.10413409024477005,
-0.028122354298830032,
-0.00903583038598299,
-0.07240671664476395,
-0.05533132702112198,
0.014221210964024067,
-0.002271353965625167,
0.03345915302634239,
-0.016943303868174553,
-0.026346687227487564,
0.025648992508649826,
0.015770329162478447,
0.028100376948714256,
-0.04830203205347061,
0.024033334106206894,
0.03583095595240593,
0.013763687573373318,
0.04674791917204857,
0.02140369825065136,
-0.04883439093828201,
-0.039063215255737305,
-0.0150840412825346,
-0.005891638807952404,
0.043744634836912155,
-0.014791903086006641,
0.04585384950041771,
-0.06696126610040665,
0.013432520441710949,
0.05746804177761078,
-0.08027266710996628,
0.053285613656044006,
0.0755024179816246,
-0.022633656859397888,
-0.047666557133197784,
-0.08147115260362625,
0.05083661898970604,
-0.11616335809230804,
-0.0235943291336298,
0.04440391808748245,
-0.11970394849777222,
-0.12518343329429626,
-0.1129959374666214,
-0.037425022572278976,
0.03154153749346733,
-1.5613154739911633e-33,
0.09294470399618149,
0.06220071762800217,
-0.01903814822435379,
-0.00788866262882948,
-0.056967221200466156,
-0.11316743493080139,
-0.042851198464632034,
0.03458059951663017,
-0.010779019445180893,
-0.03749674931168556,
-0.011868731118738651,
-0.04331301525235176,
0.04251551255583763,
0.011577936820685863,
0.06457842886447906,
-0.03591403737664223,
-0.028301406651735306,
-0.04050903394818306,
-0.012053576298058033,
0.007408943492919207,
0.049824245274066925,
0.058447837829589844,
0.012904619798064232,
-0.02407960593700409,
-0.14445814490318298,
0.08969567716121674,
-0.01777421496808529,
0.028355566784739494,
-0.007508379872888327,
-0.005244548432528973,
0.04953385517001152,
-0.09119313210248947,
-0.042983829975128174,
-0.0020314869470894337,
-0.06200576201081276,
0.013868630863726139,
0.040748849511146545,
-0.008963021449744701,
0.030213845893740654,
0.0336541049182415,
0.07011368870735168,
-0.053091246634721756,
-0.07115668058395386,
0.09919703751802444,
-0.06281168758869171,
0.007136092986911535,
-0.04023630917072296,
0.03180879354476929,
-0.020233141258358955,
0.04594112187623978,
-0.067421555519104,
0.08568042516708374,
-0.04253600537776947,
-0.07428386807441711,
0.004831375554203987,
-0.09284607321023941,
-0.09065664559602737,
-0.01654621586203575,
0.10893219709396362,
0.030077362433075905,
-0.06278249621391296,
-0.04754400998353958,
-0.027852684259414673,
-0.0008198239956982434,
0.01767922006547451,
-0.012650853022933006,
-0.04210508614778519,
-0.02801508642733097,
-0.015157414600253105,
0.019131973385810852,
0.05723249539732933,
-0.004662170074880123,
-0.038578104227781296,
0.08885020762681961,
-0.01312602311372757,
-0.03219481557607651,
0.001885231933556497,
0.03219405934214592,
0.12130258232355118,
-0.016866330057382584,
0.042801592499017715,
-0.053860269486904144,
0.05773403123021126,
0.06726352870464325,
0.13554903864860535,
0.09393258392810822,
-0.07844199240207672,
0.03528442233800888,
-0.03244069591164589,
0.05568111687898636,
-0.015797458589076996,
0.07485546916723251,
-0.005328505299985409,
0.0067322710528969765,
-0.011685199104249477,
-3.312784002673652e-8,
-0.07177560031414032,
0.0775909349322319,
0.0471978485584259,
-0.004801301751285791,
-0.033362578600645065,
-0.013796366751194,
-0.024087844416499138,
0.11634565144777298,
-0.04986809194087982,
0.006432188209146261,
0.04385136067867279,
-0.08185675740242004,
-0.018378961831331253,
-0.0004833627899643034,
-0.022681187838315964,
0.09283854067325592,
-0.034405604004859924,
-0.0084916353225708,
-0.02356693334877491,
0.024654224514961243,
-0.01927829347550869,
0.020507683977484703,
0.01137175876647234,
-0.002324313623830676,
0.016395870596170425,
-0.036118704825639725,
-0.015441984869539738,
0.04412786662578583,
0.04111984372138977,
0.03184971585869789,
-0.0048125134781003,
0.03533204272389412,
0.00031672028126195073,
-0.0038356338627636433,
0.007761932443827391,
-0.0423365980386734,
-0.018353667110204697,
-0.023445138707756996,
0.01756630651652813,
-0.11768710613250732,
0.05182179808616638,
0.07525254040956497,
-0.0033117872662842274,
0.02668978087604046,
0.061189454048871994,
0.01914340816438198,
0.012775876559317112,
-0.0636955052614212,
-0.008010393008589745,
-0.08068935573101044,
-0.04824390634894371,
-0.010236977599561214,
-0.02050473913550377,
0.04719220846891403,
0.012739015743136406,
0.0329296849668026,
-0.020189274102449417,
0.004071781411767006,
0.07126140594482422,
0.09046489745378494,
-0.08637136965990067,
0.04947185516357422,
-0.05020822957158089,
-0.016249218955636024
] | 0 | 0 | 0 | 0 |
|
An Asynchronous Bundle Method for Distributed Learning Problems | https://openreview.net/forum?id=Kwo20MWWCb | [
"Daniel Cederberg",
"Xuyang Wu",
"Stephen P. Boyd",
"Mikael Johansson"
] | Poster | We propose a novel asynchronous bundle method to solve distributed learning problems. Compared to existing asynchronous methods, our algorithm computes the next iterate based on a more accurate approximation of the objective function and does not require any prior information about the maximal information delay in the system. This makes the proposed method fast and easy to tune. We prove that the algorithm converges in both deterministic and stochastic (mini-batch) settings, and quantify how the convergence times depend on the level of asynchrony. The practical advantages of our method are illustrated through numerical experiments on classification problems of varying complexities and scales. | Distributed optimization; asynchronous optimization; model-based optimization | null | 12,605 | null | [
-0.06162860989570618,
-0.004608245100826025,
-0.029005948454141617,
0.0010960224317386746,
0.04633272811770439,
-0.006130949128419161,
-0.02824901044368744,
-0.017177406698465347,
-0.03145246207714081,
-0.02264920249581337,
-0.0008085480076260865,
0.06419934332370758,
-0.014722364954650402,
0.0012228781124576926,
-0.04107044264674187,
0.010852108709514141,
0.02751593105494976,
0.07407624274492264,
-0.059654053300619125,
-0.06576791405677795,
-0.06673678010702133,
0.03462185710668564,
0.025361614301800728,
0.025735076516866684,
0.08439735323190689,
-0.06873567402362823,
-0.003686969168484211,
-0.07294661551713943,
-0.005306742154061794,
0.03836760297417641,
0.12118802964687347,
-0.025634681805968285,
-0.02527783438563347,
-0.006192195229232311,
-0.06252763420343399,
0.06974871456623077,
0.027486637234687805,
0.02763848751783371,
0.0020556009840220213,
0.039346881210803986,
0.044751472771167755,
0.03668594732880592,
-0.024540262296795845,
0.033454444259405136,
0.03925029933452606,
-0.05070875585079193,
-0.012128415517508984,
-0.03806106373667717,
0.022231632843613625,
0.013813690282404423,
-0.032544344663619995,
-0.02441021241247654,
-0.12174482643604279,
0.03666877746582031,
-0.024272285401821136,
-0.03366463631391525,
0.04141436517238617,
-0.0007821127073839307,
-0.02186785265803337,
0.06037410721182823,
0.07519540935754776,
-0.16724617779254913,
-0.0498761422932148,
0.0481334887444973,
0.013170501217246056,
0.022151127457618713,
0.05804739147424698,
0.001586221274919808,
0.05518227443099022,
-0.03698456287384033,
-0.0005986212054267526,
0.0443013459444046,
-0.006115326192229986,
0.0816328302025795,
0.05399361252784729,
-0.059324637055397034,
0.08382945507764816,
0.021245259791612625,
-0.007183410692960024,
-0.007831352762877941,
-0.0428156740963459,
0.09683137387037277,
0.004857617896050215,
-0.08241119235754013,
0.1034279316663742,
0.002044511027634144,
-0.03949793055653572,
0.014734635129570961,
-0.003271625144407153,
-0.05696354806423187,
0.023761654272675514,
0.019599078223109245,
-0.08276393264532089,
-0.009335323236882687,
-0.05472142994403839,
-0.05275127291679382,
0.053688470274209976,
0.038963474333286285,
-0.0008700994658283889,
0.0803748294711113,
-0.06101062148809433,
-0.008758162148296833,
-0.008355461061000824,
-0.02906932681798935,
0.05365129932761192,
-0.03774269297719002,
0.02021906152367592,
-0.005821287631988525,
-0.01797105371952057,
-0.12223149836063385,
0.012025611475110054,
-0.0022720498964190483,
-0.05402914062142372,
0.058381929993629456,
-0.022668076679110527,
-0.019070390611886978,
0.07528786361217499,
-0.006667290814220905,
0.04527418687939644,
0.041734300553798676,
0.03168810158967972,
0.0026182103902101517,
0.10642117261886597,
-0.01402168720960617,
-0.018627366051077843,
0.04096289351582527,
-0.07217613607645035,
2.695551442467527e-33,
-0.032475799322128296,
-0.03133644536137581,
0.07369168847799301,
-0.08505048602819443,
-0.0005388879217207432,
-0.0003245158295612782,
0.08131959289312363,
0.035826705396175385,
-0.044943779706954956,
0.03163095563650131,
-0.0012414844240993261,
0.005090710241347551,
0.003965845797210932,
0.03832165524363518,
0.02572186104953289,
-0.11741182953119278,
-0.020096946507692337,
0.04104506969451904,
0.017516108229756355,
-0.11053197830915451,
0.09999088197946548,
-0.10555806010961533,
0.011107703670859337,
-0.0476222038269043,
-0.04885774478316307,
-0.06706120818853378,
0.06318391859531403,
0.030937593430280685,
-0.0024923430755734444,
0.09507425129413605,
-0.0067203049547970295,
0.0006719744997099042,
0.0021612823475152254,
0.031163616105914116,
0.019412510097026825,
0.007138602901250124,
-0.056356292217969894,
0.04195350036025047,
-0.04577745869755745,
-0.03752858564257622,
-0.03244486078619957,
-0.027253221720457077,
0.012309936806559563,
-0.06746670603752136,
-0.11444730311632156,
-0.05222274735569954,
0.018725251778960228,
-0.01697932742536068,
0.028116116300225258,
-0.08049491047859192,
0.013995297253131866,
-0.03950966149568558,
-0.05899116024374962,
-0.0210043266415596,
0.017688555642962456,
0.0551493838429451,
0.04767458513379097,
0.021758388727903366,
-0.02888939529657364,
0.05033707991242409,
0.007153092417865992,
-0.04997813329100609,
-0.03302893787622452,
0.026777978986501694,
0.07907555252313614,
0.0341537781059742,
-0.006093365140259266,
0.006889983080327511,
0.06352771073579788,
-0.0647445097565651,
0.04018351063132286,
0.06553694605827332,
-0.04689343646168709,
-0.05839410424232483,
0.08023044466972351,
-0.04045441374182701,
0.03938107565045357,
0.057860855013132095,
-0.10571017861366272,
-0.021603086963295937,
-0.019040262326598167,
-0.018597206100821495,
-0.08397562056779861,
-0.013540123589336872,
-0.028124244883656502,
0.03288928046822548,
0.0036863316781818867,
0.03893235698342323,
-0.05505884066224098,
0.022796601057052612,
-0.1446262151002884,
0.04900695011019707,
0.046735405921936035,
-0.015766216441988945,
0.06662192195653915,
-4.1064558021152554e-33,
-0.049625057727098465,
-0.01852068305015564,
-0.023589178919792175,
0.05149294063448906,
0.07540527731180191,
-0.026195021346211433,
-0.07998759299516678,
0.020574187859892845,
-0.06716541200876236,
0.0006531578837893903,
-0.037603430449962616,
-0.032522741705179214,
-0.0930204838514328,
0.04069146141409874,
0.015880851075053215,
0.09428514540195465,
0.04787776246666908,
-0.0020490321330726147,
0.09416422247886658,
0.06271842122077942,
-0.08455783873796463,
0.04359499737620354,
-0.03406999260187149,
0.010805309750139713,
0.013855370692908764,
-0.03149871900677681,
-0.01586456410586834,
0.0438297875225544,
0.0034667623694986105,
-0.04588064178824425,
-0.0057599483989179134,
-0.06245797127485275,
-0.049247659742832184,
-0.011974046938121319,
-0.04540510103106499,
0.07383234798908234,
-0.032139379531145096,
0.010593372397124767,
-0.016942381858825684,
0.03234224021434784,
0.08595699071884155,
-0.05997627228498459,
-0.057375628501176834,
-0.0723852813243866,
0.09032955020666122,
-0.05414091423153877,
-0.06706370413303375,
0.054928913712501526,
-0.048029299825429916,
0.14028100669384003,
0.008026150055229664,
0.040556490421295166,
-0.04588127136230469,
-0.03595440089702606,
-0.031771134585142136,
0.0346887968480587,
-0.06903461366891861,
-0.0027204074431210756,
0.04015609249472618,
-0.0283784419298172,
0.002537915948778391,
-0.10001599043607712,
0.07565125077962875,
-0.012151011265814304,
0.061702653765678406,
-0.05033139884471893,
0.026167312636971474,
0.03174937516450882,
-0.005403835792094469,
0.09483333677053452,
0.07806552946567535,
0.06876953691244125,
-0.03825840353965759,
-0.022097205743193626,
-0.015013318508863449,
-0.0348886214196682,
0.00029793501016683877,
0.0015347301959991455,
-0.04779784753918648,
0.03257760405540466,
0.02434806153178215,
-0.0018472203519195318,
0.03168554604053497,
0.04695948213338852,
0.0645560771226883,
0.06390847265720367,
0.19238844513893127,
0.06995604187250137,
0.0974111557006836,
-0.06812690198421478,
0.013207433745265007,
0.028985124081373215,
-0.005859620403498411,
-0.0062149097211658955,
-0.04761844128370285,
-4.155664967697703e-8,
-0.028643261641263962,
-0.00034014537231996655,
-0.008982643485069275,
0.005978565197438002,
0.10541386157274246,
0.0956181138753891,
-0.037058256566524506,
0.025828389450907707,
-0.04215158522129059,
-0.025969674810767174,
0.010460684075951576,
-0.005682010669261217,
0.03248191997408867,
-0.029406333342194557,
-0.08372074365615845,
-0.03772897645831108,
-0.06194739043712616,
-0.025420431047677994,
0.03239510953426361,
0.013448413461446762,
0.032692037522792816,
-0.018356746062636375,
0.07922756671905518,
-0.005078439135104418,
-0.007530767470598221,
-0.03697097301483154,
0.060622721910476685,
0.031554725021123886,
0.06940735876560211,
-0.027307748794555664,
-0.06830132007598877,
0.040865570306777954,
-0.06933867186307907,
0.004931936971843243,
0.0041200085543096066,
0.016068261116743088,
0.09311109036207199,
-0.004816100932657719,
0.004321255721151829,
-0.02517741732299328,
0.03654981404542923,
-0.04642917960882187,
0.011798650957643986,
0.027663832530379295,
0.0139188626781106,
0.028682397678494453,
0.006903061643242836,
-0.006366205867379904,
0.022803103551268578,
-0.007726170122623444,
0.002667021472007036,
0.004198579583317041,
-0.07856320589780807,
0.0802241861820221,
0.0965195968747139,
-0.06987710297107697,
-0.015314595773816109,
-0.11039908975362778,
0.012569425627589226,
0.03520432487130165,
-0.059838876128196716,
-0.03716687858104706,
-0.09573113918304443,
0.03228909894824028
] | 0 | 0 | 0 | 0 |
|
On Linear Representations and Pretraining Data Frequency in Language Models | https://openreview.net/forum?id=EDoD3DgivF | [
"Jack Merullo",
"Noah A. Smith",
"Sarah Wiegreffe",
"Yanai Elazar"
] | Poster | Pretraining data has a direct impact on the behaviors and quality of language models (LMs), but we only understand the most basic principles of this relationship. While most work focuses on pretraining data's effect on downstream task behavior, we investigate its relationship to LM representations. Previous work has discovered that, in language models, some concepts are encoded "linearly" in the representations, but what factors cause these representations to form (or not)? We study the connection between pretraining data frequency and models' linear representations of factual relations (e.g., mapping France to Paris in a capital prediction task). We find evidence that the formation of linear representations is strongly connected to pretraining term frequencies; specifically for subject-relation-object fact triplets, both subject-object co-occurrence frequency and in-context learning accuracy for the relation are highly correlated with linear representations. This is the case across all phases of pretraining, i.e., it is not affected by the model's underlying capability. In OLMo-7B and GPT-J (6B), we discover that a linear representation consistently (but not exclusively) forms when the subjects and objects within a relation co-occur at least 1k and 2k times, respectively, regardless of when these occurrences happen during pretraining (and around 4k times for OLMo-1B). Finally, we train a regression model on measurements of linear representation quality in fully-trained LMs that can predict how often a term was seen in pretraining. Our model achieves low error even on inputs from a different model with a different pretraining dataset, providing a new method for estimating properties of the otherwise-unknown training data of closed-data models. We conclude that the strength of linear representations in LMs contains signal about the models' pretraining corpora that may provide new avenues for controlling and improving model behavior: particularly, manipulating the models' training data to meet specific frequency thresholds. We release our code to support future work. | pretraining data, pretraining, linear, linear features, interpretability, linear representations, corpus frequency | Language Models (LMs) form linear representations of relations when the average subject-object frequency surpasses a certain threshold | 12,604 | null | [
0.011611981317400932,
-0.10453369468450546,
0.039061903953552246,
0.06360649317502975,
0.04554258659482002,
0.06184697523713112,
-0.036116454750299454,
-0.03310523182153702,
0.10312560200691223,
-0.028710246086120605,
-0.051419083029031754,
0.0006696747150272131,
0.03866063430905342,
0.06369246542453766,
0.059875212609767914,
-0.046109553426504135,
0.08122473210096359,
0.023371737450361252,
-0.10665244609117508,
-0.043789319694042206,
0.07345057278871536,
-0.035567160695791245,
0.020345143973827362,
0.013315284624695778,
0.06691368669271469,
-0.01986026205122471,
-0.00957438163459301,
-0.026274921372532845,
0.058913830667734146,
-0.015228352509438992,
0.026109592989087105,
0.10841289162635803,
0.022709697484970093,
0.04586266726255417,
-0.09581650048494339,
0.04819906875491142,
-0.04587729647755623,
0.023682976141572,
0.014031641185283661,
0.011966695077717304,
0.010061081498861313,
-0.045987289398908615,
0.056266069412231445,
0.07667519152164459,
0.06063176691532135,
0.006917670369148254,
0.033965103328228,
-0.018446313217282295,
-0.10005511343479156,
0.031700391322374344,
-0.1315506249666214,
0.01258398313075304,
0.01990276202559471,
0.07932047545909882,
-0.04382765665650368,
0.02919670194387436,
0.035131849348545074,
0.09310761094093323,
-0.028902381658554077,
0.02602231688797474,
-0.1117219552397728,
-0.04066602140665054,
-0.10736469179391861,
-0.012666222639381886,
-0.03657637909054756,
0.022177783772349358,
-0.018579579889774323,
0.0584804005920887,
-0.08975255489349365,
0.07054409384727478,
0.0006135855801403522,
0.08137019723653793,
-0.03321491554379463,
0.0427449606359005,
0.060623668134212494,
0.029574288055300713,
-0.02407263219356537,
0.00891690794378519,
0.07276271283626556,
-0.04287578538060188,
0.019656790420413017,
0.03720241039991379,
0.023045580834150314,
-0.06242363527417183,
0.0866706445813179,
0.01895715668797493,
0.05264940485358238,
0.011130230501294136,
-0.010486728511750698,
-0.024979181587696075,
-0.02708076313138008,
-0.08067124336957932,
0.07230648398399353,
0.06419935077428818,
-0.018223002552986145,
0.06637492030858994,
0.036931637674570084,
0.02762480638921261,
0.03632548823952675,
0.020683398470282555,
0.06377916038036346,
0.0459228940308094,
-0.02467016689479351,
0.09446407109498978,
-0.04209250584244728,
-0.022808030247688293,
-0.008450500667095184,
0.017933767288923264,
-0.033328160643577576,
-0.014093144796788692,
-0.024971336126327515,
0.023252399638295174,
-0.01905289851129055,
0.019163059070706367,
-0.03393685445189476,
-0.028807029128074646,
0.005616902839392424,
-0.10238749533891678,
0.008460719138383865,
0.04712811857461929,
0.0065107038244605064,
0.05700497701764107,
0.056753043085336685,
-0.0012041248846799135,
-0.04856378585100174,
-0.10435516387224197,
-0.07444629073143005,
-5.8094992524526196e-34,
0.023181283846497536,
0.013127722777426243,
0.022764550521969795,
0.04490293189883232,
-0.027796003967523575,
-0.05193645507097244,
-0.07052155584096909,
0.024150053039193153,
0.05013482645153999,
0.019899750128388405,
-0.032175686210393906,
0.06445091217756271,
-0.036034874618053436,
0.07628428936004639,
-0.028360430151224136,
0.07186359912157059,
-0.03976420685648918,
-0.008049621246755123,
-0.05251751467585564,
-0.028014926239848137,
-0.00932716578245163,
0.01276710070669651,
-0.0025204778648912907,
-0.01935671456158161,
0.01812807284295559,
0.04251708462834358,
0.05505470931529999,
-0.07867839187383652,
-0.09986764937639236,
0.009557499550282955,
-0.07217878103256226,
0.012245520018041134,
0.016327321529388428,
-0.017121030017733574,
-0.020854122936725616,
-0.052594151347875595,
0.006584554444998503,
-0.038784850388765335,
0.04904010146856308,
-0.02013741061091423,
0.010427326895296574,
0.01588001288473606,
0.030696818605065346,
-0.035749081522226334,
-0.0667123794555664,
-0.0783049687743187,
0.03644808381795883,
-0.03343815356492996,
-0.03444135934114456,
0.011670156382024288,
0.03713288530707359,
0.02316756546497345,
-0.039358191192150116,
-0.013928735628724098,
0.006807769648730755,
0.07702883332967758,
-0.024195179343223572,
0.025832703337073326,
-0.025720277801156044,
0.005313340108841658,
-0.002531317761167884,
0.040621038526296616,
0.07924113422632217,
0.05383852496743202,
0.029492028057575226,
0.038425154983997345,
-0.05437649413943291,
0.07935839146375656,
0.03662791848182678,
0.002316197147592902,
-0.026439685374498367,
-0.021704712882637978,
-0.0495389848947525,
-0.05427602306008339,
0.07747521996498108,
0.021787455305457115,
0.026562761515378952,
-0.11871306598186493,
-0.08861109614372253,
0.05572890490293503,
-0.0021259544882923365,
-0.058193814009428024,
0.053985968232154846,
-0.061733514070510864,
-0.12223898619413376,
-0.03716457635164261,
0.0513351745903492,
-0.1365509331226349,
0.034286171197891235,
-0.0425858199596405,
0.020337259396910667,
-0.03416701778769493,
-0.05569729208946228,
0.031057599931955338,
-0.006850574165582657,
-1.4742339341272867e-33,
-0.010020850226283073,
0.01883731409907341,
-0.009444944560527802,
0.00011677030852297321,
-0.045870035886764526,
-0.0593680776655674,
0.056244172155857086,
0.022574277594685555,
-0.018863912671804428,
-0.03448791801929474,
-0.02688576467335224,
-0.1376802921295166,
0.0070393383502960205,
0.05101380869746208,
0.0896085724234581,
-0.05486151948571205,
-0.020226530730724335,
0.04358023777604103,
0.03081958182156086,
0.02066139131784439,
0.028048787266016006,
0.06504281610250473,
-0.11095863580703735,
0.026693295687437057,
-0.004304044414311647,
-0.01285785436630249,
-0.055127888917922974,
0.02293241396546364,
-0.01308692991733551,
0.03700654208660126,
-0.0606551393866539,
0.019230946898460388,
0.025973014533519745,
-0.05065307393670082,
-0.08477866649627686,
0.05888383463025093,
0.02383158914744854,
-0.030778754502534866,
-0.025862842798233032,
0.02358061634004116,
0.059996750205755234,
-0.04608983173966408,
0.005680846516042948,
-0.0633547306060791,
-0.05106442794203758,
-0.04527559131383896,
-0.030266156420111656,
-0.051249176263809204,
0.050195615738630295,
-0.020616034045815468,
0.021432289853692055,
0.06792522221803665,
-0.0707816407084465,
-0.0006385686574503779,
-0.029396751895546913,
-0.04418636113405228,
0.006910202093422413,
-0.07789753377437592,
0.0388481467962265,
-0.03240758925676346,
-0.020611532032489777,
0.037154652178287506,
0.024176528677344322,
-0.09752892702817917,
0.021680723875761032,
-0.10866469889879227,
-0.06488397717475891,
-0.047171831130981445,
0.0809745341539383,
-0.03816039487719536,
-0.023818304762244225,
-0.011124307289719582,
-0.034321852028369904,
0.007037709932774305,
-0.051792506128549576,
-0.01710362546145916,
-0.03727585822343826,
-0.06767652183771133,
-0.07992634922266006,
-0.011017999611794949,
-0.03592566400766373,
-0.020616278052330017,
0.00016821190365590155,
0.03733392804861069,
-0.019680367782711983,
0.07823832333087921,
-0.008138717152178288,
0.04029902443289757,
0.0410296656191349,
-0.01963220350444317,
-0.0224660262465477,
0.021440347656607628,
-0.1093229204416275,
0.09715192019939423,
-0.0472024567425251,
-4.8531006768826046e-8,
-0.06212334707379341,
0.0016488538822159171,
-0.020657334476709366,
0.06952669471502304,
0.0041737803258001804,
-0.05522869899868965,
0.003941784147173166,
0.08801019191741943,
-0.03901590406894684,
0.018379192799329758,
0.06727692484855652,
0.03072567842900753,
-0.0676615759730339,
-0.07658819854259491,
0.03571731597185135,
0.06756338477134705,
0.025726230815052986,
0.012125460430979729,
0.023090576753020287,
0.01031481008976698,
0.1258680671453476,
0.028145279735326767,
-0.037493746727705,
0.02347683720290661,
0.053021740168333054,
-0.06413985788822174,
-0.03403150662779808,
0.1940622180700302,
0.0563492588698864,
0.023463066667318344,
-0.007472930010408163,
0.033966995775699615,
0.003745801979675889,
-0.09736170619726181,
-0.022077519446611404,
0.057317104190588,
0.07878727465867996,
-0.061562784016132355,
-0.007885001599788666,
-0.004345254972577095,
0.05746671184897423,
0.0743020549416542,
-0.08220670372247696,
0.06332080811262131,
0.05908570811152458,
0.03295048326253891,
-0.022470567375421524,
-0.07281150668859482,
0.024730410426855087,
-0.08519139885902405,
-0.03387821465730667,
0.06218913570046425,
-0.03393590450286865,
0.07035782188177109,
0.060433290898799896,
0.03142145276069641,
-0.01762857846915722,
0.009716506116092205,
-0.012722262181341648,
-0.01653684489428997,
0.03140333667397499,
0.03825617954134941,
0.025741517543792725,
-0.023321714252233505
] | 0 | 0 | 0 | 0 |
|
CoRNStack: High-Quality Contrastive Data for Better Code Retrieval and Reranking | https://openreview.net/forum?id=iyJOUELYir | [
"Tarun Suresh",
"Revanth Gangi Reddy",
"Yifei Xu",
"Zach Nussbaum",
"Andriy Mulyar",
"Brandon Duderstadt",
"Heng Ji"
] | Poster | Effective code retrieval plays a crucial role in advancing code generation, bug fixing, and software maintenance, particularly as software systems increase in complexity. While current code embedding models have demonstrated promise in retrieving code snippets for small-scale, well-defined tasks, they often underperform in more demanding real-world applications such as bug localization within GitHub repositories. We hypothesize that a key issue is their reliance on noisy and inconsistent datasets for training, which impedes their ability to generalize to more complex retrieval scenarios. To address these limitations, we introduce CoRNStack, a large-scale, high-quality contrastive training dataset for code that spans multiple programming languages. This dataset is curated using consistency filtering to eliminate noisy positives and is further enriched with mined hard negatives, thereby facilitating more effective learning. We demonstrate that contrastive training of embedding models using CoRNStack leads to state-of-the-art performance across a variety of code retrieval tasks. Furthermore, the dataset can be leveraged for training code reranking models, a largely underexplored area compared to text reranking. Our finetuned code reranking model significantly improves the ranking quality over the retrieved results. Finally, by employing our code retriever and reranker together, we demonstrate significant improvements in function localization for GitHub issues, an important
component of real-world software development. | code representation learning, code re-ranking, contrastive learning | We introduce CoRNStack, a large-scale, high-quality contrastive training dataset curated using consistency filtering to eliminate noisy positives and further enriched with mined hard negatives. | 12,599 | 2412.01007 | [
-0.10003440827131271,
-0.11807938665151596,
-0.0014361587818711996,
0.051219016313552856,
0.04343922436237335,
-0.009016552940011024,
-0.024108776822686195,
0.003348148660734296,
-0.05391032248735428,
-0.07700435817241669,
-0.03212108463048935,
-0.041302379220724106,
0.061344411224126816,
-0.04004376381635666,
-0.14154483377933502,
0.04014134779572487,
0.10395484417676926,
0.0983329638838768,
-0.02842903696000576,
-0.09088993072509766,
-0.04053354263305664,
0.01957053691148758,
0.016572745516896248,
-0.005877564661204815,
0.09378325939178467,
0.029560893774032593,
-0.05050380155444145,
-0.010241111740469933,
0.014541289769113064,
-0.0482020378112793,
0.004960000514984131,
0.04170643538236618,
-0.04423823207616806,
0.08814901858568192,
0.030102431774139404,
0.12902645766735077,
-0.0673479288816452,
-0.011454300954937935,
-0.033182524144649506,
0.0003441518056206405,
-0.06904856860637665,
0.08312614262104034,
0.02515815757215023,
0.0014457724755629897,
0.039920564740896225,
-0.1032862737774849,
0.00995061919093132,
-0.03249365836381912,
0.027927391231060028,
-0.07676635682582855,
-0.06549721956253052,
-0.018230492249131203,
-0.06599174439907074,
0.016628818586468697,
0.038153424859046936,
-0.034559592604637146,
0.08078989386558533,
0.011943500488996506,
-0.018489159643650055,
-0.058848969638347626,
-0.00014629482757300138,
-0.006845833268016577,
0.013567393645644188,
-0.024021022021770477,
0.006461090873926878,
-0.05481424927711487,
0.00007778307917760685,
0.06641647219657898,
-0.04536781832575798,
0.003140835789963603,
-0.053617093712091446,
0.055062852799892426,
-0.04288239777088165,
0.12825624644756317,
-0.021275019273161888,
0.06924349814653397,
0.009803914465010166,
0.03840058296918869,
0.055453769862651825,
-0.08403661847114563,
0.016705704852938652,
-0.018226558342576027,
0.03488573431968689,
0.08168041706085205,
0.11933083087205887,
0.045807648450136185,
0.09799189120531082,
0.04938412457704544,
0.0023580791894346476,
0.07271554321050644,
0.006176536902785301,
-0.024416588246822357,
0.07415204495191574,
0.009341774508357048,
0.061394065618515015,
0.007046185899525881,
0.0808793306350708,
-0.04608413577079773,
-0.02340082637965679,
0.10837510973215103,
-0.04741581156849861,
0.07368817180395126,
0.06368482112884521,
-0.06109216809272766,
-0.012862136587500572,
0.05537029728293419,
0.08809550106525421,
0.030711932107806206,
0.03802204877138138,
-0.07490915060043335,
-0.011447563767433167,
0.09494299441576004,
-0.037763286381959915,
-0.04883091151714325,
0.03264094144105911,
0.001153024029918015,
0.017484167590737343,
-0.017960872501134872,
-0.006354532670229673,
0.007179684937000275,
0.06299827992916107,
0.05467227101325989,
-0.04332338273525238,
-0.04728911444544792,
0.03509417921304703,
0.019111644476652145,
-0.012332403101027012,
1.593703098421202e-33,
0.0707014873623848,
0.00040319067193195224,
-0.02133479155600071,
0.000910507224034518,
0.07522822171449661,
-0.0793190747499466,
0.01983930915594101,
0.05883955955505371,
-0.09209546446800232,
-0.03258030116558075,
0.04650070518255234,
0.06363598257303238,
-0.04688870534300804,
0.06163308396935463,
-0.02511923387646675,
0.004505837801843882,
-0.05477442964911461,
-0.02716396376490593,
-0.06887968629598618,
0.03217608109116554,
-0.05028850957751274,
-0.07370693236589432,
0.010371181182563305,
-0.08433989435434341,
-0.00282529485411942,
-0.011383315548300743,
0.03482743352651596,
0.015254800207912922,
-0.05868452042341232,
0.011578353121876717,
-0.07353465259075165,
-0.0041313618421554565,
0.051519107073545456,
-0.00029387164977379143,
0.04249011352658272,
-0.06075924634933472,
-0.014810878783464432,
-0.0598781518638134,
0.0023069309536367655,
0.03808781877160072,
-0.00041025399696081877,
0.0019746236503124237,
-0.07511059194803238,
-0.0158299058675766,
0.07482818514108658,
-0.08192586153745651,
-0.01382917258888483,
-0.03373183310031891,
-0.016983257606625557,
-0.022783959284424782,
0.0862187072634697,
0.051938652992248535,
0.03841335326433182,
-0.011783363297581673,
-0.013819878920912743,
-0.0007696171524003148,
0.030096348375082016,
-0.032114144414663315,
0.040382254868745804,
0.019342686980962753,
-0.06788936257362366,
-0.007831885479390621,
0.017031360417604446,
0.0289029348641634,
0.00984762329608202,
0.04502522572875023,
0.011700593866407871,
0.08245965093374252,
0.0638013556599617,
-0.004315622616559267,
0.008715729229152203,
0.015557030215859413,
-0.004265953786671162,
-0.0017492086626589298,
0.024766692891716957,
-0.009653996676206589,
0.025878772139549255,
-0.06659036874771118,
0.012677468359470367,
-0.1069701686501503,
0.0009505355847068131,
-0.12478634715080261,
0.007886252366006374,
-0.043691813945770264,
-0.08843571692705154,
-0.0044083851389586926,
0.03284641355276108,
-0.07793455570936203,
-0.009710762649774551,
0.036034539341926575,
-0.044004205614328384,
-0.07879563421010971,
-0.017995037138462067,
-0.1064770296216011,
0.057370010763406754,
-2.2355946456571743e-33,
0.030021531507372856,
-0.01827043853700161,
-0.055685292929410934,
0.013029888272285461,
-0.07824120670557022,
-0.015251385979354382,
0.020984284579753876,
0.03834439069032669,
-0.03279556334018707,
-0.002704562386497855,
-0.017575271427631378,
-0.04742838069796562,
-0.009900697506964207,
-0.035322945564985275,
0.07654391974210739,
0.013791875913739204,
-0.037540823221206665,
0.0499144047498703,
-0.02032736875116825,
0.1036706492304802,
-0.04069866985082626,
0.07741931825876236,
-0.020784446969628334,
0.017519699409604073,
-0.010850731283426285,
0.05443428456783295,
-0.0631350576877594,
0.014052676036953926,
-0.018181173130869865,
-0.1438930779695511,
0.028999926522374153,
0.01861443743109703,
-0.07371611893177032,
-0.06288550049066544,
-0.027998516336083412,
0.03380455821752548,
0.03758964315056801,
-0.06850770860910416,
-0.01328396238386631,
0.1081133559346199,
0.08142714947462082,
-0.0007176845683716238,
-0.025534730404615402,
0.03598453477025032,
0.009150972589850426,
0.0279237013310194,
-0.07603264600038528,
0.003362464951351285,
0.01268488448113203,
-0.0609046034514904,
-0.005757441278547049,
-0.00022188923321664333,
-0.027239270508289337,
0.01881933957338333,
0.009433073922991753,
-0.032447610050439835,
-0.04393471032381058,
0.02404113858938217,
0.010199733078479767,
0.09094460308551788,
-0.06744515150785446,
-0.08269508183002472,
0.017600057646632195,
-0.055968012660741806,
-0.031122298911213875,
-0.07587838917970657,
-0.04607078805565834,
0.0246305912733078,
-0.04601125046610832,
-0.015788668766617775,
0.026945175603032112,
-0.061135660856962204,
-0.027464943006634712,
-0.021755021065473557,
0.001765415770933032,
0.019286317750811577,
0.020872268825769424,
-0.0030443735886365175,
-0.04370873421430588,
-0.03310580924153328,
0.052104514092206955,
-0.021179068833589554,
0.005647413432598114,
-0.007687858771532774,
0.035909466445446014,
0.10090024024248123,
0.018084147945046425,
0.05731407180428505,
0.03853936493396759,
-0.0340968482196331,
-0.042440734803676605,
-0.007408155128359795,
0.023681798949837685,
-0.02688818797469139,
0.021477511152625084,
-3.6645719347916383e-8,
-0.051334962248802185,
0.06758856773376465,
-0.009055369533598423,
0.00248711253516376,
-0.01968351937830448,
-0.028556212782859802,
-0.03592316433787346,
0.15168340504169464,
-0.07544019818305969,
0.03252045065164566,
0.04501228407025337,
0.03296789154410362,
-0.018460102379322052,
-0.04682401195168495,
-0.01830630749464035,
0.11765307188034058,
-0.05943994224071503,
0.019022805616259575,
0.011206468567252159,
0.0001263278245460242,
0.018712248653173447,
0.10102105140686035,
0.0624862015247345,
0.018083781003952026,
-0.07167883962392807,
-0.022402102127671242,
0.04589034616947174,
0.11378712207078934,
0.10677782446146011,
-0.0009012487716972828,
0.09169294685125351,
0.0534517876803875,
0.009979895316064358,
-0.10637086629867554,
-0.01611679419875145,
0.0013743514427915215,
-0.009793322533369064,
-0.05635382980108261,
0.03652619570493698,
-0.051737748086452484,
0.05483744293451309,
0.010438327677547932,
0.04369576647877693,
0.020246943458914757,
-0.048733703792095184,
-0.01042775809764862,
0.001978752901777625,
-0.04381096363067627,
0.014592385850846767,
-0.05110630393028259,
-0.006755760405212641,
-0.07130495458841324,
-0.10628926753997803,
0.02753681316971779,
0.050399549305438995,
0.04985662177205086,
-0.0829743891954422,
0.01867140829563141,
0.014440620318055153,
0.019620118662714958,
-0.028833305463194847,
0.020670587196946144,
0.011027907021343708,
0.05775013193488121
] | https://github.com/gangiswag/cornstack | 29 | 0 | 0 | 0 |
BingoGuard: LLM Content Moderation Tools with Risk Levels | https://openreview.net/forum?id=HPSAkIHRbb | [
"Fan Yin",
"Philippe Laban",
"XIANGYU PENG",
"Yilun Zhou",
"Yixin Mao",
"Vaibhav Vats",
"Linnea Ross",
"Divyansh Agarwal",
"Caiming Xiong",
"Chien-Sheng Wu"
] | Poster | Malicious content generated by large language models (LLMs) can pose varying degrees of harm.
Although existing LLM-based moderators can detect harmful content, they struggle to assess risk levels and may miss lower-risk outputs.
Accurate risk assessment allows platforms with different safety thresholds to tailor content filtering and rejection. In this paper, we introduce per-topic severity rubrics for 11 harmful topics and build BingoGuard, an LLM-based moderation system designed to predict both binary safety labels and severity levels.
To address the lack of annotations on levels of severity, we propose a scalable generate-then-filter framework that first generates responses across different severity levels and then filters out low-quality responses. Using this framework, we create BingoGuardTrain, a training dataset with 54,897 examples covering a variety of topics, response severity, styles, and BingoGuardTest, a test set with 988 examples explicitly labeled based on our severity rubrics that enables fine-grained analysis on model behaviors on different severity levels. Our BingoGuard-8B, trained on BingoGuardTrain, achieves the state-of-the-art performance on several moderation benchmarks, including WildGuardTest and HarmBench, as well as BingoGuardTest, outperforming best public models, WildGuard, by 4.3\%. Our analysis demonstrates that incorporating severity levels into training significantly enhances detection performance and enables the model to effectively gauge the severity of harmful responses. Warning: this paper includes red-teaming examples that may be harmful in nature. | LLM, safety guardrail, content moderator | We build a LLM moderator that can express risk levels. This moderator achieves the best performance on public benchmarks and our new test set. | 12,595 | 2503.06550 | [
-0.0470593199133873,
0.004983632359653711,
-0.03796292841434479,
0.01247241161763668,
0.10471278429031372,
0.03912153095006943,
0.07061010599136353,
-0.03944447264075279,
-0.0433778390288353,
-0.0529923252761364,
-0.05078340694308281,
-0.0909508764743805,
0.10923617333173752,
0.017638646066188812,
-0.03244058042764664,
0.05755145475268364,
0.06232522800564766,
0.08633080124855042,
-0.060036949813365936,
0.016828635707497597,
0.043081797659397125,
0.08425451070070267,
0.053587593138217926,
-0.022186608985066414,
-0.06428559869527817,
-0.026069626212120056,
-0.04504929482936859,
0.020988482981920242,
-0.012810717336833477,
-0.03654724359512329,
0.01960740052163601,
0.011825420893728733,
-0.013413960114121437,
0.03325055539608002,
-0.021157830953598022,
-0.026575442403554916,
0.020947003737092018,
-0.021302349865436554,
-0.04839262738823891,
-0.08311643451452255,
-0.04072510078549385,
0.014127327129244804,
-0.025665638968348503,
0.03166279196739197,
-0.0036051233764737844,
-0.07189168781042099,
-0.037150442600250244,
-0.03321161866188049,
-0.04742405563592911,
-0.009837957099080086,
-0.053802892565727234,
-0.011939138174057007,
0.030887281522154808,
0.009246204048395157,
-0.03887126222252846,
-0.06751985847949982,
0.014167280867695808,
-0.008089885115623474,
-0.05036609619855881,
-0.030624089762568474,
-0.07907532900571823,
-0.09819517284631729,
-0.04902857169508934,
-0.036220792680978775,
-0.027966992929577827,
-0.047921355813741684,
-0.015095503069460392,
0.09058865904808044,
0.007900585420429707,
0.05947067216038704,
-0.04747559875249863,
0.030256858095526695,
-0.016194546595215797,
0.07577653229236603,
-0.011586910113692284,
0.03732592612504959,
-0.050036631524562836,
0.010422301478683949,
0.0880838930606842,
-0.04138651862740517,
-0.017680728808045387,
-0.04424364119768143,
0.06451202929019928,
0.02851843647658825,
0.03928441181778908,
-0.015181692317128181,
0.0033892695792019367,
0.017164751887321472,
-0.021689629182219505,
0.11026646196842194,
-0.04141620546579361,
0.029977675527334213,
0.15304888784885406,
0.0527888685464859,
0.0028198633808642626,
-0.025105085223913193,
-0.017149480059742928,
-0.15523025393486023,
0.04748200997710228,
0.10593771934509277,
-0.0132847735658288,
0.0619935467839241,
-0.013706373050808907,
-0.08479306101799011,
0.07240249961614609,
-0.02221670374274254,
0.01293158158659935,
0.017366468906402588,
-0.03021574765443802,
-0.02538461983203888,
-0.012393169105052948,
0.12489613890647888,
0.042890675365924835,
-0.11811994016170502,
0.09992241114377975,
0.000766972079873085,
0.0012208130210638046,
-0.001224058330990374,
0.01722700148820877,
0.03884490579366684,
0.01622934266924858,
0.038428790867328644,
0.04351181909441948,
0.019074643030762672,
0.07464911043643951,
0.03566628322005272,
-0.05953404679894447,
1.5207791941515299e-33,
0.13270491361618042,
0.0007481278153136373,
-0.049740858376026154,
0.04541792348027229,
0.02220175229012966,
0.0018648741533979774,
-0.02189786173403263,
0.05518410727381706,
0.01935841701924801,
-0.05424799025058746,
-0.013988985680043697,
-0.051467303186655045,
-0.04692366346716881,
0.049263108521699905,
0.07241751998662949,
0.0073334453627467155,
-0.05063847079873085,
0.0008035727660171688,
-0.05751648172736168,
-0.012147168628871441,
-0.0097317760810256,
-0.06839196383953094,
0.033852469176054,
-0.03155900165438652,
0.04303435608744621,
0.034939270466566086,
0.060136936604976654,
-0.00942147709429264,
0.008385895751416683,
0.0401201993227005,
-0.1026356890797615,
-0.035864394158124924,
0.0658181831240654,
0.001875568530522287,
0.04609810933470726,
0.02590961754322052,
-0.0706278458237648,
0.007918545044958591,
0.02127961814403534,
-0.010305398143827915,
-0.05798061937093735,
0.007195224519819021,
0.031104039400815964,
0.017245497554540634,
-0.004259063396602869,
-0.03240596503019333,
-0.008190572261810303,
-0.01742830127477646,
0.004724485799670219,
-0.06358998268842697,
0.07655680924654007,
-0.021779300644993782,
0.08347810059785843,
0.01569719798862934,
-0.1323288381099701,
0.010337822139263153,
-0.014382172375917435,
-0.0162038691341877,
-0.02610146813094616,
0.016564415767788887,
0.02062424272298813,
0.03542127087712288,
-0.021548140794038773,
-0.03754135221242905,
0.05419813469052315,
-0.023641273379325867,
0.02415238693356514,
0.04733003303408623,
-0.016537975519895554,
-0.04228559881448746,
-0.01719009131193161,
0.006188458763062954,
0.026955094188451767,
-0.039531122893095016,
-0.08569887280464172,
0.020901553332805634,
0.04033007472753525,
-0.03204004094004631,
0.0371658056974411,
-0.03417107090353966,
0.09177990257740021,
-0.03820377215743065,
0.06195184215903282,
-0.08538457006216049,
-0.05670728161931038,
-0.0157448910176754,
0.0718136578798294,
-0.0541752353310585,
-0.04466591402888298,
-0.026301056146621704,
-0.010361452586948872,
-0.03338032588362694,
0.051647622138261795,
-0.010965125635266304,
0.04355723038315773,
-1.2810955286768808e-33,
-0.08498584479093552,
-0.034158896654844284,
-0.0806991383433342,
0.06497195363044739,
-0.036480799317359924,
-0.06851743161678314,
-0.017851009964942932,
0.04898549243807793,
0.03223082795739174,
-0.026204103603959084,
-0.08288413286209106,
-0.061220984905958176,
0.06648124754428864,
-0.036594681441783905,
0.026343073695898056,
-0.028107058256864548,
-0.05258810520172119,
0.057073839008808136,
-0.058893635869026184,
0.04207749664783478,
-0.03519226610660553,
0.1145857647061348,
-0.09094058722257614,
0.034052979201078415,
-0.08782907575368881,
-0.017239155247807503,
-0.006619622465223074,
0.0010366322239860892,
0.08789906650781631,
0.024084262549877167,
0.005094702355563641,
0.08452761173248291,
0.008366665802896023,
0.007384857628494501,
-0.013825316913425922,
-0.038029737770557404,
0.07551998645067215,
-0.03062313422560692,
0.005923230201005936,
0.06735219806432724,
0.0502038449048996,
0.0000033043938856280874,
-0.11478601396083832,
0.01347887422889471,
-0.0034106916282325983,
0.011534308083355427,
-0.0017639625584706664,
-0.00870048813521862,
0.03850146755576134,
-0.021862192079424858,
-0.0743255540728569,
-0.01713397726416588,
-0.016621313989162445,
0.03304193913936615,
-0.008014998398721218,
-0.14384222030639648,
-0.06812756508588791,
-0.0767800509929657,
0.023369908332824707,
0.057032451033592224,
-0.08051514625549316,
0.04748285934329033,
-0.02441825345158577,
-0.013527123257517815,
0.07690554112195969,
-0.05396322160959244,
-0.013841934502124786,
0.02329828217625618,
-0.015246463008224964,
-0.01955343224108219,
0.04181183502078056,
0.0016765693435445428,
-0.001959308283403516,
-0.02074064128100872,
-0.00457687396556139,
-0.03990746662020683,
0.02975347451865673,
-0.02161009982228279,
-0.025923829525709152,
0.01957922801375389,
0.09818378835916519,
0.014235648326575756,
0.003961898386478424,
0.0310021061450243,
0.05144478380680084,
0.015973398461937904,
0.055260561406612396,
0.09703391790390015,
0.014799278229475021,
0.08441456407308578,
-0.03646667301654816,
0.003851730842143297,
-0.045953843742609024,
0.010271359235048294,
-0.02548515796661377,
-3.8443872085736075e-8,
0.002190243685618043,
-0.015795810148119926,
-0.09721130132675171,
0.07124416530132294,
0.04499354213476181,
0.12979093194007874,
-0.06572137773036957,
0.0602031908929348,
-0.00019474020518828183,
-0.03410385921597481,
0.050629857927560806,
0.044121213257312775,
-0.09218621253967285,
-0.10048386454582214,
-0.01950916461646557,
0.039831891655921936,
-0.04518831521272659,
0.09537019580602646,
-0.040881700813770294,
0.0163725633174181,
0.036124054342508316,
0.07641907036304474,
0.038846950978040695,
-0.029609793797135353,
0.024883737787604332,
-0.03889999911189079,
0.023991821333765984,
0.04353491961956024,
0.060361310839653015,
0.08125170320272446,
-0.00346768950112164,
-0.037651654332876205,
-0.09503377974033356,
-0.008628369309008121,
-0.015019413083791733,
0.15500357747077942,
-0.012614618986845016,
-0.01831737905740738,
0.04406438767910004,
0.005310511216521263,
0.022333383560180664,
0.01716947741806507,
-0.018973125144839287,
0.01612180657684803,
-0.011132081970572472,
-0.04517211392521858,
-0.07042161375284195,
-0.08477866649627686,
0.11311472952365875,
-0.06049620360136032,
-0.014781036414206028,
-0.019148144870996475,
0.01959965005517006,
0.03723914921283722,
0.021089620888233185,
0.03632821887731552,
-0.029635965824127197,
-0.0226246640086174,
0.07058852165937424,
0.05911262705922127,
0.07082347571849823,
-0.040482960641384125,
0.07851943373680115,
0.015204019844532013
] | 0 | 0 | 0 | 0 |
|
Learn hybrid prototypes for multivariate time series anomaly detection | https://openreview.net/forum?id=8TBGdH3t6a | [
"Ke-Yuan Shen"
] | Poster | In multivariate time series anomaly detection (MTSAD), reconstruction-based models reconstruct testing series with learned knowledge of only normal series and identify anomalies with higher reconstruction errors. In practice, over-generalization often occurs with unexpectedly well reconstruction of anomalies. Although memory banks are employed by reconstruction-based models to fight against over-generalization, these models are only efficient to detect point anomalies since they learn normal prototypes from time points, leaving interval anomalies and periodical anomalies to be discovered. To settle this problem, this paper propose a hybrid prototypes learning model for MTSAD based on reconstruction, named as H-PAD. First, normal prototypes are learned from different sizes of the patches for time series to discover interval anomalies. These prototypes in different sizes are integrated together to reconstruct query series so that any anomalies would be smoothed off and high reconstruction errors are produced. Furthermore, period prototypes are learned to discover periodical anomalies. One period prototype is memorized for one variable of the query series. Finally, extensive experiments on five benchmark datasets show the effectiveness of H-PAD. | prototypes;time series;anomaly detection | This paper propose a hybrid prototypes learning model for MTSAD based on reconstruction to fight against over-generalization. | 12,594 | null | [
-0.09922835230827332,
-0.057764772325754166,
0.03150862455368042,
0.028027834370732307,
-0.011617513373494148,
-0.03692276030778885,
-0.002240534871816635,
-0.026092279702425003,
0.03823741152882576,
-0.02729373425245285,
-0.03523009642958641,
0.020183902233839035,
0.032439179718494415,
0.009950172156095505,
-0.07494305074214935,
-0.06279661506414413,
-0.0012105759233236313,
0.037199053913354874,
0.035404104739427567,
-0.010796241462230682,
-0.03361072391271591,
0.07665310055017471,
-0.05518614500761032,
0.035767726600170135,
0.0030207987874746323,
0.044839758425951004,
0.06500660628080368,
0.013411472551524639,
-0.03176235035061836,
0.019076256081461906,
0.00899585522711277,
0.016082612797617912,
-0.0909322202205658,
-0.03683611378073692,
0.03548261150717735,
0.02330375276505947,
-0.007351200096309185,
0.02724325656890869,
-0.006592040881514549,
-0.03767668455839157,
0.044221267104148865,
-0.04032008722424507,
0.006449675653129816,
0.023879917338490486,
0.05076246336102486,
-0.01226415392011404,
0.032270222902297974,
-0.07151733338832855,
0.02058698982000351,
-0.0015625833766534925,
0.018010277301073074,
-0.029505521059036255,
-0.05573621019721031,
0.030338266864418983,
0.04145398363471031,
-0.005168200470507145,
-0.07964546978473663,
-0.04343443736433983,
0.022176172584295273,
0.006052400451153517,
-0.05869721621274948,
-0.03366712108254433,
0.01326773315668106,
-0.04407700151205063,
-0.008141566067934036,
0.001813385752029717,
-0.006054317578673363,
0.0802423283457756,
0.10282435268163681,
0.022662721574306488,
-0.10104002803564072,
0.10336218029260635,
-0.05354565009474754,
0.0014691043179482222,
-0.00754723371937871,
0.07713563740253448,
0.054174281656742096,
-0.05367153137922287,
0.04430435597896576,
-0.06386367231607437,
-0.0522906631231308,
0.03588132560253143,
0.050199393182992935,
-0.02527451142668724,
0.1173744723200798,
0.0035846277605742216,
0.06887964904308319,
0.09526921808719635,
0.003419151296839118,
-0.04717884957790375,
0.010517730377614498,
-0.012049006298184395,
-0.010634738951921463,
-0.03774159029126167,
-0.028922418132424355,
-0.01487993635237217,
0.022133762016892433,
0.010538912378251553,
0.12455767393112183,
0.08730413764715195,
0.014031370170414448,
0.07225148379802704,
0.054446492344141006,
-0.007412684615701437,
-0.005269902292639017,
-0.011897100135684013,
0.011430357582867146,
-0.023548603057861328,
0.08304501324892044,
-0.06815534830093384,
-0.04170583561062813,
0.024415522813796997,
0.01777811348438263,
-0.01994507759809494,
0.04625523462891579,
-0.007710662204772234,
-0.015457932837307453,
-0.02859293855726719,
-0.04342731088399887,
0.03632020950317383,
0.012015281245112419,
-0.03378148749470711,
0.029354410246014595,
0.025601163506507874,
0.025438454002141953,
0.024921389296650887,
-0.03879804164171219,
4.99231936515447e-33,
0.04396975785493851,
0.054681360721588135,
0.004885009955614805,
0.008471560664474964,
0.0017041987739503384,
-0.002602013759315014,
-0.037753358483314514,
0.04736872389912605,
0.08684127032756805,
0.034918736666440964,
-0.04028031975030899,
-0.0015209709526970983,
-0.04696771875023842,
0.0011871774913743138,
0.006136631593108177,
0.008672093972563744,
-0.05137316882610321,
0.024468187242746353,
-0.020647747442126274,
-0.03197230398654938,
-0.0355660542845726,
-0.06062287092208862,
0.05611618235707283,
-0.02463771216571331,
0.02885681577026844,
0.004741949494928122,
0.06585007905960083,
-0.002944955602288246,
-0.019613953307271004,
0.026407156139612198,
-0.04904865473508835,
-0.00975730363279581,
-0.028170965611934662,
0.04491834342479706,
0.015238556079566479,
-0.04789862409234047,
0.02648860588669777,
-0.060674190521240234,
0.04288407787680626,
-0.046597566455602646,
-0.06883040815591812,
-0.042823124676942825,
0.004554831422865391,
-0.06385250389575958,
-0.018555335700511932,
-0.045302219688892365,
0.10671040415763855,
-0.03813515976071358,
-0.028823159635066986,
-0.057346194982528687,
-0.03535108640789986,
-0.03751827031373978,
-0.017856748774647713,
-0.1150955930352211,
-0.027238402515649796,
0.00045237320591695607,
-0.031887829303741455,
-0.016217906028032303,
0.05866265669465065,
0.15698382258415222,
0.00716116838157177,
0.032100144773721695,
0.02562817558646202,
0.030044788494706154,
-0.04516930505633354,
0.04044486954808235,
0.06393877416849136,
0.05554109439253807,
0.06660458445549011,
0.05058250203728676,
0.09515656530857086,
-0.04818430542945862,
-0.024915015324950218,
-0.037769805639982224,
0.04174903407692909,
-0.06367620825767517,
0.150880366563797,
0.043869536370038986,
0.0058254217728972435,
-0.04498951509594917,
-0.012902434915304184,
-0.09488837420940399,
0.09938965737819672,
-0.07732418924570084,
-0.059673670679330826,
-0.039761222898960114,
0.03443891555070877,
-0.026737429201602936,
-0.09343628585338593,
-0.07708162814378738,
-0.0597616508603096,
0.03411391004920006,
-0.013748159632086754,
-0.04149126261472702,
-0.004049957264214754,
-5.742810060017113e-33,
0.017991559579968452,
0.021031618118286133,
-0.06701276451349258,
-0.05179792270064354,
-0.05812571942806244,
-0.10081511735916138,
-0.08459785580635071,
0.0388575941324234,
-0.05256031081080437,
-0.043060433119535446,
-0.021039754152297974,
-0.049664780497550964,
0.007904240861535072,
-0.09224428236484528,
-0.038091693073511124,
0.01514839194715023,
-0.029175899922847748,
-0.10707072168588638,
0.002548747695982456,
0.019350433722138405,
0.027015240862965584,
0.005232708994299173,
-0.07861977070569992,
-0.05664672330021858,
-0.03677471727132797,
0.03709225356578827,
-0.05967707931995392,
0.08689562976360321,
-0.019182177260518074,
0.023736534640192986,
-0.06766650825738907,
-0.07209678739309311,
0.039181940257549286,
0.026187710464000702,
-0.03346621245145798,
0.05059017241001129,
0.06541553139686584,
-0.042340558022260666,
0.013818240724503994,
0.03872202709317207,
0.04793649911880493,
0.07357757538557053,
0.03313419595360756,
0.00030450860504060984,
-0.006141678895801306,
0.01950719580054283,
-0.0356832779943943,
0.13671599328517914,
0.06494119018316269,
-0.018025752156972885,
-0.04941317066550255,
0.041672222316265106,
0.0003133064310532063,
-0.014923793263733387,
-0.08327295631170273,
0.027600586414337158,
0.05477723106741905,
-0.05311403051018715,
-0.057901229709386826,
0.044763870537281036,
-0.04582758620381355,
-0.08108052611351013,
0.021806124597787857,
-0.009842636063694954,
0.021259186789393425,
0.0005363717209547758,
0.028128767386078835,
-0.0375186912715435,
-0.05883482098579407,
-0.005415944382548332,
0.04921913146972656,
0.007386419456452131,
-0.03772785887122154,
-0.07234559953212738,
0.024817651137709618,
-0.057043664157390594,
-0.023312749341130257,
-0.03777008503675461,
-0.036062926054000854,
-0.01884246990084648,
-0.05810875818133354,
-0.010115090757608414,
0.07797599583864212,
0.1250399947166443,
-0.012447413057088852,
0.142644003033638,
-0.0045766681432724,
-0.004151951987296343,
0.014630033634603024,
-0.09188102185726166,
-0.01358980592340231,
0.05705434828996658,
-0.08115038275718689,
0.06715685874223709,
0.010707898996770382,
-4.466208025633023e-8,
-0.14264684915542603,
0.08166616410017014,
0.016865914687514305,
-0.08690893650054932,
0.10581883043050766,
-0.07584355026483536,
-0.0844380259513855,
0.0772329792380333,
0.016401460394263268,
-0.058096352964639664,
0.04753497615456581,
-0.1223207637667656,
-0.01360208448022604,
-0.04122043401002884,
0.0050720153376460075,
0.020857352763414383,
0.028877589851617813,
0.016890065744519234,
-0.0878007635474205,
0.003929500933736563,
0.016566202044487,
0.04308805242180824,
0.04842406511306763,
0.0033906218595802784,
0.09136056154966354,
-0.022357042878866196,
-0.02895091101527214,
0.09908389300107956,
0.0920693427324295,
0.02352898009121418,
0.010287893936038017,
0.0009958266746252775,
0.04843025282025337,
-0.023249316960573196,
0.05207153782248497,
0.11040493100881577,
0.05232706293463707,
-0.0005052081542089581,
-0.07127457857131958,
-0.03590092435479164,
-0.03509373217821121,
-0.012179411947727203,
-0.006275822874158621,
0.0614754781126976,
-0.015369433909654617,
-0.031113745644688606,
-0.009958969429135323,
-0.06935662776231766,
0.0567016676068306,
-0.03612714260816574,
0.08567523211240768,
-0.005511273164302111,
0.06686672568321228,
0.04140089824795723,
0.09074323624372482,
0.021666724234819412,
0.018667371943593025,
-0.0510682612657547,
0.03973284736275673,
0.029127901419997215,
0.06401962041854858,
-0.005012965761125088,
-0.0989421010017395,
0.052230317145586014
] | 0 | 0 | 0 | 0 |
|
Reasoning with Latent Thoughts: On the Power of Looped Transformers | https://openreview.net/forum?id=din0lGfZFd | [
"Nikunj Saunshi",
"Nishanth Dikkala",
"Zhiyuan Li",
"Sanjiv Kumar",
"Sashank J. Reddi"
] | Poster | Large language models have shown remarkable reasoning abilities and scaling laws suggest that large parameter count, especially along the depth axis, is the primary driver. In this work, we make a stronger claim --- many reasoning problems require a large depth but not necessarily many parameters. This unlocks a novel application of looped models for reasoning. Firstly, we show that for many synthetic reasoning problems like addition, $p$-hop induction, and math problems, a $k$-layer transformer looped $L$ times nearly matches the performance of a $kL$-layer non-looped model, and is significantly better than a $k$-layer model. This is further corroborated by theoretical results showing that many such reasoning problems can be solved via iterative algorithms, and thus, can be solved effectively using looped models with nearly optimal depth. Perhaps surprisingly, these benefits also translate to practical settings of language modeling --- on many downstream reasoning tasks, a language model with $k$-layers looped $L$ times can be competitive to, if not better than, a $kL$-layer language model. In fact, our empirical analysis reveals an intriguing phenomenon: looped and non-looped models exhibit scaling behavior that depends on their effective depth, akin to the inference-time scaling of chain-of-thought (CoT) reasoning. We further elucidate the connection to CoT reasoning by proving that looped models implicitly generate latent thoughts and can simulate $T$ steps of CoT with $T$ loops. Inspired by these findings, we also present an interesting dichotomy between reasoning and memorization, and design a looping-based regularization that is effective on both fronts. | looped models, reasoning, language model, iterative algorithm, inductive bias, scaling | Looped models can solve many reasoning problems and have an inductive bias towards improving reasoning of language models | 12,582 | 2502.17416 | [
-0.09021521359682083,
-0.043564584106206894,
0.004637864883989096,
0.0005220180610194802,
0.01766127720475197,
-0.02438374236226082,
0.01642068289220333,
0.07042758911848068,
0.08496642112731934,
0.010511117987334728,
-0.03437831997871399,
-0.0468263179063797,
0.04518341273069382,
0.09217628091573715,
0.041552938520908356,
0.03146374598145485,
0.020794522017240524,
0.10020139068365097,
-0.07002558559179306,
-0.12627287209033966,
-0.0047812024131417274,
-0.05470021814107895,
0.017445307224988937,
-0.02942095510661602,
-0.00512506952509284,
0.004656679928302765,
0.003597421571612358,
-0.07828857749700546,
0.015762289986014366,
0.03606816381216049,
-0.010541426949203014,
0.07527543604373932,
-0.03886287286877632,
0.04478520154953003,
-0.00019190045713912696,
0.07347020506858826,
-0.05864850431680679,
0.005820442922413349,
-0.0626554861664772,
0.05717520788311958,
0.023176370188593864,
-0.041822806000709534,
0.024918146431446075,
-0.04092036187648773,
0.03828423470258713,
-0.010755771771073341,
-0.022985514253377914,
0.03437355160713196,
-0.02278180420398712,
-0.015754127874970436,
-0.07412754744291306,
-0.02142416313290596,
-0.08388246595859528,
0.02462010644376278,
-0.011792750097811222,
-0.038397204130887985,
0.056159500032663345,
0.04395001381635666,
-0.019665105268359184,
-0.019872067496180534,
-0.013494441285729408,
-0.09447138756513596,
-0.00779027258977294,
-0.015425198711454868,
-0.04450133442878723,
-0.0017443933757022023,
-0.021079987287521362,
0.06947851926088333,
-0.035692647099494934,
0.09887097775936127,
-0.02471916377544403,
0.0816001296043396,
-0.07818672060966492,
0.00477990647777915,
0.06979233771562576,
0.01825820654630661,
0.08054941892623901,
-0.07060323655605316,
0.027079984545707703,
-0.10614661872386932,
-0.01287802029401064,
0.007130472920835018,
-0.07471846044063568,
-0.011401137337088585,
0.0985657200217247,
-0.006439534481614828,
-0.020116977393627167,
0.03755645453929901,
0.023921212181448936,
-0.01432440523058176,
-0.024333227425813675,
-0.06982409954071045,
0.004199109971523285,
0.07026664167642593,
0.11921679973602295,
0.09057558327913284,
0.015336628071963787,
-0.013189881108701229,
-0.05026431009173393,
0.058245908468961716,
-0.0021322784014046192,
0.08683689683675766,
0.0335865318775177,
-0.04714534431695938,
0.005488153547048569,
-0.07801488041877747,
0.01100097969174385,
0.034391675144433975,
-0.005420781206339598,
-0.1296028345823288,
-0.003881220007315278,
0.03582585230469704,
0.08834431320428848,
0.02176331728696823,
0.03702668845653534,
-0.0725020244717598,
0.07315811514854431,
-0.021692480891942978,
-0.03225087374448776,
-0.040511321276426315,
0.02729867585003376,
0.06520188599824905,
-0.009126748889684677,
0.04382112994790077,
0.05980812385678291,
-0.014645078219473362,
-0.0033739095088094473,
3.08962251559826e-34,
0.028402259573340416,
0.11176732182502747,
0.10774777084589005,
0.07567243278026581,
0.054395221173763275,
0.01107257790863514,
0.06743089109659195,
-0.004672511015087366,
0.005510515533387661,
0.03225364908576012,
-0.08318112045526505,
0.028208132833242416,
-0.08664409816265106,
0.028226641938090324,
0.02319454774260521,
0.012020601890981197,
-0.026051001623272896,
0.027619071304798126,
-0.04353664070367813,
-0.06776110082864761,
0.11066452413797379,
-0.0718802958726883,
0.011956057511270046,
0.031961943954229355,
-0.004927926231175661,
-0.020437104627490044,
0.06242953613400459,
-0.06137385591864586,
-0.06826737523078918,
-0.0340174064040184,
-0.10817326605319977,
-0.04694418981671333,
-0.035408832132816315,
0.0734240859746933,
-0.01790671981871128,
-0.06781477481126785,
0.053466636687517166,
-0.10469352453947067,
0.02800721675157547,
0.001158075756393373,
-0.05214071646332741,
0.03498026728630066,
0.010878326371312141,
0.07026076316833496,
-0.034186869859695435,
-0.04348498210310936,
-0.04434303566813469,
-0.02908504754304886,
-0.06424429267644882,
0.00485461950302124,
0.060553379356861115,
0.04947442188858986,
-0.014860201627016068,
-0.07157912105321884,
0.05738276615738869,
0.008154908195137978,
0.03501315414905548,
0.04013921692967415,
0.07164829224348068,
0.12135583162307739,
0.0033158008009195328,
0.03667375072836876,
-0.05910114198923111,
0.08369728177785873,
0.008349097333848476,
0.03296094387769699,
-0.007516392972320318,
0.06261501461267471,
0.07527212798595428,
-0.009353495202958584,
-0.05426205322146416,
0.06090501695871353,
-0.062147993594408035,
-0.025637105107307434,
0.06222468242049217,
-0.06796915084123611,
-0.0034333046060055494,
-0.1354653686285019,
0.04286791384220123,
-0.011861134320497513,
0.013796553947031498,
0.006682770326733589,
-0.05700830742716789,
-0.0025679890532046556,
-0.0389242023229599,
-0.027086550369858742,
-0.016666023060679436,
-0.09422556310892105,
-0.00940315704792738,
-0.01110958680510521,
-0.059619661420583725,
-0.026456374675035477,
0.02176317200064659,
-0.05797490477561951,
0.03141186758875847,
-1.3177922582576448e-33,
-0.010611684992909431,
-0.0813041478395462,
-0.032337892800569534,
0.09044641256332397,
-0.03125496208667755,
-0.07198897004127502,
0.02142375148832798,
-0.014852332882583141,
-0.03826341778039932,
-0.037529975175857544,
-0.03469967842102051,
0.009045753628015518,
0.05250925198197365,
0.004159220028668642,
0.0935865119099617,
-0.05132581666111946,
-0.02901916205883026,
-0.022763749584555626,
0.030291976407170296,
0.04606272280216217,
0.01805855706334114,
0.07252370566129684,
-0.11669857054948807,
-0.02497977949678898,
-0.06093284487724304,
0.03867919370532036,
-0.09737907350063324,
0.03497510030865669,
0.05051052197813988,
0.036850061267614365,
0.023218030110001564,
-0.08260460197925568,
-0.02967146970331669,
0.007837367244064808,
-0.055149052292108536,
0.03954092785716057,
0.0468725860118866,
0.0016440630424767733,
-0.0511055514216423,
0.021889105439186096,
-0.005665960256010294,
-0.12700079381465912,
-0.015895767137408257,
-0.03376712277531624,
-0.06804775446653366,
-0.015602211467921734,
-0.03632950410246849,
-0.018156832084059715,
-0.008938487619161606,
-0.004425569903105497,
0.018862100318074226,
0.002549341879785061,
-0.014152668416500092,
0.008030004799365997,
0.007957248017191887,
0.00794712919741869,
-0.07713734358549118,
-0.02273837849497795,
0.05755781754851341,
-0.04717721417546272,
-0.06487713754177094,
-0.02964644692838192,
0.04035906493663788,
-0.07198850810527802,
-0.021444784477353096,
-0.014273595996201038,
-0.031004533171653748,
-0.019273150712251663,
0.01636626198887825,
0.023109974339604378,
0.05042019486427307,
-0.026492906734347343,
-0.05443282425403595,
0.025430867448449135,
0.02496432140469551,
0.004890599753707647,
-0.0017203579191118479,
0.04054497182369232,
0.01965581439435482,
-0.05863432586193085,
-0.046194400638341904,
-0.030631234869360924,
0.05922988802194595,
-0.001575158559717238,
0.021218720823526382,
-0.020919976755976677,
0.05901995301246643,
0.07659897208213806,
0.019743457436561584,
0.012228126637637615,
0.007525161374360323,
-0.004624862223863602,
-0.015960628166794777,
0.09002839028835297,
-0.04028157517313957,
-4.146081522549139e-8,
-0.08958465605974197,
0.08717388659715652,
0.03201182931661606,
-0.03429228439927101,
0.10316276550292969,
-0.10757862031459808,
0.018009500578045845,
0.06150605529546738,
-0.0023037672508507967,
0.021719394251704216,
0.08098851889371872,
-0.037295930087566376,
0.018792979419231415,
0.030083067715168,
-0.0014941766858100891,
0.08598477393388748,
-0.0014158107805997133,
-0.01386132463812828,
0.011685332283377647,
-0.016647297888994217,
-0.00924756657332182,
0.1130436584353447,
-0.09032613039016724,
0.035798221826553345,
0.02288384921848774,
-0.03215571492910385,
-0.027118554338812828,
0.1292160302400589,
0.08351761102676392,
0.04101363196969032,
-0.00925133191049099,
-0.03319963067770004,
-0.0025201179087162018,
0.027013584971427917,
0.09729709476232529,
0.07319634407758713,
-0.008769504725933075,
0.006487267091870308,
-0.03644833341240883,
-0.07502220571041107,
0.01645502634346485,
0.04560406133532524,
-0.06699024140834808,
0.03742066025733948,
0.05662769079208374,
-0.019347507506608963,
-0.07521089166402817,
-0.09234116971492767,
-0.05068136379122734,
0.012781145982444286,
-0.01789754256606102,
-0.036826103925704956,
-0.03335872292518616,
0.023498909547924995,
0.0226326622068882,
0.0837707668542862,
-0.01931637153029442,
-0.05531800165772438,
0.009643006138503551,
0.04116328805685043,
-0.032951027154922485,
0.019128065556287766,
-0.013341298326849937,
0.0783531442284584
] | 0 | 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.