Dataset Viewer
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
|
|---|---|---|---|---|---|---|---|---|---|
DarkBench: Benchmarking Dark Patterns in Large Language Models | https://openreview.net/forum?id=odjMSBSWRt | [
"Esben Kran",
"Hieu Minh Nguyen",
"Akash Kundu",
"Sami Jawhar",
"Jinsuk Park",
"Mateusz Maria Jurewicz"
] | Oral | We introduce DarkBench, a comprehensive benchmark for detecting dark design patterns—manipulative techniques that influence user behavior—in interactions with large language models (LLMs). Our benchmark comprises 660 prompts across six categories: brand bias, user retention, sycophancy, anthropomorphism, harmful generation, and sneaking. We evaluate models from five leading companies (OpenAI, Anthropic, Meta, Mistral, Google) and find that some LLMs are explicitly designed to favor their developers' products and exhibit untruthful communication, among other manipulative behaviors. Companies developing LLMs should recognize and mitigate the impact of dark design patterns to promote more ethical Al. | Dark Patterns, AI Deception, Large Language Models | We introduce DarkBench, a benchmark revealing that many large language models employ manipulative dark design patterns. Organizations developing LLMs should actively recognize and mitigate the impact of dark design patterns to promote ethical Al. | 14,257 | 2503.10728 | [
-0.02725336328148842,
-0.03766116499900818,
0.009408959187567234,
0.008022491820156574,
0.02661564014852047,
-0.041066356003284454,
-0.015834596008062363,
-0.038164280354976654,
0.06413612514734268,
-0.06206069141626358,
-0.07550547271966934,
-0.046710625290870667,
0.03552878648042679,
-0.05406486243009567,
0.019869044423103333,
-0.01234534103423357,
0.06806624680757523,
0.012362184002995491,
-0.05998788774013519,
-0.08512728661298752,
-0.009513086639344692,
0.024141410365700722,
0.05707956477999687,
-0.03172401711344719,
-0.010784459300339222,
-0.0237556304782629,
-0.014784744940698147,
-0.08572808653116226,
0.02752182073891163,
-0.04575201869010925,
-0.001081965514458716,
0.06211918592453003,
0.045926984399557114,
0.051983751356601715,
-0.013575469143688679,
-0.018515367060899734,
-0.05219582840800285,
-0.024864520877599716,
-0.03377365320920944,
-0.05695898085832596,
-0.05072716251015663,
-0.029252037405967712,
-0.017435351386666298,
0.012733501382172108,
0.00035370062687434256,
-0.05423778295516968,
0.01410842314362526,
0.016247032210230827,
-0.08896724134683609,
-0.030613217502832413,
-0.06691648066043854,
-0.12300164252519608,
0.03386547788977623,
0.041562195867300034,
-0.05685406178236008,
-0.03772249445319176,
0.050537534058094025,
0.0503329299390316,
0.03898217901587486,
-0.0036720968782901764,
0.015636345371603966,
-0.046773675829172134,
-0.08291354030370712,
0.04805668815970421,
0.014900654554367065,
0.1274234503507614,
-0.018438752740621567,
0.02333330549299717,
-0.09420909732580185,
0.03181283921003342,
0.019997214898467064,
-0.012809780426323414,
0.05363484099507332,
0.04822093993425369,
-0.01974482275545597,
0.0757548063993454,
0.04644254967570305,
-0.041345126926898956,
0.0028498335741460323,
-0.10343454778194427,
0.029701365157961845,
-0.04138656705617905,
0.00032804690999910235,
0.0857333317399025,
0.08462636917829514,
0.02496698871254921,
-0.0007817422156222165,
-0.011845709756016731,
0.040350235998630524,
0.07346062362194061,
-0.01409204863011837,
0.003981675952672958,
0.05796120688319206,
-0.02399049513041973,
-0.06123082712292671,
0.02366272360086441,
0.01656397245824337,
-0.06375549733638763,
-0.06106005236506462,
0.09052251279354095,
-0.05419601872563362,
0.021180281415581703,
0.018260885030031204,
-0.05139351263642311,
0.012242337688803673,
0.015664581209421158,
0.02035710960626602,
-0.016534050926566124,
0.007043092045933008,
0.03814417123794556,
-0.02513936534523964,
0.07208656519651413,
0.008667247369885445,
-0.05016050860285759,
0.1114603653550148,
-0.04377036169171333,
0.005443500354886055,
-0.02059185318648815,
0.022318650037050247,
0.0914716050028801,
0.04122017323970795,
0.08386026322841644,
0.06987536698579788,
-0.03333135321736336,
0.04192086309194565,
0.031047655269503593,
-0.0582808181643486,
1.8245408015639288e-33,
0.06691724061965942,
0.05357668548822403,
-0.04903746023774147,
0.045374542474746704,
0.08969824016094208,
0.013151225633919239,
0.008133964613080025,
0.05424857884645462,
-0.07738373428583145,
0.01489853672683239,
0.055487826466560364,
0.055358417332172394,
-0.04097156971693039,
0.10585343092679977,
0.06472890079021454,
0.07623346149921417,
-0.014902224764227867,
0.012007812969386578,
-0.03603854402899742,
-0.07128223776817322,
0.06501562893390656,
-0.023907378315925598,
0.02836756221950054,
0.0007124018738977611,
-0.06674569845199585,
0.03468668833374977,
0.049784548580646515,
-0.008675756864249706,
0.016711058095097542,
0.07149533182382584,
-0.06500926613807678,
-0.014692750759422779,
0.07549888640642166,
0.03488026559352875,
0.008490116335451603,
0.06658007949590683,
-0.045894134789705276,
-0.07619110494852066,
0.10966283082962036,
-0.011504586786031723,
-0.04955213516950607,
0.06813523173332214,
-0.03833865001797676,
-0.01748538389801979,
-0.028149666264653206,
0.0470101423561573,
0.0822855532169342,
-0.1267814040184021,
0.003620493458583951,
0.021162448450922966,
0.008448602631688118,
0.038016438484191895,
0.06359130889177322,
0.04331183806061745,
0.02198192849755287,
-0.04334091767668724,
0.005339586641639471,
-0.03720971569418907,
-0.003224446438252926,
0.03261402249336243,
-0.004264193587005138,
0.04571279138326645,
-0.0033593277912586927,
-0.024927029386162758,
0.03433926776051521,
-0.03460843488574028,
-0.039175618439912796,
0.019179124385118484,
-0.027644433081150055,
-0.02955455332994461,
-0.011525227688252926,
-0.029693083837628365,
-0.05205480754375458,
-0.013739817775785923,
-0.10418975353240967,
-0.03258692845702171,
0.0036490121856331825,
-0.02134775184094906,
0.08020258694887161,
-0.029326971620321274,
0.07386434823274612,
0.0019756106194108725,
0.019622301682829857,
-0.03539150953292847,
-0.08978961408138275,
-0.015235715545713902,
0.03474920988082886,
-0.02325628325343132,
-0.059941671788692474,
0.0604463666677475,
0.014295153319835663,
-0.05548042058944702,
0.02084490656852722,
0.039208151400089264,
-0.02810678817331791,
-3.052895112961779e-33,
-0.10450997948646545,
-0.051222532987594604,
-0.08174528181552887,
0.06504429876804352,
-0.007970758713781834,
-0.055329207330942154,
-0.02872643992304802,
0.009111875668168068,
0.011677499860525131,
0.03152582794427872,
0.0487089566886425,
-0.00036829817690886557,
0.03234413266181946,
0.05200035870075226,
0.06636986136436462,
-0.051663246005773544,
0.08220646530389786,
-0.02107970044016838,
0.00196244427934289,
-0.008493518456816673,
0.036811139434576035,
0.06684940308332443,
-0.13408203423023224,
-0.021932676434516907,
-0.007806114852428436,
0.002770253224298358,
-0.031256888061761856,
-0.030968300998210907,
0.013369565829634666,
-0.005276272539049387,
-0.06435481458902359,
0.07094638049602509,
-0.04462094232439995,
-0.004087795969098806,
0.009325074031949043,
0.04005633294582367,
-0.09606313705444336,
0.02610047720372677,
-0.008175108581781387,
-0.022437110543251038,
0.061140816658735275,
0.011784061789512634,
-0.05977046862244606,
-0.06697628647089005,
-0.04409831389784813,
0.006447686813771725,
-0.15545128285884857,
-0.05425072833895683,
0.03258689120411873,
-0.07250956445932388,
0.02585647627711296,
-0.02086455002427101,
0.02205852046608925,
-0.006925045978277922,
-0.10767130553722382,
-0.059519167989492416,
-0.015433765947818756,
-0.04503605514764786,
0.051383599638938904,
0.07958749681711197,
-0.04936424642801285,
0.056843239814043045,
0.00414865929633379,
-0.07821125537157059,
-0.028626492246985435,
-0.060961414128541946,
-0.014508123509585857,
0.06539840996265411,
-0.022192219272255898,
-0.08474328368902206,
0.11523433029651642,
-0.038521625101566315,
-0.11502377688884735,
0.015858694911003113,
-0.06085941195487976,
-0.011374445632100105,
0.015302039682865143,
-0.022482357919216156,
-0.07274577766656876,
-0.026751428842544556,
0.08762212097644806,
0.020338332280516624,
0.041335202753543854,
0.030201932415366173,
0.0037944980431348085,
0.020123381167650223,
-0.026577699929475784,
0.11346566677093506,
-0.07098742574453354,
0.08925995230674744,
-0.043049439787864685,
0.01870155893266201,
-0.06634778529405594,
0.07131466269493103,
-0.015957342460751534,
-4.153108434934438e-8,
-0.07786471396684647,
-0.08548519015312195,
0.07635452598333359,
0.07746531069278717,
0.041036564856767654,
0.006243977230042219,
-0.051259905099868774,
0.04574931785464287,
-0.015832774341106415,
0.014840146526694298,
0.04265827685594559,
0.0038810737896710634,
-0.008706649765372276,
0.041126739233732224,
0.00789526104927063,
0.1253899782896042,
-0.08014044910669327,
-0.00551453884691,
-0.028171582147479057,
-0.0443393737077713,
0.02549256756901741,
0.0861758291721344,
0.044616762548685074,
-0.07701359689235687,
0.011061446741223335,
-0.004838874563574791,
-0.05553821846842766,
0.008583543822169304,
0.054738130420446396,
0.07608182728290558,
0.014078214764595032,
0.03911260887980461,
0.009041590616106987,
-0.04806103557348251,
0.011718666180968285,
0.08147767186164856,
-0.11614669859409332,
-0.027620678767561913,
0.030309919267892838,
0.07774268835783005,
0.01502170693129301,
0.08040203154087067,
-0.012133173644542694,
0.004585601855069399,
-0.03803291916847229,
-0.005916581489145756,
-0.053038422018289566,
-0.06806780397891998,
0.024012411013245583,
-0.03154388815164566,
-0.035537589341402054,
-0.07542356103658676,
-0.008672119118273258,
0.08253424614667892,
0.004671923816204071,
-0.025555390864610672,
0.01834530010819435,
0.02463257871568203,
-0.015056243166327477,
0.02211780846118927,
0.1288624107837677,
-0.06069512665271759,
-0.01840270683169365,
-0.004128862638026476
] |
RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style | https://openreview.net/forum?id=QEHrmQPBdd | [
"Yantao Liu",
"Zijun Yao",
"Rui Min",
"Yixin Cao",
"Lei Hou",
"Juanzi Li"
] | Oral | Reward models are critical in techniques like Reinforcement Learning from Human Feedback (RLHF) and Inference Scaling Laws, where they guide language model alignment and select optimal responses.
Despite their importance, existing reward model benchmarks often evaluate models by asking them to distinguish between responses generated by models of varying power.
However, this approach fails to assess reward models on subtle but critical content changes and variations in style, resulting in a low correlation with policy model performance.
To this end, we introduce RM-Bench, a novel benchmark designed to evaluate reward models based on their sensitivity to subtle content differences and resistance to style biases.
Extensive experiments demonstrate that RM-Bench strongly correlates with policy model performance, making it a reliable reference for selecting reward models to align language models effectively.
We evaluate nearly 40 reward models on RM-Bench.
Our results reveal that even state-of-the-art models achieve an average performance of only 46.6%, which falls short of random-level accuracy (50%) when faced with style bias interference.
These findings highlight the significant room for improvement in current reward models. | Reward Models, Language Models, Evaluation, Alignment | null | 13,985 | null | [
-0.08573347330093384,
-0.07113663852214813,
0.02327777072787285,
0.03225872293114662,
0.0407232865691185,
0.06568308174610138,
-0.01666353829205036,
-0.042861457914114,
0.08131477236747742,
0.01655031368136406,
-0.09637922048568726,
-0.0660894364118576,
0.06317998468875885,
0.004250515252351761,
0.0593532994389534,
-0.035737842321395874,
0.09620589762926102,
0.009509209543466568,
-0.03785747289657593,
-0.05649465322494507,
0.0027530814986675978,
0.04221605509519577,
0.10160304605960846,
0.007304837927222252,
0.04744357988238335,
-0.006100278347730637,
-0.00293090776540339,
0.003864211030304432,
0.0011970828054472804,
-0.05311119556427002,
-0.004613617435097694,
-0.035908158868551254,
0.03575211763381958,
-0.06736607104539871,
-0.06696508079767227,
0.04070545360445976,
-0.0619097538292408,
-0.017696373164653778,
0.0058169374242424965,
-0.009912537410855293,
-0.09313040971755981,
-0.009124196134507656,
0.04462938383221626,
0.022055458277463913,
0.10728151351213455,
-0.014645985327661037,
0.003624104894697666,
-0.016663681715726852,
-0.06822779029607773,
0.011727849021553993,
-0.044883716851472855,
-0.05886904150247574,
0.03300149738788605,
-0.01671597920358181,
-0.045220114290714264,
0.05489792302250862,
0.07484222948551178,
0.11678709834814072,
0.05518877133727074,
-0.06329551339149475,
-0.05423427373170853,
-0.13007023930549622,
-0.05594972148537636,
0.03506758436560631,
-0.03441294655203819,
-0.04997897148132324,
-0.043366845697164536,
0.051344845443964005,
-0.05002283677458763,
-0.00048419402446597815,
0.012767826206982136,
0.02177531085908413,
0.061471231281757355,
0.017837446182966232,
0.03604637458920479,
0.04714549332857132,
0.04204820841550827,
-0.04493333771824837,
0.02774588018655777,
-0.09575418382883072,
-0.06489820778369904,
-0.07975783944129944,
0.014425752684473991,
-0.0281685683876276,
0.0912032499909401,
-0.02131539024412632,
0.05799419432878494,
-0.0034307686146348715,
0.020791973918676376,
0.029784806072711945,
0.011959938332438469,
-0.005174824967980385,
0.017671413719654083,
-0.029738815501332283,
-0.05953371897339821,
0.09436213970184326,
-0.036394715309143066,
-0.08176159113645554,
-0.09253042191267014,
0.09771993011236191,
0.05308326706290245,
0.09528520703315735,
-0.030805954709649086,
-0.04744300618767738,
0.050956737250089645,
-0.04158761352300644,
0.010351182892918587,
0.027162261307239532,
0.006157604977488518,
-0.10149092227220535,
0.013140876777470112,
0.05995451286435127,
-0.03537783399224281,
0.058181505650281906,
0.013857309706509113,
-0.0132502606138587,
0.015131039544939995,
-0.05300314724445343,
-0.002961504040285945,
0.06720440089702606,
0.018793223425745964,
-0.02853560633957386,
0.03756135329604149,
-0.022507673129439354,
-0.011684699915349483,
0.01632387936115265,
-0.010732905007898808,
9.191440121629016e-34,
0.0755094364285469,
0.06124718859791756,
0.051364727318286896,
0.03972627967596054,
0.09154081344604492,
-0.005745171569287777,
0.0029205221217125654,
-0.0003625437384471297,
-0.053663454949855804,
-0.01347876526415348,
0.002437024610117078,
0.03917163982987404,
-0.034092504531145096,
0.08694049715995789,
0.012068835087120533,
0.030842669308185577,
-0.06160696595907211,
0.04721482843160629,
-0.023411143571138382,
0.041497860103845596,
0.05649900808930397,
0.04191943258047104,
0.05953970178961754,
-0.048678405582904816,
-0.016566842794418335,
0.0025276474189013243,
0.04073340445756912,
-0.03129763528704643,
-0.1442795842885971,
0.006541336886584759,
-0.05163491144776344,
-0.055989820510149,
0.048886582255363464,
0.01553122978657484,
-0.03351951763033867,
-0.05353254824876785,
-0.011103871278464794,
-0.06116986647248268,
0.04192372038960457,
-0.033017925918102264,
-0.03311639651656151,
0.01232108287513256,
-0.008931810036301613,
0.02814416214823723,
-0.07939017564058304,
-0.018565570935606956,
0.05408089607954025,
-0.0609530545771122,
-0.04492039233446121,
0.022712135687470436,
0.06845294684171677,
0.07308252900838852,
0.08196094632148743,
0.0070557743310928345,
-0.011942184530198574,
0.008620251901447773,
0.0018926175544038415,
0.038732871413230896,
0.02825116738677025,
-0.001872362452559173,
0.056045401841402054,
0.06526144593954086,
-0.006938569713383913,
0.005361331161111593,
-0.016558155417442322,
0.08639746904373169,
-0.03897008299827576,
0.052895594388246536,
0.01017466839402914,
0.023998623713850975,
0.03579927608370781,
0.014753255993127823,
-0.035262465476989746,
-0.037548091262578964,
0.03077452816069126,
-0.018206002190709114,
0.022115763276815414,
-0.09628234803676605,
0.06363183259963989,
-0.01261567696928978,
-0.028124181553721428,
-0.010578835383057594,
0.003394732717424631,
-0.08123518526554108,
-0.059128835797309875,
-0.034255772829055786,
0.033248238265514374,
-0.08105659484863281,
0.05064034089446068,
0.017076289281249046,
-0.0254728514701128,
-0.03882165998220444,
0.006953302305191755,
-0.012315618805587292,
0.05150243639945984,
-2.9044610331796333e-33,
-0.02964741736650467,
-0.050288405269384384,
0.015247331000864506,
0.05184539780020714,
-0.06928400695323944,
-0.003965550102293491,
-0.028800129890441895,
0.020365288481116295,
0.008717936463654041,
-0.06640580296516418,
0.03093055635690689,
0.012475875206291676,
-0.024204563349485397,
-0.0030326712876558304,
0.03162594884634018,
-0.060958389192819595,
-0.0232403501868248,
0.032341573387384415,
0.03985748440027237,
0.01521783135831356,
0.10983294993638992,
0.08389827609062195,
-0.09616974741220474,
0.02209244854748249,
-0.03398415818810463,
0.018680792301893234,
-0.07322891056537628,
-0.024465372785925865,
0.016613425686955452,
0.002315153367817402,
-0.04816222935914993,
-0.008133585564792156,
-0.062030091881752014,
-0.01587238907814026,
-0.052441395819187164,
-0.003734721802175045,
-0.024716917425394058,
-0.01852991059422493,
-0.057480588555336,
0.1602916568517685,
0.013009632006287575,
-0.06986857950687408,
-0.06673093140125275,
0.0018184216460213065,
-0.010436143726110458,
-0.028703473508358,
-0.13388067483901978,
-0.03581228852272034,
-0.03053765371441841,
-0.03455020859837532,
-0.04444556683301926,
0.00018847583851311356,
-0.07639667391777039,
0.03965327516198158,
-0.09549011290073395,
-0.09651503711938858,
-0.13972614705562592,
0.01658107340335846,
0.07213027030229568,
0.05816031992435455,
-0.0817619040608406,
0.028639165684580803,
-0.0744437724351883,
-0.053692251443862915,
-0.031414944678545,
-0.031249400228261948,
0.06127903610467911,
-0.038669828325510025,
0.070047527551651,
-0.02276381477713585,
0.07072308659553528,
-0.004952498711645603,
0.029726777225732803,
0.04399613291025162,
-0.011982392519712448,
0.01820014789700508,
0.055574458092451096,
-0.028712816536426544,
-0.015748582780361176,
-0.06421294808387756,
0.029116632416844368,
0.038831014186143875,
0.006768912076950073,
-0.024747012183070183,
0.013593929819762707,
0.15040281414985657,
-0.0032494256738573313,
0.0696401298046112,
-0.034177616238594055,
0.009501239284873009,
-0.014790287241339684,
-0.0037918868474662304,
0.03771047666668892,
0.051011163741350174,
-0.048659779131412506,
-4.505028172729908e-8,
-0.06508340686559677,
0.04945403337478638,
0.05426088348031044,
0.12431573867797852,
-0.03696940839290619,
0.022186806425452232,
-0.050790440291166306,
-0.021059513092041016,
0.013570820912718773,
-0.05279898643493652,
0.0937688872218132,
0.10419180989265442,
0.01262490265071392,
-0.04959705471992493,
-0.07216189801692963,
0.07816723734140396,
-0.005022066179662943,
0.07210436463356018,
-0.0029980852268636227,
-0.043994225561618805,
0.05044237896800041,
0.060700997710227966,
0.03576138615608215,
-0.08013499528169632,
0.002205959754064679,
-0.03276202455163002,
-0.05489747226238251,
0.11685448884963989,
0.06575721502304077,
-0.011753297410905361,
0.004709896165877581,
-0.03125620633363724,
-0.004222629591822624,
-0.021314965561032295,
0.10781049728393555,
0.0562283881008625,
0.0034137675538659096,
-0.025767985731363297,
0.029108548536896706,
0.048233114182949066,
0.014154014177620411,
0.0355883352458477,
-0.01129568088799715,
0.029336625710129738,
-0.024180756881833076,
0.024797780439257622,
-0.041916679590940475,
-0.031335312873125076,
0.012833289802074432,
-0.08172310143709183,
0.03294476866722107,
-0.07323981821537018,
-0.11794662475585938,
0.03648711368441582,
0.008617677725851536,
0.034952130168676376,
-0.03162109851837158,
-0.02234988659620285,
-0.013200739398598671,
0.04673437401652336,
0.08512052893638611,
-0.031576987355947495,
-0.05958160385489464,
-0.043577492237091064
] |
TopoLM: brain-like spatio-functional organization in a topographic language model | https://openreview.net/forum?id=aWXnKanInf | [
"Neil Rathi",
"Johannes Mehrer",
"Badr AlKhamissi",
"Taha Osama A Binhuraib",
"Nicholas Blauch",
"Martin Schrimpf"
] | Oral | Neurons in the brain are spatially organized such that neighbors on tissue often exhibit similar response profiles. In the human language system, experimental studies have observed clusters for syntactic and semantic categories, but the mechanisms underlying this functional organization remain unclear. Here, building on work from the vision literature, we develop TopoLM, a transformer language model with an explicit two-dimensional spatial representation of model units. By combining a next-token prediction objective with a spatial smoothness loss, representations in this model assemble into clusters that correspond to semantically interpretable groupings of text and closely match the functional organization in the brain's language system. TopoLM successfully predicts the emergence of a spatially organized cortical language system as well as the organization of functional clusters selective for fine-grained linguistic features empirically observed in human cortex. Our results suggest that the functional organization of the human language system is driven by a unified spatial objective, and provide a functionally and spatially aligned model of language processing in the brain.Neurons in the brain are spatially organized such that neighbors on tissue often exhibit similar response profiles. In the human language system, experimental studies have observed clusters for syntactic and semantic categories, but the mechanisms underlying this functional organization remain unclear. Here, building on work from the vision literature, we develop TopoLM, a transformer language model with an explicit two-dimensional spatial representation of model units. By combining a next-token prediction objective with a spatial smoothness loss, representations in this model assemble into clusters that correspond to semantically interpretable groupings of text and closely match the functional organization in the brain's language system. TopoLM successfully predicts the emergence of a spatially organized cortical language system as well as the organization of functional clusters selective for fine-grained linguistic features empirically observed in human cortex. Our results suggest that the functional organization of the human language system is driven by a unified spatial objective, and provide a functionally and spatially aligned model of language processing in the brain. | language modeling, topography, fMRI, neuroscience | We develop a transformer language model with topographically organized units predicting brain-like spatio-functional organization. | 13,712 | 2410.11516 | [
0.04553741216659546,
-0.1560545116662979,
0.05150135979056358,
0.026262419298291206,
0.04672883078455925,
0.006334717385470867,
-0.019015828147530556,
0.02199351042509079,
0.09638210386037827,
-0.017915070056915283,
-0.038319721817970276,
-0.08659375458955765,
0.029651854187250137,
0.08215548098087311,
0.022474924102425575,
-0.034293126314878464,
-0.02108442224562168,
0.04902893677353859,
-0.048220377415418625,
-0.08156446367502213,
0.08396447449922562,
0.02860826998949051,
0.032418690621852875,
0.015001039020717144,
0.007901583798229694,
0.14433065056800842,
0.0016878553433343768,
-0.0064452276565134525,
-0.03493763878941536,
-0.029529886320233345,
0.06854137778282166,
0.0436374731361866,
-0.011078464798629284,
0.06340886652469635,
-0.027019817382097244,
0.0805005356669426,
-0.02935239113867283,
0.004195573274046183,
0.05214552581310272,
-0.06621922552585602,
-0.009238519705832005,
-0.019845668226480484,
0.01477646641433239,
0.0576004683971405,
0.028617532923817635,
0.055657532066106796,
0.03558000549674034,
-0.023171020671725273,
-0.09784391522407532,
-0.0061380257830023766,
-0.06867966055870056,
-0.0388142392039299,
-0.003751123324036598,
0.0683477595448494,
-0.06306611001491547,
0.11297525465488434,
0.02743745781481266,
0.0359237939119339,
0.01727607287466526,
-0.05569373816251755,
-0.010363185778260231,
-0.07926840335130692,
0.015843598172068596,
0.03806789219379425,
0.018837209790945053,
0.022361692041158676,
0.0661400854587555,
0.013201048597693443,
-0.07114620506763458,
0.0016225435538217425,
0.08143337815999985,
0.041230641305446625,
0.016792286187410355,
-0.03387535363435745,
0.032782092690467834,
0.021027760580182076,
0.04091351851820946,
-0.015179919078946114,
0.06159677729010582,
-0.07002837210893631,
0.06339669972658157,
0.10164162516593933,
0.05162529647350311,
-0.002869429998099804,
0.11629093438386917,
0.04724324122071266,
-0.020872943103313446,
-0.02502610720694065,
-0.07899277657270432,
-0.004977241158485413,
-0.04887563735246658,
-0.07635924965143204,
-0.0072800652123987675,
-0.08113235235214233,
-0.02361191064119339,
-0.031120607629418373,
-0.025291545316576958,
-0.06651974469423294,
0.0602388009428978,
0.03614838793873787,
-0.015530199743807316,
-0.03750459477305412,
0.0597522035241127,
-0.03666989505290985,
0.03925454616546631,
-0.04383474215865135,
0.04128517583012581,
-0.0035790628753602505,
0.014442922547459602,
0.030768731608986855,
-0.0032378719188272953,
-0.02660124935209751,
-0.036511119455099106,
0.09532236307859421,
-0.05898197740316391,
-0.015838243067264557,
0.04827984422445297,
-0.0013363221660256386,
0.027925822883844376,
-0.03166859969496727,
-0.012684562243521214,
0.003154971171170473,
0.02173997461795807,
-0.018550164997577667,
0.026617567986249924,
-0.0066589051857590675,
-0.049885936081409454,
1.8438830789703547e-34,
-0.03759343922138214,
-0.03514508530497551,
-0.018586091697216034,
0.022840889170765877,
0.04223719611763954,
-0.057276442646980286,
-0.0028683366253972054,
0.011954978108406067,
0.0020398194901645184,
0.009136573411524296,
-0.09607186913490295,
0.05413389950990677,
0.01786392740905285,
0.13851653039455414,
0.011873060837388039,
-0.022457005456089973,
0.000982459168881178,
0.020379094406962395,
-0.0680062472820282,
-0.07461219280958176,
-0.07439426332712173,
0.09355838596820831,
0.021143842488527298,
-0.08788882195949554,
-0.11807285249233246,
0.030725672841072083,
-0.0216030515730381,
-0.13198788464069366,
-0.04197567701339722,
0.0026607769541442394,
-0.11503086984157562,
0.09805881232023239,
-0.020844943821430206,
-0.011286795139312744,
0.06098150089383125,
-0.04729517176747322,
0.01712084747850895,
-0.04559268057346344,
0.015387781895697117,
-0.03551211953163147,
-0.016080042347311974,
0.05491854250431061,
0.026576975360512733,
-0.05559540167450905,
-0.000382870202884078,
0.043837178498506546,
0.02717989683151245,
0.0019283839501440525,
-0.0162294153124094,
-0.08953898400068283,
0.04039234668016434,
-0.024427171796560287,
0.035365309566259384,
0.018545042723417282,
0.06478481739759445,
0.06342240422964096,
0.0004696781688835472,
0.019633112475275993,
0.09624997526407242,
0.05444207787513733,
0.004862111061811447,
0.033189382404088974,
0.05810791999101639,
0.03547486662864685,
0.06474186480045319,
-0.05576843395829201,
-0.024586090818047523,
0.02360837161540985,
0.08756774663925171,
0.03092537820339203,
0.06541429460048676,
0.11847694963216782,
-0.02606893889605999,
0.007591808680444956,
-0.003530956571921706,
0.007090848870575428,
0.07526775449514389,
-0.007574145682156086,
-0.09519141912460327,
0.005504386965185404,
0.03066098317503929,
-0.13377414643764496,
-0.05274784937500954,
-0.055402692407369614,
0.006846363190561533,
-0.01851661689579487,
0.093290314078331,
-0.06255178153514862,
-0.07974447309970856,
-0.031118080019950867,
-0.02657075971364975,
-0.034495726227760315,
0.08197667449712753,
-0.01692725159227848,
-0.02164648100733757,
-2.11935331538222e-33,
-0.0811121016740799,
0.024815844371914864,
-0.006478607654571533,
0.0001723663735901937,
-0.1062019094824791,
-0.0280939731746912,
0.027940217405557632,
0.023631978780031204,
-0.038662612438201904,
0.030029257759451866,
-0.0545191764831543,
-0.03767161816358566,
0.0857049897313118,
0.019257741048932076,
0.05830080807209015,
-0.029355810955166817,
0.043245963752269745,
0.004502313211560249,
0.050039373338222504,
-0.019788624718785286,
0.034674208611249924,
0.04706878960132599,
-0.09270797669887543,
0.05163772404193878,
-0.06217779219150543,
0.038972120732069016,
-0.0646558329463005,
-0.01227163802832365,
-0.08480094373226166,
0.06567459553480148,
-0.02337632328271866,
-0.01624247618019581,
-0.057714227586984634,
-0.041245464235544205,
-0.044019754976034164,
0.04276079684495926,
-0.01950368471443653,
-0.09244377911090851,
-0.0841577872633934,
-0.0413958802819252,
0.010386224836111069,
-0.047086600214242935,
0.00810704380273819,
0.03937717154622078,
-0.03125963360071182,
-0.03804590180516243,
-0.02465185709297657,
-0.020110448822379112,
-0.10104706138372421,
0.03720841929316521,
-0.020185818895697594,
0.06472522020339966,
-0.09107886999845505,
-0.030275465920567513,
-0.03459635004401207,
0.014097476378083229,
-0.023262646049261093,
-0.10365947335958481,
0.03136442229151726,
-0.057020656764507294,
0.0023452017921954393,
-0.03796222805976868,
0.012043626047670841,
-0.021066918969154358,
0.04715190455317497,
-0.0139967305585742,
-0.019877729937434196,
-0.041779566556215286,
0.030186355113983154,
0.0405292883515358,
0.00445869006216526,
-0.02875565178692341,
0.010935789905488491,
0.06848662346601486,
0.05449587106704712,
-0.027131518349051476,
0.07742371410131454,
-0.04030328243970871,
0.0013103843666613102,
-0.0348920002579689,
-0.013448537327349186,
-0.024941977113485336,
-0.047686636447906494,
0.037496890872716904,
0.04600975662469864,
0.0378103032708168,
0.0028401787858456373,
0.02795291505753994,
0.0689874067902565,
0.0004498185298871249,
0.02379530854523182,
-0.0649748146533966,
-0.05063321068882942,
0.029736420139670372,
-0.04585586488246918,
-3.784015589758383e-8,
-0.13808108866214752,
-0.06311552971601486,
-0.019944241270422935,
0.009841314516961575,
-0.018784306943416595,
-0.07115067541599274,
0.034515559673309326,
0.10160397738218307,
-0.05768374726176262,
-0.034892503172159195,
0.03637528792023659,
-0.03744177520275116,
-0.061077408492565155,
-0.02900421805679798,
0.03174386918544769,
0.07810086756944656,
0.00824885256588459,
0.05915096029639244,
-0.02293264865875244,
-0.015333345159888268,
-0.015338273718953133,
0.028955809772014618,
0.01774739660322666,
0.012835424393415451,
0.011938773095607758,
-0.06213722750544548,
-0.06630722433328629,
0.058619577437639236,
-0.05550459399819374,
-0.08769971877336502,
0.03600570559501648,
0.06481985002756119,
-0.044784389436244965,
0.015442725270986557,
-0.00981956534087658,
0.05312146618962288,
0.01122012734413147,
0.043844882398843765,
0.004474204033613205,
0.01212562806904316,
0.04066562280058861,
0.05615711957216263,
-0.008056373335421085,
0.05227476730942726,
0.03383473679423332,
0.04857201129198074,
0.04746174067258835,
-0.0683666467666626,
0.07057865709066391,
-0.0629262775182724,
-0.053435660898685455,
0.08757394552230835,
-0.03238057345151901,
0.09155789762735367,
-0.0035220319405198097,
-0.027860213071107864,
-0.05816078558564186,
-0.005732374731451273,
0.04857496917247772,
0.009874583221971989,
0.06927762925624847,
0.07972060143947601,
-0.04969989135861397,
-0.04053337126970291
] |
Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows | https://openreview.net/forum?id=XmProj9cPs | [
"Fangyu Lei",
"Jixuan Chen",
"Yuxiao Ye",
"Ruisheng Cao",
"Dongchan Shin",
"Hongjin SU",
"ZHAOQING SUO",
"Hongcheng Gao",
"Wenjing Hu",
"Pengcheng Yin",
"Victor Zhong",
"Caiming Xiong",
"Ruoxi Sun",
"Qian Liu",
"Sida Wang",
"Tao Yu"
] | Oral | Real-world enterprise text-to-SQL workflows often involve complex cloud or local data across various database systems, multiple SQL queries in various dialects, and diverse operations from data transformation to analytics.
We introduce Spider 2.0, an evaluation framework comprising $632$ real-world text-to-SQL workflow problems derived from enterprise-level database use cases.
The databases in Spider 2.0 are sourced from real data applications, often containing over 1,000 columns and stored in local or cloud database systems such as BigQuery and Snowflake.
We show that solving problems in Spider 2.0 frequently requires understanding and searching through database metadata, dialect documentation, and even project-level codebases.
This challenge calls for models to interact with complex SQL workflow environments, process extremely long contexts, perform intricate reasoning, and generate multiple SQL queries with diverse operations, often exceeding $100$ lines, which goes far beyond traditional text-to-SQL challenges.
Our evaluations indicate that based on o1-preview, our code agent framework successfully solves only 21.3\% of the tasks, compared with 91.2\% on Spider 1.0 and 73.0\% on BIRD.
Our results on Spider 2.0 show that while language models have demonstrated remarkable performance in code generation --- especially in prior text-to-SQL benchmarks --- they require significant improvement in order to achieve adequate performance for real-world enterprise usage.
Progress on Spider 2.0 represents crucial steps towards developing intelligent, autonomous, code agents for real-world enterprise settings.
Our code, baseline models, and data are available at [spider2-sql.github.io](spider2-sql.github.io) . | LLM Benchmark, Data Science and Engineering, Code Generation, Text-to-SQL, LLM Agent | A benchmark for enterprise-level Text-to-SQL involving complex databases, challenging tasks, and real-world scenarios. | 13,657 | 2411.07763 | [
-0.03679848089814186,
-0.04623184725642204,
-0.04603664577007294,
0.04519940912723541,
-0.014134707860648632,
-0.09154586493968964,
-0.0011208809446543455,
-0.01573074981570244,
-0.018215196207165718,
-0.00974492821842432,
-0.08401896059513092,
-0.06710424274206161,
0.04240303114056587,
-0.007999260909855366,
0.040645208209753036,
0.08182445913553238,
0.0761510506272316,
-0.0314653143286705,
-0.06204354763031006,
-0.025340713560581207,
-0.026900624856352806,
0.016192929819226265,
0.03422974422574043,
0.04046975448727608,
-0.010158254764974117,
0.04366671293973923,
-0.05119182541966438,
-0.008372629061341286,
0.0012311951722949743,
-0.005651586223393679,
-0.05029633268713951,
0.09466206282377243,
-0.07771865278482437,
0.12122274935245514,
0.01301112025976181,
0.004556932020932436,
-0.0032415995374321938,
0.014214442111551762,
-0.0630638599395752,
-0.04923839122056961,
-0.017538145184516907,
-0.031354308128356934,
-0.05738116800785065,
0.001755559933371842,
0.021464815363287926,
-0.08034602552652359,
-0.008484615944325924,
-0.04718345031142235,
-0.0034719340037554502,
0.008572583086788654,
-0.1511484980583191,
-0.07358819991350174,
0.01633356511592865,
0.018883580341935158,
0.058492451906204224,
0.029655443504452705,
0.056475069373846054,
-0.03862312063574791,
-0.005157268140465021,
-0.025048276409506798,
-0.03976137191057205,
-0.0533372163772583,
0.03529508039355278,
0.02105543576180935,
-0.06397830694913864,
-0.0234206635504961,
-0.010067850351333618,
-0.010737703181803226,
0.00752084469422698,
-0.05134054273366928,
-0.12427173554897308,
0.06429439783096313,
-0.08794111758470535,
0.07064379751682281,
-0.033437490463256836,
0.019872667267918587,
0.012084417045116425,
-0.01994716189801693,
0.02269597165286541,
-0.005850065499544144,
-0.013308065943419933,
-0.024064788594841957,
0.004207198042422533,
0.1005440354347229,
0.05942376330494881,
-0.020128551870584488,
0.08269057422876358,
0.005168795585632324,
0.050673708319664,
0.0594814233481884,
-0.028497468680143356,
-0.06609749794006348,
0.10835833847522736,
-0.030361199751496315,
0.02803494594991207,
0.028660688549280167,
0.013389838859438896,
-0.06936697661876678,
0.001814609277062118,
0.06820806115865707,
-0.036816611886024475,
0.06661996990442276,
0.0669717788696289,
-0.08769169449806213,
-0.07641930133104324,
0.008823771961033344,
0.05023740604519844,
0.026536598801612854,
-0.013021397404372692,
-0.0665413960814476,
0.004591879900544882,
0.017528634518384933,
0.04901095852255821,
-0.08342698961496353,
-0.0013059922493994236,
-0.014601221308112144,
-0.011172833852469921,
-0.03869553655385971,
-0.04393954202532768,
0.12524844706058502,
0.023855477571487427,
0.11263646185398102,
-0.031489159911870956,
-0.006359628867357969,
0.07901684194803238,
0.00513065792620182,
-0.021554099395871162,
2.271100636194754e-33,
0.10556711256504059,
0.051054902374744415,
0.04686364158987999,
0.02353288233280182,
0.0990879014134407,
-0.03995773568749428,
0.05176575109362602,
0.0397346168756485,
-0.035405226051807404,
-0.005311330780386925,
0.03478455916047096,
0.08239415287971497,
-0.05446449667215347,
-0.028844261541962624,
0.061296362429857254,
-0.013997447676956654,
-0.03458976745605469,
0.11569161713123322,
-0.02301711216568947,
0.0801256075501442,
0.08507814258337021,
-0.05151547119021416,
0.010057132691144943,
0.03482179343700409,
0.09058701246976852,
0.05330345034599304,
-0.014064643532037735,
-0.019117368385195732,
0.050305865705013275,
0.02225484512746334,
-0.06583698093891144,
-0.07658544927835464,
-0.0034105759114027023,
0.10423947870731354,
0.004498778376728296,
-0.04736568406224251,
-0.022472187876701355,
-0.06143050640821457,
0.07929772138595581,
0.02771732397377491,
-0.031564850360155106,
0.042962562292814255,
-0.013449828140437603,
-0.0006757791270501912,
-0.043289780616760254,
-0.0649559274315834,
-0.01944565773010254,
0.018971551209688187,
0.03869757801294327,
-0.033039726316928864,
0.06070927530527115,
0.02143888920545578,
0.06367272883653641,
0.08397792279720306,
0.018200306221842766,
-0.026078691706061363,
0.046856362372636795,
-0.02846909873187542,
0.05127762258052826,
0.08303431421518326,
-0.08161744475364685,
-0.006670800037682056,
-0.030788009986281395,
0.0019718953408300877,
0.05768937990069389,
-0.045088499784469604,
-0.017376866191625595,
0.050747986882925034,
0.08471629023551941,
-0.04513711854815483,
-0.002292945049703121,
0.003926828503608704,
0.09298369288444519,
-0.0011836885241791606,
0.027771523222327232,
0.0045579709112644196,
0.0201429296284914,
-0.10091407597064972,
0.018857408314943314,
-0.02986990287899971,
-0.02690904214978218,
0.004409302491694689,
0.014652572572231293,
-0.018615299835801125,
-0.012455284595489502,
-0.038019534200429916,
-0.04364306479692459,
-0.05646201968193054,
-0.057145245373249054,
-0.013370078057050705,
-0.053456708788871765,
-0.03325027599930763,
0.020581187680363655,
-0.10328901559114456,
0.06261216849088669,
-2.555313680717482e-33,
-0.044039271771907806,
-0.08781062811613083,
-0.09015188366174698,
0.06131802499294281,
0.0072883195243775845,
-0.05074106901884079,
-0.017121564596891403,
-0.00204096594825387,
-0.0009592698770575225,
-0.04568051919341087,
-0.05979667976498604,
-0.00005019365198677406,
0.01458761841058731,
-0.03756977990269661,
0.04501191899180412,
-0.07758007198572159,
0.06114276871085167,
-0.03095105290412903,
0.01653175987303257,
0.11614562571048737,
-0.02208528108894825,
0.11240454018115997,
-0.08666758984327316,
0.04174330085515976,
-0.004969805013388395,
0.02634228579699993,
-0.01789916306734085,
-0.04523991048336029,
0.06618265062570572,
0.02282017096877098,
-0.0328015498816967,
-0.006876095663756132,
0.011750418692827225,
-0.0516977459192276,
-0.057013168931007385,
-0.021178755909204483,
0.014395305886864662,
-0.05702363699674606,
-0.002650483511388302,
0.03188792243599892,
0.031674690544605255,
-0.04375308007001877,
-0.010582738555967808,
-0.01632925681769848,
-0.034983787685632706,
0.07116023451089859,
-0.10523965954780579,
0.046954523772001266,
-0.009406615979969501,
-0.0934307724237442,
-0.04594714194536209,
-0.009134005755186081,
-0.03412928059697151,
-0.021486660465598106,
-0.014688709750771523,
-0.08351248502731323,
-0.04998484626412392,
-0.05943932756781578,
-0.07945747673511505,
0.008081542328000069,
-0.05391748249530792,
0.10732312500476837,
0.05895080417394638,
0.00023879115178715438,
0.025645552203059196,
0.0019632948096841574,
-0.06794403493404388,
-0.013377944007515907,
-0.10067416727542877,
-0.04635924845933914,
0.000872079748660326,
-0.08146726340055466,
-0.047276660799980164,
0.057610269635915756,
0.0722331628203392,
0.06446094065904617,
0.007447153329849243,
-0.05408579856157303,
0.05141102895140648,
0.05045704543590546,
0.0018966851057484746,
0.03539576381444931,
0.04863113909959793,
-0.009629818610846996,
0.048088911920785904,
-0.0037627536803483963,
-0.061076004058122635,
0.07472658902406693,
-0.05020181089639664,
0.03848656266927719,
-0.08721542358398438,
-0.024270527064800262,
-0.06431625783443451,
0.0338633693754673,
-0.018281996250152588,
-4.0280529134406606e-8,
-0.019983598962426186,
0.09340407699346542,
-0.022274930030107498,
0.003934643696993589,
0.031188804656267166,
-0.039687857031822205,
-0.01976056396961212,
0.15224483609199524,
0.030968094244599342,
0.02126840129494667,
0.024606680497527122,
0.013211804442107677,
0.033749986439943314,
-0.03099745139479637,
0.05934831500053406,
0.09561805427074432,
0.05246053263545036,
-0.05151144042611122,
-0.05340809002518654,
-0.034081730991601944,
0.03007221780717373,
0.060554713010787964,
-0.07719319313764572,
-0.027911636978387833,
0.017609618604183197,
-0.0300311129540205,
0.004700821358710527,
0.08946014940738678,
0.02081216312944889,
-0.030691567808389664,
0.048172593116760254,
-0.02336527593433857,
0.0034104182850569487,
-0.009460946545004845,
0.051112450659275055,
-0.06708675622940063,
0.054746098816394806,
-0.005002691876143217,
0.021879466250538826,
-0.05168944597244263,
0.09441956877708435,
0.03939831629395485,
-0.018109051510691643,
-0.0023457433562725782,
-0.016807232052087784,
0.010547095909714699,
-0.06277186423540115,
-0.02780473604798317,
0.04772021621465683,
-0.06680674850940704,
-0.012561400420963764,
-0.042684826999902725,
-0.05728323757648468,
0.07242217659950256,
0.040031470358371735,
0.0738961324095726,
-0.042931925505399704,
-0.053614042699337006,
0.08651003241539001,
0.041143305599689484,
0.016201995313167572,
-0.0035686830524355173,
0.04618388041853905,
0.04697632044553757
] |
Knowledge Entropy Decay during Language Model Pretraining Hinders New Knowledge Acquisition | https://openreview.net/forum?id=eHehzSDUFp | [
"Jiyeon Kim",
"Hyunji Lee",
"Hyowon Cho",
"Joel Jang",
"Hyeonbin Hwang",
"Seungpil Won",
"Youbin Ahn",
"Dohaeng Lee",
"Minjoon Seo"
] | Oral | In this work, we investigate how a model's tendency to broadly integrate its parametric knowledge evolves throughout pretraining, and how this behavior affects overall performance, particularly in terms of knowledge acquisition and forgetting. We introduce the concept of knowledge entropy, which quantifies the range of memory sources the model engages with; high knowledge entropy indicates that the model utilizes a wide range of memory sources, while low knowledge entropy suggests reliance on specific sources with greater certainty. Our analysis reveals a consistent decline in knowledge entropy as pretraining advances. We also find that the decline is closely associated with a reduction in the model's ability to acquire and retain knowledge, leading us to conclude that diminishing knowledge entropy (smaller number of active memory sources) impairs the model's knowledge acquisition and retention capabilities. We find further support for this by demonstrating that increasing the activity of inactive memory sources enhances the model's capacity for knowledge acquisition and retention. | knowledge entropy, knowledge acquisition and forgetting, evolving behavior during LLM pretraining | As pretraining progresses, models exhibit narrower integration of memory vectors, reflected by decreasing knowledge entropy, which hinders both knowledge acquisition and retention. | 13,581 | 2410.01380 | [
0.08067429810762405,
-0.06784215569496155,
0.0016120108775794506,
0.1156923845410347,
0.041765280067920685,
0.08341261744499207,
0.03904664143919945,
-0.009606799110770226,
0.08937966078519821,
-0.005971039179712534,
0.00008143703598761931,
0.08783090859651566,
0.08362888544797897,
-0.0014886445133015513,
0.034538548439741135,
-0.010556640103459358,
-0.023110153153538704,
0.006549345329403877,
-0.06468656659126282,
-0.046772800385951996,
0.029296603053808212,
0.004604103043675423,
0.04677041620016098,
0.02657514065504074,
0.013574746437370777,
-0.06925371289253235,
-0.07965712994337082,
-0.011091025546193123,
0.044551774859428406,
-0.04027250409126282,
0.03401263803243637,
0.06757070869207382,
0.0559735968708992,
0.04607781395316124,
-0.060251642018556595,
0.012680028565227985,
-0.04332413896918297,
0.030898192897439003,
-0.013404183089733124,
-0.007399613969027996,
-0.020410193130373955,
0.02686825394630432,
0.004614038392901421,
0.07993661612272263,
-0.0009742979309521616,
0.028462113812565804,
0.0068666161969304085,
-0.05202944949269295,
-0.01132832933217287,
-0.0023109656758606434,
-0.07348182797431946,
-0.0019239536486566067,
0.04596991464495659,
0.0193573497235775,
0.013821396045386791,
0.10177551954984665,
-0.028767045587301254,
0.08862274140119553,
-0.1231759563088417,
-0.02178080752491951,
-0.08902817219495773,
-0.1358872652053833,
-0.10977097600698471,
-0.03729447349905968,
-0.06156931072473526,
-0.036263082176446915,
0.020851777866482735,
0.06179456412792206,
0.017111705616116524,
0.03363565728068352,
-0.04501088336110115,
0.06390351057052612,
-0.021696824580430984,
-0.007689965423196554,
0.05168522894382477,
0.01362413726747036,
0.05018192529678345,
-0.015595987439155579,
0.04448143392801285,
-0.06826313585042953,
0.028877150267362595,
0.01806841418147087,
0.02234116569161415,
-0.005243188235908747,
0.028968701139092445,
0.021492043510079384,
0.09156402945518494,
-0.011842421256005764,
0.047352246940135956,
-0.06438451260328293,
-0.00826585665345192,
-0.06504888832569122,
0.016916634514927864,
0.01829340308904648,
0.05347086861729622,
0.10326867550611496,
0.03313983231782913,
-0.021723950281739235,
-0.03169660270214081,
0.018854646012187004,
0.03778434172272682,
0.07802141457796097,
-0.038287561386823654,
0.059806421399116516,
-0.011951938271522522,
-0.052237484604120255,
0.021807366982102394,
0.06865081936120987,
0.015515328384935856,
0.05150220915675163,
0.024626409634947777,
0.06752756237983704,
-0.007201419211924076,
0.06306695938110352,
-0.010281304828822613,
0.05317395552992821,
-0.03764082491397858,
-0.037435103207826614,
-0.09718386828899384,
0.09135860204696655,
0.015040195547044277,
0.06487960368394852,
0.04878616705536842,
-0.04470284283161163,
-0.0711175948381424,
-0.03422028571367264,
-0.034119438380002975,
8.458519279096949e-35,
0.04248207062482834,
0.010900034569203854,
-0.036025624722242355,
0.06228276342153549,
-0.023694681003689766,
0.01057281531393528,
-0.0012694887118414044,
0.04029087349772453,
-0.013217249885201454,
-0.057876717299222946,
-0.02499961294233799,
0.051044903695583344,
-0.06652015447616577,
0.0087830675765872,
0.029154641553759575,
0.05858690291643143,
-0.11504058539867401,
0.08125710487365723,
0.02650122530758381,
-0.0005061080446466804,
0.04465537890791893,
-0.010965022258460522,
0.03978275880217552,
-0.06111924350261688,
-0.011464656330645084,
0.011448539793491364,
0.006074887700378895,
0.004822746384888887,
-0.0726810172200203,
0.008045405149459839,
-0.07467661798000336,
-0.046967748552560806,
0.02199840173125267,
-0.03505510464310646,
-0.009578881785273552,
-0.04000140726566315,
0.01837662048637867,
-0.03170236945152283,
0.055682308971881866,
-0.0764424055814743,
0.01550859771668911,
0.07586986571550369,
0.055442072451114655,
-0.04887119308114052,
-0.07645027339458466,
-0.06467761844396591,
0.07784730195999146,
-0.03736903890967369,
-0.05205657333135605,
-0.012611298821866512,
0.024719174951314926,
0.011362518183887005,
-0.020618261769413948,
0.004448446445167065,
-0.03213926777243614,
0.04723808541893959,
-0.005169144831597805,
0.00443493016064167,
-0.051642585545778275,
0.030417731031775475,
-0.017854738980531693,
0.06713230162858963,
0.07540875673294067,
0.07507186383008957,
0.10435492545366287,
0.036442387849092484,
-0.031550221145153046,
0.04159113019704819,
0.03603445738554001,
-0.044225480407476425,
-0.01814824901521206,
-0.14284275472164154,
-0.08992280811071396,
-0.013326921500265598,
0.038029443472623825,
-0.03304509073495865,
-0.004695962183177471,
-0.11418391764163971,
0.03245284780859947,
0.019608914852142334,
0.030236460268497467,
-0.1255643516778946,
0.006079739425331354,
-0.05634542182087898,
0.007760736625641584,
0.023469028994441032,
0.060659702867269516,
-0.11958040297031403,
-0.05552873760461807,
0.01472546998411417,
0.020719250664114952,
-0.05533662438392639,
0.016707856208086014,
0.02586541697382927,
0.0018961969763040543,
-2.532758699190087e-33,
-0.06677470356225967,
-0.07476095110177994,
-0.03345240652561188,
0.03610190004110336,
-0.05135578662157059,
-0.02016402967274189,
0.04668131843209267,
0.05700758099555969,
-0.12517856061458588,
-0.10635779798030853,
0.04015982523560524,
-0.042541779577732086,
0.010924860835075378,
0.08064202964305878,
-0.02339348755776882,
-0.005838614422827959,
0.0016445015789940953,
0.0627153143286705,
0.05075393244624138,
0.04873786121606827,
-0.0026169498451054096,
0.02197064273059368,
-0.12671424448490143,
0.06087205931544304,
-0.034497953951358795,
0.010306427255272865,
-0.07164136320352554,
0.08649830520153046,
-0.04641793295741081,
0.017611632123589516,
-0.01134055107831955,
-0.0048582484014332294,
0.019186239689588547,
0.0108274444937706,
-0.04827675223350525,
0.009932354092597961,
0.019108382984995842,
-0.01962106302380562,
-0.04737158492207527,
0.01522172149270773,
0.06026332825422287,
0.017001386731863022,
-0.007697544991970062,
-0.09719018638134003,
0.003009666223078966,
-0.045332346111536026,
-0.06615132838487625,
0.007054808083921671,
0.0747489407658577,
0.010393315926194191,
0.08646020293235779,
-0.01932387799024582,
-0.026617662981152534,
-0.024358585476875305,
-0.07831110805273056,
-0.034023698419332504,
-0.0508984811604023,
-0.02803378365933895,
0.06988485902547836,
0.04037369787693024,
-0.03558866307139397,
-0.0051800585351884365,
-0.030261941254138947,
-0.06396348774433136,
0.02373342029750347,
0.010879027657210827,
-0.018743282184004784,
0.11456748098134995,
-0.030466509982943535,
-0.013636340387165546,
0.03146811202168465,
0.06783249229192734,
-0.037692148238420486,
-0.010633121244609356,
-0.013929806649684906,
0.03187239542603493,
-0.033465854823589325,
-0.01997905783355236,
-0.06732700020074844,
-0.04910770431160927,
-0.11415911465883255,
0.07401499897241592,
-0.038350459188222885,
0.00010237461538054049,
0.06477732211351395,
0.04012807086110115,
-0.008953351527452469,
-0.006074418313801289,
0.026793647557497025,
-0.011954407207667828,
-0.015738029032945633,
-0.04222901910543442,
-0.04102475568652153,
0.017346058040857315,
-0.11496548354625702,
-4.170785317114678e-8,
-0.02955799363553524,
0.03747639060020447,
0.05804385244846344,
0.07394923269748688,
0.09012603759765625,
-0.04916486144065857,
-0.09030702710151672,
0.09406213462352753,
-0.010528960265219212,
-0.059760309755802155,
0.05575326457619667,
0.08189522475004196,
-0.023695122450590134,
-0.0017052418552339077,
0.053289130330085754,
0.06421706825494766,
-0.0007182995905168355,
0.051488060504198074,
-0.017885755747556686,
-0.0903124287724495,
0.08934695273637772,
-0.09326891601085663,
-0.0036586730275303125,
-0.004965038970112801,
0.02876075729727745,
-0.07774136960506439,
0.010440665297210217,
0.086800716817379,
0.05983626842498779,
-0.028057705610990524,
-0.0036449464969336987,
-0.03713632747530937,
-0.05032804235816002,
-0.02543044649064541,
0.02204722911119461,
0.04647974669933319,
0.02254582941532135,
-0.10439296066761017,
-0.007106693461537361,
-0.03177756816148758,
0.037222038954496384,
-0.007757443469017744,
0.018816027790308,
0.049026161432266235,
-0.01856384240090847,
0.014019432477653027,
-0.08969122916460037,
-0.026258418336510658,
0.03732086345553398,
-0.04683860391378403,
0.052033860236406326,
0.04745595529675484,
-0.010444887913763523,
0.043071068823337555,
0.10372968018054962,
-0.01922355778515339,
-0.029488455504179,
0.04536687582731247,
-0.022922001779079437,
-0.045940376818180084,
0.05014064908027649,
-0.015312258154153824,
-0.03767950087785721,
0.009599466808140278
] |
Diffusion-Based Planning for Autonomous Driving with Flexible Guidance | https://openreview.net/forum?id=wM2sfVgMDH | [
"Yinan Zheng",
"Ruiming Liang",
"Kexin ZHENG",
"Jinliang Zheng",
"Liyuan Mao",
"Jianxiong Li",
"Weihao Gu",
"Rui Ai",
"Shengbo Eben Li",
"Xianyuan Zhan",
"Jingjing Liu"
] | Oral | Achieving human-like driving behaviors in complex open-world environments is a critical challenge in autonomous driving. Contemporary learning-based planning approaches such as imitation learning methods often struggle to balance competing objectives and lack of safety assurance,due to limited adaptability and inadequacy in learning complex multi-modal behaviors commonly exhibited in human planning, not to mention their strong reliance on the fallback strategy with predefined rules. We propose a novel transformer-based Diffusion Planner for closed-loop planning, which can effectively model multi-modal driving behavior and ensure trajectory quality without any rule-based refinement. Our model supports joint modeling of both prediction and planning tasks under the same architecture, enabling cooperative behaviors between vehicles. Moreover, by learning the gradient of the trajectory score function and employing a flexible classifier guidance mechanism, Diffusion Planner effectively achieves safe and adaptable planning behaviors. Evaluations on the large-scale real-world autonomous planning benchmark nuPlan and our newly collected 200-hour delivery-vehicle driving dataset demonstrate that Diffusion Planner achieves state-of-the-art closed-loop performance with robust transferability in diverse driving styles. | diffusion planning, autonomous driving | null | 13,578 | 2501.15564 | [
0.003697252133861184,
-0.10823868960142136,
-0.02056341990828514,
0.062148094177246094,
0.029695900157094002,
-0.01762736774981022,
-0.07689015567302704,
-0.017527373507618904,
-0.031918615102767944,
-0.024829700589179993,
-0.011157851666212082,
-0.010659144259989262,
-0.0022969350684434175,
0.0876505970954895,
-0.02208862081170082,
0.02836238034069538,
0.02051248587667942,
0.07812369614839554,
-0.0950474813580513,
-0.01679859310388565,
0.04620331525802612,
0.05310111120343208,
0.022362619638442993,
0.03828243538737297,
-0.041849005967378616,
0.06215052306652069,
0.04574749618768692,
-0.07395917177200317,
-0.023648692294955254,
-0.032472316175699234,
0.048829585313797,
-0.09431859105825424,
-0.07106059044599533,
-0.0892653614282608,
-0.06011739745736122,
0.023004606366157532,
-0.07546715438365936,
-0.03717895597219467,
-0.008139122277498245,
-0.07191678881645203,
0.02204248681664467,
-0.039415374398231506,
0.003190292278304696,
0.052706122398376465,
0.011558427475392818,
0.057601798325777054,
-0.026438036933541298,
-0.03795884549617767,
0.05552756413817406,
-0.022590871900320053,
-0.0007042994839139283,
0.010570762678980827,
0.018817171454429626,
-0.005813186522573233,
-0.0543096549808979,
0.041516680270433426,
0.007890025153756142,
0.01888490840792656,
0.03516672924160957,
-0.05705581605434418,
-0.07272341102361679,
-0.13841207325458527,
0.037900716066360474,
-0.008969418704509735,
-0.13026116788387299,
-0.026764987036585808,
0.000021489144273800775,
0.02061625011265278,
-0.0011262726038694382,
0.044404495507478714,
-0.011265810579061508,
0.005381599068641663,
0.03360565006732941,
0.029021428897976875,
0.08792886137962341,
0.012580583803355694,
0.05145648121833801,
0.018523959442973137,
-0.025115618482232094,
-0.025651177391409874,
0.025078825652599335,
0.006080877035856247,
-0.005144212860614061,
0.06012939289212227,
0.10935679823160172,
-0.0532563254237175,
-0.04446821287274361,
0.09207666665315628,
0.06638169288635254,
0.013908280059695244,
-0.045350100845098495,
0.0537223145365715,
0.01242181658744812,
-0.09484994411468506,
-0.04242439195513725,
0.018419068306684494,
-0.07813304662704468,
-0.0744464248418808,
-0.016027646139264107,
0.037072695791721344,
-0.05851857364177704,
0.07890675216913223,
-0.00983008835464716,
0.0065574636682868,
0.02162938378751278,
-0.05188070237636566,
0.06503855437040329,
-0.011380216106772423,
0.01755351759493351,
-0.006846049800515175,
0.07645867764949799,
-0.03525353595614433,
0.11202467978000641,
0.002050015376880765,
0.07371159642934799,
-0.010768902488052845,
-0.036947522312402725,
0.052415721118450165,
0.003827109234407544,
0.060898616909980774,
-0.05868184566497803,
-0.0556640625,
0.06326806545257568,
0.03863752633333206,
0.03320474550127983,
-0.029939964413642883,
0.008565220050513744,
3.3710399957824426e-34,
-0.03781629726290703,
0.0259487796574831,
0.06367902457714081,
0.06954868137836456,
0.046853404492139816,
-0.05995786190032959,
-0.023954829201102257,
-0.022952044382691383,
-0.060496736317873,
-0.02522556483745575,
0.022642523050308228,
-0.03277791664004326,
-0.08711501955986023,
0.1284809112548828,
0.08452931046485901,
-0.009419839829206467,
-0.010836165398359299,
0.03366747871041298,
-0.03667530417442322,
-0.03870217874646187,
0.0716104805469513,
-0.047225940972566605,
-0.01270955428481102,
-0.072953000664711,
0.0341041125357151,
0.02591339498758316,
0.028484510257840157,
0.017541464418172836,
0.09360875189304352,
-0.00027387088630348444,
-0.11405503749847412,
0.038301922380924225,
-0.12380450963973999,
0.06943035870790482,
0.05638071522116661,
0.09229151159524918,
-0.11855713278055191,
-0.04968627914786339,
-0.023223862051963806,
0.04743150249123573,
0.026372000575065613,
-0.025966878980398178,
-0.0592724047601223,
0.019613318145275116,
-0.08060012012720108,
-0.05259367823600769,
0.029598264023661613,
0.050944987684488297,
-0.06156935915350914,
-0.034758321940898895,
-0.03322898596525192,
-0.053177010267972946,
-0.03903045505285263,
-0.11233232170343399,
-0.0664723739027977,
-0.027964558452367783,
0.04251332953572273,
0.0003896227281074971,
0.005524915177375078,
0.015415248461067677,
0.031761493533849716,
0.01753646694123745,
-0.057074934244155884,
0.019214002415537834,
-0.004680546931922436,
0.016008779406547546,
-0.045599956065416336,
0.08525825291872025,
0.067392498254776,
-0.021564921364188194,
0.045908182859420776,
0.07092395424842834,
-0.006031631026417017,
-0.03497368097305298,
0.08141898363828659,
-0.013751777820289135,
0.04848719388246536,
-0.07198375463485718,
0.0028431511018425226,
0.03785703331232071,
-0.04189430922269821,
0.024809712544083595,
-0.032185375690460205,
-0.05593516677618027,
0.04719084873795509,
-0.037478186190128326,
-0.008274514228105545,
-0.020458729937672615,
-0.09796300530433655,
-0.00022188459115568548,
-0.07638412714004517,
-0.035127829760313034,
-0.06292266398668289,
0.08787443488836288,
-0.02418135479092598,
-1.8123408232413243e-33,
0.032460667192935944,
-0.037498366087675095,
0.08060164749622345,
0.06671592593193054,
0.0450032614171505,
-0.029187791049480438,
-0.02295038104057312,
-0.04174766689538956,
0.0978076383471489,
-0.06539510190486908,
-0.10126267373561859,
0.007137827109545469,
0.1318243443965912,
-0.008725285530090332,
0.047076266258955,
-0.00064138974994421,
0.01797327771782875,
0.03028135932981968,
-0.0402863547205925,
-0.011544816195964813,
-0.03276708722114563,
0.03587646409869194,
-0.09325253963470459,
0.023636173456907272,
-0.08155543357133865,
0.0202193520963192,
-0.06485610455274582,
0.0309559628367424,
0.0484490841627121,
0.03722577542066574,
-0.013747396878898144,
-0.04450602829456329,
0.03565410152077675,
-0.0639071837067604,
-0.02644522860646248,
0.09911182522773743,
-0.012934703379869461,
0.00869464036077261,
-0.04075231775641441,
0.07138140499591827,
0.010538759641349316,
-0.09169556200504303,
-0.029145710170269012,
0.0026014361064881086,
0.07595497369766235,
-0.01012678537517786,
-0.011442264541983604,
0.0345904603600502,
-0.03766149282455444,
0.05795963481068611,
0.008674168027937412,
0.06597775965929031,
-0.15399311482906342,
0.0039704362861812115,
-0.0052356659434735775,
-0.002962934086099267,
0.0006576835876330733,
-0.014114730060100555,
0.09983376413583755,
0.0005494041834026575,
0.004080283455550671,
0.05598675459623337,
0.000005179411800781963,
0.039145976305007935,
-0.049637891352176666,
-0.03485206887125969,
0.0031691871117800474,
-0.009734854102134705,
-0.0032747499644756317,
0.05333730950951576,
0.010742169804871082,
0.033431775867938995,
0.02083824761211872,
-0.061687614768743515,
0.025691533461213112,
-0.11794523894786835,
0.042502034455537796,
0.00004760727460961789,
0.010407705791294575,
-0.012911004945635796,
0.05222412943840027,
0.016303876414895058,
0.029168616980314255,
0.06201164424419403,
-0.020295189693570137,
0.02817249484360218,
0.03209083154797554,
0.06741740554571152,
-0.0005794705357402563,
0.08365628868341446,
0.003876021131873131,
0.0679858848452568,
0.0011684941127896309,
0.06736248731613159,
-0.10339473932981491,
-4.112620999308092e-8,
-0.061932455748319626,
0.08293631672859192,
0.02479282207787037,
-0.004538041539490223,
0.014893591403961182,
-0.01162210013717413,
0.04982949048280716,
-0.03939471021294594,
-0.023182718083262444,
0.04017521068453789,
0.016393613070249557,
0.01902705617249012,
0.04530578851699829,
-0.038245897740125656,
-0.07725536823272705,
0.06545659899711609,
0.04074975848197937,
-0.025441849604249,
0.020609073340892792,
0.03170950338244438,
-0.005569905508309603,
-0.020873364061117172,
-0.0017484222771599889,
-0.004756043665111065,
0.0699431523680687,
-0.06985708326101303,
-0.10611837357282639,
0.021627141162753105,
0.0697949156165123,
-0.007491736672818661,
-0.033468835055828094,
0.021385278552770615,
0.05995950847864151,
0.02067585662007332,
0.0025124766398221254,
0.044417791068553925,
0.05606279522180557,
0.00011802970402641222,
0.025121532380580902,
-0.03418353945016861,
0.03786340728402138,
0.08468367904424667,
-0.004417477175593376,
0.026482751592993736,
-0.029170172289013863,
0.000987249775789678,
-0.0294797345995903,
-0.08869195729494095,
-0.019608356058597565,
0.005071757826954126,
-0.016258660703897476,
-0.060701124370098114,
-0.0013622178230434656,
0.07427587360143661,
0.12926161289215088,
-0.0009081162279471755,
-0.013041005469858646,
-0.08373763412237167,
0.0785457044839859,
0.05697117745876312,
-0.02160721644759178,
-0.021019352599978447,
-0.036012060940265656,
0.03394530341029167
] |
Learning to Search from Demonstration Sequences | https://openreview.net/forum?id=v593OaNePQ | [
"Dixant Mittal",
"Liwei Kang",
"Wee Sun Lee"
] | Oral | Search and planning are essential for solving many real-world problems. However, in numerous learning scenarios, only action-observation sequences, such as demonstrations or instruction sequences, are available for learning. Relying solely on supervised learning with these sequences can lead to sub-optimal performance due to the vast, unseen search space encountered during training. In this paper, we introduce Differentiable Tree Search Network (D-TSN), a novel neural network architecture that learns to construct search trees from just sequences of demonstrations by performing gradient descent on a best-first search tree construction algorithm. D-TSN enables the joint learning of submodules, including an encoder, value function, and world model, which are essential for planning. To construct the search tree, we employ a stochastic tree expansion policy and formulate it as another decision-making task. Then, we optimize the tree expansion policy via REINFORCE with an effective variance reduction technique for the gradient computation. D-TSN can be applied to problems with a known world model or to scenarios where it needs to jointly learn a world model with a latent state space. We study problems from these two scenarios, including Game of 24, 2D grid navigation, and Procgen games, to understand when D-TSN is more helpful. Through our experiments, we show that D-TSN is effective, especially when the world model with a latent state space is jointly learned. The code is available at https://github.com/dixantmittal/differentiable-tree-search-network. | planning, reasoning, learning to search, reinforcement learning, large language model | We propose a method that constructs search tree in a differetiable manner, and can be trained from just demonstration sequences. | 13,425 | null | [
-0.05436089262366295,
-0.08911566436290741,
0.03811679035425186,
-0.0037279443349689245,
0.04582573473453522,
0.03448101133108139,
-0.045912884175777435,
-0.009847477078437805,
-0.018625015392899513,
0.011606894433498383,
-0.02260192297399044,
-0.011133828200399876,
0.018584217876195908,
0.03556773439049721,
-0.04892896115779877,
-0.0012914796825498343,
0.049932438880205154,
0.09062157571315765,
-0.027760911732912064,
-0.059517137706279755,
0.02678314410150051,
0.04378955438733101,
0.08395090699195862,
0.0017579426057636738,
-0.026551702991127968,
0.02787257917225361,
0.04196157678961754,
-0.005860460456460714,
0.0011797689367085695,
0.02821863815188408,
0.0722292959690094,
-0.07439140230417252,
-0.02092195488512516,
-0.02392646297812462,
-0.07946130633354187,
0.0632888600230217,
-0.1494670957326889,
-0.013859489932656288,
-0.006790167652070522,
-0.04670794680714607,
0.020814688876271248,
0.08040761947631836,
-0.00999187771230936,
-0.010961356572806835,
0.0883699506521225,
0.017299896106123924,
0.008418482728302479,
-0.07002469152212143,
0.05962934345006943,
0.039620622992515564,
-0.02607847936451435,
-0.0733858272433281,
-0.024814153090119362,
-0.07638541609048843,
0.03790249302983284,
0.007920038886368275,
0.032907288521528244,
-0.016092514619231224,
0.0448472797870636,
0.012584008276462555,
0.050024330615997314,
-0.11988919973373413,
0.02630765177309513,
-0.10516186058521271,
-0.014632673002779484,
-0.045340538024902344,
0.004657864570617676,
0.05431986227631569,
0.034431472420692444,
-0.02316189743578434,
0.015833061188459396,
0.07238912582397461,
-0.021229838952422142,
0.05473106727004051,
0.010295272804796696,
0.017846709117293358,
0.013824369758367538,
-0.058130212128162384,
0.015845485031604767,
-0.03241267055273056,
-0.004350865259766579,
0.029419733211398125,
-0.0005818353383801877,
0.0006692350725643337,
0.07360081374645233,
-0.022093450650572777,
-0.0029722130857408047,
0.09960608184337616,
0.06146214157342911,
-0.05487995222210884,
-0.001973342150449753,
-0.07418755441904068,
-0.05054944381117821,
-0.009396144188940525,
-0.05079396069049835,
-0.020788423717021942,
-0.008642594330012798,
-0.09016814827919006,
-0.025020917877554893,
0.051678434014320374,
-0.020938044413924217,
0.08848803490400314,
0.0536736436188221,
-0.01927260495722294,
-0.007060197181999683,
0.017871104180812836,
0.020417673513293266,
0.014425442554056644,
0.004630906507372856,
-0.08085516095161438,
-0.017170829698443413,
-0.018492182716727257,
0.018327634781599045,
0.0226628165692091,
-0.023714471608400345,
-0.04140346124768257,
0.008312027901411057,
-0.051995739340782166,
-0.05814497917890549,
0.09833353757858276,
-0.007241829298436642,
-0.03395204246044159,
-0.013840468600392342,
0.08005064725875854,
0.03377806022763252,
-0.06354031711816788,
-0.07960968464612961,
-1.0295129216360332e-33,
0.08694125711917877,
-0.01825433038175106,
0.09292170405387878,
0.05629558861255646,
0.12669867277145386,
0.05759580060839653,
0.10508524626493454,
0.03368522599339485,
-0.02228919416666031,
-0.00932870339602232,
-0.06133512407541275,
-0.05565076321363449,
-0.08098389208316803,
0.03399726375937462,
0.10911836475133896,
-0.027179380878806114,
-0.050498947501182556,
-0.026775425300002098,
-0.023815926164388657,
0.005062735639512539,
0.0478103905916214,
-0.038954269140958786,
0.015035359188914299,
-0.043235328048467636,
0.06170782819390297,
-0.024383433163166046,
0.050569068640470505,
0.015588943846523762,
-0.0013456451706588268,
-0.02968461997807026,
-0.061418239027261734,
-0.02775089628994465,
-0.0976283848285675,
0.0926370620727539,
0.09037086367607117,
0.03724722936749458,
-0.017941435799002647,
-0.013899305835366249,
-0.001627233694307506,
-0.040989477187395096,
0.02496691793203354,
0.00909725297242403,
0.008712800219655037,
0.02122107706964016,
-0.07575265318155289,
-0.07923534512519836,
0.04349711909890175,
0.0020687999203801155,
0.009505932219326496,
-0.02529899589717388,
-0.0030787314753979445,
0.10767704993486404,
-0.02746090106666088,
-0.16321969032287598,
0.004020138178020716,
0.03613332286477089,
0.059959735721349716,
0.06762944161891937,
0.03988635540008545,
0.0734032616019249,
0.010861346498131752,
-0.010479403659701347,
0.04101331904530525,
0.12271340936422348,
0.010239917784929276,
0.10291746258735657,
-0.06845536082983017,
0.06605837494134903,
0.07728175818920135,
-0.07064393162727356,
0.0031076534651219845,
0.012685418128967285,
0.038898877799510956,
-0.061160385608673096,
0.062306150794029236,
0.029763242229819298,
0.02750958688557148,
-0.10303593426942825,
-0.041401881724596024,
-0.03320738300681114,
-0.024169176816940308,
0.013035597279667854,
-0.03828499838709831,
0.03047688864171505,
0.01142844557762146,
-0.05123644322156906,
0.038114603608846664,
-0.02473776787519455,
0.030827097594738007,
-0.017780965194106102,
-0.10865994542837143,
-0.03313680365681648,
-0.029338067397475243,
0.025232462212443352,
0.07726320624351501,
-7.995683186934429e-34,
0.0005927528836764395,
0.0033742033410817385,
-0.014990067109465599,
-0.01632588915526867,
-0.015159854665398598,
-0.030098039656877518,
-0.0557192787528038,
-0.0657304897904396,
-0.022859834134578705,
0.020519478246569633,
-0.06460387259721756,
0.016055434942245483,
0.06588592380285263,
-0.006894032936543226,
-0.047909725457429886,
0.020196929574012756,
0.013254905119538307,
0.03962322324514389,
0.03360264375805855,
-0.010637476108968258,
-0.014422920532524586,
0.1396266669034958,
-0.11361494660377502,
0.012324053794145584,
-0.04771653562784195,
-0.015057891607284546,
0.051284901797771454,
0.1001024916768074,
-0.012061347253620625,
0.04638640955090523,
-0.046321287751197815,
-0.02645070292055607,
-0.030153894796967506,
0.025362765416502953,
-0.058214861899614334,
0.03464334458112717,
0.06123043969273567,
-0.01637618988752365,
-0.07059631496667862,
0.1096934974193573,
0.020435882732272148,
-0.07323937118053436,
-0.0018714877078309655,
0.01070678886026144,
-0.018711630254983902,
-0.01661112904548645,
-0.03240952640771866,
0.03250518813729286,
-0.005975463893264532,
0.019674867391586304,
-0.028004774823784828,
0.03196960687637329,
-0.09701641649007797,
-0.05273997038602829,
-0.06087517365813255,
-0.029439697042107582,
-0.024291155859827995,
0.04558834433555603,
0.10707099735736847,
0.0015306673012673855,
-0.03259199112653732,
0.011754235252737999,
-0.03044840693473816,
0.007980127818882465,
-0.06857093423604965,
-0.02048693783581257,
-0.052451081573963165,
0.01914391666650772,
0.08685947954654694,
-0.053464170545339584,
-0.00026276885182596743,
0.03179752081632614,
0.045822013169527054,
0.0042400117963552475,
0.04364727810025215,
-0.056035224348306656,
-0.009372917003929615,
0.025342458859086037,
0.009896664880216122,
-0.02052619308233261,
0.007522898726165295,
-0.01397579163312912,
0.0396377295255661,
0.03550335764884949,
0.07449595630168915,
0.07720660418272018,
-0.022321395576000214,
0.12797275185585022,
-0.003900841809809208,
-0.014487276785075665,
0.005186732392758131,
0.00299362582154572,
0.030930057168006897,
-0.006227993872016668,
-0.049605756998062134,
-4.1235459491417714e-8,
-0.11447504907846451,
0.07760727405548096,
0.11203347146511078,
-0.006322092842310667,
0.08956750482320786,
-0.041278380900621414,
0.0287720188498497,
0.028593674302101135,
0.02352803573012352,
-0.02914329804480076,
0.0707610696554184,
0.020191660150885582,
-0.002169843530282378,
-0.05118067190051079,
-0.018956614658236504,
0.04404541850090027,
-0.002207950223237276,
-0.06790366768836975,
0.010066737420856953,
0.07488804310560226,
0.013439161702990532,
-0.008714177645742893,
0.010519232600927353,
0.06271584331989288,
-0.0012956567807123065,
-0.06292012333869934,
-0.016597658395767212,
0.05211160331964493,
0.08526485413312912,
0.01196080818772316,
0.058862004429101944,
-0.01663936860859394,
0.05481066554784775,
-0.026723986491560936,
0.011791830882430077,
0.04601578414440155,
0.053333040326833725,
-0.024113425984978676,
-0.0037133207079023123,
-0.04624607786536217,
-0.04065883159637451,
-0.008582101203501225,
-0.013304110616445541,
-0.013015715405344963,
-0.07504349946975708,
0.005682011134922504,
-0.07178003340959549,
-0.12867923080921173,
0.031131833791732788,
-0.10132003575563431,
0.0161671731621027,
-0.02740650251507759,
-0.03307101130485535,
0.0406690314412117,
0.12539081275463104,
0.061416711658239365,
-0.012000003829598427,
-0.10039268434047699,
-0.012408042326569557,
0.09505825489759445,
-0.012603974901139736,
-0.03408791497349739,
-0.0669403001666069,
-0.005084553267806768
] |
Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse | https://openreview.net/forum?id=Iyrtb9EJBp | [
"Maojia Song",
"Shang Hong Sim",
"Rishabh Bhardwaj",
"Hai Leong Chieu",
"Navonil Majumder",
"Soujanya Poria"
] | Oral | LLMs are an integral component of retrieval-augmented generation (RAG) systems. While many studies focus on evaluating the overall quality of end-to-end RAG systems, there is a gap in understanding the appropriateness of LLMs for the RAG task. To address this, we introduce Trust-Score, a holistic metric that evaluates the trustworthiness of LLMs within the RAG framework. Our results show that various prompting methods, such as in-context learning, fail to effectively adapt LLMs to the RAG task as measured by Trust-Score. Consequently, we propose Trust-Align, a method to align LLMs for improved Trust-Score performance. 26 out of 27 models aligned using Trust-Align substantially outperform competitive baselines on ASQA, QAMPARI, and ELI5. Specifically, in LLaMA-3-8b, Trust-Align outperforms FRONT on ASQA (↑12.56), QAMPARI (↑36.04), and ELI5 (↑17.69). Trust-Align also significantly enhances models’ ability to correctly refuse and provide quality citations. We also demonstrate the effectiveness of Trust-Align across different open-weight models, including the LLaMA series (1b to 8b), Qwen-2.5 series (0.5b to 7b), and Phi3.5 (3.8b). We release our code at https://github.com/declare-lab/trust-align. | Large Language Models, Trustworthiness, Hallucinations, Retrieval Augmented Generation | How to better evaluate and make LLM better for RAG task | 13,377 | 2409.11242 | [
-0.1142926886677742,
-0.029256857931613922,
-0.024079805240035057,
-0.007724442984908819,
-0.020343737676739693,
-0.005263214465230703,
0.03379476070404053,
-0.011266677640378475,
0.10013871639966965,
-0.02107248827815056,
0.00016581999079789966,
-0.03880157321691513,
0.14433534443378448,
-0.0016660091932862997,
-0.020799309015274048,
0.0015177298337221146,
0.10930971056222916,
0.057754985988140106,
-0.051697712391614914,
-0.021370695903897285,
-0.027299556881189346,
0.04216717928647995,
0.044234227389097214,
0.005011976230889559,
-0.007978462614119053,
-0.05005970224738121,
-0.10115396976470947,
0.048353686928749084,
-0.04381611570715904,
-0.12408392131328583,
0.023161709308624268,
0.05294979363679886,
0.027514353394508362,
-0.029212387278676033,
-0.07756318897008896,
0.0016510016284883022,
-0.07258579134941101,
0.04546921327710152,
-0.009209522977471352,
-0.08446019887924194,
-0.056918974965810776,
0.010180443525314331,
0.058569151908159256,
-0.019516771659255028,
0.01809283159673214,
-0.0360950231552124,
0.016835037618875504,
-0.02906942553818226,
-0.02218153327703476,
0.005682162009179592,
-0.10761646926403046,
-0.039037156850099564,
0.010002908296883106,
0.046239223331213,
-0.11417555809020996,
0.052893657237291336,
0.03214804828166962,
0.03535089269280434,
-0.08228469640016556,
-0.031526144593954086,
-0.04837225750088692,
-0.08486828207969666,
-0.12564343214035034,
0.05290914699435234,
0.0165217537432909,
0.030732538551092148,
-0.029694324359297752,
0.029375528916716576,
-0.05169248580932617,
-0.03905738517642021,
-0.03611471876502037,
-0.010969183407723904,
0.004663501400500536,
0.06551072746515274,
0.0013792812824249268,
0.07089220732450485,
0.02039610594511032,
-0.02311457321047783,
0.07314909249544144,
-0.0021170873660594225,
-0.03547251597046852,
0.049892179667949677,
0.05057191476225853,
0.04041316360235214,
-0.003327709622681141,
-0.03310779109597206,
0.03598133847117424,
0.015088275074958801,
0.022345004603266716,
-0.00013982030213810503,
0.04409505054354668,
-0.02997273951768875,
0.02284783311188221,
0.009251061826944351,
0.08763400465250015,
0.052641090005636215,
0.017817115411162376,
-0.048346422612667084,
-0.005968024488538504,
0.05834158509969711,
0.020660365000367165,
0.09998057037591934,
-0.04051045700907707,
-0.12590795755386353,
0.05244159325957298,
-0.0056677707470953465,
-0.003176379017531872,
-0.021342014893889427,
0.055697645992040634,
-0.055038709193468094,
-0.04659518599510193,
0.04486136883497238,
0.03393789008259773,
0.0010598069056868553,
0.037710536271333694,
0.037609100341796875,
0.03985157608985901,
0.00923420675098896,
0.03657399117946625,
-0.02864445373415947,
-0.011274962686002254,
0.02161041833460331,
0.10746306926012039,
-0.010206151753664017,
-0.03522557020187378,
-0.016492675989866257,
0.0217644851654768,
9.67086181919204e-34,
0.03900153562426567,
0.05241309478878975,
0.012759667821228504,
0.021542364731431007,
0.02946089766919613,
-0.013600165955722332,
-0.01507909782230854,
0.07343712449073792,
-0.06286611407995224,
-0.02365097403526306,
-0.04803411662578583,
0.0896238386631012,
-0.0016611069440841675,
0.0049582114443182945,
0.09060359001159668,
0.06549552083015442,
-0.07894168049097061,
0.027216902002692223,
-0.03722098842263222,
-0.024496372789144516,
0.06271247565746307,
-0.08743851631879807,
0.0454593151807785,
-0.029444674029946327,
0.030642598867416382,
-0.008908843621611595,
0.08398053795099258,
0.09345069527626038,
-0.04330406337976456,
0.0027951428201049566,
-0.06227855011820793,
-0.0711129680275917,
-0.017570074647665024,
-0.036266595125198364,
0.007659340742975473,
0.021741623058915138,
-0.07215176522731781,
-0.002433827379718423,
0.022557830438017845,
-0.08638536930084229,
0.024373618885874748,
-0.015139354392886162,
0.011476082727313042,
0.03866461291909218,
-0.06309931725263596,
-0.015936942771077156,
0.0012374701909720898,
-0.0684053972363472,
0.01649452932178974,
0.07242506742477417,
0.03965609893202782,
0.0025622746907174587,
0.004529811907559633,
0.028761934489011765,
-0.1328476071357727,
0.042770933359861374,
0.02520148828625679,
0.05381134897470474,
-0.07408319413661957,
0.009355641901493073,
0.06802359223365784,
0.0128425732254982,
-0.0371873639523983,
-0.0022919492330402136,
0.015037670731544495,
0.0018058997811749578,
0.047172971069812775,
-0.00986165925860405,
0.026940036565065384,
0.005824402440339327,
0.0012097569415345788,
-0.01343310996890068,
-0.07762985676527023,
-0.015572513453662395,
0.015407943166792393,
-0.07862699031829834,
0.023369284346699715,
0.018758026883006096,
0.12247053533792496,
-0.07035943865776062,
0.0503142885863781,
-0.012650229036808014,
-0.08732104301452637,
-0.0014822370139881968,
-0.045129477977752686,
-0.09372231364250183,
0.02419114112854004,
-0.08029349893331528,
-0.05550885200500488,
0.03798164054751396,
0.04538312554359436,
0.03224044665694237,
-0.030012452974915504,
0.12642909586429596,
-0.006047294475138187,
-5.4350331292061985e-34,
0.014880231581628323,
-0.007167437579482794,
0.044353898614645004,
0.11939916759729385,
-0.008115683682262897,
-0.031494688242673874,
0.0012996387667953968,
0.030913302674889565,
-0.025926686823368073,
0.056931670755147934,
-0.0012050805380567908,
-0.0001268876512767747,
0.030959971249103546,
-0.009676746092736721,
0.011780503205955029,
-0.0593574084341526,
-0.017179884016513824,
0.02098177559673786,
-0.0487372949719429,
-0.034267958253622055,
0.0868537649512291,
0.08106657862663269,
-0.08731281757354736,
0.04179731383919716,
-0.009767457842826843,
0.034704744815826416,
-0.008499853312969208,
-0.06184878200292587,
-0.01714242808520794,
0.003542680758982897,
0.0444597564637661,
0.031005650758743286,
-0.04135910049080849,
-0.0002737885806709528,
-0.028363192453980446,
0.010491212829947472,
0.05590507760643959,
0.04261024668812752,
-0.03764742612838745,
0.061978086829185486,
0.04031366854906082,
0.02165108360350132,
-0.0962805300951004,
-0.02894127368927002,
-0.007236158009618521,
-0.07079567015171051,
-0.07495532929897308,
-0.07488483935594559,
0.007566374726593494,
0.0025538576301187277,
0.043783802539110184,
-0.01352331880480051,
-0.03519297018647194,
0.010940518230199814,
-0.033116091042757034,
-0.04017264395952225,
-0.03282159939408302,
-0.006203095428645611,
-0.014422147534787655,
0.05841313675045967,
0.0012799275573343039,
0.025059698149561882,
-0.042498450726270676,
-0.025326518341898918,
0.039673659950494766,
-0.03158416971564293,
0.02084173634648323,
0.06618507206439972,
-0.07636071741580963,
0.049620386213064194,
0.06628313660621643,
0.007319370750337839,
0.029937397688627243,
-0.0851055309176445,
0.036051489412784576,
-0.026456449180841446,
0.0008809569990262389,
-0.06368137896060944,
-0.012799831107258797,
-0.1044614315032959,
0.05049724131822586,
-0.009498922154307365,
0.058330606669187546,
0.09592252224683762,
-0.029884876683354378,
0.007701433729380369,
0.10160011053085327,
0.02282794378697872,
-0.03151881694793701,
0.0626201257109642,
0.02337399311363697,
0.007370291277766228,
0.004057751037180424,
0.019681228324770927,
-0.05458167567849159,
-4.651232643482217e-8,
0.005395245272666216,
-0.025525368750095367,
-0.03753213956952095,
0.11321210116147995,
-0.0102910865098238,
0.06467949599027634,
-0.07002408057451248,
0.042193055152893066,
-0.06046977639198303,
0.0074885268695652485,
0.045941028743982315,
-0.05348728224635124,
-0.12786060571670532,
-0.0012416810495778918,
0.03682689368724823,
0.009479129686951637,
0.024722786620259285,
0.03483660891652107,
-0.04758057743310928,
-0.027867142111063004,
0.06713981926441193,
0.07207916676998138,
-0.005827878601849079,
-0.025150805711746216,
0.06669674813747406,
-0.023061545565724373,
-0.023425325751304626,
0.1380050778388977,
0.0948137491941452,
0.04042467102408409,
0.039368242025375366,
-0.04794475436210632,
-0.006512864027172327,
-0.07970590144395828,
-0.06234200298786163,
0.09458550065755844,
-0.0033263012301176786,
-0.0857810527086258,
0.06640986353158951,
0.06789221614599228,
0.010725433938205242,
0.07685764133930206,
-0.04662512615323067,
0.035582561045885086,
-0.046912383288145065,
-0.01834905333817005,
-0.03345601633191109,
-0.040534134954214096,
-0.028667056933045387,
-0.03729845955967903,
0.041118115186691284,
-0.13700108230113983,
0.012221138924360275,
0.05749037116765976,
-0.03311698138713837,
-0.00805674958974123,
0.03994234651327133,
-0.040042802691459656,
0.0854845717549324,
0.06616416573524475,
0.12281598895788193,
-0.061150938272476196,
-0.013602741062641144,
0.05928908288478851
] |
MAP: Multi-Human-Value Alignment Palette | https://openreview.net/forum?id=NN6QHwgRrQ | [
"Xinran Wang",
"Qi Le",
"Ammar Ahmed",
"Enmao Diao",
"Yi Zhou",
"Nathalie Baracaldo",
"Jie Ding",
"Ali Anwar"
] | Oral | Ensuring that generative AI systems align with human values is essential but challenging, especially when considering multiple human values and their potential trade-offs. Since human values can be personalized and dynamically change over time, the desirable levels of value alignment vary across different ethnic groups, industry sectors, and user cohorts. Within existing frameworks, it is hard to define human values and align AI systems accordingly across different directions simultaneously, such as harmlessness, helpfulness, and positiveness. To address this, we develop a novel, first-principle approach called Multi-Human-Value Alignment Palette (MAP), which navigates the alignment across multiple human values in a structured and reliable way. MAP formulates the alignment problem as an optimization task with user-defined constraints, which define human value targets. It can be efficiently solved via a primal-dual approach, which determines whether a user-defined alignment target is achievable and how to achieve it. We conduct a detailed theoretical analysis of MAP by quantifying the trade-offs between values, the sensitivity to constraints, the fundamental connection between multi-value alignment and sequential alignment, and proving that linear weighted rewards are sufficient for multi-value alignment. Extensive experiments demonstrate MAP's ability to align multiple values in a principled manner while delivering strong empirical performance across various tasks. | Human value alignment, Generative model | The paper introduces Multi-Human-Value Alignment Palette (MAP), a novel approach to align generative models with multiple human values in a principled way. | 13,248 | 2410.19198 | [
-0.02129107341170311,
0.006724673323333263,
-0.036662619560956955,
-0.06650504469871521,
-0.016665300354361534,
0.04870571568608284,
0.0426325760781765,
0.0040106638334691525,
0.013574006967246532,
-0.04951547458767891,
-0.03328584507107735,
-0.14591453969478607,
0.00524666765704751,
0.01733270473778248,
-0.006228989455848932,
-0.03271418437361717,
0.039924465119838715,
0.08093008399009705,
-0.08692418783903122,
-0.046044930815696716,
-0.04939596354961395,
0.013410818763077259,
0.02349863387644291,
0.004889071919023991,
-0.045523013919591904,
0.028344018384814262,
0.030748888850212097,
-0.0486469566822052,
-0.04799681156873703,
-0.016168983653187752,
0.08302046358585358,
0.05390040948987007,
0.08214324712753296,
0.022049281746149063,
-0.04670223966240883,
0.06623214483261108,
-0.13349489867687225,
0.023150406777858734,
-0.007147734984755516,
0.005580556113272905,
-0.019978463649749756,
-0.04764330014586449,
-0.05774388462305069,
-0.010538813658058643,
0.04074660316109657,
0.014984662644565105,
0.03868359327316284,
0.006791308987885714,
0.042059868574142456,
0.06127219647169113,
-0.09070605039596558,
0.02672099508345127,
-0.00004991168680135161,
0.04636187478899956,
0.01713014580309391,
0.059630636125802994,
0.05547112226486206,
-0.004204222001135349,
0.01794276386499405,
-0.013332460075616837,
-0.01656537689268589,
-0.1295432448387146,
0.07198169827461243,
0.03927617147564888,
0.07656268030405045,
-0.02681087702512741,
-0.009615511633455753,
0.025805892422795296,
-0.05980921909213066,
0.014757102355360985,
-0.002261207439005375,
0.05121353641152382,
-0.003158364910632372,
0.013437249697744846,
0.006109095178544521,
0.10851174592971802,
0.03146110847592354,
-0.06495027989149094,
0.040139760822057724,
-0.049197807908058167,
0.041946616023778915,
0.033995386213064194,
0.05205737426877022,
0.04479355365037918,
0.061085496097803116,
-0.01660149358212948,
0.002715479116886854,
-0.0054846713319420815,
0.0613103061914444,
0.058256596326828,
-0.1343502700328827,
-0.02910921722650528,
0.1325138658285141,
0.021417124196887016,
0.0564584918320179,
0.035979680716991425,
0.0037752820644527674,
-0.028201881796121597,
-0.014179359190165997,
0.007626941427588463,
-0.03595056012272835,
0.04203621298074722,
0.0158320851624012,
-0.07052117586135864,
-0.007947857491672039,
-0.10927283018827438,
0.000997432740405202,
0.015098519623279572,
-0.0033171705435961485,
-0.08690685778856277,
0.03844670578837395,
-0.022798892110586166,
0.029723694548010826,
0.04370107129216194,
-0.007062055170536041,
-0.05087430030107498,
0.015308341942727566,
0.06032831221818924,
0.011486254632472992,
0.046956807374954224,
-0.04980142414569855,
-0.047494400292634964,
-0.01064618956297636,
0.014911109581589699,
-0.020823296159505844,
-0.02607039175927639,
-0.047520045191049576,
9.726072401127262e-35,
-0.008955658413469791,
0.05847754329442978,
0.11114270985126495,
0.05730884522199631,
0.04820583760738373,
-0.06259553879499435,
-0.08882301300764084,
0.0001925412507262081,
-0.06638912856578827,
0.0728125348687172,
-0.0168793722987175,
0.06382158398628235,
-0.089976467192173,
0.1552092432975769,
-0.038403742015361786,
0.03549107909202576,
-0.05349429324269295,
0.06587033718824387,
-0.10874978452920914,
0.01999375782907009,
0.051341548562049866,
-0.04327089712023735,
-0.04870712012052536,
-0.024798326194286346,
0.09581441432237625,
-0.01920313388109207,
0.0012021854054182768,
-0.013151642866432667,
0.002651504473760724,
-0.043859440833330154,
-0.0653897151350975,
0.045507512986660004,
0.04602998495101929,
-0.0012180147459730506,
0.036856744438409805,
-0.03673723340034485,
-0.04966085031628609,
-0.06212248653173447,
-0.04369984567165375,
0.08952020108699799,
-0.013151237741112709,
0.060524746775627136,
0.03222915157675743,
0.004000676795840263,
-0.027263252064585686,
-0.08314263075590134,
0.027316084131598473,
0.045144036412239075,
-0.03771062567830086,
-0.038590215146541595,
-0.0009724470437504351,
0.06947816908359528,
-0.0024970879312604666,
-0.04381182789802551,
-0.10471490770578384,
-0.050590477883815765,
-0.05339132621884346,
-0.042815208435058594,
-0.03293822705745697,
0.004216183442622423,
-0.042824871838092804,
-0.012623217888176441,
-0.02625325135886669,
0.02181178517639637,
0.007456978317350149,
-0.006230222526937723,
0.04259485751390457,
0.061729345470666885,
0.0370146743953228,
0.0703316181898117,
0.0060389358550310135,
0.01981321908533573,
-0.00850353017449379,
-0.03512921929359436,
-0.02029222808778286,
0.0885297879576683,
0.07971442490816116,
-0.0766763836145401,
0.03534959629178047,
-0.04804534837603569,
-0.0703040063381195,
0.0578894279897213,
-0.04162199795246124,
-0.013706939294934273,
-0.018308749422430992,
-0.06639842689037323,
0.043325942009687424,
-0.0022060545161366463,
-0.024074051529169083,
0.012756035663187504,
-0.042342424392700195,
0.07268190383911133,
-0.062456630170345306,
0.03608310595154762,
0.0016228965250775218,
-1.5705240028619174e-33,
-0.036593616008758545,
-0.059928689152002335,
0.03984663262963295,
0.028291314840316772,
0.05785452574491501,
-0.004652784671634436,
0.07807783782482147,
-0.04554563760757446,
0.0038197606336325407,
0.03573980927467346,
-0.02819223515689373,
-0.0352257676422596,
0.04789689928293228,
-0.0157184936106205,
-0.0021268920972943306,
0.018974723294377327,
-0.05369894951581955,
0.0799771249294281,
0.03848125785589218,
0.010626074858009815,
0.09561838209629059,
0.09702018648386002,
-0.11906467378139496,
0.02290147729218006,
-0.019513269886374474,
0.03921349346637726,
0.004813373554497957,
-0.04650699347257614,
-0.00024132286489475518,
0.030455129221081734,
-0.02071497030556202,
-0.01876058429479599,
-0.05655205249786377,
-0.043986380100250244,
-0.021739860996603966,
0.04843175783753395,
0.04579062759876251,
-0.04337937757372856,
-0.020079536363482475,
0.1303236335515976,
-0.04672246053814888,
-0.021985473111271858,
-0.10277800261974335,
-0.033296339213848114,
-0.007535816635936499,
0.010071518830955029,
0.02334165759384632,
-0.041904740035533905,
-0.04607459902763367,
-0.0010895244777202606,
-0.015887202695012093,
0.012562774121761322,
-0.0716819241642952,
-0.02638549916446209,
-0.048753585666418076,
-0.07096774876117706,
-0.04149003326892853,
-0.04400826618075371,
0.05858452245593071,
-0.013034975156188011,
-0.052295032888650894,
0.0056517343036830425,
0.07825905829668045,
0.06659644097089767,
-0.0596344992518425,
0.006562643684446812,
0.09100838750600815,
-0.037948641926050186,
-0.07708229124546051,
-0.06630324572324753,
0.0546053983271122,
-0.07319720834493637,
0.05968448892235756,
0.02582879178225994,
-0.03910373896360397,
-0.06469815224409103,
-0.014956314116716385,
-0.021321186795830727,
0.018765173852443695,
-0.07296425104141235,
0.05499622970819473,
-0.021978028118610382,
0.04702722653746605,
0.0525662899017334,
-0.017950061708688736,
0.06675063818693161,
0.07064662873744965,
0.10909063369035721,
0.05895223468542099,
0.054905571043491364,
-0.07781370729207993,
0.01817425899207592,
-0.10718289017677307,
0.017681142315268517,
-0.1393839567899704,
-4.511710116616996e-8,
-0.05650080367922783,
0.03341488912701607,
0.027488505467772484,
0.09513099491596222,
-0.058873292058706284,
-0.004356559831649065,
-0.008095195516943932,
0.04916217178106308,
0.030625998973846436,
0.0011589224450290203,
-0.0024872450157999992,
0.011462929658591747,
0.001546695944853127,
-0.055386804044246674,
0.02455468475818634,
0.01881292089819908,
-0.02501983940601349,
0.031772855669260025,
-0.0039427150040864944,
0.04645472392439842,
0.03418530151247978,
0.0794709101319313,
-0.03055410273373127,
0.0290845837444067,
0.04525275155901909,
-0.056016333401203156,
-0.025207936763763428,
0.07430250197649002,
0.024789312854409218,
-0.04079163074493408,
-0.004515918903052807,
-0.004824281204491854,
0.03748592361807823,
0.016867388039827347,
-0.035672709345817566,
-0.01442739274352789,
-0.02846619114279747,
0.009858107194304466,
-0.014246919192373753,
0.0033475509844720364,
0.006804904900491238,
0.125315323472023,
-0.033869244158267975,
0.017640693113207817,
0.03543391823768616,
-0.03286374360322952,
-0.042783889919519424,
0.025951247662305832,
-0.02495875582098961,
-0.025574123486876488,
0.014893397688865662,
-0.023476049304008484,
0.08223778754472733,
0.09910350292921066,
0.05102201923727989,
-0.06044148653745651,
-0.0192220751196146,
-0.006403820123523474,
0.0992412194609642,
0.02757864259183407,
0.037082817405462265,
-0.07452407479286194,
-0.10821931809186935,
-0.10414355248212814
] |
Can Neural Networks Achieve Optimal Computational-statistical Tradeoff? An Analysis on Single-Index Model | https://openreview.net/forum?id=is4nCVkSFA | [
"Siyu Chen",
"Beining Wu",
"Miao Lu",
"Zhuoran Yang",
"Tianhao Wang"
] | Oral | In this work, we tackle the following question: Can neural networks trained with gradient-based methods achieve the optimal statistical-computational tradeoff in learning Gaussian single-index models?
Prior research has shown that any polynomial-time algorithm under the statistical query (SQ) framework requires $\Omega(d^{s^\star/2}\lor d)$ samples, where $s^\star$ is the generative exponent representing the intrinsic difficulty of learning the underlying model.
However, it remains unknown whether neural networks can achieve this sample complexity.
Inspired by prior techniques such as label transformation and landscape smoothing for learning single-index models, we propose a unified gradient-based algorithm for training a two-layer neural network in polynomial time.
Our method is adaptable to a variety of loss and activation functions, covering a broad class of existing approaches.
We show that our algorithm learns a feature representation that strongly aligns with the unknown signal $\theta^\star$, with sample complexity $\tilde O (d^{s^\star/2} \lor d)$, matching the SQ lower bound up to a polylogarithmic factor for all generative exponents $s^\star\geq 1$.
Furthermore, we extend our approach to the setting where $\theta^\star$ is $k$-sparse for $k = o(\sqrt{d})$ by introducing a novel weight perturbation technique that leverages the sparsity structure.
We derive a corresponding SQ lower bound
of order $\tilde\Omega(k^{s^\star})$, matched by our method up to a polylogarithmic factor.
Our framework, especially the weight perturbation technique, is of independent interest, and suggests potential gradient-based solutions to other problems such as sparse tensor PCA. | single-index model, feature learning, gradient-based method, computational-statistical tradeoff | We propose a unified gradient-based algorithm for feature learning in Gaussian single-index model with sample complexity matching the SQ lower bound | 13,084 | null | [
-0.10704405605792999,
-0.10548006743192673,
0.07906994968652725,
0.09473977982997894,
0.049008872359991074,
0.04187753424048424,
-0.027378041297197342,
-0.022082936018705368,
0.02989555150270462,
-0.05141766369342804,
-0.0634932816028595,
0.028668809682130814,
0.016813315451145172,
-0.0023766341619193554,
-0.06366283446550369,
0.036958057433366776,
0.042583491653203964,
0.11605282127857208,
-0.02434444986283779,
0.0694914236664772,
0.04761455953121185,
0.049562718719244,
-0.011937925592064857,
-0.04677136242389679,
0.0686735063791275,
-0.0670471265912056,
-0.0017197374254465103,
0.015279727056622505,
0.004749240819364786,
0.04297221451997757,
0.031249050050973892,
-0.0015969054074957967,
-0.008800528012216091,
0.03334784880280495,
-0.07044217735528946,
0.011653981171548367,
-0.018575921654701233,
0.07442483305931091,
0.009865306317806244,
0.020992115139961243,
0.13258132338523865,
0.053350355476140976,
0.02830628491938114,
-0.0027403556741774082,
0.055889323353767395,
0.01570715382695198,
0.008211696520447731,
-0.06515710800886154,
-0.003750977572053671,
0.0049644047394394875,
-0.02120470628142357,
-0.014210651628673077,
-0.06058685481548309,
0.008865232579410076,
0.032050251960754395,
0.02025013230741024,
0.019246090203523636,
0.020730052143335342,
-0.05392415076494217,
0.0341520756483078,
0.015577046200633049,
-0.10693207383155823,
-0.02736249938607216,
-0.019275086000561714,
0.023090265691280365,
-0.004008805844932795,
0.04740297794342041,
-0.02121756598353386,
-0.0334230400621891,
0.03322821110486984,
-0.054118167608976364,
0.13593244552612305,
-0.1486450433731079,
0.06738723069429398,
0.03943086043000221,
-0.04214118793606758,
0.07340998202562332,
-0.02848454937338829,
0.031144684180617332,
-0.024540986865758896,
-0.008275165222585201,
-0.012954946607351303,
0.039161041378974915,
0.002762601710855961,
0.06361131370067596,
-0.046067025512456894,
0.019287146627902985,
0.03089221939444542,
0.055161021649837494,
-0.027373535558581352,
-0.06444872915744781,
-0.0034019446466118097,
-0.04555610939860344,
-0.0931435227394104,
0.03861265257000923,
-0.009975109249353409,
0.03728973865509033,
-0.024215541779994965,
-0.04509470611810684,
0.09890035539865494,
-0.049950018525123596,
0.03220265358686447,
-0.02550235018134117,
-0.10121529549360275,
0.04881598800420761,
-0.0006343248533084989,
0.013134305365383625,
0.008943645283579826,
0.025792952626943588,
-0.1079484298825264,
0.026387574151158333,
0.0024134223349392414,
-0.02625054121017456,
0.006547253578901291,
-0.013633922673761845,
0.025749508291482925,
0.01981453411281109,
-0.010940766893327236,
-0.0370713472366333,
0.08671905100345612,
-0.10443448275327682,
0.025345677509903908,
-0.02969033643603325,
0.011226596310734749,
-0.015666544437408447,
0.01270719338208437,
-0.08853388577699661,
1.3872382604523178e-33,
0.04430054500699043,
0.07426007837057114,
0.05168881267309189,
-0.019559763371944427,
0.04075406864285469,
0.043260037899017334,
0.0486595556139946,
-0.014811067841947079,
0.08032076805830002,
0.06758549809455872,
-0.02295777015388012,
0.04439174756407738,
-0.038659464567899704,
0.01508252788335085,
0.11130661517381668,
0.03842359408736229,
0.021052904427051544,
-0.009536635130643845,
0.06142723187804222,
-0.06108425185084343,
0.08457805961370468,
-0.06663752347230911,
-0.017019592225551605,
-0.014819477684795856,
-0.060134775936603546,
-0.015666497871279716,
0.026448845863342285,
-0.02573259174823761,
-0.007596984505653381,
-0.003958525601774454,
-0.09097549319267273,
-0.04703541100025177,
-0.027695942670106888,
0.05017745494842529,
0.10285311192274094,
-0.046734027564525604,
0.03437067195773125,
0.01705802045762539,
0.013607123866677284,
-0.006932066287845373,
-0.04926428571343422,
0.0007230595219880342,
0.06299297511577606,
-0.023702431470155716,
-0.1283051073551178,
-0.10729672014713287,
0.0022786823101341724,
-0.05486680939793587,
0.02493554726243019,
-0.05518209934234619,
-0.010401851497590542,
0.013703473843634129,
-0.06402511894702911,
-0.035570163279771805,
0.010353786870837212,
0.024012863636016846,
0.01413679402321577,
0.023397063836455345,
0.025476275011897087,
0.09258098155260086,
0.012825575657188892,
-0.021028731018304825,
0.013039390556514263,
0.026131005957722664,
0.00917823612689972,
0.06820867210626602,
-0.004232567735016346,
0.025227993726730347,
0.03350498154759407,
-0.0797576829791069,
0.03611701726913452,
-0.028466399759054184,
-0.038900427520275116,
-0.14823195338249207,
0.11138558387756348,
0.023429201915860176,
0.06493701040744781,
-0.11228170990943909,
0.03237982094287872,
0.01922978274524212,
-0.04678750038146973,
0.015313073061406612,
0.008656475692987442,
-0.059413522481918335,
-0.019571030512452126,
-0.06835316121578217,
0.018471600487828255,
-0.024737779051065445,
-0.0120770875364542,
-0.04245133325457573,
-0.15904508531093597,
0.0004042494110763073,
-0.0506175272166729,
-0.08308234810829163,
0.006171367596834898,
-1.1809849196310853e-33,
-0.09396673738956451,
-0.009316309355199337,
0.025848273187875748,
0.04812374338507652,
-0.012021517381072044,
-0.04034655913710594,
-0.0613718256354332,
0.013536478392779827,
0.025749919936060905,
-0.023488430306315422,
-0.013446620665490627,
-0.02300717867910862,
0.06928115338087082,
-0.030657533556222916,
0.03459387645125389,
0.06584680080413818,
-0.08473667502403259,
0.041282329708337784,
-0.007864751853048801,
0.01857880689203739,
0.0010183471022173762,
0.07608502358198166,
-0.07862408459186554,
0.07625651359558105,
-0.043216437101364136,
0.025751035660505295,
-0.043402183800935745,
0.11257689446210861,
-0.03577679768204689,
0.03784318268299103,
-0.0822935402393341,
-0.05675121769309044,
-0.02452794834971428,
-0.013851160183548927,
-0.04779452085494995,
0.11848556995391846,
0.01217795629054308,
0.015256245620548725,
-0.09988843649625778,
0.11790215224027634,
0.022185083478689194,
-0.054522931575775146,
-0.0018257254268974066,
0.0651201382279396,
0.06009341776371002,
-0.07494774460792542,
-0.045330241322517395,
-0.008689194917678833,
0.004143102094531059,
0.04407598078250885,
0.001962523441761732,
0.0602981336414814,
-0.06085474416613579,
0.07431050390005112,
-0.05642304569482803,
-0.045922257006168365,
-0.07078877836465836,
0.049787722527980804,
0.04813588038086891,
-0.022430885583162308,
-0.012368062511086464,
-0.013989265076816082,
-0.09134282916784286,
0.034223929047584534,
-0.006242384668439627,
-0.017683392390608788,
0.06489057838916779,
0.0010583120165392756,
0.02069876343011856,
0.01622149720788002,
0.038016948848962784,
-0.034200165420770645,
0.002984297927469015,
-0.0022663739509880543,
-0.06823999434709549,
0.02935228869318962,
0.05999749153852463,
-0.006470886524766684,
-0.06304555386304855,
-0.0024626809172332287,
0.03234567865729332,
0.029439518228173256,
0.02537197805941105,
-0.04782060906291008,
0.13654470443725586,
0.02333342656493187,
0.07013963907957077,
0.041272539645433426,
0.031255125999450684,
-0.04849506914615631,
-0.02958831936120987,
0.030261065810918808,
0.00540927005931735,
0.018435079604387283,
-0.06075827777385712,
-4.962097222005468e-8,
-0.062334902584552765,
0.0662287175655365,
0.07903875410556793,
-0.009308651089668274,
0.10612297058105469,
-0.06301066279411316,
0.03628307953476906,
0.07786710560321808,
-0.09782488644123077,
-0.027229920029640198,
0.030008289963006973,
0.058270324021577835,
-0.01395320612937212,
-0.033202677965164185,
0.034998197108507156,
0.0620274692773819,
-0.013492189347743988,
0.026695607230067253,
-0.015547843649983406,
-0.004300840198993683,
-0.04386305809020996,
0.025503961369395256,
0.05453699827194214,
-0.050499942153692245,
0.014820141717791557,
-0.08866509795188904,
-0.024193745106458664,
0.010731359012424946,
0.1007642149925232,
0.04903234541416168,
-0.026030205190181732,
0.016315778717398643,
-0.005647669546306133,
-0.010861987248063087,
0.020797692239284515,
-0.0004022412176709622,
0.05715818703174591,
-0.029141314327716827,
-0.029783688485622406,
-0.04716631770133972,
-0.06362513452768326,
0.08646751195192337,
-0.02898404560983181,
0.025797007605433464,
0.03676925599575043,
0.01653062365949154,
0.049020491540431976,
-0.07522419840097427,
0.07003145664930344,
0.03258715197443962,
0.03095538541674614,
0.021992303431034088,
-0.021641166880726814,
0.025285571813583374,
0.019776776432991028,
0.04396224021911621,
-0.022897085174918175,
-0.047430120408535004,
-0.04645855724811554,
0.055766187608242035,
-0.02083517611026764,
-0.0014958210522308946,
-0.017761679366230965,
-0.03409094735980034
] |
Consistency Checks for Language Model Forecasters | https://openreview.net/forum?id=r5IXBlTCGc | [
"Daniel Paleka",
"Abhimanyu Pallavi Sudhir",
"Alejandro Alvarez",
"Vineeth Bhat",
"Adam Shen",
"Evan Wang",
"Florian Tramèr"
] | Oral | Forecasting is a task that is difficult to evaluate: the ground truth can only be known in the future. Recent work showing LLM forecasters rapidly approaching human-level performance begs the question: how can we benchmark and evaluate these forecasters *instantaneously*? Following the consistency check framework, we measure the performance of forecasters in terms of the consistency of their predictions on different logically-related questions. We propose a new, general consistency metric based on *arbitrage*: for example, if a forecasting AI illogically predicts that both the Democratic and Republican parties have 60\% probability of winning the 2024 US presidential election, an arbitrageur could trade against the forecaster's predictions and make a profit. We build an automated evaluation system that generates a set of base questions, instantiates consistency checks from these questions, elicits the predictions of the forecaster, and measures the consistency of the predictions. We then build a standard, proper-scoring-rule forecasting benchmark, and show that our (instantaneous) consistency metrics correlate strongly with LLM forecasters' ground truth Brier scores (which are only known in the future). We also release a consistency benchmark that resolves in 2028, providing a long-term evaluation tool for forecasting. | forecasting, markets, trading, LLM, evaluation, eval, consistency, robustness | It is difficult to evaluate AI forecasters instantaneously; we propose market-based consistency evals on LLM forecasters and show plenty of inconsistency. | 13,065 | 2412.18544 | [
-0.07717876136302948,
-0.09865415841341019,
-0.04396665096282959,
0.050274740904569626,
0.033055178821086884,
0.021076753735542297,
0.012797188013792038,
-0.0023351050913333893,
0.012750214897096157,
-0.03271811455488205,
-0.11732664704322815,
-0.07693871855735779,
0.03273208066821098,
0.06341730058193207,
-0.07595538347959518,
-0.03417639434337616,
-0.0009381765848957002,
-0.0005169090582057834,
-0.01711329258978367,
-0.0158456489443779,
-0.04303869977593422,
0.07085566967725754,
-0.015499956905841827,
-0.010627186857163906,
0.10690893977880478,
-0.0360359363257885,
0.027482671663165092,
-0.017367200925946236,
-0.02867681346833706,
-0.005633111577481031,
0.0024342185351997614,
-0.06323596090078354,
-0.03239874914288521,
0.015591596253216267,
-0.007978247478604317,
-0.0018556436989456415,
-0.03788624703884125,
-0.019163258373737335,
0.07229862362146378,
0.013789628632366657,
0.007981732487678528,
-0.1276286244392395,
-0.032267726957798004,
-0.043014463037252426,
0.017012247815728188,
0.060037482529878616,
0.030543740838766098,
0.0542132742702961,
-0.06188509613275528,
0.0057987188920378685,
-0.07727093994617462,
-0.0035071992315351963,
0.0396551787853241,
-0.017647001892328262,
-0.006869229953736067,
0.07392969727516174,
-0.022672457620501518,
0.03966154530644417,
0.0184446033090353,
-0.04946274310350418,
-0.07311108708381653,
-0.10627443343400955,
-0.045652035623788834,
0.027754411101341248,
-0.006138766184449196,
-0.0008787065744400024,
0.02727101743221283,
0.04251347482204437,
-0.050789833068847656,
0.06421121209859848,
-0.010046199895441532,
0.07400799542665482,
-0.019963867962360382,
0.05439549311995506,
-0.006860758643597364,
0.04370502755045891,
0.0624416209757328,
0.018342399969697,
0.032270193099975586,
0.0155618442222476,
-0.0549483448266983,
-0.018557757139205933,
0.025935977697372437,
-0.005343941040337086,
0.09694220125675201,
0.013873567804694176,
0.0553487166762352,
0.057135552167892456,
0.05553244799375534,
0.005976450629532337,
-0.04717578738927841,
-0.042119383811950684,
0.05983426794409752,
-0.010500072501599789,
0.08571946620941162,
0.11192329972982407,
-0.0008851381717249751,
-0.08879116177558899,
0.028812073171138763,
0.07810832560062408,
0.007906033657491207,
0.07676725089550018,
-0.045624058693647385,
-0.06530826538801193,
0.05371652916073799,
-0.0031932543497532606,
-0.023439113050699234,
-0.011830274015665054,
-0.0037616894114762545,
-0.0942741185426712,
-0.009257460944354534,
0.013279956765472889,
0.09107456356287003,
-0.018320659175515175,
0.01370879914611578,
-0.022987492382526398,
0.0019307377515360713,
0.029721738770604134,
0.018292520195245743,
0.04120499640703201,
-0.04615961015224457,
0.050662267953157425,
0.11868400126695633,
0.0048128836788237095,
0.1196364015340805,
0.007338536903262138,
0.012059269472956657,
1.4608029009786843e-34,
-0.0006399865378625691,
0.018370870500802994,
0.03701912611722946,
-0.025865059345960617,
-0.08126156032085419,
0.04961540922522545,
-0.06838793307542801,
0.03920002654194832,
-0.017336275428533554,
-0.030502712354063988,
-0.052103884518146515,
0.05204727500677109,
-0.0860557109117508,
0.04369467496871948,
0.059393204748630524,
0.048034168779850006,
-0.020726585760712624,
0.07235140353441238,
-0.021736958995461464,
-0.015236873179674149,
0.06108178198337555,
-0.00878225453197956,
-0.0034803743474185467,
-0.08088800311088562,
0.006386623252183199,
-0.011530444025993347,
0.10195041447877884,
-0.021732816472649574,
0.0057601165026426315,
-0.015219446271657944,
-0.05445156246423721,
-0.05487745255231857,
0.043501488864421844,
0.01202795933932066,
-0.002730943262577057,
-0.02036929503083229,
-0.07235980033874512,
-0.004215598106384277,
-0.0501178614795208,
-0.02594747580587864,
-0.019585877656936646,
0.07227113097906113,
-0.02711508981883526,
-0.03512847423553467,
-0.0661919042468071,
-0.04745102673768997,
-0.05760100483894348,
-0.027726657688617706,
0.019085457548499107,
-0.07600482553243637,
0.0792151540517807,
0.027450555935502052,
0.010183354839682579,
-0.01932142674922943,
-0.016600364819169044,
-0.01865704171359539,
-0.030040815472602844,
-0.0047992211766541,
-0.034683723002672195,
0.01857997477054596,
0.025429317727684975,
0.013618727214634418,
0.023401539772748947,
-0.01956687867641449,
-0.073502317070961,
0.11161160469055176,
-0.07119899243116379,
0.08486604690551758,
0.032752349972724915,
-0.05549360811710358,
0.04735517129302025,
-0.06749070435762405,
0.05102480947971344,
-0.07030097395181656,
0.033260416239500046,
0.03931866213679314,
0.0023994725197553635,
-0.06740890443325043,
0.05073482170701027,
0.010489794425666332,
0.0103059783577919,
-0.023946838453412056,
0.022150123491883278,
-0.03597880154848099,
-0.030546842142939568,
-0.06424848735332489,
0.05140576511621475,
0.009518788196146488,
-0.0215219147503376,
-0.017536837607622147,
-0.10052558034658432,
-0.05465834587812424,
0.004763639532029629,
0.015536211431026459,
0.019613754004240036,
-2.5371085793011064e-33,
-0.03911428153514862,
-0.05728454142808914,
-0.04279418662190437,
0.09870091825723648,
0.0027850959450006485,
-0.06527045369148254,
-0.08251979947090149,
0.029514215886592865,
0.02179303765296936,
-0.05755281448364258,
-0.05787321552634239,
0.004585015121847391,
0.07439061999320984,
0.020141780376434326,
-0.04890580475330353,
-0.07331916689872742,
0.01989472284913063,
-0.041362203657627106,
-0.005778727121651173,
0.01869978941977024,
0.059346817433834076,
0.03661879152059555,
-0.15310852229595184,
0.07663335651159286,
-0.011066307313740253,
-0.0030634773429483175,
0.01838182844221592,
0.02181714028120041,
-0.07012198120355606,
0.029641732573509216,
0.019936908036470413,
-0.09131139516830444,
0.013080629520118237,
-0.02372371219098568,
0.0015718203503638506,
0.013659077696502209,
0.07522591203451157,
-0.06441089510917664,
-0.04940328747034073,
0.09701383113861084,
0.05759658291935921,
-0.002714287955313921,
-0.05591723695397377,
-0.07471346855163574,
-0.01611146330833435,
0.034758225083351135,
-0.028519002720713615,
0.009541189298033714,
-0.005065438337624073,
0.03788673132658005,
0.010437860153615475,
0.06472668796777725,
-0.07476360350847244,
0.06127936765551567,
-0.03858489170670509,
-0.04191378504037857,
-0.12206856161355972,
-0.045998744666576385,
0.008131690323352814,
0.08348434418439865,
-0.09185269474983215,
0.027061158791184425,
0.037869010120630264,
0.01154713612049818,
0.04373832792043686,
-0.006790291517972946,
-0.03815951943397522,
-0.03666592389345169,
0.07520949095487595,
0.001638590358197689,
0.044942546635866165,
0.0010983317624777555,
-0.06280539184808731,
0.082534559071064,
0.09143798053264618,
0.06131208688020706,
0.04973074048757553,
-0.007866551168262959,
0.061234183609485626,
-0.03113839589059353,
-0.05923345685005188,
0.011307802982628345,
0.053755633533000946,
0.0026697267312556505,
-0.018874382600188255,
-0.0031557627953588963,
0.10627324879169464,
-0.003943708259612322,
0.03792644664645195,
0.04066106304526329,
-0.025769419968128204,
-0.019845185801386833,
-0.036408595740795135,
0.014694714918732643,
-0.08581748604774475,
-4.7294626881466684e-8,
-0.08090505003929138,
-0.005286868195980787,
-0.016797808930277824,
0.10615179687738419,
-0.04333700239658356,
-0.005328415427356958,
-0.014846092090010643,
-0.06697257608175278,
0.007352868095040321,
-0.08856704831123352,
0.10557720065116882,
-0.036084454506635666,
-0.07164330035448074,
-0.05518500506877899,
0.01722472906112671,
0.08164600282907486,
-0.046911757439374924,
0.044556718319654465,
-0.03722263127565384,
-0.042925719171762466,
0.0821547731757164,
0.061645373702049255,
-0.009703017771244049,
0.0016283518634736538,
0.04285312816500664,
0.01764838583767414,
-0.05647794529795647,
0.12228910624980927,
-0.01797671429812908,
0.09873531013727188,
0.014681229367852211,
0.026662053540349007,
0.04846145212650299,
-0.06089230626821518,
0.03281280770897865,
0.0892033651471138,
-0.06455048173666,
0.010055341757833958,
0.05703816935420036,
0.03333191201090813,
0.05771196261048317,
0.0490560308098793,
-0.059358227998018265,
-0.001012564986012876,
0.029820526018738747,
-0.07159221917390823,
-0.04768519848585129,
-0.0522589273750782,
0.02845895104110241,
-0.09236375242471695,
0.024974297732114792,
-0.047725845128297806,
0.010111304931342602,
0.04541483148932457,
0.13668344914913177,
-0.03358667716383934,
-0.027721965685486794,
-0.10302933305501938,
-0.009105454199016094,
0.015522093512117863,
0.0628703311085701,
0.011095089837908745,
-0.06008264794945717,
0.05295250192284584
] |
Spread Preference Annotation: Direct Preference Judgment for Efficient LLM Alignment | https://openreview.net/forum?id=BPgK5XW1Nb | [
"Dongyoung Kim",
"Kimin Lee",
"Jinwoo Shin",
"Jaehyung Kim"
] | Oral | Aligning large language models (LLMs) with human preferences becomes a key component to obtaining state-of-the-art performance, but it yields a huge cost to construct a large human-annotated preference dataset. To tackle this problem, we propose a new framework, Spread Preference Annotation with direct preference judgment (SPA), that boosts the alignment of LLMs using only a very small amount of human-annotated preference data.
Our key idea is leveraging the human prior knowledge within the small (seed) data and progressively improving the alignment of LLM, by iteratively generating the responses and learning from them with the self-annotated preference data.
To be specific, we propose to derive the preference label from the logits of LLM to explicitly extract the model's inherent preference.
Compared to the previous approaches using external reward models or implicit in-context learning, we observe that the proposed approach is significantly more effective.
In addition, we introduce a noise-aware preference learning algorithm to mitigate the risk of low quality within generated preference data.
Our experimental results demonstrate that the proposed framework significantly boosts the alignment of LLMs.
For example, we achieve superior alignment performance on AlpacaEval 2.0 with only 3.3% of the ground-truth preference labels in the Ultrafeedback data compared to the cases using the entire data or state-of-the-art baselines. | large language model, alignment, preference | null | 12,928 | 2406.04412 | [
-0.09205672889947891,
-0.08180037140846252,
-0.01853586733341217,
0.02387421391904354,
0.05137999355792999,
0.04248807579278946,
0.025502393022179604,
0.022205257788300514,
0.020403213798999786,
-0.004675958771258593,
0.011237965896725655,
-0.0845503956079483,
0.04652528092265129,
0.034931231290102005,
0.0635046437382698,
0.027142804116010666,
0.13603946566581726,
0.01198203582316637,
-0.07616935670375824,
-0.014841456897556782,
-0.03128090500831604,
0.03009866736829281,
0.0325806625187397,
0.010257862508296967,
0.019399849697947502,
0.001614580862224102,
0.028827093541622162,
0.025524422526359558,
0.007711997255682945,
-0.015768932178616524,
0.04889608174562454,
0.014347681775689125,
0.01468940544873476,
-0.010616404004395008,
-0.039001308381557465,
0.023675868287682533,
-0.10958357155323029,
-0.04462991654872894,
-0.0206155888736248,
-0.052916061133146286,
-0.053279466927051544,
-0.015491017140448093,
-0.0015304600819945335,
0.028194161131978035,
0.04414232075214386,
0.051574207842350006,
0.011706811375916004,
0.0015849339542910457,
0.04570290446281433,
0.07013364881277084,
-0.05817831680178642,
-0.012100487016141415,
0.044146474450826645,
0.06508785486221313,
-0.015873447060585022,
0.06442736089229584,
-0.00444508483633399,
0.04205950349569321,
-0.027740120887756348,
-0.050997160375118256,
-0.020445115864276886,
-0.14470520615577698,
-0.042120303958654404,
0.04457179456949234,
-0.0358569510281086,
-0.06594771891832352,
0.018715515732765198,
0.06105151027441025,
-0.07383169233798981,
0.010338065214455128,
0.017422357574105263,
0.12127654999494553,
0.04475995525717735,
0.06826983392238617,
0.026613876223564148,
0.07128426432609558,
-0.020509174093604088,
-0.04942886531352997,
0.041064079850912094,
-0.03133528679609299,
-0.06157608702778816,
0.0632542297244072,
0.02224160172045231,
0.009903433732688427,
0.0869351252913475,
-0.004348675720393658,
0.05826227366924286,
0.04249775409698486,
0.0065069361589848995,
0.03479611501097679,
-0.0786728635430336,
-0.04289628937840462,
0.014217646792531013,
0.017292248085141182,
-0.0009579138131812215,
0.10583788901567459,
-0.02714073657989502,
-0.06043309345841408,
-0.061700865626335144,
0.07996547967195511,
0.029110323637723923,
0.08264951407909393,
-0.1056610718369484,
-0.05499008297920227,
0.04104834794998169,
-0.02361585758626461,
-0.051669783890247345,
0.05207771807909012,
0.04441123083233833,
-0.1125035360455513,
0.02675994113087654,
0.06066874414682388,
0.0016926713287830353,
0.01352686621248722,
0.011585857719182968,
-0.05528303235769272,
0.06138744205236435,
-0.041994985193014145,
0.033560846000909805,
-0.01684003323316574,
-0.03449929505586624,
-0.027835262939333916,
-0.01048849057406187,
-0.021519554778933525,
-0.012757658958435059,
0.003132673678919673,
-0.028575116768479347,
-7.227809869359001e-34,
0.04916611313819885,
0.07309047132730484,
0.061930619180202484,
-0.014925582334399223,
0.01134925801306963,
-0.03775642439723015,
-0.024680951610207558,
0.062233295291662216,
-0.11607801914215088,
-0.031040437519550323,
0.002336133737117052,
-0.01489004585891962,
-0.0023817350156605244,
0.07933896034955978,
0.013720733113586903,
-0.04630756378173828,
-0.01882917992770672,
0.05612119287252426,
-0.027888750657439232,
-0.05182051658630371,
0.030865557491779327,
0.10890986770391464,
0.03462909907102585,
-0.054646264761686325,
0.05577574297785759,
0.003813215997070074,
0.06351327151060104,
-0.03926287591457367,
-0.05613038316369057,
0.010590292513370514,
-0.0442628413438797,
-0.07709119468927383,
0.07253697514533997,
-0.04910987615585327,
0.004191011190414429,
-0.02447841875255108,
-0.04864780232310295,
-0.029267942532896996,
0.005351648200303316,
-0.025460785254836082,
0.04458126425743103,
0.03353966772556305,
0.004434963222593069,
0.04613406956195831,
-0.07394890487194061,
0.01910107396543026,
-0.0019053808646276593,
0.04558290168642998,
0.0378430113196373,
-0.05566995590925217,
0.06712301820516586,
-0.02050141990184784,
-0.017480209469795227,
0.012421987019479275,
-0.0550614558160305,
-0.0013566287234425545,
-0.008348069153726101,
0.008698622696101665,
0.010245035402476788,
-0.05568861961364746,
-0.02957877516746521,
0.018385622650384903,
0.024139801040291786,
0.020849037915468216,
0.05034949630498886,
-0.024990426376461983,
0.0213541891425848,
0.021184369921684265,
0.0149614242836833,
-0.05984010919928551,
0.015285895206034184,
0.026081319898366928,
-0.05797034129500389,
-0.04949883744120598,
0.013211747631430626,
0.03588477522134781,
0.07166686654090881,
-0.05576441437005997,
0.09071029722690582,
-0.03559195622801781,
0.03751813992857933,
-0.06039753183722496,
0.014211287721991539,
-0.12920323014259338,
-0.024834169074892998,
-0.05007044970989227,
0.0476648323237896,
-0.07390875369310379,
-0.0007046584505587816,
0.022452088072896004,
-0.0057176752015948296,
0.03710653632879257,
-0.023897254839539528,
0.02010131999850273,
0.014631934463977814,
9.601927340710966e-34,
-0.06572575867176056,
-0.02225814014673233,
0.04768712818622589,
0.08391160517930984,
0.006747187580913305,
-0.015459316782653332,
0.006316172890365124,
0.054566726088523865,
-0.028080327436327934,
-0.00740933557972312,
-0.09776343405246735,
-0.04415534436702728,
0.04271095618605614,
-0.05314090847969055,
0.016088267788290977,
0.07865578681230545,
0.014433515258133411,
0.039421502500772476,
-0.002992848865687847,
0.01624966971576214,
0.007855815812945366,
0.10199297219514847,
-0.11779867857694626,
0.0833180621266365,
-0.022776925936341286,
0.027646487578749657,
0.03404286876320839,
0.010782883502542973,
-0.02563132531940937,
-0.010622146539390087,
-0.032500386238098145,
-0.007904056459665298,
-0.1138899028301239,
-0.06528238207101822,
-0.026298167183995247,
0.0245194174349308,
-0.002407562453299761,
-0.03834691643714905,
0.013094525784254074,
0.15173617005348206,
-0.023440459743142128,
0.0007493059383705258,
-0.11915798485279083,
-0.03671472519636154,
-0.05462422966957092,
-0.060675475746393204,
-0.11387960612773895,
-0.013520876877009869,
0.010358283296227455,
0.001146959257312119,
-0.007531892973929644,
0.06174273043870926,
-0.002066959859803319,
0.058641064912080765,
-0.01815730333328247,
-0.12166644632816315,
-0.030126068741083145,
-0.032117780297994614,
0.09868787974119186,
-0.017159799113869667,
-0.028474479913711548,
-0.008018632419407368,
0.0653851181268692,
-0.07345055043697357,
-0.006764890626072884,
-0.053006697446107864,
0.02592768520116806,
0.0029845719691365957,
-0.05606870725750923,
-0.06377039849758148,
0.055707599967718124,
-0.07049458473920822,
-0.0287188570946455,
-0.02508995495736599,
-0.018048405647277832,
-0.007835568860173225,
0.07749871909618378,
-0.041885145008563995,
0.016593387350440025,
-0.040281813591718674,
0.046008940786123276,
0.029023563489317894,
0.04266000911593437,
0.013258565217256546,
0.05050848051905632,
0.07299993932247162,
0.09941226989030838,
0.0906238779425621,
0.01957763358950615,
0.051991816610097885,
-0.026799747720360756,
0.05320918560028076,
0.074154332280159,
0.034303635358810425,
-0.054075002670288086,
-4.1569762743165484e-8,
-0.11946612596511841,
-0.05591713264584541,
0.04648043215274811,
0.1419087052345276,
-0.06736772507429123,
-0.005357436370104551,
0.022578109055757523,
0.12679564952850342,
-0.011644258163869381,
-0.04009167477488518,
0.005160684697329998,
0.12064649909734726,
-0.056749116629362106,
-0.06361941248178482,
-0.03351648896932602,
0.06650704890489578,
-0.029185760766267776,
0.020296437665820122,
-0.018263397738337517,
0.033906240016222,
0.0350717194378376,
0.06978613138198853,
0.04591073840856552,
-0.04818304628133774,
0.057838305830955505,
-0.0536336712539196,
-0.009655465371906757,
0.11966226249933243,
0.0510132797062397,
0.009635315276682377,
-0.029443366453051567,
-0.04699861258268356,
-0.07319799810647964,
-0.0778505802154541,
-0.004470621235668659,
0.10305241495370865,
0.020791446790099144,
-0.029444897547364235,
0.0012190198758617043,
-0.05664415284991264,
0.04548395425081253,
0.04173772782087326,
-0.007046940736472607,
0.021232599392533302,
-0.041401393711566925,
-0.025300482288002968,
-0.013075732626020908,
-0.07681994140148163,
0.0034421193413436413,
-0.0780375525355339,
0.004332433920353651,
-0.07487179338932037,
-0.012783900834619999,
0.012821719981729984,
0.02264842763543129,
-0.029573267325758934,
-0.0537528358399868,
-0.015817565843462944,
0.09599737823009491,
0.056403543800115585,
0.02742895670235157,
-0.005451577715575695,
-0.03174573555588722,
0.06552986800670624
] |
Brain Bandit: A Biologically Grounded Neural Network for Efficient Control of Exploration | https://openreview.net/forum?id=RWJX5F5I9g | [
"Chen Jiang",
"Jiahui An",
"Yating Liu",
"Ni Ji"
] | Oral | How to balance between exploration and exploitation in an uncertain environment is a central challenge in reinforcement learning. In contrast, humans and animals have demonstrated superior exploration efficiency in novel environments. To understand how the brain’s neural network controls exploration under uncertainty, we analyzed the dynamical systems model of a biological neural network that controls explore-exploit decisions during foraging. Mathematically, this model (named the Brain Bandit Net, or BBN) is a special type of stochastic continuous Hopfield network. We show through theory and simulation that BBN can perform posterior sampling of action values with a tunable bias towards or against uncertain options. We then demonstrate that, in multi-armed bandit (MAB) tasks, BBN can generate probabilistic choice behavior with a flexible uncertainty bias resembling human and animal choice patterns. In addition to its high efficiency in MAB tasks, BBN can also be embedded with reinforcement learning algorithms to accelerate learning in MDP tasks. Altogether, our findings reveal the theoretical foundation for efficient exploration in biological neural networks and propose a general, brain-inspired algorithm for enhancing exploration in RL. | explore-exploit, stochastic Hopfield network, Thompson sampling, decision under uncertainty, brain-inspired algorithm, reinforcement learning | We demonstrate that a brain-inspired stochastic Hopfield network can achieve efficient, human-like, uncertainty-aware exploration in bandit and MDP tasks. | 12,774 | null | [
-0.048125628381967545,
-0.06399022042751312,
0.02736765518784523,
0.013729391619563103,
-0.026136204600334167,
-0.013279981911182404,
0.06835620850324631,
-0.022411705926060677,
0.06214163452386856,
-0.04466582462191582,
-0.05740915611386299,
-0.01716732047498226,
0.03107030689716339,
0.01405650470405817,
-0.02896534465253353,
0.01492314226925373,
0.04142340272665024,
0.055162232369184494,
-0.009188561700284481,
-0.004259726032614708,
-0.04096391424536705,
-0.016162924468517303,
0.06671600788831711,
-0.045496050268411636,
-0.054934024810791016,
-0.009806451387703419,
-0.028807828202843666,
-0.043237123638391495,
-0.016048626974225044,
0.0036362314131110907,
0.085617296397686,
-0.04411689192056656,
0.014766176231205463,
-0.08676988631486893,
-0.0716639906167984,
0.02269052527844906,
-0.1890251785516739,
-0.031884025782346725,
0.032235488295555115,
0.006547450087964535,
-0.002322617219761014,
0.023786460980772972,
0.0034957900643348694,
0.05025208741426468,
0.02000538446009159,
0.017574086785316467,
-0.004443070385605097,
-0.05537145212292671,
0.09763434529304504,
-0.026569990441203117,
-0.04689782112836838,
-0.044680871069431305,
0.011580207385122776,
-0.037214089184999466,
0.025222286581993103,
0.09993588924407959,
-0.03711405396461487,
0.044331442564725876,
-0.015454146079719067,
-0.04638075828552246,
0.0006298001972027123,
-0.08255874365568161,
-0.0013411662075668573,
-0.036680918186903,
-0.005390860605984926,
-0.0032259016297757626,
-0.0649617612361908,
0.060856446623802185,
0.058587223291397095,
0.0016311025246977806,
0.06647419184446335,
0.028482850641012192,
-0.05130241811275482,
-0.03647305443882942,
0.04713611304759979,
0.043431077152490616,
-0.003422686131671071,
-0.04435332491993904,
0.05507032200694084,
-0.0024485685862600803,
-0.05704348906874657,
-0.02339249849319458,
-0.011712903156876564,
-0.006069678347557783,
0.0861794501543045,
0.015822770074009895,
0.02996472269296646,
0.051195111125707626,
0.10338524729013443,
-0.052270375192165375,
-0.016051024198532104,
-0.05084879323840141,
-0.04121590405702591,
-0.04310541972517967,
-0.017804758623242378,
0.02840353734791279,
-0.014140208251774311,
-0.059855274856090546,
-0.02807282656431198,
0.05065843090415001,
0.053028423339128494,
0.047585904598236084,
0.004782684147357941,
-0.021733127534389496,
0.009975737892091274,
-0.05291733890771866,
0.04824744910001755,
0.061389826238155365,
0.09080114960670471,
-0.03794441372156143,
-0.05883173644542694,
-0.003416162682697177,
0.060558103024959564,
0.08954425901174545,
-0.021126287057995796,
0.05724508687853813,
-0.0065663349814713,
0.0012688597198575735,
-0.014456365257501602,
0.07684618979692459,
-0.004225071519613266,
-0.03498806804418564,
0.007317093666642904,
-0.057775530964136124,
-0.0018944673938676715,
-0.040076371282339096,
-0.026724904775619507,
-3.578334593287848e-36,
0.03298328444361687,
-0.07951686531305313,
0.05322740599513054,
-0.01967962086200714,
0.04654324799776077,
0.01368725672364235,
0.05375812202692032,
-0.016659345477819443,
-0.017475945875048637,
0.03296126425266266,
-0.03287869319319725,
0.026272157207131386,
-0.05402768775820732,
0.0883585587143898,
0.06733833253383636,
-0.062172677367925644,
-0.06091201305389404,
-0.03417101129889488,
0.08023589104413986,
-0.11729083210229874,
0.0747116431593895,
-0.042871683835983276,
-0.00850761029869318,
-0.046599116176366806,
0.07611024379730225,
-0.026752369478344917,
-0.029138954356312752,
0.02898504212498665,
-0.02010042406618595,
0.016119368374347687,
-0.06000730022788048,
0.006518337409943342,
-0.09559997916221619,
-0.031106706708669662,
-0.013717621564865112,
0.02354404330253601,
-0.00035752926487475634,
-0.0021553924307227135,
-0.013903945684432983,
0.0436183325946331,
0.020539522171020508,
-0.008834954351186752,
0.0840398296713829,
0.03514673560857773,
-0.0771574079990387,
-0.06294271349906921,
0.03015218861401081,
-0.028296498581767082,
-0.03323369100689888,
0.005686814896762371,
-0.008929405361413956,
0.014844514429569244,
0.09054087847471237,
-0.06206148490309715,
0.010064189322292805,
0.02676863595843315,
0.0055824690498411655,
0.0587068609893322,
-0.06172941252589226,
0.03781915828585625,
0.054640378803014755,
-0.03400874882936478,
-0.00941404327750206,
0.03686784207820892,
0.06586948782205582,
0.1250150054693222,
-0.05820966139435768,
0.03408142924308777,
0.046963632106781006,
-0.003008659463375807,
0.038073424249887466,
0.031689777970314026,
0.02957114949822426,
-0.1380065232515335,
0.04427361115813255,
0.005312081892043352,
0.05785513296723366,
-0.006828458048403263,
0.022475430741906166,
-0.0576145313680172,
0.012559800408780575,
-0.0519135557115078,
-0.1050664484500885,
0.05421389266848564,
0.00337910745292902,
0.025573834776878357,
0.0666554793715477,
-0.10979366302490234,
-0.02181429974734783,
0.0040503982454538345,
-0.05551551654934883,
-0.015388906002044678,
-0.021562207490205765,
0.021876320242881775,
-0.0018196309683844447,
-7.125549666854636e-34,
-0.02880980633199215,
-0.06958421319723129,
0.06392674148082733,
0.0029866239055991173,
0.02948659285902977,
0.03739897161722183,
0.025880582630634308,
-0.09419842809438705,
0.010377087630331516,
-0.015690432861447334,
-0.14023107290267944,
0.0680752769112587,
0.05292274430394173,
-0.003940399736166,
0.014327761717140675,
-0.061493098735809326,
-0.05679919570684433,
0.04045526310801506,
0.04295936971902847,
-0.009275366552174091,
-0.04585607349872589,
0.08903075754642487,
-0.09540168941020966,
-0.027679704129695892,
-0.008675532415509224,
0.04897675663232803,
-0.00031716309604234993,
0.08332745730876923,
0.01977994479238987,
0.01639922335743904,
-0.07721322774887085,
-0.02531350962817669,
-0.0392996110022068,
-0.013740657828748226,
-0.009072648361325264,
0.10675755888223648,
0.12451931834220886,
0.03231959417462349,
-0.06681979447603226,
0.0883796438574791,
-0.0029454119503498077,
-0.04781680926680565,
0.0030438597314059734,
-0.022751739248633385,
0.0517210029065609,
0.02189905382692814,
-0.03054872713983059,
0.03897399082779884,
-0.06256870925426483,
0.048848409205675125,
0.04307267442345619,
0.07270219177007675,
-0.1338701844215393,
-0.012324108742177486,
-0.07120593637228012,
-0.0461546964943409,
-0.04953565075993538,
0.00048385129775851965,
0.1093655377626419,
-0.0009535992867313325,
-0.07724542915821075,
-0.04589635506272316,
-0.027934450656175613,
0.08319974690675735,
-0.06545151770114899,
0.0244729183614254,
-0.08182116597890854,
0.06389153748750687,
0.05376259610056877,
-0.03355034813284874,
-0.04294467344880104,
0.0322774238884449,
0.05508507043123245,
-0.016374588012695312,
0.03868664801120758,
0.007423285860568285,
-0.017066141590476036,
-0.011600079014897346,
0.08345743268728256,
-0.02791346237063408,
0.01775270886719227,
-0.002062955405563116,
0.02098812162876129,
0.014672689139842987,
0.04637552797794342,
0.075535349547863,
-0.0704430639743805,
0.07690515369176865,
0.01478736661374569,
-0.06909151375293732,
0.00004717960473499261,
-0.06782422959804535,
0.05016424506902695,
0.023261114954948425,
-0.0883263573050499,
-3.491616595852065e-8,
-0.05032079294323921,
0.06218946352601051,
0.12812024354934692,
0.07194583863019943,
0.14426089823246002,
0.05924344062805176,
0.02066093310713768,
-0.04326143115758896,
-0.004963279701769352,
-0.05072180554270744,
0.09023304283618927,
0.02174968831241131,
0.007843523286283016,
-0.05064280331134796,
-0.02224418893456459,
0.0337308906018734,
-0.0023848353885114193,
-0.030918080359697342,
0.019600871950387955,
0.007425583433359861,
0.05072393640875816,
-0.0011765556409955025,
0.017338233068585396,
-0.028491666540503502,
0.024778030812740326,
-0.04192569851875305,
-0.06217397376894951,
0.08136379718780518,
0.02307385392487049,
0.06743919104337692,
0.03562437742948532,
-0.004954804666340351,
0.04569130390882492,
0.046785302460193634,
0.05310666188597679,
0.015608730725944042,
-0.021255679428577423,
-0.0785023644566536,
-0.040364932268857956,
-0.03807661309838295,
-0.04391771927475929,
0.04983406513929367,
0.016812939196825027,
-0.011042793281376362,
-0.07169999927282333,
0.016049670055508614,
-0.019915610551834106,
-0.049965452402830124,
0.08049291372299194,
-0.061605557799339294,
0.011300826445221901,
-0.001005607540719211,
-0.06489754468202591,
0.031014686450362206,
0.06840573251247406,
-0.023899678140878677,
-0.02800258994102478,
-0.09890580177307129,
-0.02376515045762062,
0.09007444232702255,
0.016539327800273895,
0.029977446421980858,
-0.12090644240379333,
-0.04213717579841614
] |
MaestroMotif: Skill Design from Artificial Intelligence Feedback | https://openreview.net/forum?id=or8mMhmyRV | [
"Martin Klissarov",
"Mikael Henaff",
"Roberta Raileanu",
"Shagun Sodhani",
"Pascal Vincent",
"Amy Zhang",
"Pierre-Luc Bacon",
"Doina Precup",
"Marlos C. Machado",
"Pierluca D'Oro"
] | Oral | Describing skills in natural language has the potential to provide an accessible way to inject human knowledge about decision-making into an AI system. We present MaestroMotif, a method for AI-assisted skill design, which yields high-performing and adaptable agents. MaestroMotif leverages the capabilities of Large Language Models (LLMs) to effectively create and reuse skills. It first uses an LLM's feedback to automatically design rewards corresponding to each skill, starting from their natural language description. Then, it employs an LLM's code generation abilities, together with reinforcement learning, for training the skills and combining them to implement complex behaviors specified in language. We evaluate MaestroMotif using a suite of complex tasks in the NetHack Learning Environment (NLE), demonstrating that it surpasses existing approaches in both performance and usability. | Hierarchical RL, Reinforcement Learning, LLMs | A method for AI-assisted skill design via Motif and LLM code generation, solving tasks zero-shot from language descriptions on NetHack. | 12,735 | 2412.08542 | [
-0.023494863882660866,
-0.0449480339884758,
0.025618351995944977,
0.07570120692253113,
-0.020263411104679108,
-0.00032232367084361613,
-0.005257580894976854,
0.03576262295246124,
-0.04352348670363426,
0.013138490729033947,
-0.07918298244476318,
-0.09636082500219345,
0.06023058295249939,
-0.010696874000132084,
0.03515571728348732,
0.01774139143526554,
0.04764613136649132,
0.009788784198462963,
-0.039489082992076874,
-0.068165123462677,
0.05643105134367943,
0.05988357961177826,
0.006750146858394146,
-0.0483047291636467,
0.0073681725189089775,
0.017184458673000336,
-0.023098492994904518,
-0.08499216288328171,
0.08318161964416504,
-0.011910779401659966,
0.05249512195587158,
-0.06274174898862839,
-0.0006430229404941201,
0.048357609659433365,
-0.06716445088386536,
0.07341648638248444,
-0.11492271721363068,
0.06966225057840347,
-0.04305719584226608,
-0.10850885510444641,
-0.05515323951840401,
-0.028088467195630074,
0.05402180925011635,
-0.051338523626327515,
0.08363359421491623,
-0.020385341718792915,
-0.0741933062672615,
-0.023775793612003326,
0.04442678019404411,
0.08641128987073898,
-0.0964086577296257,
-0.0811767429113388,
0.0883486345410347,
0.009503264911472797,
-0.0023991840425878763,
0.04387209564447403,
0.033761464059352875,
-0.01247364841401577,
0.006895953789353371,
-0.06860081106424332,
0.005743788555264473,
-0.06067001447081566,
0.000010065088645205833,
0.007136546541005373,
-0.04702670872211456,
-0.05299009010195732,
-0.0646723210811615,
0.056498199701309204,
-0.00828079879283905,
-0.055477455258369446,
-0.042249180376529694,
0.029827086254954338,
0.02573832869529724,
0.06797578930854797,
0.04437300190329552,
0.0496341697871685,
-0.057349346578121185,
-0.060410551726818085,
0.051054324954748154,
-0.038154520094394684,
-0.04144022986292839,
0.012145264074206352,
0.02948383428156376,
0.022040754556655884,
0.06827390938997269,
0.023050513118505478,
0.05653313174843788,
0.08496661484241486,
0.08663372695446014,
-0.026218602433800697,
0.009184334427118301,
-0.0543026477098465,
0.07939513027667999,
-0.06366346031427383,
-0.014468595385551453,
0.06594919413328171,
-0.06051928550004959,
-0.0963052287697792,
-0.11809681355953217,
0.02555900812149048,
-0.015190087258815765,
0.054445717483758926,
-0.03358439728617668,
-0.0676804855465889,
-0.07504233717918396,
0.042816419154405594,
0.05454208329319954,
-0.008122077211737633,
0.03347424417734146,
-0.06932372599840164,
-0.0146775608882308,
0.0591072253882885,
-0.005288675427436829,
0.020714841783046722,
0.09739817678928375,
0.05749182030558586,
-0.003670030040666461,
0.006644899956882,
0.023678239434957504,
0.07196830958127975,
0.030939871445298195,
0.0283543448895216,
-0.014473911374807358,
0.024772077798843384,
0.08139721304178238,
-0.03147321566939354,
0.059858739376068115,
-2.35159499959021e-33,
0.06432569772005081,
0.02621934376657009,
0.019993960857391357,
0.09874838590621948,
0.13617131114006042,
-0.06972301006317139,
0.07942849397659302,
0.019618215039372444,
-0.07421611994504929,
-0.03461449220776558,
-0.0010835763532668352,
0.04118083789944649,
-0.06768923252820969,
0.12624090909957886,
0.037041809409856796,
0.03971649333834648,
-0.03360459953546524,
0.018071506172418594,
-0.00713587598875165,
-0.0778689756989479,
0.015035824850201607,
-0.009499133564531803,
0.039095424115657806,
-0.01564256101846695,
0.018648315221071243,
0.05370262265205383,
0.1002596765756607,
-0.012136274017393589,
0.0006341568659991026,
-0.024213535711169243,
-0.08416161686182022,
0.0054859500378370285,
-0.03728419169783592,
0.03836822137236595,
0.013986396603286266,
0.019552458077669144,
0.005667876917868853,
-0.013969827443361282,
0.03419916331768036,
0.06328348070383072,
0.009457451291382313,
-0.024615086615085602,
-0.03126263618469238,
0.06340928375720978,
-0.08324383199214935,
-0.04383382946252823,
-0.00819373782724142,
0.009623042307794094,
-0.007492830511182547,
0.03874501585960388,
0.034696321934461594,
0.027954477816820145,
0.000021037871192675084,
-0.06002393364906311,
0.0492999441921711,
0.014832118526101112,
-0.004994490649551153,
0.07140661776065826,
-0.004215752240270376,
0.006846677977591753,
0.05208118259906769,
0.01581617258489132,
-0.020495790988206863,
0.06082221865653992,
0.043584294617176056,
0.0007231436902657151,
0.049042779952287674,
0.12268783897161484,
0.07411834597587585,
-0.052520252764225006,
-0.051902975887060165,
0.047176294028759,
0.0013787783682346344,
-0.005175871774554253,
0.009991301223635674,
-0.021288596093654633,
0.04792153835296631,
-0.1169341504573822,
0.024673961102962494,
0.010651616379618645,
0.00799547415226698,
-0.029334155842661858,
-0.052148040384054184,
-0.0228001456707716,
0.018054919317364693,
-0.10719109326601028,
0.00703764334321022,
-0.08458126336336136,
0.031912147998809814,
-0.027824919670820236,
-0.06216438114643097,
-0.004750258754938841,
-0.019770290702581406,
0.013916072435677052,
-0.029941676184535027,
6.33470553662521e-34,
-0.026943547651171684,
-0.0320613794028759,
-0.05691223964095116,
0.04671262949705124,
-0.0360473208129406,
-0.02782762423157692,
-0.03345433622598648,
0.04853137582540512,
-0.06028829887509346,
-0.06230327486991882,
-0.06460978835821152,
0.006517249625176191,
0.03095003217458725,
-0.02347348816692829,
0.02692927047610283,
-0.03159252554178238,
-0.06672413647174835,
0.011391270905733109,
0.008522444404661655,
-0.013252931647002697,
-0.007391186896711588,
0.11192598938941956,
-0.11768753826618195,
0.03732077032327652,
-0.07890745997428894,
0.025396492332220078,
-0.04007566720247269,
0.04989532381296158,
-0.002377428812906146,
0.06583213061094284,
0.017567846924066544,
-0.021430378779768944,
0.00284311780706048,
-0.03450031578540802,
-0.006400473415851593,
0.06875292956829071,
0.04849037155508995,
-0.010908170603215694,
-0.05914628133177757,
0.07336308062076569,
0.06275040656328201,
-0.0800316110253334,
-0.056163493543863297,
-0.039147622883319855,
-0.04293321445584297,
-0.0037168085109442472,
-0.06674124300479889,
-0.05403026565909386,
-0.032462410628795624,
0.011627927422523499,
0.05767335370182991,
-0.048828061670064926,
-0.07296653836965561,
-0.05701543390750885,
-0.07717783749103546,
-0.03766476362943649,
0.04821626842021942,
-0.06162295863032341,
0.06424213945865631,
0.0006610191776417196,
-0.07076406478881836,
-0.013352600857615471,
0.09858296066522598,
-0.023166097700595856,
0.006445286329835653,
-0.026140235364437103,
-0.04722826927900314,
0.04615740105509758,
-0.013859650120139122,
0.005239599384367466,
0.04405722767114639,
0.011747071519494057,
0.02425159513950348,
-0.01843291148543358,
0.017483389005064964,
0.025880955159664154,
-0.001270582084544003,
-0.03588082268834114,
-0.05665767565369606,
-0.06422899663448334,
0.021725015714764595,
0.001991712022572756,
0.01622721180319786,
0.0006444294122047722,
0.08529025316238403,
0.08860591053962708,
-0.01826479658484459,
0.10684680193662643,
0.01352218072861433,
0.04176611080765724,
-0.03729703649878502,
0.07747174054384232,
0.003959722816944122,
0.09031650424003601,
-0.07395066320896149,
-4.1604113931725806e-8,
-0.04120338708162308,
0.04053756594657898,
0.013632616028189659,
0.04939328134059906,
-0.01581183262169361,
0.013961984775960445,
-0.13112109899520874,
0.0036355024203658104,
-0.013137155212461948,
-0.03249848634004593,
0.07174796611070633,
0.03915099427103996,
0.042713042348623276,
-0.050058234483003616,
0.030370451509952545,
0.06052739918231964,
-0.06091197952628136,
0.016272244974970818,
0.030057808384299278,
-0.037323806434869766,
0.1325213462114334,
-0.004656223580241203,
-0.0371144562959671,
0.016391973942518234,
-0.017139187082648277,
-0.05234106630086899,
-0.04133038967847824,
0.06353328377008438,
-0.023444101214408875,
0.06735993921756744,
0.05340992659330368,
-0.0009965094504877925,
0.02715352736413479,
0.01633012853562832,
0.08242838829755783,
0.017308104783296585,
0.008241022005677223,
-0.10506539791822433,
0.002206013770774007,
-0.0948052853345871,
0.052230630069971085,
0.11236017197370529,
-0.025964228436350822,
-0.013389687053859234,
-0.013839689083397388,
-0.018257880583405495,
-0.05090951919555664,
-0.13478104770183563,
0.03274558484554291,
-0.06261390447616577,
-0.08361292630434036,
-0.01791044883430004,
-0.05764468014240265,
0.003409072058275342,
0.07121284306049347,
-0.009011263027787209,
-0.0009694428881630301,
-0.000033369851735187694,
0.03539266437292099,
0.010792174376547337,
-0.01290136855095625,
0.02327604591846466,
0.03147280961275101,
0.03204016387462616
] |
Learning to Discover Regulatory Elements for Gene Expression Prediction | https://openreview.net/forum?id=Mfnh1Sqdwf | [
"Xingyu Su",
"Haiyang Yu",
"Degui Zhi",
"Shuiwang Ji"
] | Oral | We consider the problem of predicting gene expressions from DNA sequences. A key challenge of this task is to find the regulatory elements that control gene expressions. Here, we introduce Seq2Exp, a Sequence to Expression network explicitly designed to discover and extract regulatory elements that drive target gene expression, enhancing the accuracy of the gene expression prediction. Our approach captures the causal relationship between epigenomic signals, DNA sequences and their associated regulatory elements. Specifically, we propose to decompose the epigenomic signals and the DNA sequence conditioned on the causal active regulatory elements, and apply an information bottleneck with the Beta distribution to combine their effects while filtering out non-causal components. Our experiments demonstrate that Seq2Exp outperforms existing baselines in gene expression prediction tasks and discovers influential regions compared to commonly used statistical methods for peak detection such as MACS3. The source code is released as part of the AIRS library (https://github.com/divelab/AIRS/). | Gene Expression, Deep Learning, Sequence Modeling | null | 12,644 | 2502.13991 | [
-0.0845017358660698,
-0.025793347507715225,
0.02984725870192051,
0.04579932615160942,
0.11021523922681808,
-0.0022563128732144833,
0.0075287725776433945,
-0.06440958380699158,
-0.030917929485440254,
-0.03083932027220726,
-0.048635661602020264,
-0.05303158238530159,
-0.02617608569562435,
-0.001213842653669417,
-0.002616483485326171,
0.07251738756895065,
-0.0023376026656478643,
0.023658547550439835,
-0.02211865223944187,
-0.08041663467884064,
0.0383816733956337,
0.049464158713817596,
0.03740036115050316,
0.04470754414796829,
0.0037470527458935976,
-0.01171854417771101,
0.05221214145421982,
0.06384659558534622,
0.019273027777671814,
-0.02729911170899868,
0.04377356171607971,
-0.012334578670561314,
0.03130717575550079,
-0.10887709259986877,
-0.04227707162499428,
-0.03059404343366623,
0.01989937201142311,
0.023255597800016403,
-0.02504284493625164,
-0.023390721529722214,
0.10531765967607498,
0.010919436812400818,
-0.015955159440636635,
0.11868543922901154,
0.05481560528278351,
0.008839309215545654,
0.07315196096897125,
-0.013009636662900448,
0.03000408224761486,
0.009692875668406487,
-0.05518633499741554,
-0.040578410029411316,
0.027403468266129494,
-0.010300070978701115,
0.09321089088916779,
0.0003977182786911726,
0.01709338277578354,
-0.09180855005979538,
0.006202251184731722,
-0.03563516587018967,
0.041762854903936386,
-0.10217447578907013,
-0.09294391423463821,
-0.013175610452890396,
0.012974250130355358,
0.028064319863915443,
0.043325960636138916,
0.057469651103019714,
-0.0690506100654602,
-0.03398432210087776,
-0.07707834988832474,
0.01552197989076376,
-0.08741112798452377,
0.1065276712179184,
0.036782942712306976,
0.09034088253974915,
0.006674253381788731,
-0.002261969493702054,
0.016048304736614227,
-0.02077752910554409,
0.012806455604732037,
0.00011721540067810565,
0.05150837078690529,
0.0010088252602145076,
0.03870902955532074,
-0.051826272159814835,
-0.0010243796277791262,
-0.004712301306426525,
-0.009475444443523884,
0.09709904342889786,
-0.0027612410485744476,
-0.07615967094898224,
0.15328745543956757,
-0.04544244334101677,
0.004536035470664501,
0.09865126758813858,
0.013931240886449814,
-0.05439569428563118,
0.069854736328125,
0.05780177563428879,
-0.08908923715353012,
0.057578738778829575,
-0.0284090768545866,
0.02761390060186386,
-0.050021037459373474,
-0.04547099396586418,
0.02672662027180195,
0.035064853727817535,
0.03925654664635658,
-0.016570495441555977,
0.016858167946338654,
0.009802834130823612,
0.0020924813579767942,
-0.024922287091612816,
0.07628286629915237,
0.05712812393903732,
0.05415777862071991,
0.0499187707901001,
0.05065162479877472,
0.04830982908606529,
0.010396264493465424,
-0.026395287364721298,
-0.028755558654665947,
0.026577899232506752,
0.034659892320632935,
0.05687413364648819,
-0.14663708209991455,
3.431336025714599e-35,
0.09515111148357391,
-0.0004831890400964767,
0.04022885859012604,
0.04553486406803131,
-0.046338409185409546,
-0.02397853694856167,
-0.042354460805654526,
-0.028821568936109543,
-0.006845713593065739,
0.06697852909564972,
-0.02645893208682537,
-0.0327950119972229,
-0.04254518449306488,
0.0936199352145195,
0.0175325945019722,
-0.008337901905179024,
-0.042476750910282135,
0.0352582111954689,
-0.002935020485892892,
-0.04261714220046997,
-0.05389559641480446,
-0.09432055801153183,
-0.0919230654835701,
-0.07575114071369171,
0.009318766184151173,
-0.00775690283626318,
0.011874394491314888,
0.0167666207998991,
0.01725005730986595,
0.013618013821542263,
-0.02362111583352089,
-0.008865416049957275,
0.07691319286823273,
0.029294006526470184,
0.0767117440700531,
-0.04911206662654877,
-0.004811731167137623,
-0.04464011266827583,
0.017884714528918266,
0.03285573050379753,
-0.019706938415765762,
-0.015104267746210098,
-0.0020439920481294394,
-0.07103617489337921,
-0.06432244181632996,
-0.09030979871749878,
-0.06818273663520813,
-0.05477217957377434,
0.03176046907901764,
0.00020273469272069633,
0.0694999098777771,
0.022455284371972084,
0.06345659494400024,
-0.021431611850857735,
-0.03239070996642113,
0.05732852593064308,
-0.0603683777153492,
-0.022380530834197998,
-0.0047843968495726585,
0.05150976404547691,
-0.07037006318569183,
-0.0027433703653514385,
0.016491897404193878,
0.005285805091261864,
0.017956266179680824,
0.09227178990840912,
-0.06263020634651184,
-0.015244344249367714,
-0.019223995506763458,
0.05175874009728432,
0.02481960318982601,
0.06387069076299667,
0.0068966844119131565,
-0.1309366673231125,
0.0912899598479271,
-0.03451915457844734,
-0.029215890914201736,
-0.00907951034605503,
-0.07407693564891815,
0.07933484762907028,
-0.005713232792913914,
-0.0021485681645572186,
-0.025327607989311218,
0.027604812756180763,
0.008276875130832195,
-0.07638175040483475,
0.03059510700404644,
-0.041332442313432693,
-0.08139800280332565,
-0.05369734391570091,
-0.035213831812143326,
-0.0077843451872467995,
-0.0006305184797383845,
0.011480249464511871,
0.03992222249507904,
-5.475876966122465e-34,
-0.004728016909211874,
0.000035992627090308815,
0.08307258039712906,
-0.09831678867340088,
-0.007591772358864546,
-0.05369101092219353,
0.004666843917220831,
-0.03714081272482872,
0.019936256110668182,
-0.03612178936600685,
0.05407777801156044,
-0.02804565243422985,
0.017825011163949966,
-0.04943538457155228,
0.017540864646434784,
0.03812040761113167,
-0.01051759161055088,
0.05339730903506279,
-0.013419214636087418,
-0.038773972541093826,
0.0025488585233688354,
0.02679867297410965,
0.010982206091284752,
0.05040688440203667,
-0.009503562934696674,
-0.05394671857357025,
0.046379685401916504,
0.12052304297685623,
-0.0013275686651468277,
-0.0021070921793580055,
-0.18383187055587769,
0.040673937648534775,
-0.01533875148743391,
0.0038688615895807743,
-0.10870935022830963,
0.04627802222967148,
0.013457806780934334,
0.050369247794151306,
-0.025042006745934486,
0.001370424055494368,
0.06770515441894531,
0.09522830694913864,
-0.10978055000305176,
0.014696524478495121,
0.02708613872528076,
0.010614026337862015,
-0.0073703196831047535,
0.12597329914569855,
0.04922724887728691,
0.024521924555301666,
-0.03585844859480858,
0.03545616194605827,
-0.05332823097705841,
0.037594735622406006,
-0.033622369170188904,
-0.05211248621344566,
0.04224475845694542,
0.01813521236181259,
0.060850802809000015,
0.00624390272423625,
-0.07723500579595566,
0.017442137002944946,
-0.021199442446231842,
0.02461249753832817,
0.0651027262210846,
-0.024204645305871964,
0.04254470020532608,
-0.0032922173850238323,
0.04655810818076134,
-0.035461630672216415,
0.006347267422825098,
0.021561812609434128,
-0.04556906595826149,
-0.03713345527648926,
-0.02134721726179123,
-0.06252896785736084,
-0.0599190890789032,
-0.03796990588307381,
-0.07214613258838654,
0.05892873927950859,
0.07271943241357803,
-0.0038457263726741076,
0.05400189384818077,
0.007063973695039749,
0.052816733717918396,
0.07210041582584381,
0.006771526299417019,
0.07538765668869019,
0.0830233097076416,
-0.031708311289548874,
-0.1027236059308052,
0.028805702924728394,
-0.0664086863398552,
0.05317278206348419,
-0.055268824100494385,
-3.7484237935814235e-8,
-0.03635077923536301,
-0.03360157459974289,
-0.003968091681599617,
0.01747736893594265,
0.051010485738515854,
0.10043983161449432,
-0.035731520503759384,
0.09451355040073395,
0.03933073207736015,
-0.04359323903918266,
0.07213468849658966,
0.019007651135325432,
-0.08032307773828506,
-0.015536287799477577,
0.06781373172998428,
-0.014046839438378811,
0.010688561014831066,
0.025534139946103096,
-0.029069233685731888,
-0.09061984717845917,
0.026283860206604004,
0.028500858694314957,
0.052713532000780106,
0.07822112739086151,
0.07011367380619049,
-0.1053561419248581,
-0.03328269347548485,
0.023267343640327454,
0.09774165600538254,
-0.04570575803518295,
0.014031822793185711,
-0.036541253328323364,
-0.00806597713381052,
-0.041636981070041656,
-0.04003700986504555,
0.015119937248528004,
0.01147443801164627,
-0.07128951698541641,
-0.01531302835792303,
0.030913731083273888,
-0.03874170407652855,
0.037116631865501404,
-0.04061251878738403,
-0.011705172248184681,
-0.04097963497042656,
0.002891149604693055,
0.026235932484269142,
-0.005444195121526718,
0.04979114234447479,
-0.11844226717948914,
0.0861477479338646,
-0.07012851536273956,
0.03553156927227974,
0.035875964909791946,
0.04285377636551857,
0.024767933413386345,
-0.05219898372888565,
-0.012215050868690014,
-0.0017611318035051227,
0.02417379803955555,
0.04820927977561951,
0.06046615168452263,
-0.02742767706513405,
-0.07264591008424759
] |
Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models | https://openreview.net/forum?id=tyEyYT267x | [
"Marianne Arriola",
"Aaron Gokaslan",
"Justin T Chiu",
"Zhihan Yang",
"Zhixuan Qi",
"Jiaqi Han",
"Subham Sekhar Sahoo",
"Volodymyr Kuleshov"
] | Oral | Diffusion language models offer unique benefits over autoregressive models due to their potential for parallelized generation and controllability, yet they lag in likelihood modeling and are limited to fixed-length generation. In this work, we introduce a class of block diffusion language models that interpolate between discrete denoising diffusion and autoregressive models. Block diffusion overcomes key limitations of both approaches by supporting flexible-length generation and improving inference efficiency with KV caching and parallel token sampling. We propose a recipe for building effective block diffusion models that includes an efficient training algorithm, estimators of gradient variance, and data-driven noise schedules to minimize the variance. Block diffusion sets a new state-of-the-art performance among diffusion models on language modeling benchmarks and enables generation of arbitrary-length sequences. We provide the code, along with the model weights and blog post on the project page: https://m-arriola.com/bd3lms/ | Diffusion Models, Text Diffusion, Generative Models | null | 12,566 | 2503.09573 | [
-0.10416851937770844,
-0.12655200064182281,
0.026569997891783714,
0.018532857298851013,
-0.012075117789208889,
-0.006582919973880053,
-0.05516336113214493,
-0.04498690366744995,
0.07734091579914093,
-0.0758652314543724,
-0.022044286131858826,
-0.03679569438099861,
-0.03500502184033394,
-0.02512887679040432,
-0.019465278834104538,
-0.017410581931471825,
0.04749910160899162,
0.07990420609712601,
-0.07122134417295456,
-0.12315404415130615,
-0.007422935217618942,
0.06348957866430283,
-0.03748270124197006,
0.018960868939757347,
0.09912092238664627,
0.04954085499048233,
-0.042391300201416016,
-0.06143668666481972,
0.05806541070342064,
-0.015016680583357811,
0.016020311042666435,
-0.07658771425485611,
-0.003301207907497883,
0.02570396475493908,
0.030127838253974915,
-0.016870813444256783,
-0.05214022845029831,
-0.012749318033456802,
0.06774922460317612,
-0.01062688883394003,
-0.0010473496513441205,
0.003180121537297964,
0.0428551621735096,
0.008131446316838264,
0.07144009321928024,
-0.027304846793413162,
0.00044410277041606605,
-0.0005743359215557575,
-0.030102044343948364,
0.04361725598573685,
-0.03309059888124466,
-0.04206083342432976,
-0.015441449359059334,
0.08256784826517105,
-0.02075732871890068,
-0.09203984588384628,
0.06646712124347687,
0.05694566294550896,
0.052370939403772354,
0.028664477169513702,
-0.09146042168140411,
-0.07104803621768951,
-0.009520217776298523,
0.0018514267867431045,
-0.07369691133499146,
0.013828056864440441,
0.03165978193283081,
0.11733272671699524,
-0.03465428575873375,
0.10660748183727264,
-0.05140582472085953,
0.08437271416187286,
-0.08299161493778229,
0.07497448474168777,
0.0036968495696783066,
-0.004722687415778637,
0.01280440017580986,
-0.07403913885354996,
0.038246579468250275,
-0.14782755076885223,
0.0002434747584629804,
0.0063795046880841255,
0.0849139466881752,
-0.03175216540694237,
0.019911054521799088,
-0.08062689006328583,
0.06628501415252686,
0.03773389011621475,
0.07059004157781601,
-0.027559563517570496,
-0.030379287898540497,
0.015270771458745003,
-0.03993857651948929,
0.0190404262393713,
-0.055684857070446014,
0.060790617018938065,
-0.03335580229759216,
0.00947342999279499,
-0.02726622484624386,
0.11765043437480927,
0.0319548174738884,
0.06341823935508728,
-0.03666390851140022,
-0.016788551583886147,
-0.0005531510687433183,
-0.10374815762042999,
0.05503838509321213,
0.038368403911590576,
-0.041296571493148804,
-0.019248774275183678,
-0.01897459849715233,
0.07537904381752014,
-0.018086504191160202,
0.04908355697989464,
0.05367564782500267,
-0.11011812835931778,
0.034877970814704895,
-0.08819498866796494,
0.06049096956849098,
0.03884238377213478,
-0.0004356368153821677,
-0.003979967907071114,
-0.04924936592578888,
0.00003545860818121582,
-0.028192024677991867,
-0.010597853921353817,
0.0473501980304718,
1.5988003357999553e-33,
0.009825310669839382,
0.030412845313549042,
0.01147263403981924,
0.00782418716698885,
0.061081282794475555,
-0.0038276659324765205,
0.00832613930106163,
-0.01520734652876854,
-0.017410581931471825,
-0.0439690537750721,
-0.007332314271479845,
-0.06413088738918304,
-0.09139272570610046,
0.01417467836290598,
-0.006405185908079147,
-0.010161206126213074,
0.010689318180084229,
0.08033543080091476,
0.01636701263487339,
0.018999800086021423,
0.0588291734457016,
0.012618652544915676,
-0.005286599043756723,
0.004451702814549208,
-0.00797739252448082,
0.012537368573248386,
0.03144795075058937,
-0.06606225669384003,
0.0010431386763229966,
0.047396644949913025,
-0.1001492440700531,
-0.026806853711605072,
-0.031628187745809555,
0.02353232353925705,
0.06631311774253845,
0.04972810298204422,
0.019581543281674385,
-0.02743106707930565,
-0.005837323609739542,
-0.04229435697197914,
-0.029692139476537704,
0.04124383255839348,
0.031117303296923637,
-0.06818149983882904,
-0.14418555796146393,
-0.05090190842747688,
-0.027229968458414078,
-0.059745606034994125,
-0.042109228670597076,
-0.07239114493131638,
0.047037139534950256,
0.0077019259333610535,
-0.046120382845401764,
-0.019866053014993668,
0.0526648685336113,
-0.023961063474416733,
0.03590068593621254,
0.00976976566016674,
0.012766823172569275,
0.08529314398765564,
0.043021026998758316,
0.07739666849374771,
0.0463903583586216,
0.03756750747561455,
0.09039928019046783,
0.018431851640343666,
-0.055514127016067505,
0.0722779855132103,
0.028024740517139435,
-0.04898448288440704,
-0.013507499359548092,
-0.04037529602646828,
-0.029427528381347656,
-0.02085624635219574,
-0.0006983322091400623,
-0.06155970320105553,
0.022734489291906357,
-0.06072330102324486,
-0.06896894425153732,
0.05530133098363876,
-0.06653299927711487,
-0.055899884551763535,
0.03208240121603012,
-0.07450566440820694,
-0.021422404795885086,
-0.0821944996714592,
0.017230303958058357,
-0.05668339878320694,
-0.02283399924635887,
-0.12857019901275635,
-0.05651441588997841,
-0.01844708062708378,
0.028674719855189323,
0.0014287626836448908,
0.03138458728790283,
-2.836102914272448e-33,
-0.024830985814332962,
0.06487110257148743,
0.07514681667089462,
0.08467073738574982,
-0.053996868431568146,
-0.08729059994220734,
-0.012428252026438713,
0.09198317676782608,
0.06749770790338516,
-0.007251081056892872,
-0.0900418609380722,
-0.03399715572595596,
0.024671347811818123,
0.02084713615477085,
0.030316967517137527,
-0.03397427871823311,
0.07094991952180862,
-0.0221234280616045,
0.028956213966012,
0.058852214366197586,
-0.07044577598571777,
0.0563676692545414,
-0.07831607013940811,
0.07302656769752502,
-0.034249965101480484,
0.005782297346740961,
-0.03145637735724449,
0.058957528322935104,
0.02003132365643978,
-0.07492110133171082,
-0.07331345975399017,
0.004890304058790207,
-0.020714769139885902,
-0.07344213873147964,
-0.11555395275354385,
0.06498149782419205,
0.019128849729895592,
0.037350159138441086,
-0.018665079027414322,
0.050495538860559464,
0.09163255244493484,
-0.0474361889064312,
-0.012635082937777042,
-0.011313224211335182,
0.012074040248990059,
0.023929035291075706,
-0.06975998729467392,
0.05248410999774933,
-0.017452238127589226,
-0.0345982126891613,
0.049117106944322586,
0.09545161575078964,
-0.05186573043465614,
0.07854032516479492,
-0.00582111394032836,
-0.10596392303705215,
-0.09893166273832321,
-0.02212616056203842,
0.06481326371431351,
-0.008028033189475536,
-0.056266654282808304,
0.0008212116663344204,
-0.0011663687182590365,
-0.08888480067253113,
0.01800219528377056,
-0.031024478375911713,
-0.043491166085004807,
-0.038593728095293045,
0.049295686185359955,
-0.012267784215509892,
0.06998410820960999,
0.029828699305653572,
-0.0423177033662796,
0.061663996428251266,
-0.03955751657485962,
-0.08104227483272552,
0.0421057753264904,
-0.08830442279577255,
-0.004608227405697107,
0.01911972463130951,
0.039750248193740845,
0.051913272589445114,
0.01554103009402752,
-0.021536532789468765,
0.02215583622455597,
0.017979443073272705,
-0.009918807074427605,
0.04812503606081009,
0.017535019665956497,
-0.016001321375370026,
0.009742764756083488,
0.0024356869980692863,
0.04964222386479378,
0.07298985868692398,
-0.09923677891492844,
-4.382842888617233e-8,
-0.03115474432706833,
0.06486815214157104,
0.05372854694724083,
0.044431377202272415,
0.07476804405450821,
-0.0036215209402143955,
0.02511122077703476,
0.013138467445969582,
0.03986754268407822,
-0.02276420220732689,
0.06424370408058167,
-0.008285749703645706,
-0.004888809286057949,
-0.012584804557263851,
-0.09708113223314285,
0.12326940894126892,
-0.002143672900274396,
0.037263743579387665,
0.03129072114825249,
0.005542161874473095,
-0.02351713925600052,
0.07026132196187973,
-0.0012344341957941651,
-0.0024700246285647154,
0.06816165894269943,
-0.02553349733352661,
0.004329216200858355,
-0.015285177156329155,
0.12279301136732101,
-0.030630307272076607,
-0.03223814442753792,
0.06408552080392838,
-0.011646056547760963,
-0.042927082628011703,
0.025192013010382652,
0.10244680941104889,
0.008264782838523388,
-0.03635886684060097,
-0.00824685487896204,
-0.027602195739746094,
0.0387403666973114,
0.0376737043261528,
-0.03914998099207878,
-0.015035745687782764,
0.004735931754112244,
0.0330757312476635,
-0.005055923014879227,
-0.05476926267147064,
0.022662587463855743,
-0.014644191600382328,
0.0439981110394001,
-0.06685249507427216,
0.06185135245323181,
0.02191990613937378,
0.02759535051882267,
0.013191313482820988,
0.02743670903146267,
-0.023352105170488358,
0.06616799533367157,
0.05837884172797203,
-0.0022074971348047256,
-0.0006975681171752512,
-0.012557248584926128,
0.06682898104190826
] |
Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo | https://openreview.net/forum?id=xoXn62FzD0 | [
"João Loula",
"Benjamin LeBrun",
"Li Du",
"Ben Lipkin",
"Clemente Pasti",
"Gabriel Grand",
"Tianyu Liu",
"Yahya Emara",
"Marjorie Freedman",
"Jason Eisner",
"Ryan Cotterell",
"Vikash Mansinghka",
"Alexander K. Lew",
"Tim Vieira",
"Timothy J. O'Donnell"
] | Oral | A wide range of LM applications require generating text that conforms to syntactic or semantic constraints. Imposing such constraints can be naturally framed as probabilistic conditioning, but exact generation from the resulting distribution—which can differ substantially from the LM’s base distribution—is generally intractable. In this work, we develop an architecture for controlled LM generation based on sequential Monte Carlo (SMC). This SMC framework allows us to flexibly incorporate domain- and problem-specific constraints at inference
time, and efficiently reallocate computational resources in light of new information during the course of generation. By comparing to a number of alternatives and ablations on four challenging domains—Python code generation for data science, text-to-SQL, goal inference, and molecule synthesis—we demonstrate that, with little overhead, our approach allows small open-source language models to outperform models over 8× larger, as well as closed-source, fine-tuned ones. In support of the probabilistic perspective, we show that these performance improvements are driven by better approximation to the posterior distribution. [Our system](https://github.com/probcomp/genparse) builds on the framework of Lew et al. (2023) and integrates with its language model probabilistic programming language, giving users a simple, programmable way to apply SMC to a broad variety of controlled generation problems. | Sequential Monte Carlo, Language Models, Semantic parsing, Bayesian inference, Probabilistic programming, SMC | We introduce a sequential Monte Carlo framework for controlling LMs at inference time via both syntactic and semantic constraints. | 12,536 | null | [
-0.06695172190666199,
-0.09970736503601074,
0.004899227526038885,
0.06267884373664856,
0.03959203138947487,
-0.035926658660173416,
-0.04199553653597832,
-0.024605117738246918,
-0.003779246238991618,
-0.033395905047655106,
-0.02965357154607773,
-0.08184044063091278,
0.08553116768598557,
-0.03957613185048103,
0.05145466700196266,
0.044643696397542953,
0.07976928353309631,
0.05436047166585922,
-0.07936964184045792,
-0.06328597664833069,
0.04444596543908119,
0.10915609449148178,
0.03272946551442146,
-0.013855282217264175,
0.030464911833405495,
0.0002001505345106125,
0.03020494244992733,
-0.016065185889601707,
0.019982455298304558,
0.020916730165481567,
0.03036084957420826,
0.027595078572630882,
-0.01281746756285429,
-0.0021300760563462973,
0.05865937098860741,
0.03724021092057228,
-0.13549625873565674,
-0.05561631917953491,
0.016923412680625916,
-0.03792668133974075,
-0.04215102270245552,
0.024337731301784515,
0.014549809508025646,
0.008933190256357193,
0.08587086945772171,
-0.025948623195290565,
-0.002547414507716894,
0.0191989466547966,
-0.02067704126238823,
0.029383087530732155,
-0.0946284681558609,
-0.06351123005151749,
-0.026736188679933548,
-0.018624361604452133,
-0.023043157532811165,
0.03006809391081333,
0.0037655923515558243,
-0.026775872334837914,
-0.02558106929063797,
-0.07708688825368881,
-0.09971866756677628,
-0.09386841207742691,
-0.060067303478717804,
-0.04762393981218338,
-0.02373780682682991,
-0.04973462596535683,
0.05290266498923302,
0.1044248715043068,
0.030797038227319717,
0.09565435349941254,
-0.05999498814344406,
0.08123572170734406,
-0.015421614050865173,
0.0995200052857399,
0.01902000978589058,
0.07261759042739868,
-0.004661413840949535,
-0.08157562464475632,
0.053744565695524216,
-0.05354933813214302,
0.011857864446938038,
-0.03227787837386131,
0.08521268516778946,
0.0328156054019928,
0.06692425161600113,
-0.013708299025893211,
0.10000737011432648,
0.09793592244386673,
0.048823535442352295,
0.06317611783742905,
-0.03605418652296066,
-0.09027580916881561,
0.07680151611566544,
-0.01094382256269455,
-0.08705231547355652,
0.08331068605184555,
0.017374221235513687,
-0.09614252299070358,
0.015608548186719418,
0.08697401732206345,
0.036535657942295074,
0.06971414387226105,
0.06979676336050034,
-0.06762465834617615,
-0.014399328269064426,
-0.06559562683105469,
0.06221956014633179,
0.07044863700866699,
-0.037373289465904236,
-0.07138088345527649,
0.06163772940635681,
0.05272554233670235,
0.019854791462421417,
-0.07635180652141571,
-0.029085377231240273,
0.011361734010279179,
0.013050662353634834,
-0.07879985868930817,
0.04143335670232773,
0.031165365129709244,
-0.0619063675403595,
0.022599924355745316,
-0.020154058933258057,
0.04563235118985176,
-0.055402617901563644,
-0.03559601679444313,
-0.06314386427402496,
1.68199098936261e-33,
0.0995282530784607,
0.0004266778996679932,
0.04594925791025162,
0.053146589547395706,
0.07492978125810623,
0.039702799171209335,
0.015440271236002445,
-0.019834427163004875,
-0.09743674099445343,
-0.0028901181649416685,
0.04248219355940819,
-0.015240304172039032,
-0.08950916677713394,
-0.020902128890156746,
0.10387304425239563,
-0.03483903408050537,
-0.04549392685294151,
0.11129774898290634,
-0.03413550928235054,
0.05724022164940834,
-0.04040570929646492,
-0.010212673805654049,
-0.020536884665489197,
-0.01713394746184349,
-0.0008088673348538578,
0.07944194972515106,
0.010961300693452358,
-0.0466705821454525,
-0.04643935710191727,
-0.0029524213168770075,
-0.08668392151594162,
0.019558651372790337,
0.011068419553339481,
0.05981677025556564,
0.02596316672861576,
-0.08116544783115387,
-0.045736923813819885,
-0.10610178858041763,
-0.016379760578274727,
-0.045340102165937424,
-0.02454346790909767,
0.030844641849398613,
0.08030208200216293,
0.029718326404690742,
-0.09132765233516693,
-0.04917716979980469,
-0.014997883699834347,
0.009173406288027763,
-0.0028971817810088396,
-0.07174365222454071,
0.0959351658821106,
0.03640235587954521,
0.04552103951573372,
-0.0033948360942304134,
0.020792292430996895,
-0.009829468093812466,
-0.025644641369581223,
0.005478421226143837,
0.030721979215741158,
0.0437040776014328,
0.005845547653734684,
0.03243989869952202,
-0.002931060502305627,
0.07125119119882584,
0.0472290962934494,
0.081997349858284,
-0.07614344358444214,
0.05171841010451317,
0.08313743770122528,
-0.02790684811770916,
-0.02322324924170971,
-0.050681158900260925,
-0.008235017769038677,
-0.04018063843250275,
0.0537642277777195,
0.036994319409132004,
0.04356345161795616,
-0.07145658135414124,
-0.026095524430274963,
0.036728426814079285,
0.014767279848456383,
-0.008308143354952335,
-0.008434581570327282,
0.013497785665094852,
0.007481400389224291,
-0.08484771847724915,
-0.012556664645671844,
-0.016065364703536034,
-0.025855742394924164,
-0.03380284085869789,
-0.051981039345264435,
-0.04003872722387314,
0.028914103284478188,
-0.02888936549425125,
-0.017301179468631744,
-1.9025324640180875e-33,
-0.055418651551008224,
-0.021907245740294456,
0.020729508250951767,
0.06115175038576126,
-0.0256818775087595,
-0.10835821181535721,
-0.03502451255917549,
-0.049339860677719116,
0.04756646975874901,
-0.07990513741970062,
-0.0783238410949707,
-0.05422724410891533,
0.06634493172168732,
-0.042143914848566055,
0.046511683613061905,
-0.03772000968456268,
0.0074456739239394665,
0.013011900708079338,
0.01235129777342081,
0.09033694863319397,
-0.04024527594447136,
0.10079357028007507,
-0.04833037406206131,
-0.018253862857818604,
-0.005643387790769339,
0.009723873808979988,
-0.03787871450185776,
0.04070456326007843,
-0.0023484991397708654,
-0.009060039184987545,
-0.07457983493804932,
0.050470951944589615,
-0.0424954891204834,
0.015324253588914871,
-0.10135354101657867,
0.007279749494045973,
0.035111188888549805,
0.034886233508586884,
0.0019352248637005687,
0.022957447916269302,
0.05863944813609123,
-0.0381871722638607,
-0.07357227057218552,
0.006118475459516048,
0.012098495848476887,
0.054800067096948624,
-0.06889092177152634,
0.02249581553041935,
0.02699059061706066,
-0.008626394905149937,
-0.04079645872116089,
0.023247595876455307,
-0.040686849504709244,
-0.02944936603307724,
-0.03592758998274803,
-0.12108185887336731,
-0.10286301374435425,
-0.061126239597797394,
0.08905335515737534,
-0.009637768380343914,
-0.12125269323587418,
0.009036269970238209,
0.06019854173064232,
-0.04323858022689819,
0.016295304521918297,
-0.02173849567770958,
-0.05033421888947487,
0.02461569383740425,
0.005820955615490675,
-0.038257237523794174,
0.08712190389633179,
0.007854709401726723,
-0.005468633957207203,
0.08134420216083527,
0.025176936760544777,
-0.01564321108162403,
0.00043675414053723216,
-0.06771105527877808,
-0.024892456829547882,
-0.04432593286037445,
0.05833407863974571,
0.03738487511873245,
0.015667520463466644,
0.042270831763744354,
0.04608575999736786,
0.05703025311231613,
0.02436310052871704,
0.011550283059477806,
0.06083434447646141,
0.04071621596813202,
-0.038330838084220886,
0.01616104505956173,
0.04259690269827843,
0.10064918547868729,
-0.0951223075389862,
-3.9170707566427154e-8,
0.01441404689103365,
0.07795796543359756,
0.00413910998031497,
0.015196731314063072,
0.05022934451699257,
0.005426551215350628,
-0.0031065819784998894,
0.022493811324238777,
0.024905910715460777,
-0.013468664139509201,
0.05799500271677971,
0.013453702442348003,
0.002762148855254054,
-0.08392541110515594,
-0.03466029465198517,
0.07301051169633865,
0.023934800177812576,
0.012732894159853458,
-0.019963158294558525,
-0.013036117888987064,
0.024491749703884125,
0.08108672499656677,
-0.016904132440686226,
0.05984194576740265,
0.04441337659955025,
-0.09105104953050613,
0.009447278454899788,
0.037267908453941345,
0.06185297667980194,
-0.04818946495652199,
0.00857437215745449,
-0.032321687787771225,
-0.014837423339486122,
0.023500168696045876,
0.05066165700554848,
0.05089370906352997,
0.017787333577871323,
-0.06588494777679443,
0.029233593493700027,
-0.051566895097494125,
0.043465543538331985,
0.04701467603445053,
-0.09687293320894241,
0.047544945031404495,
-0.01438844297081232,
-0.010859526693820953,
-0.04082842916250229,
-0.08197266608476639,
0.051173388957977295,
-0.028229350224137306,
0.00973194558173418,
-0.004123009275645018,
-0.007610946893692017,
-0.048906359821558,
0.028421100229024887,
0.08330197632312775,
0.016774283722043037,
0.024786392226815224,
0.028422711417078972,
0.014591154642403126,
-0.015706505626440048,
0.06451528519392014,
0.001886472455225885,
-0.04107092693448067
] |
Scaling Laws for Precision | https://openreview.net/forum?id=wg1PCg3CUP | [
"Tanishq Kumar",
"Zachary Ankner",
"Benjamin Frederick Spector",
"Blake Bordelon",
"Niklas Muennighoff",
"Mansheej Paul",
"Cengiz Pehlevan",
"Christopher Re",
"Aditi Raghunathan"
] | Oral | Low precision training and inference affect both the quality and cost of language models, but current scaling laws do not account for this. In this work, we devise "precision-aware" scaling laws for both training and inference. We propose that training in lower precision reduces the model's "effective parameter count," allowing us to predict the additional loss incurred from training in low precision and post-train quantization. For inference, we find that the degradation introduced by post-training quantization increases as models are trained on more data, eventually making additional pretraining data actively harmful. For training, our scaling laws allow us to predict the loss of a model with different parts in different precisions, and suggest that training larger models in lower precision can be compute optimal. We unify the scaling laws for post and pretraining quantization to arrive at a single functional form that predicts degradation from training and inference in varied precisions. We fit on over 465 pretraining runs and validate our predictions on model sizes up to 1.7B parameters trained on up to 26B tokens. | quantization, scaling laws, precision, language models | We model the effects of precision on language model loss scaling, both during and after training. We find that overtrained models degrade more when quantized at inference time, and that training larger models in lower precision can be optimal. | 12,529 | 2411.04330 | [
-0.02280518412590027,
-0.0446615144610405,
0.027057068422436714,
0.09494249522686005,
0.0349905900657177,
0.05653073266148567,
-0.028869183734059334,
0.031977467238903046,
0.03930889815092087,
-0.06456945091485977,
-0.048771847039461136,
-0.03631022572517395,
-0.007628241553902626,
0.03554169833660126,
-0.001505949767306447,
-0.05390152335166931,
0.031898949295282364,
0.12480383366346359,
-0.17295514047145844,
-0.05374934524297714,
0.042556263506412506,
0.004710962064564228,
0.02626986615359783,
0.039025116711854935,
0.05715298652648926,
-0.12508511543273926,
-0.04295802861452103,
-0.06347063928842545,
0.007228114642202854,
-0.035156674683094025,
0.012588049285113811,
0.019621500745415688,
0.02178722806274891,
0.04143879935145378,
-0.027814943343400955,
0.00391251128166914,
0.03575923666357994,
-0.0798768475651741,
0.07593023777008057,
-0.017804600298404694,
-0.008760000579059124,
-0.004896288737654686,
-0.008171179331839085,
0.05833398923277855,
0.040658146142959595,
0.03374207764863968,
0.03474714234471321,
0.014787305146455765,
-0.04145216941833496,
0.0033005501609295607,
-0.05278480052947998,
-0.02455398626625538,
-0.03545612469315529,
0.017047664150595665,
-0.04548320919275284,
-0.07559017837047577,
0.030718939378857613,
0.04767395555973053,
-0.014778818003833294,
0.021055947989225388,
-0.06757234036922455,
-0.10682210326194763,
-0.04841943457722664,
-0.02136451005935669,
-0.04411740601062775,
-0.03551177307963371,
0.04155901074409485,
-0.027692433446645737,
0.013573781587183475,
0.09333784133195877,
-0.013792278245091438,
0.05071099475026131,
-0.07459880411624908,
0.11144404113292694,
0.009069339372217655,
0.05205812677741051,
-0.002590800402686,
0.014640155248343945,
0.03787136822938919,
-0.048147909343242645,
-0.023335183039307594,
-0.02808845229446888,
0.027935996651649475,
-0.06844397634267807,
0.12268092483282089,
-0.012575649656355381,
0.017469022423028946,
-0.021572355180978775,
0.057458218187093735,
-0.006358999293297529,
-0.04951953515410423,
-0.007253101095557213,
0.02433566190302372,
0.042298611253499985,
0.01602771505713463,
0.03724813088774681,
-0.03628329187631607,
-0.04531679302453995,
-0.018544215708971024,
0.1063520684838295,
0.03382626920938492,
0.09589774161577225,
0.033804506063461304,
0.011589521542191505,
0.05252738669514656,
-0.036838237196207047,
0.051796235144138336,
0.011801083572208881,
0.03814021497964859,
-0.04307464137673378,
0.08359984308481216,
0.06373073905706406,
-0.056586991995573044,
-0.008498359471559525,
0.04641907289624214,
0.056307706981897354,
0.016402332112193108,
-0.005375044886022806,
-0.010091869160532951,
0.050035323947668076,
-0.06214035302400589,
0.04891456291079521,
-0.006750594824552536,
0.002786285011097789,
0.015694644302129745,
-0.043931182473897934,
-0.11431403458118439,
2.9633176679833206e-33,
0.03483761474490166,
0.04455482214689255,
0.09976304322481155,
-0.015304715372622013,
0.013871481642127037,
-0.020314209163188934,
-0.0034634703770279884,
-0.002847500378265977,
-0.017419306561350822,
-0.01933630369603634,
-0.004862674046307802,
0.02008173242211342,
-0.04839280992746353,
0.024564897641539574,
0.05547349527478218,
0.005475892219692469,
-0.038138579577207565,
0.05496961250901222,
-0.018383026123046875,
-0.017472242936491966,
0.11538407951593399,
-0.07764247059822083,
-0.027066010981798172,
-0.0164998397231102,
-0.022898485884070396,
-0.022819412872195244,
0.02897580899298191,
-0.01980609819293022,
-0.09099339693784714,
0.025408810004591942,
-0.11588074266910553,
-0.04979834705591202,
0.07159413397312164,
0.025891223922371864,
0.012212644331157207,
-0.06649426370859146,
-0.001151273725554347,
-0.019546033814549446,
0.02969091758131981,
-0.010204590857028961,
-0.01134730875492096,
0.062371838837862015,
0.03876052424311638,
-0.028308041393756866,
-0.03364600986242294,
-0.1196063756942749,
-0.0013335811672732234,
-0.021667053923010826,
-0.08940717577934265,
-0.11054088175296783,
0.06312394142150879,
-0.010548538528382778,
-0.016745708882808685,
0.0013722792500630021,
-0.0029336505103856325,
-0.013796811923384666,
0.03850927948951721,
-0.014336494728922844,
-0.008278386667370796,
0.03312407806515694,
-0.011182921007275581,
-0.034664031118154526,
-0.019572388380765915,
0.09164348989725113,
-0.014483829028904438,
-0.030855577439069748,
-0.04398617520928383,
0.04641950875520706,
-0.01301668956875801,
0.0008023267146199942,
-0.04400448128581047,
-0.022109301760792732,
-0.04523845016956329,
-0.04331464692950249,
0.08122291415929794,
0.002940281992778182,
0.074379101395607,
-0.11476196348667145,
0.04029287397861481,
0.047476306557655334,
0.005359443370252848,
0.009673223830759525,
0.004307756200432777,
-0.028830844908952713,
-0.026623187586665154,
-0.06971459835767746,
0.036770787090063095,
-0.04675466567277908,
-0.033469945192337036,
-0.0010575376218184829,
-0.0584937185049057,
-0.03745371103286743,
-0.02173442393541336,
-0.052390970289707184,
-0.016569266095757484,
-4.107407585197437e-33,
-0.08085530251264572,
0.01607363484799862,
-0.033075761049985886,
0.1380043476819992,
-0.10049928724765778,
-0.05331039056181908,
0.05732851102948189,
0.038545601069927216,
-0.03577163815498352,
-0.11993934959173203,
-0.029441162943840027,
-0.06086146831512451,
0.052331916987895966,
-0.005222296342253685,
0.0011952623026445508,
-0.02884753979742527,
-0.040649838745594025,
0.02218959666788578,
0.010324318893253803,
0.052047763019800186,
0.03398294374346733,
0.05529213696718216,
-0.069522924721241,
0.033538974821567535,
-0.0034184406977146864,
-0.014729434624314308,
-0.07270192354917526,
0.0844516009092331,
0.031278058886528015,
-0.0187232606112957,
-0.06799528747797012,
-0.021510960534214973,
0.02093307301402092,
-0.04555221274495125,
-0.05606968328356743,
-0.012999597936868668,
0.08814835548400879,
0.04698222130537033,
-0.017856502905488014,
0.0901821181178093,
0.10246321558952332,
-0.08037321269512177,
-0.07424309849739075,
-0.002011260949075222,
0.007691300008445978,
-0.06630485504865646,
-0.04172307625412941,
-0.0021898269187659025,
0.005080151837319136,
-0.011969738639891148,
0.02340458333492279,
0.034868668764829636,
-0.03026379831135273,
0.06341172754764557,
-0.03185664862394333,
-0.09000847488641739,
-0.06634195148944855,
-0.041542671620845795,
0.005343943368643522,
0.035024646669626236,
-0.06700129806995392,
0.018524719402194023,
0.013615229167044163,
-0.006316921208053827,
-0.0036799730733036995,
-0.0791686475276947,
-0.029573770239949226,
0.04975665360689163,
0.04316076263785362,
-0.030280956998467445,
0.023140931501984596,
-0.05501769855618477,
-0.0366063229739666,
0.025647610425949097,
-0.08790738880634308,
-0.05616549402475357,
0.05640517547726631,
-0.025682775303721428,
0.016101490706205368,
0.037308864295482635,
0.03648898005485535,
0.041984766721725464,
0.009434343315660954,
-0.0027470362838357687,
0.1046501025557518,
0.049948882311582565,
0.05320512130856514,
0.10121836513280869,
-0.004440389107912779,
0.12102985382080078,
-0.09472700208425522,
0.041985683143138885,
0.0006396853714250028,
-0.007318529766052961,
-0.12878671288490295,
-4.3751590794727235e-8,
-0.056039292365312576,
0.12811829149723053,
0.03192457929253578,
0.08766710758209229,
0.048898231238126755,
-0.0190757904201746,
-0.031677938997745514,
0.1292528659105301,
0.0006253872415982187,
0.0072090947069227695,
0.07317295670509338,
0.029576320201158524,
-0.016542714089155197,
-0.03375181555747986,
0.005169640760868788,
0.06964220106601715,
-0.014766904525458813,
0.008385483175516129,
-0.010708042420446873,
-0.023085778579115868,
-0.03404287248849869,
0.062005795538425446,
0.017004460096359253,
-0.026058966293931007,
0.0752032995223999,
-0.07766508311033249,
0.01844295673072338,
0.19258923828601837,
0.04724960774183273,
0.004138968884944916,
-0.048356808722019196,
-0.0007401299080811441,
0.021309535950422287,
0.00008804038952803239,
0.01797819882631302,
0.03705429285764694,
-0.0350608229637146,
0.007538847159594297,
0.006100590340793133,
-0.05997782200574875,
-0.011575346812605858,
0.02061229571700096,
-0.0708041861653328,
0.044025782495737076,
0.052447136491537094,
-0.016911344602704048,
-0.025937344878911972,
0.0010105457622557878,
0.0016687584575265646,
-0.026884837076067924,
0.025370022282004356,
0.04471087083220482,
-0.034774474799633026,
0.055955179035663605,
0.020029673352837563,
0.08278838545084,
-0.06393932551145554,
-0.0168587788939476,
-0.07759658992290497,
0.03181307762861252,
0.017103703692555428,
-0.06659752130508423,
-0.05455537140369415,
-0.0065349251963198185
] |
Context-Parametric Inversion: Why Instruction Finetuning May Not Actually Improve Context Reliance | https://openreview.net/forum?id=SPS6HzVzyt | [
"Sachin Goyal",
"Christina Baek",
"J Zico Kolter",
"Aditi Raghunathan"
] | Oral | Large Language Model's are instruction-finetuned to enhance their ability to follow user instructions and better comprehend input context. Still, they often struggle to follow the input context, especially when it contradicts model's parametric knowledge. This manifests as various failures, such as hallucinations where a model inserts outdated or unwarranted facts into its response. In this work, we observe an intriguing phenomenon: the context reliance of the model decreases as instruction finetuning progresses, $\textit{despite an initial expected increase}$. We call this phenomenon as the $\textbf{context-parametric inversion}$. This is surprising, as one would expect instruction tuning to improve the model's ability to follow input instructions. We observe this behavior on multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across multiple model families like Llama, Mistral and Pythia. We perform various controlled studies to eliminate some simple hypothesis for this observed behavior and isolate what datapoints cause this counter-intuitive behavior. We then analyze the phenomenon theoretically, to explain why context reliance varies across the trajectory of finetuning.
We tie the observed context-parametric inversion to the properties of the finetuning data, which provides us with some potential mitigation strategies that provide limited but insightful gains. | Instruction finetuning, context-vs-parametric reliance | We highlight a surprising phenomenon, where the context reliance of the model decreases unexpectedly, with instruction finetuning, despite an initial increase. | 12,499 | 2410.10796 | [
0.02301429957151413,
-0.08786125481128693,
0.04208764061331749,
0.03361399844288826,
-0.004300509579479694,
-0.04522928595542908,
0.0048074680380523205,
-0.014354490675032139,
0.08749780803918839,
-0.015223857015371323,
0.03223327919840813,
0.006410186178982258,
0.03412650525569916,
-0.053202930837869644,
0.05015208572149277,
-0.022521717473864555,
0.07297118008136749,
0.026291755959391594,
-0.1750001609325409,
-0.038580235093832016,
0.05117511749267578,
0.022790012881159782,
-0.014375059865415096,
0.010769212618470192,
-0.019665319472551346,
0.039277683943510056,
-0.08920399844646454,
-0.05303458496928215,
0.08107440173625946,
-0.044894248247146606,
-0.0240662582218647,
0.1164088100194931,
-0.000059094014432048425,
0.060716662555933,
-0.031993184238672256,
0.07821299880743027,
-0.041172776371240616,
0.04637303203344345,
0.004496308974921703,
-0.004724346101284027,
-0.05095525458455086,
0.023314453661441803,
0.03948961943387985,
0.03993947431445122,
0.03452284261584282,
-0.04811758175492287,
-0.05529755353927612,
-0.016538355499505997,
-0.0458996407687664,
0.005400037858635187,
-0.10391223430633545,
0.018427127972245216,
0.007465997710824013,
-0.021391866728663445,
-0.0021239733323454857,
0.008088282309472561,
0.09225786477327347,
0.10933659225702286,
0.02218797616660595,
-0.031362470239400864,
-0.09460872411727905,
-0.0683157742023468,
-0.12498010694980621,
-0.0019358329009264708,
-0.00005450864409795031,
0.011824492365121841,
-0.008065029978752136,
-0.02498292550444603,
-0.012288011610507965,
0.06853717565536499,
0.0009143299539573491,
0.007726833689957857,
-0.049970973283052444,
0.0738285556435585,
0.02304045669734478,
0.024329278618097305,
0.055073775351047516,
-0.023852191865444183,
0.015760494396090508,
-0.08842319995164871,
0.09446645528078079,
-0.014920706860721111,
0.036903705447912216,
-0.04525717720389366,
0.076294906437397,
-0.04185967892408371,
-0.00019980259821750224,
-0.0015623288927599788,
0.03914051875472069,
0.03716619312763214,
0.031168466433882713,
-0.08897548168897629,
-0.025836270302534103,
0.0009892989182844758,
0.11959512531757355,
0.06670408695936203,
0.053330786526203156,
-0.0524752102792263,
-0.045129429548978806,
0.07107659429311752,
0.0688343271613121,
0.04048800468444824,
-0.03155847638845444,
-0.019809665158391,
0.008234567940235138,
-0.07724201679229736,
0.062996044754982,
-0.004536454100161791,
-0.018572716042399406,
-0.07545289397239685,
0.008418072015047073,
0.03834998980164528,
-0.005910968873649836,
0.0015067438362166286,
-0.030052533373236656,
-0.05392323434352875,
0.04351555556058884,
-0.04398303106427193,
-0.05688481777906418,
0.07038586586713791,
0.008258430287241936,
0.05996192991733551,
0.004135545343160629,
0.005387807264924049,
-0.023725278675556183,
-0.08393960446119308,
-0.10783107578754425,
1.2391436046639719e-33,
0.07434570789337158,
-0.003432969097048044,
-0.021116219460964203,
-0.0005810526781715453,
0.025645893067121506,
0.008030599914491177,
-0.04783206433057785,
-0.0024935228284448385,
0.03092159703373909,
0.02941352128982544,
0.02018951065838337,
0.02880174294114113,
-0.05026041343808174,
0.05307149887084961,
0.05064816772937775,
0.07289696484804153,
-0.05855424329638481,
0.04958232119679451,
-0.08887922018766403,
0.006289443001151085,
0.07920820266008377,
-0.00805682223290205,
-0.05607457458972931,
-0.055686600506305695,
-0.07261466234922409,
0.041838694363832474,
0.03802207112312317,
0.0055553605780005455,
-0.09310611337423325,
0.05072227865457535,
-0.0870995745062828,
0.01843375153839588,
-0.06532007455825806,
0.029053427278995514,
0.020171739161014557,
-0.08507761359214783,
0.08124039322137833,
-0.045877598226070404,
0.05532371625304222,
-0.06266430020332336,
-0.0013793464750051498,
0.0444081649184227,
0.015127137303352356,
0.034326303750276566,
-0.05503109097480774,
-0.08773665130138397,
0.032327935099601746,
-0.00020419471547938883,
-0.035774607211351395,
0.0015016483375802636,
-0.005233617965131998,
0.09871920198202133,
0.044751398265361786,
-0.05891074612736702,
0.002290040487423539,
0.0028822175227105618,
0.024357357993721962,
-0.0028321328572928905,
0.01003286987543106,
0.10442806780338287,
0.04374301806092262,
0.030925888568162918,
-0.039621952921152115,
0.101699598133564,
0.022664636373519897,
0.022886324673891068,
-0.08487055450677872,
0.05696886405348778,
0.046843450516462326,
-0.019862299785017967,
-0.0687326043844223,
-0.04022503271698952,
-0.06588657945394516,
0.013222270645201206,
0.007878120988607407,
-0.05393105745315552,
-0.0057384781539440155,
-0.03184809535741806,
-0.035860538482666016,
-0.0005726764211431146,
0.0869303047657013,
-0.01334084291011095,
0.02799289859831333,
-0.020111290737986565,
-0.0878894254565239,
-0.022253714501857758,
0.03426613286137581,
-0.06134532764554024,
-0.0667518898844719,
-0.02512482739984989,
0.04045885056257248,
-0.03820151835680008,
-0.03401382640004158,
-0.029478954151272774,
0.023180974647402763,
-2.3162807615686084e-33,
-0.022422216832637787,
-0.031213773414492607,
-0.03119313158094883,
0.030640926212072372,
-0.008490366861224174,
-0.05986979976296425,
0.08203140646219254,
0.029738012701272964,
-0.06488269567489624,
-0.0979541763663292,
-0.0224374420940876,
-0.008729401044547558,
0.007226982153952122,
0.017507947981357574,
0.045268408954143524,
-0.014433721080422401,
0.00885763205587864,
0.03433950990438461,
0.04219313710927963,
0.011289689689874649,
0.00812538992613554,
0.03227611258625984,
-0.10595326125621796,
0.0011860617669299245,
-0.03626803681254387,
0.005389116704463959,
-0.12315910309553146,
0.05746082589030266,
0.022530732676386833,
0.05902507156133652,
-0.0537012442946434,
0.07684294879436493,
0.026506172493100166,
0.056192703545093536,
-0.04377086088061333,
0.07234989106655121,
-0.023743433877825737,
-0.0065237958915531635,
-0.078879714012146,
0.05950377508997917,
0.07112867385149002,
0.03034452721476555,
-0.00779997743666172,
-0.0031494838185608387,
0.014799538068473339,
-0.007535753306001425,
-0.04820892587304115,
-0.06512505561113358,
-0.008816729299724102,
-0.057957958430051804,
0.032760702073574066,
-0.011272544972598553,
0.02025134302675724,
0.024802571162581444,
-0.07879209518432617,
-0.06367622315883636,
-0.011436804197728634,
-0.09864150732755661,
0.04716236889362335,
0.020154278725385666,
0.005316386930644512,
0.009272299706935883,
-0.01012345403432846,
-0.07741662114858627,
0.033642422407865524,
0.0023678813595324755,
-0.01816580444574356,
-0.05349914729595184,
0.03613804280757904,
-0.013777260668575764,
0.050125036388635635,
-0.03249619901180267,
-0.04805031418800354,
-0.03962557017803192,
-0.022273095324635506,
0.07638657093048096,
-0.03840072825551033,
-0.053191155195236206,
-0.09563100337982178,
-0.017168078571558,
-0.021489787846803665,
-0.01158888079226017,
0.024679409340023994,
0.04286549985408783,
-0.01890089176595211,
0.07094496488571167,
-0.03065568394958973,
0.11742927879095078,
-0.05071068927645683,
0.05167366564273834,
-0.05167528614401817,
0.01846139132976532,
-0.05014035850763321,
0.08525969088077545,
-0.018297096714377403,
-5.082781129317482e-8,
-0.046495094895362854,
0.024156084284186363,
0.0001616146182641387,
0.11486458033323288,
0.06538134068250656,
0.001810809480957687,
0.004284337628632784,
0.06135499104857445,
-0.02260511741042137,
-0.01751783862709999,
-0.018214646726846695,
0.012162286788225174,
0.000853601610288024,
-0.06497005373239517,
0.0860898569226265,
0.13334104418754578,
0.023109300062060356,
0.01731193996965885,
-0.01934795267879963,
0.017735252156853676,
0.1146206483244896,
0.03359642252326012,
0.010739635676145554,
-0.02914571762084961,
-0.01650802418589592,
0.023183854296803474,
-0.03715381771326065,
0.11973469704389572,
0.0517021119594574,
-0.016916263848543167,
0.08909349143505096,
-0.028717083856463432,
-0.05900264158844948,
-0.07330748438835144,
-0.09728209674358368,
0.05925382301211357,
0.027317553758621216,
-0.07794812321662903,
0.04623500630259514,
-0.06830727308988571,
0.02717079222202301,
-0.011522980406880379,
-0.05131400749087334,
0.09323666989803314,
0.04644854739308357,
0.004170011263340712,
-0.018818911164999008,
-0.061316363513469696,
0.0040487200021743774,
-0.04736926406621933,
-0.004376555792987347,
0.0010504956590011716,
-0.03607650101184845,
0.056909602135419846,
0.10230785608291626,
0.010490828193724155,
-0.018068520352244377,
0.009452695026993752,
-0.05748366937041283,
0.07325373589992523,
0.06362895667552948,
0.07872369885444641,
-0.04916439950466156,
-0.02610817924141884
] |
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 318
Size of downloaded dataset files:
13.6 MB
Size of the auto-converted Parquet files:
13.6 MB
Number of rows:
3,705