以人为中心的大型语言模型（LLM）研究综述 -A Survey on Human-Centric LLMs

ASurveyonHuman-CentricLLMs

JINGYIWANG*,TsinghuaUniversity,China

arXiv:2411.14491v2[cs.CL]26Nov2024

NICHOLASSUKIENNIK*,TsinghuaUniversity,ChinaTONGLI,TsinghuaUniversity,China

WEIKANGSU,TsinghuaUniversity,China

QIANYUEHAO,TsinghuaUniversity,ChinaJINGBOXU,TsinghuaUniversity,China

ZIHANHUANG,TsinghuaUniversity,ChinaFENGLIXU,TsinghuaUniversity,China

YONGLI,TsinghuaUniversity,China

Therapidevolutionoflargelanguagemodels(LLMs)andtheircapacitytosimulatehumancognitionandbehaviorhasgivenrisetoLLM-basedframeworksandtoolsthatareevaluatedandappliedbasedontheirabilitytoperformtaskstraditionallyperformedbyhumans,namelythoseinvolvingcognition,decision-making,andsocialinteraction.Thissurveyprovidesacomprehensiveexaminationofsuchhuman-centricLLMcapabilities,focusingontheirperformanceinbothindividualtasks(whereanLLMactsasastand-inforasinglehuman)andcollectivetasks(wheremultipleLLMscoordinatetomimicgroupdynamics).WefirstevaluateLLMcompetenciesacrosskeyareasincludingreasoning,perception,andsocialcognition,comparingtheirabilitiestohuman-likeskills.Then,weexplorereal-worldapplicationsofLLMsinhuman-centricdomainssuchasbehavioralscience,politicalscience,andsociology,assessingtheireffectivenessinreplicatinghumanbehaviorsandinteractions.Finally,weidentifychallengesandfutureresearchdirections,suchasimprovingLLMadaptability,emotionalintelligence,andculturalsensitivity,whileaddressinginherentbiasesandenhancingframeworksforhuman-AIcollaboration.ThissurveyaimstoprovideafoundationalunderstandingofLLMsfromahuman-centricperspective,offeringinsightsintotheircurrentcapabilitiesandpotentialforfuturedevelopment.

AdditionalKeyWordsandPhrases:LargeLanguageModels,Human-CenteredComputing.

1INTRODUCTION

Aslargelanguagemodels(LLMs)

[1,

2],suchasOpenAI’sGPTfamily

[3,

4]andMeta’sLLaMA

[5,

6],continuetoevolve,theirabilitytosimulate,analyze,andinfluencehumanbehavioris

growingatanunprecedentedrate.Thesemodelscannowprocessandgeneratehuman-liketextandperformcognitivetasksatlevelscomparabletohumansinmanysituations,providingnewtoolsforunderstandinghumancognition,decision-making,andsocialdynamics.

Assuch,thissurveyaimstoprovideacomprehensiveevaluationofLLMsfromahuman-centricperspective,focusingontheirabilitytosimulate,complement,andenhancehumancognitionandbehavior,bothonanindividualandcollectivelevel.WhileLLMshavetraditionallybeenrootedincomputerscienceandengineering

[7,

8],theirincreasingsophisticationinreplicatinghuman-like

reasoning,decision-making,andsocialinteractionshasexpandedtheiruseintodomainswherehumansarethefocalpoint.Thishasallowedresearcherstoaddressquestionsthatwereoncetoointricateorabstractforcomputationalanalysis.Forexample,inpoliticalscience,LLMsareusedtoanalyzepoliticaldiscourse,detectbiases,andmodelelectionoutcomes

[9];insociology,they

assistinunderstandingsocialmediaconversations,publicsentiment,andgroupbehaviors

[10];

Authors’addresses:JingYiWang*,TsinghuaUniversity,Beijing,China,jy-w22@;NicholasSukiennik*,TsinghuaUniversity,Beijing,China,sukiennikn10@;TongLi,TsinghuaUniversity,Beijing,China,tongli@;WeikangSu,TsinghuaUniversity,Beijing,China;QianyueHao,TsinghuaUniversity,Beijing,China;JingboXu,TsinghuaUniversity,Beijing,China;ZihanHuang,TsinghuaUniversity,Beijing,China;FengliXu,TsinghuaUniversity,Beijing,China;YongLi,TsinghuaUniversity,Beijing,China,liyong07@.

J.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

2Wangetal.

andinpsychology,theyhelpmodelhumancognitionanddecision-making

[11]

.LLMshavealsorevolutionizedlinguisticsbyenablinglarge-scaleanalysisoflanguage,fromsyntaxandsemanticstopragmatics

[12],andineconomics,theyallowformodelingcomplexinteractionsbetweenpolicies

andsocietaloutcomes

[13]

.

Tostructurethisinvestigation,thesurveyisdividedintotwomainsections.First,weevaluatehuman-centricLLMs,focusingontheircognitive,perceptual,social,andculturalcompetencies.ThissectionexamineshowLLMsperformtaskscommonlyassociatedwithhumancognition,suchasreasoning,perception,emotionalawareness,andsocialunderstanding.Weassesstheirstrengthsinstructuredreasoning,patternrecognition,andcreativity,whileidentifyingtheirlimitationsinareassuchasreal-timelearning,empathy,andhandlingcomplex,multi-steplogic.BybenchmarkingLLMperformanceagainsthumanstandards,wehighlightareaswhereLLMsexcelandwherefurtherimprovementsareneeded.

Second,weexploreLLMsinhuman-centricapplieddomains,whereLLMsareusedinreal-worldscenariosthattraditionallyrequirehumaninput.Thissectionisdividedintostudiesfocusingonindividualandcollectiveapplications,whereindividual-focusedstudiesinvolveanLLMperformingtaskstypicallydonebyasinglehuman,suchasdecision-making,problem-solving,orcontentcreation,andcollective-focusedstudiesexplorehowmultipleLLMscanworktogethertosimu-lategroupbehaviors,interactions,orcollaborativetasks,offeringinsightsintosocialdynamics,organizationalbehavior,andmulti-agentcoordination.Inbothcontexts,weexaminethemethodsemployedsuchasbasicprompting,multi-agentprompting,andfine-tuning,alongwiththetheoret-icalframeworksthatguidetheseapplications,includinggametheory,sociallearningtheory,andtheoryofmind,etc.

Ultimately,thissurveyseekstoprovideadetailedunderstandingofhowLLMscanbetteralignwithhumanbehaviorsandsocialcontexts,identifyingboththeirstrengthsandareasforimprovement.Figure

1

providesanoverviewofthisframework,categorizingLLMcapabilitiesintoindividualskills,suchascognition,perception,analysis,andexecutivefunctioning,andcollectiveskillslikesocialabilities,andhighlightingtheircapabilitiesinapplyingtostudiesacrossindividualdomainslikebehavioralscience,psychology,andlinguistics,andcollectivedomainsincludingpoliticalscience,economics,andsociology.Inclassifyingresearchworkswiththisframework,weofferinsightsintohowLLMscanbemademoreeffective,ethical,andrealistictoolsforresearchandpracticalapplications,whetherinindividualorcollectivehuman-centricsettings.

Themaincontributionsofthispapercanbesummarizedasfollows.

•Weprovideanin-depthevaluationofLLMcapabilitiesinhuman-centrictasks,focusingontheircognitive,perceptual,andsocialcompetencies,andcomparingtheirperformancetohuman-likereasoning,decision-making,andemotionalunderstanding.

•WeexploreLLM’scapabilitiesinhuman-centricdomains,namelyfocusingonreal-worldapplicationsinindividualandcollectivecontexts,assessingtheirabilitytoreplicatehumanbehaviorsinfieldssuchasbehavioralscience,politicalscience,economics,andsociology,bothassingle-agentmodelsandinmulti-agentsystems.

•Weidentifykeychallengesandfutureresearchdirections,includingimprovingLLMs’real-worldadaptability,emotionalintelligence,andculturalsensitivity,whileaddressingbiasesanddevelopingmoreadvancedframeworksforhuman-AIcollaboration.

Thepaperisorganizedasfollows:Section2providesanoverviewofAI-empoweredhuman-centricstudiesandLLMs,whileSection3evaluatesLLMcompetenciesacrosscognitive,perceptual,analytical,executive,andsocialskills.Section4discusseshowLLMscanbeappliedinavarietyofinterdisciplinaryscenariostobothenhanceLLMdevelopmentandassistinhuman-centeredtasks.

Section5exploresopenchallengesandoutlinesfuturedirectionsforadvancingLLMs.SectionJ.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

ASurveyonHuman-CentricLLMs3

Individual

Collective

Domains

Skills

Cognition

LLM

BehavioralScience

ExecutiveFunction

PoliticalScience

Psychology

Linguistics

Perception

Sociability

Sociology

Economics

Analysis

Fig.1.OurframeworkdepictshowLLMsareevaluatedonfoundationalhuman-likeskills,dividedintoindi-vidual(e.g.,cognition,perception,analysis,executivefunctioning)andcollective(e.g.,sociability)levels,andappliedwithinvariousfieldsofstudysimilarlycategorizedasindividual(e.g.,BehavioralScience,Psychology,Linguistics)andcollective(e.g.,PoliticalScience,Economics,Sociology)domains.

6summarizeskeyinsightsandemphasizestheimportanceofinterdisciplinarycollaborationtoenhanceLLMs’understandingofhumanbehavior.

2OVERVIEW

2.1Human-CentricArtificalIntelligence

2.1.1TraditionalAIApproachesinHuman-CentricStudies.TheapplicationofAIinvarioushuman-centeredfieldshasundergonealongprogression,nowreachingapinnaclewiththeriseofgenerativemodels,withAImethodstobeingusedinvestigatevarioushumanphenomena.however,despitetheirrelativenaivetycomparedtoLLMs,thosetraditionalmethodshavenonethelessenabledresearcherstoaddresscomplexsocialphenomenathroughcomputation.

Foralmostaslongasithasbeeninvestigated,AIhasbeenusedinareasthatarehighlyim-pactfulonsociety

[14]

.SincethenresearchershaveevaluatedthemanywaysinwhichAIcouldemulatehumanbehaviorandthoughtprocession,forexampleincognition

[15],perception

[16],

andexecutivefunction

[17]

.Morerecently,though,withtheriseofthewebandsocialmedia,AI’susescomeclosertoourday-to-daylives.Forexample,inpoliticalcommunicationresearch,thedetectionofpoliticalbiasinnewsarticleshasemergedasacriticalareaofstudy,particularlygiventheincreasingpolarizationinmediaandonlinespaces.Traditionalmethodsforpredictingpoliticalideology,basedonstatisticalmodelingandnetworkanalysis,havebecomeanurgenttaskduetothevastamountofcontentproduceddaily.Forinstance,researchby

[18]employednetwork

analysistoestimateideologicalpreferencesofsocialmediausers.Moreover,techniquesliketopicmodelingandcontentanalysishavebeenwidelyusedtoidentifybiasandmisinformationinnewsarticlesusingdata-miningmethods

[19,

20],highlightingtheuseoftraditionalAItechniquesin

understandingpoliticaldiscourse.Otherworkstackledthetaskofstancedetectionusingmethods

likerecursiveneuralnetworks[21]andclusteringalgorithms[22].Furthermore,Dezfoulietal.

[23]

exploreadversarialvulnerabilitiesindecision-makingmodels,whichiscrucialwhenconsideringJ.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

4Wangetal.

therobustnessoftraditionalbiasdetectionsystemsunderadversarialconditions.Furthermore,Dafoeetal.

[24]emphasizetheimportanceofsystemsdesignedtonavigatesocialenvironments,

suchaspoliticaldiscourse,usingmoreestablishedmulti-agentsystemsandgametheoryframe-works.Meanwhile,machineunderstandingofhumanpreferenceshasalsobeenusedtooptimizethelearningofrewardfunctionsinreinforcementlearning

[25],showingusthatAImethodsnot

onlyhelpusexplainhumanbehavior,butcanbenefitbyunderstandingthem,highlightingtheco-evolutionarynatureofadvancementsinbothAItechniquesandhuman-centricstudies.

Overall,thevastbodyofAI-empoweredhuman-centricstudiespointtotheburgeoningpotentialofusingmoreadvancedcomputationalmethods,suchasLLMs,tobothunderstandandbettersimulatehumanbehaviorandreasoningprocesses.LLMscanpresentnewopportunitiesinthefieldbysimulatinghumanbehaviorsinareaswherereal-worlddataisscarce,aswellasfacilitateinquiryintolawsanddynamicsofhumanbehaviorbasedonLLMreplicability.

2.1.2AParadigmShiftfromTraditionalAItoLLMs.TheriseofLLMshastransformednaturallanguageprocessing(NLP)andartificialintelligenceingeneralthroughkeybreakthroughsinmodelarchitecture,scale,andcapabilities.EarlymodelslikeWord2VecandGloVeusedwordembeddings,buttheintroductionoftheTransformerin2017

[26],withitsself-attentionmechanism,enabled

deepercontextualunderstandingandmarkedaturningpoint.OpenAI’sGPTseries,beginningin2018withGPT

[3],capitalizedonthis,culminatinginGPT-3

[27]andGPT-4

[28],whichdemon

-stratedunprecedentedcapabilitiesinreasoning,textgeneration,andmultimodaltasks.Meanwhile,Google’sPaLM2

[29]advancedmultilingualismandefficiency,andopen-sourcemodelslikeFalcon

[30]andBaidu’sERNIEBot

[31]broadenedaccessandspecialization

.ThesedevelopmentsreflectthegrowingimpactofLLMsacrossdiversedomains,frominterdisciplinaryresearchtoethicalAIapplications.

TherapidadoptionofLLMsacrossacademicdisciplineshasledtovaryingpredictionsaboutwhetherthesesystemswilleventuallymatchhumancognitiveabilities.WhilesomeexpertsforeseeAIachievinghuman-likegeneralintelligenceinthenearfuture,othersremainmorecautious,doubtingwhetherAIcanfullyreplicatethecomplex,abstractreasoningandcreativitythatdefinehumancognition

[32]

.Despitethesedifferingviewpoints,AIisalreadyasignificantforceineverydaylife,influencingdecision-makingandinformationprocessingacrossnumerousdomains.However,akeydistinctionremains:humancognitionisdrivenbyforward-thinking,theory-basedreasoning,whileAIoperatesonpatternsderivedfromvastdatasets,oftenrelyingonprobabilityandpastdata

[33].ThisdifferenceunderscoresthecomplementarynatureofhumanandAIsystems,

witheachexcellingindistinctaspectsofcognitiveprocessing.

Unlikehumanintelligence,LLMsoperatewithoutinherentgoals,values,oremotionalexperi-ences.Humancognition,drivenbysurvival,socialinteraction,andcreativity,isdeeplyconnectedtoourphysicalandsocialenvironments.EvenembodiedAI,whilecapableofinteractingwithitssurroundings,lacksthenuanced,purpose-drivenintelligencethatdefineshumanthought.Incontrast,LLMsgenerateresponsesbasedonprobabilisticmodelsderivedfromlargedatasets,with-outthelivedexperiencesthatinformhumandecision-making.ThoughLLMscansimulatecertainhuman-likebehaviors,theystillfallshortoftheembodiedunderstandinghumanspossess.

ThesedistinctionsraisecriticalquestionsaboutthelimitationsandpotentialsofAI,especiallyasweconsiderthediversecapabilitiesexploredinSection

3,whichdiscussesthecapabilitiesofLLMs

includingcognitive,perceptual,social,analytical,executive,cultural,moral,andcollaborativeskills.Section

4

delvesintohowinterdisciplinaryfields,suchaspoliticalscience,economics,sociology,behavioralscience,psychology,andlinguistics,contributetoLLMdevelopment,offeringinsights

J.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

ASurveyonHuman-CentricLLMs5

intohowhumanintelligenceinformsandshapestheevolutionofartificialsystems.Thisexplo-rationemphasizestheimportanceofleveragingLLMstrengthswhilerecognizingthefundamentaldifferencesbetweenhumanandartificialcognition.

3EVALUATIONOFHUMAN-CENTRICLLMS

Toevaluatehuman-centricLLMs,weshowcaseaholisticrepresentationofLLMcompetencies,categorizedintotwodomains:individual(e.g.,cognitive,perceptual,analytical,executivefunc-tioningskills)andcollective(e.g.,socialskills),asshowninFigure

2.

ThisrepresentationincludesvariouskeyLLMskills,suchasreasoning,patternrecognition,spatialawareness,adaptability,decision-making,interpersonalcommunication,andculturalcompetency.Followingthis,Figure

3

outlinestheevaluationapproachesusedtoassessLLMs,includingbenchmarkanddatasettest-ing,human-centricevaluations,interactiveandsimulation-basedevaluations,ethicalandbiasassessments,andlastly,explainabilityandinterpretabilityevaluations.Table

1

highlightsboththestrengthsandareasforimprovementinthesedomains.Byoutliningtheseabilities,weprovideacomprehensivecomparisonofhuman-likeskills,usingbenchmarkstoassesstheirstrengthsandlimitations.Additionally,AppendixTables

2

and

3

provideacomprehensiveoverviewofkeypapers,highlightingtheircontributions,theLLMsassessed,andcomparisonstohumanperformance.Thesubsequentsectiondelvesintoeachcategory,providinganin-depthexplorationoftheskillsandbenchmarksthatdefineLLMperformanceacrossthesedomains.

cuttural

competene"

O

C入

O

Recognition

Pattern

Individual

InformationProcessing

Fig.2.OverviewofLLMCapabilitiesAcrossIndividualandCollectiveDomains.

J.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

6Wangetal.

3.1CognitiveSkills

LLMsdemonstratecognitivecompetenciesthatmirrorkeyelementsofhumanintelligence,primarilythroughreasoningandlearning.WhileLLMsshowremarkableabilityinprocessingvastamountsofinformationandgeneratingcoherentresponses,theirproficiencyvarieswhenitcomestocomplexcognitivetasks.Thesemodelsshowcaseevolvedabilitiesinstructuredreasoningandgeneralizationbutencounterchallengeswhenfacedwithintricatelogicorlearningfromreal-timeinteractions.ThissectionexploresthestrengthsandlimitationsofLLMsinreasoningandlearning,highlightingtheirprogressandareasthatrequirefurtheradvancement.

3.1.1Reasoning.Logicalreasoning,acoreelementofhumancognitionandessentialfordailyfunctioning,consistsofvarioustypesofreasoning,includingdeductive,inductive,andcausalrea-soning,eachcontributingtohowweprocessinformationandmakedecisions.Deductivereasoningappliesgeneralprinciplestoobtainspecificconclusions,whileinductivereasoningdrawsgeneral-izationsfromspecificobservations

[34],andcausalreasoninghelpstounderstandcause-and-effect

relationships

[35,

36]

.

SeveralbenchmarkdatasetshavebeendevelopedtoassessthesereasoningcapabilitiesinLLMs.Fordeductivereasoning,theLogiQA2.0dataset

[37]isanotableresource,focusingonfivetypes

ofreasoning,includingcategorical,necessaryconditional,sufficientconditional,conjunctive,anddisjunctivereasoning.PrOntoQA

[38]alsoevaluatesdeductivereasoningthroughfirst-orderlogic

taskswhereLLMsderivespecificconclusionsfromlogicalpremises.Forinductivereasoning,CommonsenseQA2.0

[39]requiresgeneralizationfromeverydayfactsandcommonsenseknowl

-edge,whereastheCreakdataset

[40]furthertestsLLMs’abilitytogeneralizefromcommonsense

knowledgetoidentifyinconsistencies.Inturn,causalreasoningisassessedusingCausalBench

[41],

whichevaluatesLLMs’abilitytoreasonaboutcause-and-effectrelationshipsacrossdiversedo-mains.ContextHub

[42],ontheotherhand,servesasanotherbenchmarkfocusingonLLMs’causal

reasoninginbothabstractandcontextualizedtasks.AdditionaldatasetslikeGSM8K

[43]and

BIG-Bench-Hard

[44]arefurthermoreemployedformathematicalreasoningandevaluatingLLM

performanceacrossvariousreasoningdomains,respectively.

AnalyzingLLMperformancewiththesedatasetshasrevealedsignificantinsightsintotheirreasoningabilitiesandlimitations.Fordeductivereasoning,althoughLLMslikeGPT-3havemadeprogress,theiraccuracyremainsat68.65%intasksinvolvinglogicalinference,whichissignificantlybelowthe90%humanbenchmark

[37]

.Thisgapindicatesongoingchallengesinmasteringcomplexlogicalstructures,especiallywhenmultiplelogicalstepsorintricatereasoningprocessesarerequired.LLMslikeGPT-3.5,PaLM,andLLaMAperformwellonsimplerdeductivereasoningtasksbutstrugglewithmorecomplexscenariosthatinvolvechainingmultiplelogicalpremisestogether

[45]

.Forinductivereasoning,ontheotherhand,GPT-4showsimprovementsinruleapplicationwithupto99.5%partialaccuracy

[46],yetstruggleswithlargerproblemsandminimal

examples.EvenwithChain-of-Thought(CoT)prompting,GPT-4andDavincifacedifficultiesinrulevalidationandintegratingcomplexrules,withDavinci’saccuracydecliningto51%innuancedtasks

[47]

.Inaddition,Hanetal.

[47]evaluateGPT-3.5andGPT-4onpropertyinductiontasks,

highlightingthatwhileGPT-4morecloselyalignswithhumanreasoningpatterns,theystillstruggletofullycapturepremisenon-monotonicity,acriticalelementofhumancognitiveprocessing.

CausalreasoningremainsasignificantchallengeforLLMslikeGPT-4andDavinci,asitrequiresadeepunderstandingofcause-and-effectacrossvariouscontexts.Althoughthesemodelsshowreasonableproficiencyinmathematicalcausaltasks,theCausalBenchbenchmarkhighlightstheirstruggleswithmorecomplextext-basedandcoding-relatedcausalproblems

[41].Interpretingcausal

structuresinnarrativesorcodesnippetsoftengoesbeyondsimpledatacorrelations,demanding

robustreasoningtoavoidproducingmisleadingoutputs.EvenwhenGPT-4initiallyperformswell,J.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

ASurveyonHuman-CentricLLMs7

Interactive&

Simulation-BasedEvaluations

.Single-AgentSimulations.Multi-AgentSimulations.Task-OrientedDialogues

Human-CentricEvaluations

.ExpertEvaluations.Crowdsourced

Evaluations

.Human-in-the-Loop

Testing

Ethical&BiasAssessments

.BiasDetection

.FairnessMetrics

.EthicalCompliance

Benchmark&

DatasetTesting

.Standardized

Benchmarks

.CustomBenchmarks.PerformanceMetrics

Explainability&Interpretability

.TransparencyofReasoning

.UserInterpretability

.TechnicalInterpretability

LLM

Evaluations

Fig.3.OverviewofLLMevaluations.

itsreasoningcapabilitiesfrequentlyweakenwhenfacedwithflawedorconflictingarguments,raisingconcernsaboutitsconsistencyincomplexscenarios

[48]

.

TheContextHubbenchmarkisdevelopedtoassessLLMslikeGPT-4,PaLM,andLLaMAinhandlingbothabstractandcontextualizedlogicalproblems

[42]

.ContextHubfocusesonthechallengesthesemodelsencounterwhentransitioningfromsimplelogictaskstonuanced,real-worldreasoning.Whilemodelsperformwellwithstraightforwardproblems,theyoftenstruggletogeneralizeincontext-richscenariosrequiringdeeperinterpretativeskills.AdditionaldatasetslikeGSM8Kemphasizedeductivereasoning,andBIG-Bench-Hardevaluatesmulti-stepreasoning,factualknowledge,andcommonsenseunderstanding

[43,

44]

.Together,thesebenchmarksrevealcriticalinsightsintothestrengthsandlimitationsofmodelslikeGPT-4andDavinci,pinpointingareasthatneedimprovementforhandlingcomplex,real-worldreasoningtasks.

Overall,thesebenchmarkdatasetsprovideacomprehensiveevaluationframeworkforassessingLLMs’reasoningcapabilities,revealingboththeiradvancementsandlimitations.WhileLLMshaveshownprogressinhandlingspecificreasoningtasks,theycontinuetofacesignificantchallengesinmulti-steplogic,contextualproblem-solving,andgeneralizingtheirreasoningabilitiesacrossdiversedomains.

3.1.2Learning.LLMs’learningabilityencompassestheircapacitytoadapt,generalize,andimproveperformancebasedonpre-existingtrainingdataandinteractionswithusersorenvironments.Unliketraditionallearningmodels,LLMsdonotupdatetheirparametersduringinference.Instead,theyrelyonpre-trainedknowledgetoperformfew-shotorzero-shottasks,highlightingtheirgeneralizationcapabilities.However,thiscomeswithsignificantlimitationswhenfacedwithevolving,real-worlddata.

RecenteffortshaveaimedatimprovingLLMadaptabilitythroughvariousstrategies.Forinstance,theRLwithGuidedFeedback(RLGF)framework

[49]optimizeslearningfromfeedback,showing

thatguidedstrategiescansignificantlyimprovetextgenerationindynamicconditions.Similarly,error-drivenlearningapproaches,likeLEMA(LearningfromMistAKes)

[50],allowmodelslike

GPT-4torefinereasoningbyidentifyingandcorrectingerrors.Theseapproacheshighlightthepotentialofleveragingfeedbackanderrorcorrectiontoboostadaptability,yettheystillrelyonstaticdataatinference.

J.ACM,Vol.V,No.N,Article.Publicationdate:November2024.

8Wangetal.

Analysis

Cognition

Perception

Sociability

High

accuracyininformation

retrieval

Structuredmetadata-based

queries

Highvolumeofideasin

structuredtasks

Nuancedemotionalregulation

ExecutiveFunction

Cognition

Real-world,dynamic

challengeadaptation

Entity-basedreasoning

with

structureddatasets

Abstractlogic

reasoninginstructured

contexts

Contextualcue-basedreasoning

Abstractcommon-sense

reasoning

Contextuallogical

reasoning

Contradictorytaskhandling

Multi-step

reasoningwithreal-world

application

Structured,predefinedtask

handling

Context-specificempathy

Complex

Perception

ExecutiveFunction

Dynamicplanning

Real-time

adjustments

Controlledvirtual

environ-

ments

understanding

Socialcontextnavigation

mentalstate

Sociability

Analysis

Basic

empathytasks

Falsebelief

andindirectcue

recognition

Moreoriginal,dive