TheRiseandPotentialofLargeLanguageModelBasedAgents:ASurveyZhihengXi∗†,WenxiangChen∗,XinGuo∗,WeiHe∗,YiwenDing∗,BoyangHong∗,MingZhang∗,JunzheWang∗,SenjieJin∗,EnyuZhou∗,RuiZheng,XiaoranFan,XiaoWang,LimaoXiong,QinLiu,YuhaoZhou,WeiranWang,ChanghaoJiang,YichengZou,XiangyangLiu,ZhangyueYin,ShihanDou‡,RongxiangWeng‡,WensenCheng‡,QiZhang†,WenjuanQin,YongyanZheng,XipengQiu,XuanjingHuangandTaoGui†FudanNLPGroup,‡miHoYoIncAbstractForalongtime,humanityhaspursuedartificialintelligence(AI)equivalenttoorsurpassingthehumanlevel,withAIagentsconsideredapromisingvehicleforthispursuit.AIagentsareartificialentitiesthatsensetheirenvironment,makedecisions,andtakeactions.ManyeffortshavebeenmadetodevelopintelligentAIagentssincethemid-20thcentury.However,theseeffortshavemainlyfocusedonadvancementinalgorithmsortrainingstrategiestoenhancespecificcapabilitiesorperformanceonparticulartasks.Actually,whatthecommunitylacksisasufficientlygeneralandpowerfulmodeltoserveasastartingpointfordesigningAIagentsthatcanadapttodiversescenarios.Duetotheversatileandremarkablecapabilitiestheydemonstrate,largelanguagemodels(LLMs)areregardedaspotentialsparksforArtificialGeneralIntelligence(AGI),offeringhopeforbuildinggeneralAIagents.ManyresearcheffortshaveleveragedLLMsasthefoundationtobuildAIagentsandhaveachievedsignificantprogress.WestartbytracingtheconceptofagentsfromitsphilosophicaloriginstoitsdevelopmentinAI,andexplainwhyLLMsaresuitablefoundationsforAIagents.Buildinguponthis,wepresentaconceptualframeworkforLLM-basedagents,comprisingthreemaincomponents:brain,perception,andaction,andtheframeworkcanbetailoredtosuitdifferentapplications.Subsequently,weexploretheextensiveapplicationsofLLM-basedagentsinthreeaspects:single-agentscenarios,multi-agentscenarios,andhuman-agentcooperation.Followingthis,wedelveintoagentsocieties,exploringthebehaviorandpersonalityofLLM-basedagents,thesocialphenomenathatemergewhentheyformsocieties,andtheinsightstheyofferforhumansociety.Finally,wediscussarangeofkeytopicsandop...
