LeiWangetal. ASurveyonLargeLanguageModelbasedAutonomousAgents 23
based agentsare provided withspecific traits such databaseandkeywordsearchstrategies,specifically
astalents,preferences,andpersonalitiestoexplore designed to mitigate the hallucination issue preva-
humaneconomicbehaviorsinsimulatedscenarios. lent in such models. In addition, this model also
employs self-attention mechanism to enhance the
Social Simulation: Previously, conducting ex-
LLM’scapability viamitigatingthe impactof refer-
perimentswithhumansocietiesisoftenexpensive,
enceinaccuracies.
unethical, or even infeasible. With the ever pros-
Research Assistant: Beyond theirapplication in
peringofLLMs,manypeopleexploretobuildvir-
specializeddomains,LLM-basedagentsareincreas-
tual environment with LLM-based agents to sim-
ingly adopted as versatile assistants in the broad
ulate social phenomena, such as the propagation
fieldofsocialscienceresearch[105,114]. In[105],
of harmful information, and so on [20,34,77,79,
LLM-based agents offer multifaceted assistance,
107–110]. For example,Social Simulacra[79] sim-
ranging from generating concise article abstracts
ulates an online social community and explores
andextractingpivotalkeywordstocraftingdetailed
the potential of utilizing agent-based simulations
scriptsforstudies,showcasingtheirabilitytoenrich
toaiddecision-makerstoimprovecommunityregu-
and streamline the research process. Meanwhile,
lations. [107,108]investigatesthepotentialimpacts
ofdifferentbehavioralcharacteristicsofLLM-based in[114],LLM-basedagentsserveasawritingassis-
tant,demonstratingtheircapabilitytoidentifynovel
agentsinsocialnetworks. GenerativeAgents[20]
researchinquiriesforsocialscientists,therebyopen-
and AgentSims[34] constructmultiple agentsin a
ingnewavenuesforexplorationandinnovationin
virtual town to simulate the human daily life. So-
the field. These examples highlight the potential
cialAISchool[109]employsLLM-basedagentsto
of LLM-based agents in enhancing the efficiency,
simulateandinvestigatethefundamentalsocialcog-
creativity,andbreadthofsocialscienceresearch.
nitiveskillsduringthecourseofchilddevelopment.
S3 [77] buildsa social networksimulator, focusing
3.2 NaturalScience
onthepropagationofinformation,emotionandatti-
tude. CGMI [111] is a framework for multi-agent
Natural science is one of the branches of science
simulation. CGMImaintainsthepersonalityofthe
concernedwiththedescription,understandingand
agentsthroughatreestructureandconstructsacog-
prediction of natural phenomena, based on empir-
nitive model. The authors simulated a classroom
ical evidence from observation and experimenta-
scenariousingCGMI.
tion. With the ever prospering of LLMs, the ap-
Jurisprudence: LLM-basedagentscanserveas plicationofLLM-basedagentsinnaturalsciences
aids in legal decision-making processes, facilitat- becomes more and more popular. In the follow-
ing more informed judgements [112,113]. Blind ing, we present many representative areas, where
Judgement [113] employs several language mod- LLM-basedagentscanplayimportantroles.
els to simulate the decision-making processes of Documentation and Data Management: Natu-
multiple judges. It gathers diverse opinions and ralscientificresearchofteninvolvesthecollection,
consolidatestheoutcomesthroughavotingmech- organization,andsynthesisofsubstantialamounts
anism. ChatLaw [112] is a prominent Chinese le- ofliterature,whichrequiresasignificantdedication
galmodelbasedonLLM.Itadeptlysupportsboth of time and human resources. LLM-based agents