Front. Comput. Sci.,2024,0(0): 1–42
https://doi.org/10.1007/s11704-024-40231-1
REVIEW ARTICLE
A Survey on Large Language Model based Autonomous
Agents
Lei Wang1, Chen Ma*1, Xueyang Feng*1, Zeyu Zhang1, Hao Yang1, Jingsen
(cid:66) (cid:66)
Zhang1, Zhi-Yuan Chen1, Jiakai Tang1, Xu Chen( )1, Yankai Lin( )1, Wayne
Xin Zhao1, Zhewei Wei1, Ji-Rong Wen1
1 GaolingSchoolofArtificialIntelligence,RenminUniversityofChina,Beijing,100872,China
©HigherEducationPress2024
Abstract Autonomous agents have long been a LLM-basedautonomousagents. Basedonthepre-
research focus in academic and industry commu- vious studies, we also present several challenges
nities. Previousresearchoftenfocusesontraining andfuturedirectionsinthisfield.
agentswithlimitedknowledgewithinisolatedenvi-
Keywords Autonomous agent, Large language
ronments,whichdivergessignificantlyfromhuman
model,Human-levelintelligence
learning processes, and makes the agents hard to
achieve human-like decisions. Recently, through
1 Introduction
theacquisitionof vastamountsof webknowledge,
largelanguagemodels(LLMs) haveshownpoten-
“Anautonomousagentisasystemsituatedwithinanda
tialinhuman-levelintelligence,leadingtoasurgein
partofanenvironmentthatsensesthatenvironmentand
researchonLLM-basedautonomousagents. Inthis
actsonit,overtime,inpursuitofitsownagendaandso
paper,wepresentacomprehensivesurveyofthese
astoeffectwhatitsensesinthefuture.”
studies, delivering a systematic review of LLM-
FranklinandGraesser(1997)
based autonomous agents from a holistic perspec-
tive. WefirstdiscusstheconstructionofLLM-based
Autonomous agents have long been recognized
autonomousagents,proposingaunifiedframework
asapromisingapproachtoachievingartificialgen-
that encompasses much of previous work. Then,
eral intelligence (AGI), which is expected to ac-
we present a overview of the diverse applications
ofLLM-basedautonomousagentsinsocialscience,
Receivedmonthdd,yyyy;acceptedmonthdd,yyyy
naturalscience,andengineering. Finally,wedelve
E-mail: xu.chen@ruc.edu.cn;yankailin@ruc.edu.cn
into the evaluation strategies commonly used for
*Bothauthorscontributeequallytothispaper.
4202
rpA
4
]IA.sc[
5v23411.8032:viXra