LeiWangetal. ASurveyonLargeLanguageModelbasedAutonomousAgents 3
summaryonthisrapidlydevelopingfieldisofgreat alsoencouragefurthergroundbreakingstudies.
significanceto comprehensivelyunderstandit and
benefittoinspirefutureresearch.
2 LLM-based Autonomous Agent Con-
In this paper, we conduct a comprehensive sur-
struction
vey of the fieldof LLM-based autonomous agents.
LLM-basedautonomousagentsareexpectedtoef-
Specifically, we organize our survey based on three
fectively perform diverse tasks by leveraging the
aspectsincludingtheconstruction,application,and
human-likecapabilitiesofLLMs. Inordertoachieve
evaluationofLLM-basedautonomousagents. For
this goal, there are two significant aspects, that is,
theagentconstruction,wefocusontwoproblems,
(1)whicharchitecture shouldbedesignedtobetter
that is, (1) how to design the agent architecture to
use LLMs and (2) give the designed architecture,
better leverage LLMs, and (2) how to inspire and
enhancetheagentcapabilitytocompletedifferent how to enable the agent to acquire capabilities for
accomplishing specific tasks. Within the context
tasks. Intuitively,thefirstproblemaimstobuildthe
of architecture design, we contribute a systematic
hardwarefundamentalsfortheagent,whilethesec-
synthesisofexistingresearch,culminatinginacom-
ondproblemfocusonprovidingtheagentwithsoft-
prehensiveunifiedframework. Asforthesecondas-
ware resources. For the first problem, we present
pect,we summarizethestrategiesforagentcapabil-
a unified agent framework, which can encompass
ityacquisitionbased onwhethertheyfine-tune the
most of the previous studies. For the second prob-
LLMs. WhencomparingLLM-basedautonomous
lem,weprovideasummaryonthecommonly-used
agents to traditional machine learning, designing
strategiesforagents’capabilityacquisition. Inaddi-
theagentarchitectureisanalogoustodetermining
tion to discussing agent construction, we also pro-
the network structure, while the agent capability
videansystematicoverviewoftheapplicationsof
acquisition is similar to learning the network pa-
LLM-based autonomous agents in social science,
rameters. Inthefollowing,weintroducethesetwo
naturalscience,andengineering. Finally,wedelve
aspectsmoreindetail.
into the strategies for evaluating LLM-based au-
tonomousagents,focusingonbothsubjectiveand
2.1 AgentArchitectureDesign
objectivestrategies.
In summary, this survey conducts a systematic RecentadvancementsinLLMshavedemonstrated
review and establishes comprehensive taxonomies their great potential to accomplish a wide range
forexistingstudiesintheburgeoningfieldofLLM- of tasks in the form of question-answering (QA).
basedautonomousagents. Ourfocusencompasses However, building autonomous agents is far from
three primary areas: construction of agents, their QA, since they need to fulfill specific roles and
applications,andmethodsofevaluation. Drawing autonomously perceiveand learnfrom the environ-
fromawealthofpreviousstudies,weidentifyvar- menttoevolvethemselveslikehumans. Tobridge
ious challenges in this field and discuss potential thegap betweentraditionalLLMsand autonomous
futuredirections. Weexpectthatoursurveycanpro- agents, a crucial aspect is to design rational agent
videnewcomersofLLM-basedautonomousagents architecturestoassistLLMsinmaximizingtheirca-
with acomprehensive background knowledge, and pabilities. Along thisdirection, previous workhas