4 Front. Comput. Sci.,2024,0(0): 1–42
developedanumberofmodulestoenhanceLLMs. profilesaremanuallyspecified. Forinstance,ifone
Inthissection,weproposeaunifiedframeworkto would like to design agents with different person-
summarize these modules. In specific, the overall alities, he can use "you are an outgoing person"
structure of our framework is illustrated Figure 2, or "you are an introverted person" to profile the
whichiscomposedofaprofilingmodule,amemory agent. Thehandcraftingmethodhasbeenleveraged
module, aplanningmodule, andan actionmodule. in alot ofpreviouswork toindicate theagent pro-
The purpose of the profiling module is to identify files. Forexample,GenerativeAgent[22]describes
the role of the agent. The memory and planning theagentbytheinformationlikename,objectives,
modules place the agent into a dynamic environ- andrelationshipswithotheragents. MetaGPT[23],
ment, enabling it to recall past behaviors and plan ChatDev[18],andSelf-collaboration[24]predefine
futureactions. Theactionmoduleisresponsiblefor variousrolesandtheircorrespondingresponsibili-
translating the agent’s decisions into specific out- ties in software development, manually assigning
puts. Within these modules, the profiling module distinct profiles to each agent to facilitate collabo-
impactsthememoryandplanningmodules,andcol- ration. PTLLM [25] aims to explore and quantify
lectively, these three modules influence the action personality traits displayed in texts generated by
module. Inthefollowing,wedetailthesemodules. LLMs. ThismethodguidesLLMsingeneratingdi-
verseresponsesbymanfullydefiningvariousagent
2.1.1 ProfilingModule charactersthroughtheuseofpersonalityassessment
tools such as IPIP-NEO [26] and BFI [27]. [28]
Autonomousagentstypicallyperformtasksbyas-
studies the toxicity of the LLM output by manu-
suming specific roles, such as coders, teachers and
allypromptingLLMswithdifferentroles,suchas
domainexperts[18,19]. Theprofilingmoduleaims
politicians,journalistsandbusinesspersons. Ingen-
to indicate the profiles of the agent roles, which
eral, thehandcrafting methodis veryflexible, since
areusuallywrittenintotheprompttoinfluencethe
onecanassignanyprofileinformationtotheagents.
LLM behaviors. Agent profiles typically encom-
However, itcan bealsolabor-intensive, particularly
pass basic information such as age, gender, and
whendealingwithalargenumberofagents.
career[20],aswellaspsychologyinformation,re-
flecting the personalities of the agent, and social LLM-generationMethod: inthismethod,agent
information, detailing the relationships between profilesareautomaticallygeneratedbasedonLLMs.
agents[21]. Thechoiceofinformationtoprofilethe Typically, it begins by indicating the profile gen-
agent is largely determinedby the specific applica- eration rules, elucidating the composition and at-
tionscenarios. Forinstance,iftheapplicationaims tributesoftheagentprofileswithinthetargetpop-
tostudyhuman cognitiveprocess, thenthepsychol- ulation. Then, one can optionally specify several
ogy informationbecomes pivotal. After identifying seed agent profiles to serve as few-shot examples.
thetypesofprofileinformation,thenextimportant Atlast,LLMsareleveragedtogeneratealltheagent
problemistocreatespecificprofilesfortheagents. profiles. For example, RecAgent [21] first creates
Existingliteraturecommonlyemploysthefollowing seed profiles for a few number of agents by man-
threestrategies. ually crafting their backgrounds like age, gender,
Handcrafting Method: in this method, agent personal traits, and movie preferences. Then, it