32 Front. Comput. Sci.,2024,0(0): 1–42
LLMs in various downstream tasks. The afore- Forexample,onecanfirstlycollectreal-humandata
mentioned research encompasses various aspects for uncommon roles or psychology characters, and
oflargemodels,includingtraining,application,and thenleverageittofine-tuneLLMs. However,how
evaluation. However, prior to this paper, no work to ensure that fine-tuned model still perform well
has specifically focused on the rapidly emerging for thecommon rolesmay posefurther challenges.
and highly promising field of LLM-based Agents. Beyond fine-tuning, one can also design tailored
Inthisstudy,wehavecompiled100relevantworks agent prompts/architectures to enhance the capa-
onLLM-based Agents,covering theirconstruction, bility of LLM on role-playing. However, finding
applications,andevaluationprocesses. the optimalprompts/architectures isnot easy, since
theirdesigningspacesaretoolarge.
6 Challenges
6.2 GeneralizedHumanAlignment
While previous work on LLM-based autonomous
Humanalignmenthasbeendiscussedalotfortradi-
agenthasobtainedmanyremarkablesuccesses,this
tionalLLMs. InthefieldofLLM-basedautonomous
fieldisstillatitsinitialstage,andthereareseveral
agent, especially when the agents are leveraged
significantchallengesthatneedtobeaddressedin
for simulation, we believe this concept should be
its development. In thefollowing, wepresent many
discussed more in depth. In order to better serve
representativechallenges.
human-beings, traditional LLMs are usually fine-
tunedtobealignedwithcorrecthumanvalues,for
6.1 Role-playingCapability
example,theagentshouldnotplantomakeabomb
DifferentfromtraditionalLLMs,autonomousagent foravengingsociety. However,whentheagentsare
usuallyhastoplayasspecificroles(e.g.,program leveragedforreal-worldsimulation,anidealsimula-
coder, researcher and chemist) for accomplishing torshouldbeabletohonestlydepictdiversehuman
different tasks. Thus, the capability of the agent traits,includingtheoneswithincorrectvalues. Ac-
forrole-playingisveryimportant. AlthoughLLMs tually, simulating the human negative aspects can
caneffectivelysimulatemanycommonrolessuch beevenmoreimportant,sinceanimportantgoalof
asmoviereviewers, therearestillvariousrolesand simulation is to discover and solve problems, and
aspects thatthey struggleto captureaccurately. To without negative aspects means no problem to be
beginwith,LLMsareusuallytrainedbasedonweb- solved. For example, to simulate the real-world
corpus, thus for the roles which are seldom dis- society, we may have to allow the agent to plan
cussed on the web or the newly emerging roles, formakingabomb,andobservehowitwillactto
LLMsmaynotsimulatethemwell. Inaddition,pre- implement the plan as well as the influence of its
viousresearch[30]hasshownthatexistingLLMs behaviors. Basedon theseobservations,people can
may not well model the human cognitive psychol- makebetteractionstostopsimilarbehaviorsinreal-
ogycharacters,leadingtothelackofself-awareness world society. Inspired by the above case, maybe
inconversationscenarios. Potentialsolutiontothese an important problem for agent-based simulation
problems may include fine-tuning LLMs or care- is how to conduct generalized human alignment,
fullydesigningtheagentprompts/architectures[183]. thatis,fordifferentpurposesandapplications,the