A Survey on LLM-based Multi-Agent System:

Recent Advances and New Frontiers in Application

Shuaihang Chen1 Yuanxing Liu1 Wei Han1 Weinan Zhang1† Ting Liu1
1
Research Center for Social Computing and Information Retrieval
Harbin Institute of Technology, China
{shchen, yxliu, whan, wnzhang, tliu}@ir.hit.edu.cn

Abstract decision-making, etc., even without training (Shinn

et al., 2023). Furthermore, generative agents of-
LLM-based Multi-Agent Systems fer natural language interfaces for interacting with
(LLM-MAS) have become a research humans, making these interactions more flexible
arXiv:2412.17481v2 [cs.CL] 7 Jan 2025

hotspot since the rise of large language

and easier to explain (Park et al., 2023). Based
models (LLMs). However, with the continuous
influx of new related works, the existing on these advantages, LLM-based Multi-Agent Sys-
reviews struggle to capture them comprehen- tems (LLM-MAS) emerged. Researchers have sur-
sively. This paper presents a comprehensive veyed these emerging works and proposed a gen-
survey of these studies. We first discuss eral framework (Guo et al., 2024). However, as the
the definition of LLM-MAS, a framework number of related studies continues to grow, some
encompassing much of previous work. We works have emerged that fall outside the scope
provide an overview of the various applications
of the original framework. In this paper, we pro-
of LLM-MAS in (i) solving complex tasks,
(ii) simulating specific scenarios, and (iii) vide a new perspective based on previous reviews
evaluating generative agents. Building on of LLM-based Multi-Agent Systems (LLM-MAS)
previous studies, we also highlight several with a focus on recent advancements and discuss
challenges and propose future directions for potential research directions. We collected 125 pa-
research in this field. pers published in top artificial intelligence confer-
ences, such as *ACL, NeurIPS, AAAI, and ICLR,
1 Introduction in 2023 and 2024, along with some unpublished
yet valuable papers from arXiv.1 Based on the
Multi-Agent Systems (MAS) have seen significant
purpose of LLM-MAS, we summarize the appli-
expansion owing to its adaptability and ability to
cation of LLM-MAS as task-solving, simulation
address complex, distributed challenges (Balaji and
for specific problems, and evaluation of genera-
Srinivasan, 2010). Compared to single-agent set-
tive agents. Figure 1 illustrates the framework we
tings (Gronauer and Diepold, 2022), MAS provide
propose for LLM-MAS application. (i) Solving
a more accurate representation of the real world, as
complex tasks. Multi-agents will naturally split
many real-world applications naturally involve mul-
tasks into subtasks, which will improve task per-
tiple decision-makers interacting simultaneously.
formance. (ii) Simulating for specific scenarios.
However, constrained by traditional reinforcement
Researchers see LLM-MAS as a sandbox for simu-
learning (RL) agent parameters and the absence of
lating problems in a specific domain. (iii) Evaluat-
general knowledge and capabilities, agents are un-
ing generative agents. Compared with traditional
able to tackle complex decision-making tasks, such
task evaluation, LLM-MAS has the capability of
as collaborating with other agents for the develop-
dynamic assessment, which is more flexible and
ment (Qian et al., 2024b). In recent years, large
harder for data leakage. For each category, we will
language models (LLMs), e.g. Llama 3 (Dubey
discuss representative LLM-MAS, resources, and
et al., 2024), and GPT-4 (OpenAI et al., 2024), have
their evaluation.
achieved notable successes, training on a massive
Compared to the previous survey (Guo et al.,
web corpus (Radford et al.). Compared with RL,
2024; Li et al., 2024d; Han et al., 2024; Gronauer
generative agents, with LLM as the core control
agents, can be better at reasoning, long-trajectory 1
The list of papers included in this survey can be found in
https://github.com/bianhua-12/Multi-generative_
†
Corresponding author. Agent_System_survey

1
 Wireless Politics
Self- Network
Refine Economics Strategy
Tiny
Collective Decision- Society
Making Emotion
Game Social
Media
Muti-stage Reinforcement Learning
Physical
Evaluating on MGAS
Social
Speed Reasoning
Optimization Framework Training
Solving Fine-Tuning on
Evaluating MGAS
complex Simulating specific
generative
Distributed Communication tasks scenarios
agents Synthesizing Data
Discussion Optimization
For Training

LLM-based Multi-
Agent Systems

Generative agent Generative agent Environment

...

LLM Profile Reasoning Memory Rules Tools Intervention

Figure 1: Overview of the application framework and relationship of LLM-MAS, generative agent, and LLM.
Dashed-bordered right-angled rectangles represent content aligned with previous surveys, while rounded rectangles
indicate original contributions introduced in this study.

and Diepold, 2022), this survey has the following Compared to traditional agents, generative
distinctive contributions: (i) A Taxonomy focusing agents need to be able to perform more complex be-
on application of LLM-MAS: we introduce a haviors, such as generating complete personalized
more recent taxonomy (taxonomy and difference blog posts based on historical information (Park
are shown in Figure 1) based on the purpose of the et al., 2022). Therefore, in addition to using LLMs
application of LLM-MAS. (ii) More Resources: as the core, generative agents also require the fol-
we analyze open-source frameworks and research lowing characteristics: (i) Profiling is used to link
works with benchmarks or datasets to facilitate the their behavior by describing roles in natural lan-
research community. (iii) Challenges and Future: guage (Gao et al., 2023b), or customizing the
we discuss the challenges in LLM-MAS, and shed prompts for each generative agent based on their
light on future research. tasks (Xu et al., 2023c). (ii) Memory is used to store
historical trajectories and retrieve relevant mem-
2 Core Components of LLM-MAS ories for subsequent agent actions, enabling the
LLM-MAS refer to a system that includes a col- ability to take long-term actions while solving the
lection of generative agents capable of interacting problem of limited LLM context windows. There
and collaborating within a shared environmental usually include three layers of memory: long-term,
setting (Wang et al., 2024c). We will discuss gener- short-term, and sensory memory (Park et al., 2023).
ative agents and the environment in the following. (iii) Planning is to formulate general behavior for
a longer period of time in the future (Yao et al.,
2.1 Generative Agents 2023). (iv) Action executes the interaction between
Generative agents refer to the components of the generative agent and the environment (Wang
LLM-MAS that have role definitions, can perceive et al., 2024a). Generative agents may be required
the environment, make decisions, and perform com- to choose one of several candidate behaviors to ex-
plex actions to change the environment (Wang et al., ecute, such as voting for whom (Xu et al., 2024), or
2024a). They can be a player in a game or a user generate behaviors without mandatory constraints,
on social media and have the role of driving the such as generating a paragraph of text (Li et al.,
development of LLM-MAS and influencing its re- 2023).
sults. Generative agents can communicate with each

2
other to achieve cooperation within the system. for external intervention systems. This intervention
The communication of generative agents can be can come from any external source, human (Wang
roughly divided into two purposes. (i) The first et al., 2024b), or a supervision model (Chen et al.,
purpose is to achieve collaboration, share the infor- 2024c), even a generative agent (Qian et al., 2024b).
mation obtained by themselves with other intelli- The purpose of an intervention may be to actively
gent agents, and to some extent, aggregate multiple read information from the system (Wang et al.,
intelligent agents into a complete system, achiev- 2024b), or passively interrupt the system to pre-
ing performance beyond independent intelligent vent uncontrolled behavior from occurring (Qian
agents (Yuan et al., 2023); (ii) The second purpose et al., 2024b).
is to achieve consensus, allowing for greater simi-
larity in behavior or strategy among some agents, 3 LLM-MAS for Solving Complex Tasks
thereby enabling faster convergence to Nash equi-
librium (Oroojlooy and Hajinezhad, 2023). Completing a complex task usually requires mul-
The type of communication content can be tiple roles, multiple steps, and so on. This is dif-
roughly divided into two types: natural language ficult for a single agent, but multiple agents work-
and custom content. Natural language forms of ing together will be well suited to this task (Islam
communication have high interpretability and flexi- et al., 2024). Further, each of these agents can be
bility. Still, they are difficult to optimize, making trained independently (Shen et al., 2024; Yu et al.,
them more suitable for pursuing consensus, such as 2024). Compared with a single agent, LLM-MAS
Chatdev (Qian et al., 2024b) and job fair systems can achieve better results. That is, the multi-agent
(Li et al., 2023). Custom content may be a vector collaboration will improve the overall performance
or a discrete signal that no one can understand ex- (Du et al., 2023).
cept for the generative agent in the system. But this
3.1 Representative LLM-MAS for Solving
form is easy to optimize using policy gradients, so
Complex Tasks
it is commonly used for achieving cooperative pur-
poses, such as the DIAL (Hausknecht and Stone, This field is currently a hot research topic. Recently,
2015) algorithm and its variables. researchers mainly focus on multi-agent reasoning
frameworks and multi-agent communication opti-
2.2 Environment mization, which will be discussed below.
Environmental settings include rules, tools, and LLM-MAS reasoning framework. We sum-
intervention interfaces: (i) Tools are responsible marize three aspects by the pipeline of reason-
for translating the agent’s action instruction into ing, including: (i) multi-stage framework, (ii) col-
specific outcomes. Generative agents send action lective decision-making framework, and (iii) self-
instructions to the environment and the environ- refine framework. That is, the multi-stage frame-
ment converts the instruction into a record that the work refers to a pipeline where agents act as se-
action was taken. There are different action spaces rial problem solvers at different stages (Qian et al.,
in different scenes. In the social media scene, the 2024b), while collective decision-making (Zhao
action space concludes “like”, “comment”, “fol- et al., 2024c) refers to different agents voting
low”, etc. (Wang et al., 2024b). In the development or debating for one goal. Self-Refine refers to
scene, the action space closes the chat chain (Qian the mechanism of self-reflection in LLM-MAS.
et al., 2024b), which is larger than social networks. Researchers propose a framework for applying
(ii) Rules define the mode of communication be- multi-agents to the natural sciences (Chen et al.,
tween generative agents or the interaction with the 2024a) to enhance data analysis, model simula-
environment, directly defining the behavioral struc- tions, and decision-making processes (Yin et al.,
ture of the entire system. Based on the scene, there 2024). Zhang et al. (2023a) propose a framework
are some special rules for the system, such as rules to achieve self-adaptation and adaptive coopera-
of the game (Xu et al., 2024; Chen et al., 2024c) and tion. Scaling law in agent cooperation is also ex-
the norm of social behavior (Park et al., 2023; Wang plored (Qian et al., 2024c), finding that there is no
et al., 2024b). Normally, a generative agent in the significant pattern.
large-scale system has a smaller action space and is LLM-MAS communication optimization. The
more easily replaced by a rule-based model (Mou fully connected communication in LLM-MAS can
et al., 2024). (iii) Intervention provides an interface lead to issues such as combinatorial explosion

3
and privacy disclosure. Based on this, we sum- are considered as indicators; in BOLAA (Liu et al.,
marize two aspects in Communication Optimiza- 2023c), the recall and QA accuracy of intelligent
tion, including: (i) speed optimization and (ii) dis- physical examination retrieval are also considered
tributed discussion. Speed optimization refers to re- as evaluation indicators; in the Werewolf game (Xu
searchers trying to speed up the communication of et al., 2023c), the win rate of virtual players is
agents, for example, with non-verbal communica- naturally also an evaluation indicator; in the job
tion (Liu et al., 2024b) or shorter generation (Chen fair system (Li et al., 2023) , the proportion of cor-
et al., 2024g). While distributed discussion refers rectly recruited target job seekers by the recruiting
to agents trying to solve tasks without enough infor- party is also an evaluation indicator; in the auction
mation (Liu et al., 2024a). Agents need to commu- system (Chen et al., 2024c), the Spearman corre-
nicate with each other to achieve their goals (Zhang lation coefficient between the predicted and actual
et al., 2023a), even without complete information prices of goods, as well as the skills of bidders, are
in one agent(Liu et al., 2024a). also measured by TrueSkill scores (Graepel et al.,
2007); in Stanford Town (Park et al., 2023), the
3.2 Resources of LLM-MAS for Solving quality of behaviors generated by virtual agents
Complex Tasks and human agents is manually sorted and evaluated
We summarize common and open-source LLM- using TrueSkill; in urban simulation systems (Xu
MAS for simulation in Table 1, including code, et al., 2023a), the success rate of completing spe-
dataset, and benchmark. cific tasks such as navigation is also an evaluation
Data set. All datasets of traditional NLP tasks are metric.
available. In addition, following ECL (Qian et al., Communication cost analysis. The paramount
2024a), Qian et al. (2024b) evaluate the quality concern lies in the operational cost of the system.
of generated software on the SRDD dataset and Given that a substantial proportion of contemporary
systematically evaluate agent capabilities in the systems incorporate LLMs as a pivotal module, the
domain of software development. additional expenditure incurred during system op-
Open source community. The open-source eration has emerged as a pivotal area of interest.
and industrial communities have also contributed As an illustrative example, Mou et al. (2024) uti-
significantly to the development of LLM-MAS. lize the actual runtime of the system as a pivotal
MetaGPT (Hong et al., 2023) assigns different metric, underscoring the significance of managing
roles to generative agents to form a collaborative this operational overhead.
entity for complex tasks. Gao et al. (2024) pro-
pose AgentScope with message exchange as its 4 LLM-MAS for Simulating Specific
core communication mechanism. In the mean- Scenarios
time, this work develops a distribution framework
that facilitates seamless switching between local This section will illustrate the application for
and distributed deployments and automatic parallel LLM-MAS in simulation. Researchers apply
optimization with minimal effort. Open AI pro- agents to simulate a certain scenario to study its
poses Swarm (Ope, 2024), an experimental multi- impact on a specific subject like social science.
agent orchestration framework that is ergonomic On the one hand, compared with rule-based meth-
and lightweight. Unlike the previously released As- ods (Chuang and Rogers, 2023), generative agents
sistants API, Swarm gives developers fine-grained with natural language communication can be more
control over context, steps, and tool calls rather intuitive for humans. On the other hand, environ-
than being hosted. ment determines the properties of the simulation,
which is the core of the entire simulation.
3.3 Evaluation of LLM-MAS for solving
complex task 4.1 Representative LLM-MAS for Simulating
Specific Scenarios
Performance on specific tasks. Shown as Table
1, the performance of LLM-MAS can be evalu- The typical scenarios for LLM-MAS simulations
ated by specific tasks, which is intuitive and con- are described as follows. We will introduce the
venient. For example, in an APP system (Zhang following work according to the subject.
et al., 2023b), the average number of steps and Social domain. Social large-scale experiments
tools used by an agent to complete a specific task in the real world have high costs, and the sheer

4
Table 1: Codes and Benchmarks in LLM-MAS for solving tasks studies. “No Code” or “No Benchmark” means the
code or benchmark is unavailable.

Field SubField Paper Code Dataset and Benchmark

(Qian et al., 2024b) Code Link SRDD
(Du et al., 2024) Code Link SRDD
(Yue et al., 2024) Code Link SMART (self)
Muti-stage (Liu et al., 2023c) Code Link WebShop
(Lin et al., 2024) Code Link FG-C, CG-O
HumanEval, EvalPlus, MBPP,
(Islam et al., 2024) Code Link
APPS, xCodeEval, CodeContest
(Shen et al., 2024) Code Link ToolBench, ToolAlpaca
Reasoning (Zhao et al., 2024c) Code Link MCQA
Framework (Cheng et al., 2024) Code Link ESConv dataset, P4G dataset
Collective (Liang et al., 2024) Code Link MT-Bench
Decision-Making (Lei et al., 2024) Code Link MATH
(Zhang et al., 2024a) Code Link MMLU, MATH, Chess Move Validity
(Wang et al., 2024d) Code Link TriviaQA
(Wang et al., 2024c) Code Link FOLIO-wiki
StrategyQA, CSQA, GSM8K, AQuA,
(Chen et al., 2024e) Code Link
Self-Refine MATH, Date Understanding, ANLI
(Chen et al., 2024a) Code Link TriviaQA
(Tang et al., 2024) Code Link Trans-Review,AutoTransform,T5-Review
(Zhang et al., 2023a) Code Link Overcooked-AI
Speed Optimization (Liu et al., 2024b) No Code HotpotQA,NarrativeQA,MultifieldQA
Communication TriviaQA, Natural Questions,
Optimization (Chen et al., 2024f) Code Link
Distributed HotpotQA, 2WikiMultiHopQA
(Liu et al., 2024a) Code Link InformativeBench

scale of social participation can sometimes esca- agent evolution method by hospital simulation. Be-
late into violence and destruction, posing poten- cause doctor agent training is both inexpensive and
tial ramifications (Mou et al., 2024). Therefore, highly effective, this work can quickly scale up
it is necessary to simulate in the virtual environ- the agent to handle tens of thousands of cases in
ment; simulation can solve the problem of exces- just a few days, a task that would take a human
sive overhead in the real environment and can sim- doctor years to complete. Pan et al. (2024) propose
ulate the process in the real world for a long time a huge scale of agent simulation, increasing the
at a faster speed (Li et al., 2024a). At the same number of agents to 106 . In social game,like Were-
time, the whole process can be easily repeated, wolf (Xu et al., 2024) , Avalon (Lan et al., 2024) ,
which is conducive to further research. Researchers and Minecraft (Gong et al., 2024) for LLM-MAS
have done a lot of work to simulate social media simulation are attempted. Further, some game com-
scenarios. Based on the social media simulation panies like Netease are also actively experimenting
archetype (Park et al., 2022), Park et al. (2023) with LLM-MAS in their games.
propose Stanford Town, which leads to a one-day
simulation of the life of 25 agents with different
occupations in a small American town. At the same
time, there was work on emotional propagation in-
fluence (Gao et al., 2023b), information cocoon Physical domain. For the physical domain,
room based on recommendation scenario research the applications for generative agent simulation in-
(Wang et al., 2024b), and study of social move- clude mobility behaviors, transportation (Gao et al.,
ments (Mou et al., 2024). Researchers propose 2023a), wireless networks, etc. However, there
Urban Generative Intelligence (UGI) (Xu et al., is limited research in the area of multi-generative
2023a) to address specific urban issues and simu- agents. Zou et al. (2023) explore the application
late complex urban systems, providing a multidis- of multiple agents in the wireless field, proposing
ciplinary approach to understanding and managing a framework where multiple on-device agents can
urban complexity. Li et al. (2024a) study doctor interact with the environment and exchange knowl-
edge to solve a complex task together.

5
4.2 Resources for LLM-MAS simulation a realistic multi-agent system should have a similar
information dissemination trend to the real world.
We summarize common and open-source
Abdelzaher et al. (2020) compare the changes in
LLM-MAS for simulation in Table 2, including
the number of events occurring each day in an on-
code and benchmarks.
line social media simulation environment; S3 (Gao
To prove the effectiveness of the simulation, that
et al., 2023b) compare the number of users who
is, to fit the reality, researchers usually evaluate the
are aware of a certain event every day, as well as
simulation system by simulating real data. There-
the changes in emotional density and support rate
fore, a realistic dataset with dense users and records
for that event every day; a similar approach is also
is very important for evaluation simulation (Mou
used in Stanford Town (Park et al., 2023).
et al., 2024). An ideal dataset will be dense: that
is, data with a smaller number of users on the same
5 LLM-MAS for Evaluating Generative
scale can better evaluate the simulation capability
Agents
of the LLM-MAS.
For Benchmark, Du and Zhang (2024) propose With LLMs prevailing in the community, how to
WWQA based on werewolf scenarios to evalu- evaluate the ability of LLMs is an open question.
ate the agent’s capability in a werewolf scenario. Existing evaluation methods suffer from the fol-
SoMoSiMu-Bench (Mou et al., 2024) provides lowing shortcomings: (i) constrained evaluation
evaluation benchmarks focusing on individual user abilities, (ii) vulnerable benchmarks, and (iii) un-
behavior and social simulation system results. objective metrics. The complexity and diversity
of LLM-MAS have indicated that LLM-MAS can
4.3 Evaluation of LLM-MAS simulation evaluate LLMs. However, how to design specific
We will discuss the evaluation based on indicators evaluation indicators and evaluation methods has
used for assessing LLM-MAS as a whole, rather puzzled researchers. Similarly, LLM-MAS can
than the capabilities of individual agents. also be used in training generative agents. We sum-
Consistency. LLM-MAS necessitate a robust marize three aspects of training: (i) Supervised
congruence with the real world to ensure the deriva- Fine-Tuning (SFT) (ii) reinforcement learning (RL)
tion of meaningful and insightful experimental out- (iii) Synthesizing data for training.
comes. In the context of simulation systems, exem-
5.1 Representative LLM-MAS for Evaluating
plified by UGI (Xu et al., 2023a), the primary objec-
Generative Agents
tive lies in faithfully replicating specific real-world
scenarios. When employed for training agents like LLM-MAS can provide rewards to agents, and
SMART (Yue et al., 2024), only those agents that these rewards can be used to evaluate or train gen-
have undergone rigorous training within a virtual erative agents, which will be discussed below.
environment that closely mirrors the real environ- Evaluation of generative agents. Researchers
ment can be deemed suitable for deployment in study generative agents by putting them into
real-world settings. Similarly, when utilized for LLM-MAS. In LLM-MAS, researchers can further
evaluation purposes, such as in AgentSims (Lin study the LLM’s strategic capabilities in different
et al., 2023), the attainment of authentic and reli- scenes, such as long strategic ability (Chen et al.,
able evaluation results is contingent upon the vir- 2024c), leadership strategy (Xu et al., 2023c) and
tual environment maintaining a high degree of con- competitiveness strategy (Zhao et al., 2024b). In
sistency with its real-world counterpart. Finally, in the emotional field, MuMA-ToM (Shi et al., 2024)
the system for collecting data such as BOLAA (Liu is used to evaluate the ability of agents to under-
et al., 2023c), consistency also ensures the validity stand and reason about human interactions in a real
of the data. Therefore, an important performance home environment through video and text descrip-
measure of LLM-MAS is its consistency with the tions.
real situation. Training on generative agents. Li et al. (2024c)
Information dissemination. Compare the dif- enhance the data to Supervised Fine-Tuning
ferences between information dissemination behav- (SFT) generative agents with LLM-MAS. Xu et al.
ior in the system and reality using time series anal- (2023c) have created generative agents to over-
ysis methods. Information dissemination can to come the intrinsic bias from LLMs by proposing
some extent reflect the nature of media; therefore, a novel framework that powers generative agents

6
Table 2: Codes and Benchmarks in LLM-MAS for simulation studies. “No Code” or “No Benchmark” means the
code or benchmark is unavailable.

Domain Subdomain Paper Code Dataset and Benchmark

(Huang et al., 2024b) No Code AdaSociety
(Chen et al., 2024b) Code Link AgentCourt
Tiny Society (Park et al., 2023) Code Link No Benchmark or Dataset
(Piatti et al., 2024) Code Link No Benchmark
(Chuang et al., 2024) Code Link No Benchmark or Dataset
Social Economics (Li et al., 2024b) Code Link No Benchmark or Dataset
(Wang et al., 2024b) Code Link Movielens-1M
Social Media (Gao et al., 2023b) No Code Blog Authorship Corpus
(Mou et al., 2024) Code Link SoMoSiMu-Bench(self)
(Du and Zhang, 2024) Code Link WWQA
Game
(Pan et al., 2024) Code Link No Benchmark or Dataset
Physical Wireless (Zou et al., 2023) No Code No Benchmark or Dataset

with multi-agent reinforcement learning (Xu et al., tics of the base model LLMs, which will be care-
2023c). For LLM-MAS, Yue et al. (2024) split fully discussed below.
complex trajectories in knowledge-intensive tasks Challenges. (i) Generalized alignment for simu-
into subtasks, proposing a co-training paradigm lation (Liu et al., 2023a). When the agents are lever-
of the multi-agent framework, Long- and Short- aged for real-world simulation, a perfect generative
Trajectory Learning, which ensures synergy while agent should be able to depict diverse traits (Wang
keeping the fine-grained performance of each agent. et al., 2024a) honestly. However, due to the train-
RLHF has been criticized for its high cost. Liu ing method of the foundation model (OpenAI et al.,
et al. (2023a) propose an alignment scheme based 2024), generative agents usually cannot be aligned
on a multi-agent system, effectively addressing in- with mock objects. (ii) Hallucination. Generative
stability and reward gaming concerns associated agents have a certain probability of hallucination
with reward-based RL optimization. Either way, in their interaction with other agents (Du et al.,
LLM-MAS are essentially viewed as an environ- 2023). Various enhancement methods can allevi-
ment in RL with different ways of getting rewards ate this problem but cannot solve it (iii) Lack of
from the environment. sufficient long text capability. When processing
complex information, generative agents forget the
5.2 Resources of LLM-MAS for evaluations input information because of the lack of long-text
Table 3 shows the work with code, dataset, and ability (Zhao et al., 2024a).
benchmark we summarize, serving as a reference Future directions. The improvement of the abil-
for future researchers. ity of a single agent or the ability of the base model
has always been a hot topic. Researchers have
6 Challenges and Future Directions focused on enhancing alignment, reducing halluci-
nation, and improving the ability of long text. The
While previous work on LLM-MAS has obtained proposal of the new generation of Open AI model
many remarkable successes, this field is still at its o1 (Int, 2024), provides researchers with new ideas,
initial stage, and there are several significant chal- that is, to use more complex reasoning to enhance
lenges that need to be addressed in its development. the ability of the model.
In the following, we outline several key challenges
along with potential future directions. 6.2 Challenges posed by interactions
Due to the complexity, autoregressive, and other
6.1 Challenges posed by generative agents
characteristics of LLM-MAS, there are many prob-
Generative agents are an integral part of lems in the practical application of the system. Two
LLM-MAS. However, the generative agents have main problems, Efficiency explosion, and Accumu-
some shortcomings due to the inherent characteris- lative Effect, are listed in the following.

7
Table 3: Codes and Benchmarks in LLM-MAS for evaluation studies. “No Code” or “No Benchmark” means the
code or benchmark is unavailable.

Domain Subdomain Paper Code Dataset and Benchmark

(Liu et al., 2023b) Code Link AGENTBENCH
(Bandi and Harrasse, 2024) No Code MT-Bench
(Chan et al., 2023) Code Link ChatEval
Strategy (Chen et al., 2024d) Code Link LLMARENA
(Xu et al., 2023b) Code Link MAgIC
Evaluation of
(Huang et al., 2024a) Code Link MLAgentBench
generative agents
(Chen et al., 2024c) Code Link AUCARENA
(Zhang et al., 2024b) Code Link PsySafe
Emotion
(Shi et al., 2024) Code Link MuMA-ToM
SFT on LLM-MAS (Li et al., 2024c) Code Link MT-Bench, AlpacaEval
Training on MARL on LLM-MAS (Xu et al., 2023c) No Code No dataset or benchmark
generative agents
HH, Moral Stories, MIC,
Synthesized Ddata (Liu et al., 2023a) Code Link
ETHICS-Deontology, TruthfulQA

Efficiency explosion. Due to their autoregres- ronments, it is difficult to obtain sufficiently de-
sive architecture, LLMs typically have slow infer- tailed, specific, and direct system evaluation indi-
ence speeds. However, generative agents need to cators from current work at the population level.
query LLMs for each action multiple times, such At present, researchers mainly compare the dis-
as extracting information from memory, making tribution of the system and real environments to
plans before taking actions, and so on. When the evaluate LLM-MAS, which lacks details for the
LLM-MAS scales up, this problem will be magni- LLM-MAS running process.
fied, especially for generative agents that have a Automated evaluation and benchmark. Dif-
large action space. SoMoSiMu-Bench (Mou et al., ferent LLM-MAS of the same kind cannot be com-
2024) replaces the edge generative agents with rule- pared because of the lack of a benchmark for
based agents, alleviating this problem. However, LLM-MAS. Further, there is a lack of a common
for LLM-MAS with a complex action space in gen- benchmark framework for both individual and total-
erative agents, this problem remains unsolved. based evaluation, that can be used to evaluate most
Accumulative Effect. Since each round of LLM-MAS.
LLM-MAS is based on the results of the previ- Future directions. Studying large-scale
ous round, and LLM-MAS have a great impact LLM-MAS will be a new research hotspot, from
on the subsequent results. Researchers have used which researchers will evaluate and discover new
a rule-based model for intermediate error correc- scale effects. In the meantime, common test bench-
tion (Chen et al., 2024c), but there is still a lot of marks and evaluation methods will also emerge in
room for improvement. IOA (Chen et al., 2024f) future research.
proposed an Internet-like communication architec-
ture, which made LLM-MAS more scalable and
enhanced the adaptability to dynamic tasks.
7 Conclusion
Future directions. Industry academia has been
In this survey, we systematically summarize ex-
making efforts to reduce the communication cost
isting research in the LLM-based Multi-Agent
of LLM-MAS, such as alignment-based method
Systems (LLM-MAS) field. We present and re-
OPTIMA (Chen et al., 2024g) and Industrialized
view these studies from three application aspects:
parallel message method AgentScope (Gao et al.,
task-solving, simulation, and evaluation of the
2024), but it is still in the basic stage and has a
LLM-MAS. We provide a detailed taxonomy to
large research space.
draw connections among the existing research,
summarizing the major techniques and their de-
6.3 Challenges of Evaluating for LLM-MAS
velopment histories for each of these aspects. In
Lack of Objective metrics for group behavior. addition to reviewing the previous work, we also
As shown in Section 4.3, due to the diversity, com- propose several challenges in this field, which are
plexity, and unpredictability of multi-agent envi- expected to guide potential future directions.

8
Limitations Yemin Shi. 2024a. AutoAgents: A Frame-
work for Automatic Agent Generation. Preprint,
We have made our best effort, but some lim- arXiv:2309.17288.
itations may still exist. Due to page limita-
tions, we can only provide a brief summary of Guhong Chen, Liyang Fan, Zihan Gong, Nan Xie, Zix-
uan Li, Ziqiang Liu, Chengming Li, Qiang Qu, Shi-
each method without exhaustive technical details. wen Ni, and Min Yang. 2024b. AgentCourt: Simulat-
On the other hand, we primarily collect studies ing Court with Adversarial Evolvable Lawyer Agents.
from *ACL, NeurIPS, ICLR, AAAI, and arXiv, Preprint, arXiv:2408.08089.
and there is a chance that we may have missed
Jiangjie Chen, Siyu Yuan, Rong Ye, Bodhisattwa Prasad
some important work published in other venues. Majumder, and Kyle Richardson. 2024c. Put Your
In application, we primarily list representative Money Where Your Mouth Is: Evaluating Strategic
LLM-MAS resources with open code in Table 1, Planning and Execution of LLM Agents in an Auc-
Table 2, and Table 3. More complete papers can tion Arena. Preprint, arXiv:2310.05746.
be found in https://github.com/bianhua-12/ Junzhe Chen, Xuming Hu, Shuodi Liu, Shiyu Huang,
Multi-generative_Agent_System_survey. We Wei-Wei Tu, Zhaofeng He, and Lijie Wen. 2024d.
recognize the timeliness of our work, and we will LLMArena: Assessing Capabilities of Large Lan-
stay abreast of discussions within the research com- guage Models in Dynamic Multi-Agent Environ-
ments. In Proceedings of the 62nd Annual Meet-
munity, updating opinions and supplementing over-
ing of the Association for Computational Linguistics
looked work in the future. (Volume 1: Long Papers), pages 13055–13077.

Acknowledgments Justin Chen, Swarnadeep Saha, and Mohit Bansal.

2024e. ReConcile: Round-Table Conference Im-
This research was supported by the National proves Reasoning via Consensus among Diverse
Key Research and Development Program (No. LLMs. In Proceedings of the 62nd Annual Meeting
2022YFF0902100), and the Nature Scientific Foun- of the Association for Computational Linguistics (Vol-
ume 1: Long Papers), pages 7066–7085, Bangkok,
dation of Heilongjiang Province (YQ2021F006). Thailand. Association for Computational Linguistics.

Weize Chen, Ziming You, Ran Li, Yitong Guan, Chen

References Qian, Chenyang Zhao, Cheng Yang, Ruobing Xie,
Zhiyuan Liu, and Maosong Sun. 2024f. Inter-
2024. Introducing OpenAI o1. https://openai.com/o1/. net of Agents: Weaving a Web of Heterogeneous
Agents for Collaborative Intelligence. Preprint,
2024. Openai/swarm. OpenAI. arXiv:2407.07061.
Tarek Abdelzaher, Jiawei Han, Yifan Hao, Andong Jing,
Dongxin Liu, Shengzhong Liu, Hoang Hai Nguyen, Weize Chen, Jiarui Yuan, Chen Qian, Cheng Yang,
David M Nicol, Huajie Shao, Tianshi Wang, et al. Zhiyuan Liu, and Maosong Sun. 2024g. Optima: Op-
2020. Multiscale online media simulation with so- timizing Effectiveness and Efficiency for LLM-Based
cialcube. Computational and Mathematical Organi- Multi-Agent System. Preprint, arXiv:2410.08115.
zation Theory, 26:145–174.
Yi Cheng, Wenge Liu, Jian Wang, Chak Tou Leong,
P. G. Balaji and D. Srinivasan. 2010. An Introduction Yi Ouyang, Wenjie Li, Xian Wu, and Yefeng Zheng.
to Multi-Agent Systems. In Dipti Srinivasan and 2024. Cooper: Coordinating Specialized Agents
Lakhmi C. Jain, editors, Innovations in Multi-Agent towards a Complex Dialogue Goal. Proceedings
Systems and Applications - 1, pages 1–27. Springer, of the AAAI Conference on Artificial Intelligence,
Berlin, Heidelberg. 38(16):17853–17861.

Chaithanya Bandi and Abir Harrasse. 2024. Ad- Yun-Shiuan Chuang, Agam Goyal, Nikunj Harlalka,
versarial Multi-Agent Evaluation of Large Lan- Siddharth Suresh, Robert Hawkins, Sijia Yang, Dha-
guage Models through Iterative Debates. Preprint, van Shah, Junjie Hu, and Timothy Rogers. 2024.
arXiv:2410.04663. Simulating Opinion Dynamics with Networks of
LLM-based Agents. In Findings of the Association
Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, for Computational Linguistics: NAACL 2024, pages
Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan 3326–3346, Mexico City, Mexico. Association for
Liu. 2023. ChatEval: Towards Better LLM-based Computational Linguistics.
Evaluators through Multi-Agent Debate. Preprint,
arXiv:2308.07201. Yun-Shiuan Chuang and Timothy T. Rogers. 2023.
Computational Agent-based Models in Opinion Dy-
Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, namics: A Survey on Social Simulations and Empiri-
Jaward Sesay, Börje F. Karlsson, Jie Fu, and cal Studies. Preprint, arXiv:2306.03446.

9
Silin Du and Xiaowei Zhang. 2024. Helmsman of the Shanshan Han, Qifan Zhang, Yuhang Yao, Weizhao
Masses? Evaluate the Opinion Leadership of Large Jin, Zhaozhuo Xu, and Chaoyang He. 2024. LLM
Language Models in the Werewolf Game. In First Multi-Agent Systems: Challenges and Open Prob-
Conference on Language Modeling. lems. Preprint, arXiv:2402.03578.

Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Matthew Hausknecht and Peter Stone. 2015. Deep
Tenenbaum, and Igor Mordatch. 2023. Im- recurrent q-learning for partially observable mdps.
proving Factuality and Reasoning in Language Computer Science.
Models through Multiagent Debate. Preprint,
Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu
arXiv:2305.14325.
Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang,
Zhuoyun Du, Chen Qian, Wei Liu, Zihao Xie, Yifei Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang
Wang, Yufan Dang, Weize Chen, and Cheng Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu,
Yang. 2024. Multi-Agent Software Develop- and Jürgen Schmidhuber. 2023. MetaGPT: Meta Pro-
ment through Cross-Team Collaboration. Preprint, gramming for A Multi-Agent Collaborative Frame-
arXiv:2406.08979. work. Preprint, arXiv:2308.00352.
Qian Huang, Jian Vora, Percy Liang, and Jure
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Leskovec. 2024a. MLAgentBench: Evaluating Lan-
Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, guage Agents on Machine Learning Experimentation.
Akhil Mathur, et al. 2024. The Llama 3 Herd of Preprint, arXiv:2310.03302.
Models. Preprint, arXiv:2407.21783.
Yizhe Huang, Xingbo Wang, Hao Liu, Fanqi Kong,
Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Aoyang Qin, Min Tang, Xiaoxi Wang, Song-Chun
Ding, Zhilun Zhou, Fengli Xu, and Yong Li. 2023a. Zhu, Mingjie Bi, Siyuan Qi, and Xue Feng. 2024b.
Large Language Models Empowered Agent-based AdaSociety: An Adaptive Environment with So-
Modeling and Simulation: A Survey and Perspec- cial Structures for Multi-Agent Decision-Making.
tives. Preprint, arXiv:2312.11970. Preprint, arXiv:2411.03865.
Chen Gao, Xiaochong Lan, Zhihong Lu, Jinzhu Mao, Md. Ashraful Islam, Mohammed Eunus Ali, and
Jinghua Piao, Huandong Wang, Depeng Jin, and Md Rizwan Parvez. 2024. MapCoder: Multi-Agent
Yong Li. 2023b. S3: Social-network Simulation Sys- Code Generation for Competitive Problem Solving.
tem with Large Language Model-Empowered Agents. In Proceedings of the 62nd Annual Meeting of the
Preprint, arXiv:2307.14984. Association for Computational Linguistics (Volume 1:
Long Papers), pages 4912–4944, Bangkok, Thailand.
Dawei Gao, Zitao Li, Xuchen Pan, Weirui Kuang, Zhi- Association for Computational Linguistics.
jian Ma, Bingchen Qian, Fei Wei, Wenhao Zhang,
Yuexiang Xie, Daoyuan Chen, Liuyi Yao, Hongyi Yihuai Lan, Zhiqiang Hu, Lei Wang, Yang Wang, De-
Peng, Zeyu Zhang, Lin Zhu, Chen Cheng, Hongzhu heng Ye, Peilin Zhao, Ee-Peng Lim, Hui Xiong, and
Shi, Yaliang Li, Bolin Ding, and Jingren Zhou. 2024. Hao Wang. 2024. LLM-Based Agent Society Inves-
AgentScope: A Flexible yet Robust Multi-Agent Plat- tigation: Collaboration and Confrontation in Avalon
form. Preprint, arXiv:2402.14034. Gameplay. Preprint, arXiv:2310.14985.

Ran Gong, Qiuyuan Huang, Xiaojian Ma, Yusuke Noda, Bin Lei, Yi Zhang, Shan Zuo, Ali Payani, and Caiwen
Zane Durante, Zilong Zheng, Demetri Terzopoulos, Ding. 2024. MACM: Utilizing a Multi-Agent System
Li Fei-Fei, Jianfeng Gao, and Hoi Vo. 2024. MindA- for Condition Mining in Solving Complex Mathemat-
gent: Emergent Gaming Interaction. In Findings ical Problems. Preprint, arXiv:2404.04735.
of the Association for Computational Linguistics: Junkai Li, Siyu Wang, Meng Zhang, Weitao Li, Yungh-
NAACL 2024, pages 3154–3183, Mexico City, Mex- wei Lai, Xinhui Kang, Weizhi Ma, and Yang Liu.
ico. Association for Computational Linguistics. 2024a. Agent Hospital: A Simulacrum of Hos-
pital with Evolvable Medical Agents. Preprint,
Thore Graepel, Tom Minka, and R TrueSkill Herbrich. arXiv:2405.02957.
2007. A bayesian skill rating system. Advances
in Neural Information Processing Systems, 19(569- Nian Li, Chen Gao, Mingyu Li, Yong Li, and Qingmin
576):7. Liao. 2024b. EconAgent: Large Language Model-
Empowered Agents for Simulating Macroeconomic
Sven Gronauer and Klaus Diepold. 2022. Multi-agent Activities. In Proceedings of the 62nd Annual Meet-
deep reinforcement learning: A survey. Artificial ing of the Association for Computational Linguis-
Intelligence Review, 55(2):895–943. tics (Volume 1: Long Papers), pages 15523–15536,
Bangkok, Thailand. Association for Computational
Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Linguistics.
Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xian-
gliang Zhang. 2024. Large Language Model Based Renhao Li, Minghuan Tan, Derek F. Wong, and Min
Multi-agents: A Survey of Progress and Challenges. Yang. 2024c. CoEvol: Constructing Better Re-
In Thirty-Third International Joint Conference on sponses for Instruction Finetuning through Multi-
Artificial Intelligence, volume 9, pages 8048–8057. Agent Cooperation. Preprint, arXiv:2406.07054.

10
Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. Bangkok, Thailand and virtual meeting. Association
2024d. A survey on LLM-based multi-agent sys- for Computational Linguistics.
tems: Workflow, infrastructure, and challenges. Vici-
nagearth, 1(1):9. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal,
Lama Ahmad, Ilge Akkaya, Florencia Leoni Ale-
Yuan Li, Yixuan Zhang, and Lichao Sun. 2023. MetaA- man, Diogo Almeida, Janko Altenschmidt, Sam Alt-
gents: Simulating Interactions of Human Behav- man, et al. 2024. GPT-4 Technical Report. Preprint,
iors for LLM-based Task-oriented Coordination arXiv:2303.08774.
via Collaborative Generative Agents. Preprint,
arXiv:2310.06500. Afshin Oroojlooy and Davood Hajinezhad. 2023. A
review of cooperative multi-agent deep reinforcement
Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, learning. Applied Intelligence, 53(11):13677–13722.
Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi,
and Zhaopeng Tu. 2024. Encouraging Divergent Xuchen Pan, Dawei Gao, Yuexiang Xie, Yushuo
Thinking in Large Language Models through Multi- Chen, Zhewei Wei, Yaliang Li, Bolin Ding, Ji-Rong
Agent Debate. Preprint, arXiv:2305.19118. Wen, and Jingren Zhou. 2024. Very Large-Scale
Multi-Agent Simulation in AgentScope. Preprint,
Jiaju Lin, Haoran Zhao, Aochi Zhang, Yiting Wu, Huqi- arXiv:2407.17789.
uyue Ping, and Qin Chen. 2023. AgentSims: An
Open-Source Sandbox for Large Language Model Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai,
Evaluation. Preprint, arXiv:2308.04026. Meredith Ringel Morris, Percy Liang, and Michael S.
Bernstein. 2023. Generative Agents: Interac-
Leilei Lin, Yumeng Jin, Yingming Zhou, Wenlong Chen, tive Simulacra of Human Behavior. Preprint,
and Chen Qian. 2024. MAO: A Framework for Pro- arXiv:2304.03442.
cess Model Generation with Multi-Agent Orchestra-
tion. Preprint, arXiv:2408.01916. Joon Sung Park, Lindsay Popowski, Carrie J. Cai,
Meredith Ringel Morris, Percy Liang, and Michael S.
Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Diyi
Bernstein. 2022. Social Simulacra: Creating Pop-
Yang, and Soroush Vosoughi. 2023a. Training So-
ulated Prototypes for Social Computing Systems.
cially Aligned Language Models on Simulated Social
Preprint, arXiv:2208.04024.
Interactions. In The Twelfth International Conference
on Learning Representations. Giorgio Piatti, Zhijing Jin, Max Kleiman-Weiner, Bern-
Wei Liu, Chenxi Wang, Yifei Wang, Zihao Xie, Rennai hard Schölkopf, Mrinmaya Sachan, and Rada Mihal-
Qiu, Yufan Dang, Zhuoyun Du, Weize Chen, Cheng cea. 2024. Cooperate or Collapse: Emergence of
Yang, and Chen Qian. 2024a. Autonomous Agents Sustainable Cooperation in a Society of LLM Agents.
for Collaborative Task under Information Asymme- Preprint, arXiv:2404.16698.
try. Preprint, arXiv:2406.14928.
Chen Qian, Yufan Dang, Jiahao Li, Wei Liu, Zihao Xie,
Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu YiFei Wang, Weize Chen, Cheng Yang, Xin Cong,
Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Xiaoyin Che, Zhiyuan Liu, and Maosong Sun. 2024a.
Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Ao- Experiential Co-Learning of Software-Developing
han Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Agents. In Proceedings of the 62nd Annual Meeting
Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie of the Association for Computational Linguistics (Vol-
Huang, Yuxiao Dong, and Jie Tang. 2023b. Agent- ume 1: Long Papers), pages 5628–5640, Bangkok,
Bench: Evaluating LLMs as Agents. Preprint, Thailand. Association for Computational Linguistics.
arXiv:2308.03688.
Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan
Yuhan Liu, Esha Choukse, Shan Lu, Junchen Jiang, Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng
and Madan Musuvathi. 2024b. DroidSpeak: En- Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu,
hancing Cross-LLM Communication. Preprint, and Maosong Sun. 2024b. ChatDev: Communicative
arXiv:2411.02820. Agents for Software Development. In Proceedings
of the 62nd Annual Meeting of the Association for
Zhiwei Liu, Weiran Yao, Jianguo Zhang, Le Xue, Computational Linguistics (Volume 1: Long Papers),
Shelby Heinecke, Rithesh Murthy, Yihao Feng, pages 15174–15186, Bangkok, Thailand. Association
Zeyuan Chen, Juan Carlos Niebles, Devansh Arpit, for Computational Linguistics.
Ran Xu, Phil Mui, Huan Wang, Caiming Xiong, and
Silvio Savarese. 2023c. BOLAA: Benchmarking and Chen Qian, Zihao Xie, Yifei Wang, Wei Liu, Yu-
Orchestrating LLM-augmented Autonomous Agents. fan Dang, Zhuoyun Du, Weize Chen, Cheng Yang,
Preprint, arXiv:2308.05960. Zhiyuan Liu, and Maosong Sun. 2024c. Scaling
Large-Language-Model-based Multi-Agent Collabo-
Xinyi Mou, Zhongyu Wei, and Xuanjing Huang. 2024. ration. Preprint, arXiv:2406.07155.
Unveiling the Truth and Facilitating Change: To-
wards Agent-based Large-scale Social Movement Alec Radford, Jeffrey Wu, Rewon Child, David Luan,
Simulation. In Findings of the Association for Com- Dario Amodei, and Ilya Sutskever. Language Models
putational Linguistics ACL 2024, pages 4789–4809, are Unsupervised Multitask Learners.

11
Weizhou Shen, Chenliang Li, Hongzhan Chen, Ming Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiao-
Yan, Xiaojun Quan, Hehong Chen, Ji Zhang, and Fei long Wang, Weidong Liu, and Yang Liu. 2024. Ex-
Huang. 2024. Small LLMs Are Weak Tool Learners: ploring Large Language Models for Communication
A Multi-LLM Agent. Preprint, arXiv:2401.07324. Games: An Empirical Study on Werewolf. Preprint,
arXiv:2309.04658.
Haojun Shi, Suyu Ye, Xinyu Fang, Chuanyang Jin,
Leyla Isik, Yen-Ling Kuo, and Tianmin Shu. 2024. Zelai Xu, Chao Yu, Fei Fang, Yu Wang, and Yi Wu.
MuMA-ToM: Multi-modal Multi-Agent Theory of 2023c. Language Agents with Reinforcement Learn-
Mind. Preprint, arXiv:2408.12574. ing for Strategic Play in the Werewolf Game.
Noah Shinn, Federico Cassano, Ashwin Gopinath, Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak
Karthik Narasimhan, and Shunyu Yao. 2023. Re- Shafran, Karthik Narasimhan, and Yuan Cao. 2023.
flexion: Language agents with verbal reinforcement ReAct: Synergizing Reasoning and Acting in Lan-
learning. Advances in Neural Information Process- guage Models. Preprint, arXiv:2210.03629.
ing Systems, 36:8634–8652.
Xiangyu Yin, Chuqiao Shi, Yimo Han, and Yi Jiang.
Xunzhu Tang, Kisub Kim, Yewei Song, Cedric Lothritz, 2024. PEAR: A Robust and Flexible Automa-
Bei Li, Saad Ezzini, Haoye Tian, Jacques Klein, and tion Framework for Ptychography Enabled by Mul-
Tegawende F. Bissyande. 2024. CodeAgent: Au- tiple Large Language Model Agents. Preprint,
tonomous Communicative Agents for Code Review. arXiv:2410.09034.
Preprint, arXiv:2402.02172.
Xiaoyan Yu, Tongxu Luo, Yifan Wei, Fangyu Lei, Yim-
Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao
ing Huang, Hao Peng, and Liehuang Zhu. 2024.
Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang,
Neeko: Leveraging Dynamic LoRA for Efficient
Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei,
Multi-Character Role-Playing Agent. Preprint,
and Jirong Wen. 2024a. A survey on large language
arXiv:2402.13717.
model based autonomous agents. Frontiers of Com-
puter Science, 18(6):186345. Haoqi Yuan, Chi Zhang, Hongcheng Wang, Feiyang
Lei Wang, Jingsen Zhang, Hao Yang, Zhiyuan Chen, Xie, Penglin Cai, Hao Dong, and Zongqing Lu.
Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Rui- 2023. Skill Reinforcement Learning and Planning
hua Song, Wayne Xin Zhao, Jun Xu, Zhicheng Dou, for Open-World Long-Horizon Tasks. Preprint,
Jun Wang, and Ji-Rong Wen. 2024b. User Behav- arXiv:2303.16563.
ior Simulation with Large Language Model based
Shengbin Yue, Siyuan Wang, Wei Chen, Xuanjing
Agents. Preprint, arXiv:2306.02552.
Huang, and Zhongyu Wei. 2024. Synergistic
Qineng Wang, Zihao Wang, Ying Su, Hanghang Tong, Multi-Agent Framework with Trajectory Learn-
and Yangqiu Song. 2024c. Rethinking the Bounds of ing for Knowledge-Intensive Tasks. Preprint,
LLM Reasoning: Are Multi-Agent Discussions the arXiv:2407.09893.
Key? In Proceedings of the 62nd Annual Meeting of
the Association for Computational Linguistics (Vol- Ceyao Zhang, Kaijie Yang, Siyi Hu, Zihao Wang,
ume 1: Long Papers), pages 6106–6131, Bangkok, Guanghe Li, Yihang Sun, Cheng Zhang, Zhaowei
Thailand. Association for Computational Linguistics. Zhang, Anji Liu, Song-Chun Zhu, Xiaojun Chang,
Junge Zhang, Feng Yin, Yitao Liang, and Yaodong
Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Yang. 2023a. ProAgent: Building Proactive Cooper-
Tao Ge, Furu Wei, and Heng Ji. 2024d. Unleash- ative AI with Large Language Models. CoRR.
ing the Emergent Cognitive Synergy in Large Lan-
guage Models: A Task-Solving Agent through Multi- Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin
Persona Self-Collaboration. In Proceedings of the Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2023b.
2024 Conference of the North American Chapter of AppAgent: Multimodal Agents as Smartphone Users.
the Association for Computational Linguistics: Hu- Preprint, arXiv:2312.13771.
man Language Technologies (Volume 1: Long Pa-
pers), pages 257–279, Mexico City, Mexico. Associ- Jintian Zhang, Xin Xu, Ningyu Zhang, Ruibo Liu,
ation for Computational Linguistics. Bryan Hooi, and Shumin Deng. 2024a. Exploring
Collaboration Mechanisms for LLM Agents: A So-
Fengli Xu, Jun Zhang, Chen Gao, Jie Feng, and Yong cial Psychology View. In Proceedings of the 62nd
Li. 2023a. Urban Generative Intelligence (UGI): A Annual Meeting of the Association for Computational
Foundational Platform for Agents in Embodied City Linguistics (Volume 1: Long Papers), pages 14544–
Environment. Preprint, arXiv:2312.11813. 14607, Bangkok, Thailand. Association for Compu-
tational Linguistics.
Lin Xu, Zhiyuan Hu, Daquan Zhou, Hongyu Ren,
Zhen Dong, Kurt Keutzer, See Kiong Ng, and Ji- Zaibin Zhang, Yongting Zhang, Lijun Li, Jing Shao,
ashi Feng. 2023b. MAgIC: Investigation of Large Hongzhi Gao, Yu Qiao, Lijun Wang, Huchuan Lu,
Language Model Powered Multi-Agent in Cognition, and Feng Zhao. 2024b. PsySafe: A Comprehensive
Adaptability, Rationality and Collaboration. Preprint, Framework for Psychological-based Attack, Defense,
arXiv:2311.08562. and Evaluation of Multi-agent System Safety. In

12
 Proceedings of the 62nd Annual Meeting of the As-
sociation for Computational Linguistics (Volume 1:
Long Papers), pages 15202–15231, Bangkok, Thai-
land. Association for Computational Linguistics.
Jun Zhao, Can Zu, Hao Xu, Yi Lu, Wei He, Yiwen Ding,
Tao Gui, Qi Zhang, and Xuanjing Huang. 2024a.
LongAgent: Scaling Language Models to 128k Con-
text through Multi-Agent Collaboration. Preprint,
arXiv:2402.11550.
Qinlin Zhao, Jindong Wang, Yixuan Zhang, Yiqiao Jin,
Kaijie Zhu, Hao Chen, and Xing Xie. 2024b. Com-
peteAI: Understanding the Competition Dynamics
in Large Language Model-based Agents. Preprint,
arXiv:2310.17512.
Xiutian Zhao, Ke Wang, and Wei Peng. 2024c.
An Electoral Approach to Diversify LLM-based
Multi-Agent Collective Decision-Making. Preprint,
arXiv:2410.15168.
Hang Zou, Qiyang Zhao, Lina Bariah, Mehdi Bennis,
and Merouane Debbah. 2023. Wireless Multi-Agent
Generative AI: From Connected Intelligence to Col-
lective Intelligence. Preprint, arXiv:2307.02757.

13