LLMs have achieved remarkable success across a wide array of tasks. Due to the impressive planning and reasoning abilities of LLMs, they have been used as autonomous agents to do many tasks automatically. Recently, based on the development of using one LLM as a single planning or decision-making agent, LLM-based multi-agent systems have achieved considerable progress in complex problem-solving and world simulation. To provide the community with an overview of this dynamic field, we present this survey to offer an in-depth discussion on the essential aspects of multi-agent systems based on LLMs, as well as the challenges. Our goal is for readers to gain substantial insights on the following questions: What domains and environments do LLM-based multi-agents simulate? How are these agents profiled and how do they communicate? What mechanisms contribute to the growth of agents' capacities? For those interested in delving into this field of study, we also summarize the commonly used datasets or benchmarks for them to have convenient access. To keep researchers updated on the latest studies, we maintain an open-source GitHub repository, dedicated to outlining the research on LLM-based multi-agent systems.
The paper surveys the development and application of Large Language Model based Multi-Agent (LLM-MA) systems, illustrating their role in simulating complex environments and facilitating sophisticated agent interactions for problem-solving and planning.
It explore the core components of LLM-MA systems, including agent-environment interface, agent profiling, communication, and capability acquisition, exploring how these elements contribute to the system's functionality.
The applications of LLM-MA systems span across problem solving in software development and world simulation in economic models, showcasing the versatility and potential of these systems in various domains.
Despite significant progress, the paper identifies challenges such as agent hallucination and scaling, and suggests future research directions, including the integration of cognitive science and symbolic AI, to enhance the efficacy and applicability of LLM-MA systems.
Recent advancements have led to the emergence of LLMs based Multi-Agent (LLM-MA) systems, showcasing notable progress in simulating complex real-world environments through multiple autonomous agents. These systems leverage the comprehensive knowledge base and generative capabilities of LLMs, enabling agents to engage in sophisticated interactions, problem-solving, and planning, akin to human group dynamics. This survey systematically reviews the current landscape of LLM-MA systems, focusing on their application domains, the mechanisms underlying agent interaction and evolution, and the challenges faced in scaling and effectively deploying these systems.
To understand the intricacies of LLM-MA systems, it is imperative to examine their core components and mechanisms. This includes how agents perceive and interact with their environment, the methods for profiling agents to exhibit specific behaviors, the structure, and paradigms of agent communication, and the strategies for capability acquisition and enhancement.
LLM-MA systems find applications in diverse domains, each presenting unique challenges and benefiting differently from multi-agent collaborations.
The development of LLM-MA systems is supported by various frameworks and tools designed to facilitate agent modeling, interaction, and simulation. These include MetaGPT, CAMEL, and Autogen, each offering different approaches to integrating LLMs into multi-agent settings. Ongoing efforts in the community aim to develop more comprehensive datasets and benchmarks tailored to specific applications, addressing the need for standardized evaluation metrics across diverse domains.
Despite the promising developments in LLM-MA systems, significant challenges remain, including addressing agent hallucination, acquiring collective intelligence, scaling up systems to accommodate larger numbers of agents, and developing more nuanced evaluation methods. Future research is expected to explore these areas, pushing the boundaries of what LLM-MA systems can achieve and expanding their applicability across even broader spectrums of tasks and simulations. Additionally, theoretical perspectives from cognitive science, symbolic AI, and collective intelligence offer fertile grounds for enhancing our understanding and leveraging LLM-based multi-agents more effectively.
In conclusion, the exploration of LLM-based Multi-Agents stands at a fascinating juncture, with immense potential for advancing our capabilities in simulating, understanding, and solving complex problems. This survey underscores the significant progress made thus far, highlights the challenges ahead, and calls for continued interdisciplinary collaboration to unlock the full potential of LLM-MA systems.