Key Points

- Large Language Models (LLMs) have demonstrated remarkable potential in reasoning and planning capabilities, aligning with the expectations for autonomous agents able to perceive surroundings, make decisions, and take actions.

- Multi-agent systems, leveraging LLMs, excel in problem-solving and world simulation, contributing to specialized profiles, collective decision-making, and cooperative behaviors.

- LLM-based multi-agent systems have been utilized in diverse applications such as software development, multi-robot systems, society simulation, policy simulation, and game simulation, attracting interest from interdisciplinary researchers.

- Critical elements of LLM-MA systems include agents-environment interface, agent profiling, agent communication, and agent capability acquisition, which shape their interactions and effectiveness within different environments.

- Challenges in LLM-MA systems include the need for multi-modal environments, mitigating the hallucination problem, developing efficient learning methods, managing scalability, orchestrating advanced agents, and developing comprehensive benchmarks for evaluation.

- There are opportunities for future research in LLM-MA systems, including broadening their applications to diverse fields, exploring theoretical perspectives, and further advancements in LLM technology to address current limitations.

- The survey also systematically reviews the development of LLM-MA systems, including their positioning, differentiation, connection from various aspects, summarizing their applications for problem-solving and world simulation, and discussing commonly used datasets and benchmarks.

- LLM-MA systems have shown inspiring collective intelligence, attracting increasing interest among researchers, and the survey aims to serve as a resource for inspiring future research and exploring the potential of LLM-based Multi-Agents.

Summary

The paper presents a survey on the use of Large Language Models (LLMs) as the basis for multi-agent systems. It discusses the domains and environments simulated by LLM-based multi-agents, their profiling and communication, mechanisms contributing to their growth in capacities, and commonly used datasets or benchmarks. The research covers LLM-based multi-agent applications in problem solving, world simulation, societal simulation, gaming, psychology, economy, recommender systems, policy making, and disease propagation simulation. The paper also introduces several open-source multi-agent frameworks and addresses the challenges and opportunities in the field.

Key Challenges and Future Opportunities
The survey identifies key challenges for LLM-based multi-agent systems, including the lack of multi-modal capabilities, the hallucination problem in text generation, and scalability issues. It also highlights the need for comprehensive benchmarks and datasets across diverse research fields. The paper discusses the potential of LLM-MA systems in various theoretical perspectives, such as Cognitive Science, Cybernetics, and Collective Intelligence, and outlines future research opportunities in exploring advanced methodologies, applications, and theoretical perspectives. The authors aim to provide a useful resource for researchers working in interdisciplinary fields and to inspire future exploration and innovation in the rapidly evolving field of LLM-based multi-agent systems.


Reference: https://arxiv.org/abs/2402.01680