Preprint
Large Language Models

A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges

Xinyi Li(Wuhan University), S. Wang(Wuhan University), Siqi Zeng(Wuhan University), Yu Wu(Wuhan University), Yi Yang(Zhejiang University)
October 8, 2024Vicinagearth.569 citations
PDFDOI

569

Citations

30

Influential Citations

Vicinagearth.

Venue

2024

Year

Abstract

Abstract The pursuit of more intelligent and credible autonomous systems, akin to human society, has been a long-standing endeavor for humans. Leveraging the exceptional reasoning and planning capabilities of large language models (LLMs), LLM-based agents have been proposed and have achieved remarkable success across a wide array of tasks. Notably, LLM-based multi-agent systems (MAS) are considered a promising pathway towards realizing general artificial intelligence that is equivalent to or surpasses human-level intelligence. In this paper, we present a comprehensive survey of these studies, offering a systematic review of LLM-based MAS. Adhering to the workflow of LLM-based multi-agent systems, we synthesize a general structure encompassing five key components: profile, perception, self-action, mutual interaction, and evolution. This unified framework encapsulates much of the previous work in the field. Furthermore, we illuminate the extensive applications of LLM-based MAS in two principal areas: problem-solving and world simulation. Finally, we discuss in detail several contemporary challenges and provide insights into potential future directions in this domain.

Analysis

Why This Paper Matters

This survey addresses the rapidly growing field of LLM-based multi-agent systems (MAS), which are seen as a promising path toward artificial general intelligence. By systematically reviewing the literature and proposing a unified five-component framework, the paper provides a much-needed structure for researchers and practitioners. The framework—covering profile, perception, self-action, mutual interaction, and evolution—offers a common language to describe and compare different MAS designs.

The paper also highlights two major application areas: problem-solving and world simulation. This categorization helps clarify where LLM-based MAS have been most effective and where future efforts might focus. Given the high citation count (569) and recent publication date, this survey is already influencing the field.

Technical Contributions

  • Unified Framework: The five-component structure (profile, perception, self-action, mutual interaction, evolution) synthesizes diverse prior work into a coherent whole.
  • Systematic Review: The authors comprehensively cover workflow, infrastructure, and challenges, providing a holistic view of the field.
  • Application Taxonomy: Clear categorization into problem-solving and world simulation, with examples from the literature.
  • Challenge Identification: Detailed discussion of contemporary challenges, offering insights into future research directions.

Results

The paper does not present experimental results or quantitative benchmarks. Its main output is a structured taxonomy and framework that organizes existing research. The value lies in the synthesis and clarity it brings to a complex, fast-moving area.

Significance

This survey serves as a foundational reference for researchers entering the field of LLM-based multi-agent systems. By providing a unified framework, it enables more systematic comparisons and accelerates progress. The identification of key challenges—such as scalability, coordination, and robustness—guides future work. As LLM-based MAS continue to evolve, this paper will likely be cited as a key organizing resource.

large-language-modelsmulti-agent-systemsautonomous-agentsworkflow-frameworkproblem-solvingworld-simulationsurvey
Large Language ModelsReasoning

Related Papers