LLM-Powered Multi-Agent Systems: A Survey of Collaboration and Learning Strategies

With the progress in Large Language Models (LLMs), research in this area has grown rapidly. This has led to the development of LLM-based Multi-Agent Systems (MAS), which use multiple advanced models trained on large amounts of data. The main goal is to mimic human-like teamwork, allowing different agents to work together to solve complex problems. By splitting tasks among specialized agents, these systems can overcome challenges that a single LLM agent might struggle with, making them more efficient and effective. In this paper, we provide a comprehensive review of LLM based MAS, offering a systematic analysis of their frameworks. First, we explore the construction of MAS, introducing a categorization of collaboration structures based on existing research. Second, we examine the communication mechanisms within MAS, which define the strategies agents use to exchange information and make decisions. Next, we review the learning methods and evaluation metrics employed in MAS, highlighting how agents adapt their responses and optimize their performance for improved results, and how we can evaluate them based on the domain, and whether it is single or multi-agents’ structure. Finally, to conclude this review on LLM based MAS, we propose a comprehensive road map (Figure 4), intended to support researchers and practitioners in efficiently designing and implementing multi-agent solutions tailored to domain applications.