Conversation-based Multi-agent Reinforcement Learning with LLM as a Policy

ACL ARR 2025 February Submission5039 Authors

Abstract: Reinforcement learning (RL) benefits from Large language models (LLMs) for improved reasoning and planning, but their application in Multi-agent reinforcement learning (MARL) remains challenging due to communication conflicts. We propose a novel framework where agents engage in structured multi-round conversations before taking actions, ensuring better coordination and decision-making. By leveraging LLMs’ reasoning capabilities and integrating techniques like Chain of Thought reasoning, our approach enhances collaboration in MARL. Experimental results show improved efficiency and scalability, bridging the gap between single-agent and multi-agent LLM applications.

Paper Type: Long

Research Area: Computational Social Science and Cultural Analytics

Research Area Keywords: multi-agent reinforcement learning, large language model, inter-agent communication

Contribution Types: NLP engineering experiment

Languages Studied: english

Submission Number: 5039