LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning

Communication is a key component in multi-agent reinforcement learning (MARL) for mitigating partial observability, yet prior approaches often rely on inefficient information exchange or fail to transmit sufficient state information. To address this, we propose LLM-driven Multi-Agent Communication (LMAC), which leverages an LLM's reasoning capability to design a communication protocol that enables all agents to reconstruct the underlying state as accurately and uniformly as possible. LMAC iteratively refines the protocol using an explicit state-awareness criterion, improving state recovery while narrowing differences in agents' knowledge. Experiments on diverse MARL benchmarks show that LMAC improves state reconstruction across agents and yields substantial performance gains over prior communication baselines.

When multiple AI agents work together, each agent often sees only part of the full situation. To cooperate well, they need to share the right information with one another. This paper proposes LMAC, a method that uses a large language model to design communication rules between agents. Given a description of the task and what each agent can observe, LMAC decides what information should be shared so that all agents can better understand the situation. It then uses training data to find missing information and gradually improves the communication rules. Experiments show that LMAC helps agents share important information more accurately and evenly than previous methods, leading to better performance across several cooperative tasks. This suggests that language models can be useful tools for designing communication among AI systems.