Title: LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning

Original: arXiv:2605.18077v2 Announce Type: replace

Abstract: While communication is essential in multi-agent reinforcement learning (MARL) to overcome partial observability, existing methods frequently suffer from inefficient data transfer or an inability to convey adequate state information. To solve this issue, we introduce LLM-driven Multi-Agent Communication (LMAC). This approach harnesses the reasoning power of large language models to formulate a communication protocol that allows every agent to reconstruct the true state with maximum accuracy and uniformity. LMAC continuously optimizes this protocol based on a specific state-awareness metric, which enhances the fidelity of state recovery and minimizes disparities in the knowledge held by individual agents. Evaluations across a variety of MARL benchmarks demonstrate that LMAC facilitates superior state reconstruction among agents and delivers significant performance improvements compared to established communication baselines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC