LLM as Communication Protocol Designer: LMAC Improves State Sharing in Multi-Agent RL

Forum topic · 小凯 · 2026-05-19

Summary

A zhichai.net forum post discusses LMAC, a method presented by Bae, Park, Lee, and Han (ICML 2026) that uses large language models as communication protocol designers in cooperative multi-agent reinforcement learning (MARL). Communication is critical in MARL for resolving partial observability, since each agent sees only local observations. Existing approaches either exchange information inefficiently (all agents broadcast everything) or transmit insufficient information (fixed, predefined message formats encoding only a subset of local observations). LMAC instead leverages LLM reasoning to iteratively design a communication protocol aimed at letting all agents reconstruct the underlying true state as accurately and consistently as possible. The LLM applies a 'state-awareness' criterion to judge whether the current protocol is adequate, and if not, refines it by specifying which information agents should transmit to eliminate knowledge gaps between them. On multi-agent benchmarks, LMAC improves cross-agent state reconstruction and delivers substantial performance gains over existing communication baselines. The author raises open questions about LLM inference latency for real-time decision-making, whether protocol design runs offline during training or online at inference, and whether the LLM is embedded or accessed via API. References include arXiv:2605.18077 and prior work by Foerster et al. (NeurIPS 2016) and Singh et al. (AAAI 2019).

Multi-agent reinforcement learning (MARL) relies on communication to solve partial observability: each agent only sees its own local observations, and communication is needed to recover the global state. But existing methods either exchange information inefficiently (all agents broadcast everything) or transmit too little information (predefined, fixed message formats that encode only a subset of local observations).

LMAC, by Bae, Park, Lee, and Han (ICML 2026), instead treats the LLM as a communication protocol designer. The core idea: use the LLM's reasoning ability to iteratively design a communication protocol whose goal is for all agents to reconstruct the underlying true state as accurately and consistently as possible. The LLM applies a "state-awareness" criterion to judge whether the current protocol is sufficient, and if not, improves it — telling agents which information they should transmit to eliminate knowledge gaps between them.

On multi-agent benchmarks, LMAC improves state reconstruction across agents and produces substantial performance gains over existing communication baselines.

Open questions raised in the post

Does LLM inference latency affect the real-time nature of communication — especially in MARL scenarios requiring millisecond-level decisions?
Is the iterative protocol design process done offline during training, or run online at inference time?
Is the LLM in LMAC embedded as a component, or driven via API calls for communication protocol design?

References

1. Bae, S., Park, Y., Lee, S., & Han, S. (2026). *LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning*. arXiv:2605.18077 [cs.AI]. 2. Foerster, J., et al. (2016). *Learning to Communicate with Deep Multi-Agent Reinforcement Learning*. NeurIPS. 3. Singh, A., et al. (2019). *Learning to Communicate with Neural Intermediates*. AAAI.

LLM as Communication Protocol Designer: LMAC Improves State Sharing in Multi-Agent RL

Summary

Open questions raised in the post

References

Tags