LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning
The paper presents a novel approach to communication in multi-agent reinforcement learning (MARL) called LLM-driven Multi-Agent Communication (LMAC). This method utilizes a large language model (LLM) to enhance the efficiency and effectiveness of information exchange among agents. Experimental results indicate that LMAC significantly improves state reconstruction and overall performance compared to existing communication methods.
- ▪The proposed LMAC protocol helps agents reconstruct the underlying state more accurately.
- ▪LMAC iteratively refines communication based on an explicit state-awareness criterion.
- ▪Experiments demonstrate substantial performance gains over previous communication baselines.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.18077 (cs) [Submitted on 18 May 2026] Title:LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning Authors:Sangjun Bae, Yisak Park, Sanghyeon Lee, Seungyul Han View a PDF of the paper titled LLM-Guided Communication for Cooperative Multi-Agent Reinforcement Learning, by Sangjun Bae and 3 other authors View PDF HTML (experimental) Abstract: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.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
