LLM-Guided Role Assignment and Coordination in Multi-Agent Reinforcement Learning

Effective coordination in cooperative Multi-Agent Reinforcement Learning (MARL) fundamentally relies on the emergence of specialized roles. However, existing role-based methods typically induce such structures implicitly through reward maximization, making them data-inefficient and often incapable of uncovering semantically meaningful roles from sparse and low-level interaction signals. To address this limitation, we propose LLM-Guided Role Assignment and Coordination in Multi-Agent Reinforcement Learning (LGRA), a framework that leverages a Large Language Model (LLM) as an online semantic teacher to provide high-level role assignments and natural-language rationales. LGRA integrates two key mechanisms to distill these semantic priors into decentralized policies: an adaptive role router that aligns its role distribution with the LLM’s guidance through a KL-regularized objective, and a contrastive semantic alignment module that grounds role-specific trajectory encoders in the LLM’s reasoning space by maximizing mutual information between agent behaviors and textual rationales. Extensive experiments across the LBF, RWARE, and SMAC benchmarks demonstrate that LGRA achieves state-of-the-art performance in 13 out of 14 tasks, with particularly large gains in sparse-reward and highly coordinated scenarios, highlighting the benefits of infusing abstract reasoning and world knowledge into cooperative MARL.