LLM-guided graph neural coordination framework for cooperative multi-agent reinforcement learning

Abstract

Cooperative multi-agent reinforcement learning requires accurate modeling of interactions between agents. This work introduces an LLM-guided graph neural coordination framework that uses language-model reasoning to construct dynamic graph structures and extract strategic semantics, then combines them with graph-based policy learning. The framework also adds LLM-empowered latent reward shaping and a two-stage training strategy to improve coordination and policy optimization in sparse-reward environments.