MARLIN: Guiding Multi-Agent Reinforcement Learning with Language Negotiation

MARLIN proposes a hybrid framework in which large language models guide policy learning by generating high-level action plans through inter-robot language negotiation before the reinforcement learning policy has learned effective behaviors, and it dynamically switches between reinforcement learning and language-model negotiation during training. The authors evaluated it on simulated and physical robots with local and remote language models; results show higher early-training performance than standard multi-agent reinforcement learning, with limited impact on final performance. The code is publicly available, and the paper was accepted to ROBIO2026.