Reward and Behavioral Measures of LLM Coordination in Dining Philosophers
The study investigated if fine-tuning large language models using Group Relative Policy Optimization (GRPO) could improve their ability to coordinate as multiple agents solving the dining philosophers problem. Results showed that training methodology, specifically the reward formula's degeneracy, was the bottleneck; naive fine-tuning did not close the performance gap between frontier models and Qwen3-14B. The key takeaway is that better coordination requires modifying reward shaping and checkpoint selection strategies to prevent models from collapsing into no-action states.