Researchers publishing on Hugging Face have introduced DMM (Decentralized Master-Mind), a method for decentralized multi-agent path finding (MAPF) that addresses what the authors call the "decentralized factorization gap." The paper argues that even when agents learn to communicate, each agent still selects its final action independently, allowing locally valid choices to recombine into incompatible joint actions that cause collisions or deadlocks.
DMM replaces one-shot action sampling with discrete, iterative refinement of action intents across communication rounds, inspired by denoising in diffusion models. Agents begin with random action intents and refine them through local communication, coupling their choices before committing to an action.
The method is pretrained with imitation learning on expert MAPF solutions and then optimized with MICPO, a critic-free group-relative reinforcement-learning method designed for multi-agent, multi-round action refinement.
According to the paper, DMM generally achieves higher success rates and lower solution costs than the evaluated learnable baselines. On 1,600 MovingAI tasks, DMM fine-tuned with MICPO solves 1,598 — the highest coverage among the evaluated methods — while achieving solution costs close to the strongest baselines. The authors also report that DMM scales to over one million simultaneously acting agents in obstacle-rich environments.
In a symmetric corridor test, the paper states that DMM selects a valid joint action 95–99.9% of the time, compared with roughly 50% for prior state-of-the-art methods including LC-MAPF, MAGAT+, and HMAGAT. The authors also claim DMM outperforms centralized search methods LG-LaCAM and MAPF-LNS2 on the MovingAI benchmark.