Can diffusion models plan for thousands of agents at once?

Planning for many agents simultaneously hits a wall: the joint state space explodes exponentially. MF-Diffuser sidesteps this by planning in the space of agent distributions rather than individual trajectories, using mean-field theory to ensure a small subset captures population behavior. The method includes theoretical bounds showing approximation error shrinks with population size and offline data shift doesn't worsen with more agents. Tested on three benchmarks up to 1000+ agents, it outperforms baselines especially on suboptimal offline data.