CoFlow shows that few-step multi-agent generation does not have to trade away coordination. By placing inter-agent coupling directly inside the averaged velocity field, the model can generate coordinated trajectories in one to a few denoising steps without distilling a joint teacher into independent agents.
Coordination Lives in the Model
CVA and adaptive gates keep cross-agent information flow active during generation instead of relying on a long sampling loop to recover cooperation.
Few Steps Are Enough
Across MPE, MA-MuJoCo, and SMAC, CoFlow reaches strong return quality with a one-to-three-step inference budget under both centralized and decentralized execution.
Training Stays Practical
The finite-difference consistency surrogate avoids memory-heavy JVP backpropagation, making consistency-regularized joint flow training feasible at multi-agent scale.