CoFlow combines a joint averaged-velocity model, cross-agent attention, and efficient consistency training to support cooperative one-step and few-step generation in offline multi-agent reinforcement learning.
Joint Generation
Learn joint trajectories directly, without a separate teacher-distillation stage.
Cooperation
CVA uses teammate trajectory information during each velocity prediction.
Efficient Training
Finite differences lower the consistency-training cost of the joint model.
Experiments cover MPE and SMAC under centralized and decentralized execution. Larger teams and broader domains remain directions for future evaluation.