CoFlow: Coordinated Few-Step Flow for Offline Multi-Agent Decision Making

Guowei Zou, Haitao Wang, Beiwen Zhang, Boning Zhang, Hejun Wu

Sun Yat-sen University

Paper Code Checkpoints

Selected Rollout Loops

Tag · Expert

Coordinated pressure around the prey.

World · Expert

Structured movement through food and forests.

5m_vs_6m · Good

Coordinated micro overcomes the extra enemy.

2-Ant · Good

Coordinated legs sustain locomotion.

TL;DR: CoFlow learns a natively joint-coupled averaged velocity field for offline multi-agent reinforcement learning. It combines Coordinated Velocity Attention, adaptive coordination gating, and a finite-difference consistency surrogate so coordinated multi-agent trajectories can be generated in 1--3 denoising steps without distilling a joint teacher into independent agents.

Overview

CoFlow targets the quality-efficiency dilemma in offline multi-agent trajectory generation. Existing diffusion methods preserve coordination but require many denoising steps; existing few-step routes accelerate inference but weaken cross-agent coupling. CoFlow occupies the Pareto region where few-step inference and coordination preservation coexist.

Citation

@misc{zou2026coflowcoordinatedfewstepflow,
      title={CoFlow: Coordinated Few-Step Flow for Offline Multi-Agent Decision Making},
      author={Guowei Zou and Haitao Wang and Beiwen Zhang and Boning Zhang and Hejun Wu},
      year={2026},
      eprint={2605.01457},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2605.01457},
}