TL;DR: CoFlow directly learns a joint averaged-velocity field for cooperative one-step and few-step generation. Coordinated Velocity Attention incorporates teammate information, while finite-difference consistency training reduces training cost. The updated experiments cover MPE and SMAC under centralized and decentralized execution.
CoFlow: Coordinated Few-Step Flow for Offline Multi-Agent Decision Making
Sun Yat-sen University
Actual Rollout Trajectories
Actual rollout trajectories for three MPE tasks and four SMAC tasks.
Multi-Particle Environments (MPE)
Tag
World
StarCraft Multi-Agent Challenge (SMAC)
3m
8m
2s3z
5m_vs_6m
Overview
Multi-step generation incurs repeated model evaluations. Distillation reduces sampling steps but may weaken cross-agent dependencies. CoFlow directly learns a coupled joint velocity field for few-step generation, with CVA for cooperation and finite differences for efficient training.
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},
}