---
license: mit
library_name: pytorch
tags:
  - reinforcement-learning
  - multi-agent-reinforcement-learning
  - offline-rl
  - flow-matching
  - generative-models
  - pytorch
  - arxiv:2605.01457
---

# CoFlow Checkpoints

Official checkpoints for **CoFlow: Coordinated Few-Step Flow for Offline Multi-Agent Decision Making**.

<p align="center">
  <a href="https://arxiv.org/abs/2605.01457"><img src="https://img.shields.io/badge/Paper-arXiv%3A2605.01457-b31b1b?style=for-the-badge&logo=arxiv" alt="Paper"></a>
  <a href="https://guowei-zou.github.io/coflow/"><img src="https://img.shields.io/badge/Project-Page-4169e1?style=for-the-badge&logo=githubpages" alt="Project Page"></a>
  <a href="https://github.com/Guowei-Zou/coflow-release"><img src="https://img.shields.io/badge/Code-GitHub-111111?style=for-the-badge&logo=github" alt="Code"></a>
  <a href="#citation"><img src="https://img.shields.io/badge/Citation-BibTeX-ff8c00?style=for-the-badge" alt="Citation"></a>
</p>

CoFlow is a coordinated few-step generative model for offline multi-agent reinforcement learning. It combines Coordinated Velocity Attention with adaptive coordination gating so multi-agent actions can be generated in one to a few model calls while preserving inter-agent coordination.

## Repository Contents

This repository contains the 120 checkpoints used in the paper:

- 30 task-quality configurations
- 4 model variants per configuration

The task-quality configurations cover:

- MPE: Spread, Tag, and World with `expert`, `medium-replay`, `medium`, and `random` data qualities
- SMAC: `3m`, `8m`, `2s3z`, and `5m_vs_6m` with `Good`, `Medium`, and `Poor` data qualities
- MA-MuJoCo: `2xAnt` and `4xAnt` with `Good`, `Medium`, and `Poor` data qualities

Model variants:

- `coflow-c`: CoFlow with centralized execution
- `coflow-d`: CoFlow with decentralized execution
- `coflow-base-c`: CoFlow-base with centralized execution
- `coflow-base-d`: CoFlow-base with decentralized execution

Each leaf directory contains one paper-used `state_*.pt` checkpoint. See `MANIFEST.tsv` for the mapping from paper configuration to source run, seed, checkpoint step, and file size.

## Download

Download the full checkpoint release:

```bash
hf download Guowei-Zou/CoFlow-checkpoints --local-dir CoFlow-checkpoints
```

Download one configuration, for example MPE Spread Expert with CoFlow-C:

```bash
hf download Guowei-Zou/CoFlow-checkpoints \
  --include "mpe/simple_spread/expert/coflow-c/*" \
  --local-dir CoFlow-checkpoints
```

## Usage

The checkpoints are intended to be used with the official code release:

```bash
git clone https://github.com/Guowei-Zou/coflow-release.git
```

Please follow the setup, evaluation, and configuration instructions in the GitHub repository. The directory structure in this checkpoint repository is aligned with the paper task names and model variants.

## Citation

```bibtex
@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},
}
```