Reproduce “Multi-agent Coordination via Flow Matching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dongsu Lee, Daehee Lee, Amy Zhang
MAC-Flow, a new offline multi-agent reinforcement learning framework developed by researchers at the University of Texas at Austin and Sungkyunkwan University, leverages flow matching to learn expressive joint policies before distilling them into efficient, decentralized one-step policies. The method achieves strong multi-agent coordination performance across diverse benchmarks while offering a 14.5x faster inference speed compared to diffusion-based approaches.
Key insights
- Flow matching provides a robust and expressive mechanism for modeling complex, multi-modal joint action distributions in offline multi-agent datasets.
- Distilling an expressive, centralized flow-based policy into decentralized, one-step policies can significantly reduce inference time without compromising coordination performance.
- The Individual-Global-Max (IGM) principle is crucial for stable critic optimization and consistent Q-value estimation, especially when combined with flow-based policy learning, ensuring effective value guidance for multi-agent coordination.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.