Reproduce “Reinforcement Learning via Fixed-point 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
Denis Blessing, Onur Celik, Gerhard Neumann
Fixed-point flow matching (FFM) trains expressive flow policies for continuous-action reinforcement learning without samples from the improved policy. It generates actions from the current policy and regresses the flow using critic-derived action gradients. On OGBench, its mean success rate rises from 0.63 offline to 0.95 after online fine-tuning, exceeding the next-best 0.93; it requires differentiable critics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.