Reproduce “Elucidating the Design Space of Regression-based Diffusion Reinforcement Learning” 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

Toyota Li, David Zhao, Alan Zhao

DiffusionRFT frames diffusion-RL methods as reward optimization over generated samples, with a divergence penalty limiting movement from the generating model. Reverse KL yields exponential weights, while Pearson chi-squared requires sparsemax projection to keep linear weights nonnegative. On Stable Diffusion 3.5-Medium, sparsemax with a variance-minimizing baseline reached 0.983 GenEval evaluation reward; evidence is limited to this setup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.