Reproduce “When Anchors Shift the Target: Advantage-Gated Adaptation for 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

Pu Cao, Min Zhou, Ying Shu, Tiezheng Ge, Qing Song, Lu Yang

AGA adapts a text-to-image model by scoring generated images, re-noising them, and fitting velocity targets without differentiating through the sampling chain. It fits positive examples to their targets and uses a moving-policy tangent to stabilize negative updates, with a frozen pretrained reference for retention. On SD3.5-Medium, it reports 0.94 OCR accuracy and 0.83 ImageReward in one OCR setting, across three tasks.

Summary by alphaXiv.

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