Reproduce “Post-Training Frontier Text-to-Image Models by Composing Preference and Rubric Rewards” 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

Yuanhao Ban, I-Hung Hsu, Anastasios Angelopoulos, Wei-Lin Chiang, Ion Stoica, Cho-Jui Hsieh

The authors combine human preference with checks for requested details and unwanted additions, applying constraints only to prompts requiring strict adherence. A one-sided veto removes preference credit when detectors flag known failures. On 1,000 validation prompts, Stage 3 reaches a 0.6598 pairwise win rate against the frozen base, versus 0.589 for Arena-T2I-Hard; this is not live human Elo.

Summary by alphaXiv.

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