Reproduce “GenRM-Flow: Generators are Process-aware Reward Models in 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

Siming Fu, Ruizhe He, Zheming Fu, Zeyue Xue, Fengkai Liu, Yufei Jiao, Jie Huang, Xiaoxiao Ma and 2 more

GenRM-Flow uses a preference-tuned video generator’s velocity predictions to estimate how well generated videos match human preferences. It compares prediction errors from the tuned generator and a reference, giving an approximate reward without pixel decoding or an auxiliary scoring head. On held-out pairs, it achieves 88.96% accuracy versus PAVRM’s 82.64%, but downstream rewards assess forward-noised final outputs, not sampled paths.

Summary by alphaXiv.

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