RewardVerse: Rubric-Guided Policy Optimization for Video Reward Modeling
The paper’s claim is that video reward models score more reliably when they first build an explicit rubric for the prompt.
RewardVerse inserts generated evaluation criteria between the query and the scorer, aiming to stop scalar scores from drifting across prompts. Its RGPO training process first warms up the scorer with self-evolving seed rubrics, then jointly trains the rubric generator and scorer against human ratings. The authors report state-of-the-art pointwise and pairwise results on EvalVerse and external datasets, with a reward signal they describe as more stable and interpretable for video-generation RL. HF Daily Papers' note
RewardVerse inserts generated evaluation criteria between the query and the scorer, aiming to stop scalar scores from drifting across prompts. Its RGPO training process first warms up the scorer with self-evolving seed rubrics, then jointly trains the rubric generator and scorer against human ratings. The authors report state-of-the-art pointwise and pairwise results on EvalVerse and external datasets, with a reward signal they describe as more stable and interpretable for video-generation RL. HF Daily Papers' note
score 5