RewardVerse introduces a rubric-based reward framework for RL-trained video generation models, addressing scalar drift in reward models that map subjective video quality to an unstable single score. A two-stage Rubric-Guided Policy Optimization algorithm first warms up a scorer on self-evolving seed rubrics, then jointly trains a rubric generator to produce query-adaptive evaluation criteria aligned with human ratings. On the 16-dimensional EvalVerse benchmark, RewardVerse achieves state-of-the-art pointwise and pairwise evaluation performance and a more interpretable RL reward signal.
