A new paper proposes grading AI-generated video with written rubrics instead of a single gut-check score.
Researchers behind RewardVerse built a reward model for video generation that first drafts explicit evaluation criteria for a given prompt, then scores the video against those criteria rather than guessing a number outright. The training process, called Rubric-Guided Policy Optimization, runs in two stages: it warms up the scorer using self-generated seed rubrics, then jointly trains a rubric generator and scorer so the criteria adapt to each query while staying aligned with human ratings. The team tested the system on EvalVerse, a benchmark spanning 16 quality dimensions, plus outside datasets. They report the rubric-based approach curbs what they call scalar drift - the tendency for a reward model's scoring scale to shift or collapse across different prompts - and claim state-of-the-art results on both single-video and head-to-head video comparisons.
Reward models are the referee for reinforcement learning in video generation: get the scoring wrong and training drifts toward whatever the model can game, not what actually looks good. A rubric that adapts per query is a sensible answer to a problem RLHF researchers have wrestled with for years in text models, where unstable or exploitable reward signals produce over-optimized, hollow outputs.
The paper does not name the specific prior reward models it claims to beat or publish the score margins on EvalVerse, so "state-of-the-art" is a claim to take on faith until the numbers or code show up.