← News

ResearchSources: arXiv · 19 Nov 2026

EVA-RSI trains video evaluators to catch motion flaws through self-improvement

A research team describes a multi-agent setup that lets a video-judging model improve itself without a built-in verifier, reporting stronger motion scoring and top benchmark results among 21 models.

In plain words

Models that generate video need other models to judge whether the motion looks right, but current judges often miss motion problems. The paper's authors propose a framework where several AI agents improve such a judge step by step, using edited videos as a test instead of human labels. They report the improved judge scores motion better and transfers to flaw types it never saw during training.

What was announced

The authors frame two obstacles to applying recursive self-improvement to video judging: there is no automatic verifier, and quality spans several dimensions, such as physics, smoothness and dynamics, where improving one can hurt another. Their answer is EVA-RSI, a multi-agent framework in which each dimension gets an LLM agent proposing updates, a leader agent coordinates them, and a Nash bargaining rule picks changes that balance gains across dimensions. Supervision comes from a fixed proxy exam built by editing videos, for example reversing them in time or dropping frames, with a small human-labeled set anchoring the exam to human judgment. Humans still design the exam, review candidate changes and revise the curriculum between stages.

Starting from Qwen3-VL-8B, the team reports that EVA-RSI improves motion evaluation, with the biggest gain on smoothness, and that the improvements transfer to defect types not seen during the run. On dynamics subsets of public benchmarks, they say the evaluator posts the highest observed scores among 21 tested models. Used as a reward signal for reinforcement learning, it reportedly raises generated motion and earns human preference, and a data-selection pipeline built around it also increases motion. These are the authors' own reported results, not independent confirmations.

What it does not show

  • The source is an abstract only; full method details, experimental setup and caveats are not available for verification.
  • Benchmark scores and human-preference results come from the paper's authors and have not been independently confirmed.
  • The abstract does not state whether code, models or the proxy exam are publicly released.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.