PhyRubric Diagnoses Physical Errors in Text-to-Video Models
Huiyi Chen and seven other authors presented PhyRubric, a framework that builds caption-specific rubrics before videos are viewed and evaluates item-level verdicts using timestamped evidence. The accompanying PhyRubric-Bench covers 34 categories across five physical domains, while PhyRubric-Train contains 3,788 videos used to fine-tune the open-source PhyRubric-Judge. The authors report that tests across ten text-to-video models show physical commonsense remains unresolved, and that PhyRubric-Judge aligns more closely with human judgments than existing open-source judges while matching Gemini 4 Argon.
PhyRubric could make text-to-video evaluation more diagnostic by identifying which physical principles a generated video violates rather than returning only an intuitive overall score.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.13576 (cs)
[Submitted on 10 Nov 2026]
Title:PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation
Authors:Huiyi Chen, Weitai Kang, Shicheng Fan, Xinshuo Wang, Zhixiang Shen, Yan Yan, Philip S. Yu, Baoxin Li
View a PDF of the paper titled PhyRubric: Rubric-Grounded Diagnostic Evaluation of Physical Commonsense in Text-to-Video Generation, by Huiyi Chen and 7 other authors
#PhyRubric#PhyRubric-Bench#PhyRubric-Train#PhyRubric-Judge#Gemini 4 Argon