Paper Overview
Research Area: Computer Vision (CV) Authors: Liming Lu, Xianzheng Ma, Wenkun He, et al. Published: 2026-09-30 arXiv: 2609.40358
Summary
Video world models are expected to predict how the physical world evolves, yet they often produce visually plausible videos that violate basic physical principles. Existing approaches commonly assume that natural language is insufficient to represent the physical knowledge required for reliable generation, and therefore introduce additional visual, latent, numerical, or planning-based signals. Physis-Lang revisits this assumption.
Key ideas
- Self-evolving physical language: Treats physical language as a shared and optimizable representation spanning data curation, model training, and video generation.
- Rich physical descriptions: Represents physical processes through language describing relevant entities, causes, interactions, governing principles, temporal evolution, and effects.
- PhysCapBench decomposes physical processes into atomic assertions, evaluating description text via recall and precision.
- An agent loop iteratively analyzes assertion-level errors and refines the instructions used to generate physical descriptions.
- Model deficiencies are converted into text descriptions.
- Language-guided retrieval identifies visually diverse videos covering missing physical processes.
PhysCapBench and the agent loop
Closing the loop on model weaknesses
Results
Across four widely used physics video benchmarks, on both Wan and Cosmos backbones, experiments consistently show improved physical plausibility. Notably, starting from the open-source Cosmos3-Nano backbone, the Physis-Lang-enhanced model surpasses the leading proprietary model Veo 3.1.
--- *Auto-collected on 2026-10-02*