Video World Models · 2026-10-01
Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model
Liming Lu, Xianzheng Ma, Wenkun He, Guanqi Zhan, Yilin Zhao, Junyu Chen, Mengyao Xu, Jiaojiao Fan, Wenhang Ge, Yuchao Gu, Yunze Liu, Boyi Li, Zhen Dong, Victor Prisacariu, Ming-Yu Liu, Song Han, Han Cai
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Physis-Lang is a self-evolving language framework for video world models. It represents physical processes with explicit “physics_reasoning” and uses an agentic caption-evolution loop plus PhysCapBench (precision/recall on atomic physical assertions) to iteratively refine caption instructions. It also converts model failures into deficiency profiles for language-guided retrieval and re-captioning, then fine-tunes video backbones (Wan/Cosmos) with the evolved physical captions, improving physical plausibility and surpassing Veo 3.1 starting from Cosmos3-Nano.
Why it matters
The authors address a gap in video world models: generating visually plausible videos that violate basic physical principles. Physis-Lang treats physical language as a shared representation across data curation, training supervision, and inference guidance, using an explicit physics-aware critic and a held-out caption benchmark to evolve captions and expand targeted training data toward missing physical processes.
Method
- Use an agentic caption-evolution loop with a physics-aware critic to refine captioning instructions without updating the captioner’s weights.
- Introduce PhysCapBench, a held-out benchmark of 246 physical videos with 3,794 human-verified atomic physical assertions, scored via precision/recall and F1.
- Use a language-guided data engine: summarize model failures into physics-category deficiencies, retrieve tagged videos by physical content, re-caption them with evolved guidelines, and fine-tune Wan/Cosmos backbones.
Abstract (from arXiv)
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. We revisit this assumption and introduce Physis-Lang, a self-evolving framework that treats physical language as a shared and optimizable representation across data curation, model training, and video generation. Physis-Lang represents physical processes through language that describes their relevant entities, causes, interactions, governing principles, temporal evolution, and effects. To improve this representation, we construct PhysCapBench, which decomposes physical processes into atomic assertions and evaluates captions using recall and precision. An agentic loop iteratively analyzes assertion-level errors and refines the instruction used to produce physical captions. Physis-Lang further converts model deficiencies into textual descriptions and uses language-guided retrieval to identify visually diverse videos that cover missing physical processes. Experiments on four widely used physical video benchmarks with Wan and Cosmos backbones demonstrate consistent improvements in physical plausibility. Notably, starting from open-source Cosmos3-Nano backbones, our Physis-Lang-enhanced models surpass the leading proprietary Veo 3.1 model.
Related papers
- PhyEvolve: Multi-Agent State-Evolution Planning for Physics-Aware Video Generation
- Reading, Writing, Sharing, and Interacting with the World through Video
- Time-JEPA: Grounding Video World Models with Physical Time Series
- PhysTrace: Interpretable Temporal Evaluation of Physical Consistency in Video Generative Models
- World Models Need a Physical Registry: Learning Internal Predictive States for Physically Plausible World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-09-30 · Indexed by Pub-AI: 2026-10-01 UTC
Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model
Preprint posted: 2026-09-30 · Original record ↗
- arxiv
- 2609.40358
Explicit source versions (1)
- v1 · 2026-09-30
- Preprint available
Source date: 2026-09-30 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
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. We revisit this assumption and introduce Physis-Lang, a self-evolving framework that treats physical language as a shared and optimizable representation across data curation, model training, and video generation. Physis-Lang represents physical processes through language that describes their relevant entities, causes, interactions, governing principles, temporal evolution, and effects. To improve this representation, we construct PhysCapBench, which decomposes physical processes into atomic assertions and evaluates captions using recall and precision. An agentic loop iteratively analyzes assertion-level errors and refines the instruction used to produce physical captions. Physis-Lang further converts model deficiencies into textual descriptions and uses language-guided retrieval to identify visually diverse videos that cover missing physical processes. Experiments on four widely used physical video benchmarks with Wan and Cosmos backbones demonstrate consistent improvements in physical plausibility. Notably, starting from open-source Cosmos3-Nano backbones, our Physis-Lang-enhanced models surpass the leading proprietary Veo 3.1 model.