Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model

arXiv 2026 34 method

TLDR

Physis-Lang uses self-evolving physical language as a shared representation for video world models, improving physical plausibility via benchmark-driven caption refinement and data curation.

Reasoning

The paper presents a novel framework that revisits language's role in physical video modeling, supported by a new benchmark and experiments on multiple backbones. Weaknesses include limited visible details on quantitative results and potential limitations of language for precise physical dynamics.

Read-first score

Read-first score 34, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,experiment

Topical relevance 25%

58.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 315.

Keyword Scores

video world model

world model

world dynamics prediction

generative world model

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces Physis-Lang, a self-evolving framework that treats physical language as a shared and optimizable representation across data curation, model training, and video generation.
Challenges the assumption that natural language is insufficient to represent the physical knowledge needed for reliable video generation.
Introduces PhysCapBench, which decomposes physical processes into atomic assertions and evaluates captions using recall and precision.
Uses an agentic loop that iteratively analyzes assertion-level caption errors and refines the instruction used to produce physical captions.
Converts model deficiencies into textual descriptions and uses language-guided retrieval to identify visually diverse videos covering missing physical processes.

Methodology

Physis-Lang represents physical processes through language describing entities, causes, interactions, governing principles, temporal evolution, and effects, and treats this language as an optimizable representation across data curation, training, and generation. PhysCapBench evaluates physical captions by decomposing processes into atomic assertions scored with recall and precision, and an agentic loop refines captioning instructions based on assertion-level errors. Model deficiencies are converted into textual descriptions to guide retrieval of diverse training videos, with experiments using Wan and Cosmos backbones on four physical video benchmarks.

Key Results

Physis-Lang-enhanced models demonstrate consistent improvements in physical plausibility across four physical video benchmarks with Wan and Cosmos backbones. Starting from open-source Cosmos3-Nano backbones, the enhanced models surpass the leading proprietary Veo 3.1 model.

Resources

Paper