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

Physis-Lang. Self-Evolving Language as a Physical Representation for Video World Model. By Liming Lu and sixteen others, from NVIDIA, MIT, and one other institution. Published on the thirtieth of September. No code repository is linked to it.

A video world model is supposed to predict how the physical world evolves. In practice these models often produce videos that look plausible and still break basic physical principles. The common response assumes that natural language cannot carry the physical knowledge reliable generation needs, and so it brings in extra signals: visual ones, latent ones, numerical ones, or signals based on planning. Each of those adds another channel next to the text, on the premise that words alone run out before the physics does.

Physis-Lang revisits that assumption. It treats physical language as the representation, one that is shared and can be optimised across the whole pipeline: curating data, training the model, and generating video. A physical process is described in language that covers its relevant entities, its causes, the interactions involved, the principles that govern it, how it unfolds over time, and its effects. The bet is that a description rich enough along those lines can carry what generation needs, and that the description itself can be made better over time, which is where the self-evolving part of the name comes from.

Because the representation is meant to improve, it needs a measure. That is PhysCapBench, which breaks a physical process down into atomic assertions and scores captions by recall and precision over them. An agentic loop then analyses errors at the level of those assertions and iteratively refines the instruction used to produce physical captions. The same language also points the data in the right direction. Physis-Lang converts the model's deficiencies into textual descriptions, and uses retrieval guided by that language to find visually diverse videos covering the physical processes that are missing.

Experiments run on four widely used physical video benchmarks, with Wan and with Cosmos backbones, and show consistent improvements in physical plausibility. Starting from open-source Cosmos3-Nano backbones, the models enhanced with Physis-Lang surpass the leading proprietary model, Veo three point one.

Where others add new kinds of signal, this work asks more of the one already there. Language describes the physics, a benchmark grades the description, a loop rewrites the captioning instruction, and the model's own failures, put into words, decide which videos it sees next. Every turn of that cycle is written in words, and the words are what keep getting better.