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

What it claims

Physis-Lang (NVIDIA, MIT, Oxford) argues that video world models do not need extra visual, latent, numerical or simulator signals to become more physically plausible. Its position is that the caption language can carry the physics if it is made explicit and optimized. Two pieces are concrete enough to reuse:

• PhysCapBench: a held-out set of 246 physical videos for scoring physics-aware captions.
• PhysThinker: two distilled 4B vision-language models. PhysThinker-C writes physics-aware captions and PhysThinker-U upsamples prompts at inference. They replace the costly GPT-based captioner and upsampler.

The headline is that Cosmos3-Nano trained this way beats Veo 3.1 under the same evaluation protocols. The fetched text does not give the numbers behind that.

How it works

Caption format. Each training caption has three parts:

• the base caption;
• a natural-language physics_reasoning field covering causes, governing laws, dynamics and effects;
• at inference, a scene-specific physics_negative_prompt describing likely physically implausible evolutions, used as negative conditioning.

Instruction evolution. The captioner's weights stay fixed throughout; only the captioning instruction changes.

• The loop runs on a fixed 20-video development set. For each video there are human-curated reference assertions organized as cause, physical law and effect.
• A physics-aware critic scores each caption. Recall measures coverage of the reference assertions. Precision measures whether each claim is faithful to the video.
• An evolution agent reads the critic's failure report (omissions, unsupported claims, vague causal links) and rewrites the instruction.
• After each iteration the instruction is scored on PhysCapBench. The loop stops when the gain saturates, and the best instruction is used to re-caption the training corpus.

flowchart LR
  P[Instruction P_t] --> C[Fixed VLM captioner]
  C --> K[Critic on 20-video dev set]
  K --> A[Evolution agent]
  A --> B{PhysCapBench gain saturated?}
  B -- No --> P
  B -- Yes --> R[Best instruction: re-caption corpus]

Targeted data retrieval.

1. A GPT-5.5-based diagnosis agent assigns the generator's failures to physics categories, such as rigid-body motion, collision and fluid dynamics.
2. The results are aggregated into a category-level deficiency profile.
3. Videos in a large gallery carry physics-domain tags derived from their existing captions. Videos are matched to the deficient categories by these tags, not by visual similarity.
4. The retrieved videos are re-captioned with the evolved instruction and added to training.

Evidence

• Benchmarks: PhyGround, PhyGenBench, Physics-IQ (Verified) and VideoPhy-2.
• Backbones: Wan and Cosmos.
• Comparators shown in Figure 1: CogVideoX1.5-5B, Wan2.2-TI2V-5B, HunyuanVideo-1.5, Cosmos3-Nano and Veo 3.1.

The fetched excerpt states "consistent improvements" and the Veo 3.1 result, but includes no per-benchmark scores, deltas or compute figures.

Limits

• The fetched text cuts off at Sec. 4.1, so no results tables, ablations or stated limitations are available here.
• No links to code, weights or licence appear. The PhysThinker models and PhysCapBench may or may not be released; check before planning around them.
• The original pipeline depends on proprietary models: a GPT-based captioner and upsampler, plus GPT-5.5 for diagnosis. The 4B distillations address cost, but how much quality they lose relative to the GPT versions is not shown in this excerpt.
• Instruction evolution is driven by only 20 development videos, and stopping is decided on the same 246-video PhysCapBench that is also used as a benchmark. The PhysCapBench caption scores are therefore not independent of the tuning.
• Physics-domain tags come from existing gallery captions, so retrieval is only as good as those captions.