Slop TVNewsLatest
Explainers

PhysResponse finds video models do not follow the physics you set

The paper tested ten image-to-video models and found even the best reproduces both outcomes of a physics transition pair less than half the time.

Illustration: PhysResponse finds video models do not follow the physics you set
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • PhysResponse, posted to arXiv on November 18, 2026, evaluated ten image-to-video models and reports that even the lowest parameter errors exceed those of physically correct videos by over an order of magnitude.
  • Across the ten models, the rank correlation between the gravity a model was prompted with and the gravity recovered from its motion runs from minus 0.13 to plus 0.16, against plus 0.77 for ground-truth videos.
  • MiniMax-H3 leads the ten on gravity and friction, but under a Moon-gravity prompt its motion matched an effective gravity of 11.8 meters per second squared, and no model clears 50% pair accuracy on transition pairs.
  • The authors fine-tuned Wan2.2-TI2V-5B on 72,000 simulated videos; held-out speed error fell from 93.1% to between 13.0% and 16.4%, the paper says.

Video generation companies are selling their models as world simulators, systems that answer to the physics you describe. A new benchmark tested whether they actually do, and found they mostly do not.

PhysResponse, posted to arXiv on November 18, 2026, measures whether a generated clip's motion matches the gravity, friction and initial speed set in the prompt, rather than only whether the motion looks plausible. Its authors, led by Yuhan Li of Sun Yat-sen University with Xiaodan Liang as last author, evaluated ten image-to-video models and report that "even the lowest parameter errors exceed those of physically correct videos by over an order of magnitude." The paper is a preprint. It has not been peer-reviewed and no independent group has replicated it.

What PhysResponse measures

The benchmark works at two levels. Scene-conditioned simulator inversion replays the source scene while varying one parameter and recovers the effective gravity, friction and initial speed of what the model actually generated. Event evaluation tests whether the model produces the right outcome near and across a critical transition, such as whether a moving object stops before it reaches another or knocks it over instead.

The authors built the test around 20 everyday interaction types, scenarios T01 to T20, that span four mechanisms: inclined-surface motion, collision-mediated transfer, changes of support, and multistage cascades. Gravity affects acceleration and fall timing, friction affects sliding and transfer through contact, and initial speed decides whether two objects meet at all. The paper edits gravity across five levels from 1.62 to 24 meters per second squared.

What the benchmark found

The results are blunt. The paper titles its results table "No model follows the prompted physics." Within a scene, the rank correlation between the gravity a model was prompted with and the gravity recovered from its motion runs from minus 0.13 to plus 0.16 across the ten models, against plus 0.77 for ground-truth videos. Fitting the reference videos through the same pipeline returns gravity, friction and speed errors of 3.6%, 1.4% and 1.1%, so the models' errors are roughly an order of magnitude worse.

MiniMax-H3 leads the ten on gravity and friction, and it still shows the failure plainly. Under a Moon-gravity prompt, the paper reports its motion matched an effective gravity of 11.8 meters per second squared, not the value it was asked for. Moon gravity is about 1.62.

On the event side, high accuracy near a transition does not carry across it. MiniMax-H3 has the best near-accuracy of the ten at 67.0% and 69.6%, but its pair accuracy, whether it reproduces both outcomes of a transition pair, falls to 23.1% and 18.7%. Wan2.2-I2V-A14B leads pair accuracy at 30.7% and 42.1%. No model clears 50%, and the physics post-trained model PhysRVG scores 0% on pair accuracy under both prompt formats.

Why the usual metrics miss it

The two levels catch different failures, and the paper reports that visual metrics such as PSNR miss both. Cosmos3-Nano recovers gravity within 9.6% of the specified value yet misses a complete domino chain, an event failure the pixel-level score does not see. Cosmos-Predict2 matches both contact outcomes while both of its speed fits are rejected. A video can look right and answer the wrong physics.

That is the gap the paper aims at. Leaderboards for image-to-video models score appearance and motion quality. The rank correlation numbers above, close to zero, measure a different axis: how much the specified number controls the result.

The ten models it tested

The paper evaluated ten image-to-video models on its Core100 set across four prompt formats. They are CogVideoX-1.5, Cosmos-Predict2, Cosmos3-Nano, HunyuanVideo-1.5, MiniMax-H3, Wan2.2-I2V-A14B and Wan2.2-TI2V-5B, along with three physics post-trained variants built on the Wan2.2 bases: PhysAlign, PhysRVG and VideoGPA.

Cosmos3-Nano is NVIDIA's. NVIDIA's own Cosmos page describes the model family as a way to "Run as a controllable, physics-grounded world simulator to predict single-view or multi-view world states." The paper's critical-boundary table places Cosmos3-Nano at 12.3% and 14.8% pair accuracy. A model marketed as a physics-grounded simulator sits in the middle of a test built to measure exactly that.

The training fix the authors built

The paper's second half is a remedy. The authors generated 72,000 simulated videos across the 20 interaction types with controlled parameters, from 24,000 simulation runs of three synchronized viewpoints each, 3,600 videos per type.

Fine-tuning Wan2.2-TI2V-5B on the data lowers parameter errors, including on held-out interaction types the model was not trained on, the paper says. Held-out gravity error falls from 140.3% to between 39.8% and 68.6%, and speed error from 93.1% to between 13.0% and 16.4%. Adding a numerical encoder that reads the prompted values directly and continued training lowers all three errors further on held-out types. The gains do not automatically improve event scores: the authors report that one variant improves all ten event metrics while another lowers three.

The fix is a training recipe, not a product, and it lands on one model in one lab. What it shows is that controlled physics data moves the number.

What it means for your prompts today

If you are prompting for a specific physical behavior, a heavy object, a slippery surface, a fast approach, treat the number as a description, not a control. The benchmark says the model's motion barely tracks the value you set, and the quality score you are watching will not warn you either.

The practical move is to check the behavior you care about by eye, the timing of the fall, whether the object stops where it should, and re-roll rather than trusting the prompt to hold. The same gap runs through the physics post-trained models in the test: the paper reports their gains over the Wan2.2 bases are uneven.

Weigh the paper before you hang a decision on any single number in it. It is 11 pages, main paper and references only, with supplementary material promised later, and the held-out results rest on four interaction types.

Coverage of PhysResponse so far is thin: the paper's own pages and aggregator listings, with no independent write-up that SLOP TV found. Related work points the same way from a different angle. A separate October 2026 benchmark, World Models' Last Exam in Physics, tested eight video generation models across 40 physical tasks and reported "persistent physical inconsistencies," with the best model scoring 57.76 out of 100. Neither result has been replicated.

SLOP TV's November 19 story on StrideWorld covers a neighboring failure: that benchmark found 14 of 15 interactive world models deform the scene as the camera moves. StrideWorld is about geometry holding still. PhysResponse is about whether a model answers the physics you set. The category background lives in SLOP TV's world models guide.

The paper is at arxiv.org/abs/2611.24419; read the results tables before quoting any single model's score, and until someone replicates it, prompt for physical behavior and check the motion yourself.

Sources

  1. arxiv.org - the paper's abstract, authors and submission date
  2. arxiv.org - the paper's body: the model roster, the parameter and event results, the fine-tuning study
  3. nvidia.com - NVIDIA's own framing of Cosmos 3 as a physics-grounded world simulator; Cosmos3-Nano is one of the ten models tested
  4. arxiv.org - an independent benchmark on physical consistency in video world models, for the wider gap
  5. sloptvnews.com - SLOP TV's related coverage, for the differentiation