arXiv:2609.03153 · September 2026

VeriPhy

Agentic physical reasoning for world-model evaluation and refinement

Video models make clips that look right and behave wrong — a frisbee turns into a ball, a pot melts instead of breaking. VeriPhy measures what went wrong, says exactly what it was, and has the model generate again.

Baseline
After VeriPhy
0.0s / 0.0s

The prompt

What changed in the text

What the critic caught

    Show the full measurements
    The method

    The critic plans before it looks.

    The order matters more than any single component. VeriPhy commits to what the video must show while it still has only the sentence, so a convincing picture never gets to negotiate the standard it will be judged against.

    01

    It writes down what the video will owe

    From the prompt alone, before any frame exists, the critic produces a list of obligations: how many objects should be present, which way each should travel, what should stay upright, what should remain legible, what should be audible and when. Because the list comes from the text, a fluent video cannot talk the critic out of it.

    02

    It measures instead of guessing

    Each obligation goes to a frozen specialist tool, never to a general opinion. A tracker follows objects across frames, a counter counts them, a depth estimator says what is in front of what, a reader reads text on screen, a listener checks whether a promised sound occurs. None of these tools is ever asked whether the video looks good.

    03

    Every verdict keeps its receipt

    An obligation comes back supported, contradicted, or unknown. The third answer is what buys the other two their credibility: when the tracker loses an object, the critic abstains rather than inventing a score. Each contradiction is stored with the measurement, the time window and the rule that produced it.

    supportedcontradictedunknown
    04

    It rewrites the prompt and rolls again

    The violated obligations are folded back into the prompt as explicit instructions, and the same generator runs again with the same seed and the same control inputs. No fine-tuning, no second model, no post-processing, no frame surgery. The sentence is the only variable — which is what makes the comparison at the top of this page a fair one.

    The run

    What moved, and by how much.

    One round of evaluate, rewrite, regenerate, closed end to end on Cosmos3-Nano — a generator held out from VeriPhy's development — across the 600 VideoPhy-2 test prompts. Read every number below as VeriPhy's own plausibility score: it reflects the obligations this critic checks, and nothing else.

    The critic against human-written flaws

    On a 149-clip core carrying 304 human-written, prompt-grounded flaw records, VeriPhy accounts for more defects than a published question-decomposition evaluator given the same clips and the same claims — and each of its findings keeps the measurement, time window and rule that produced it.

    Distilling lessons from experience clips and loading them as readable context lifts held-out recall from 340 to 375 of 502 (67.7% to 74.7%), with no retraining.

    The independent rater

    The same clips scored by VideoPhy-2's official AutoRater, which shares no component with this critic.

    Baseline n = 600, rewrite n = 589 (clips completing the loop with a paired score). Hard subset follows VideoPhy-2's designation.