Physical reasoning requires adapting familiar expectations to the dynamics of the observed world. Existing benchmarks report an overall score but do not reveal whether answers follow visual evidence, especially when familiar rules no longer apply.
We introduce InfiniteWorld, a benchmark that pairs physical‑reasoning questions with simulator videos whose physical parameters are systematically altered. The generation pipeline connects simulation, scene screening, paired video translation, question construction, and input checks.
The benchmark defines five tasks: law identification, parameter estimation, temporal reasoning, analytic counterfactual comparison, and law induction.
Evaluation uses five matched input settings for each question: text‑only, ordered or shuffled observations from the simulator, and the corresponding translated videos. We evaluate a diverse set of video‑language models under these settings.
Results expose a persistent preference for familiar physical rules. Frames improve performance for most models, but gains differ across tasks. Models achieve strong accuracy on Earth‑normal questions yet show poor recognition when dynamics change, often still providing Earth‑normal answers. Task scores and semantic agreement further distinguish matching physical factors from matching their amount, and correct answers from repeated errors.
InfiniteWorld makes these behaviors measurable by providing paired changes in the evidence supplied for the same question, separating visual evidence from familiar physics.
Review