PhysVista: Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop

PhysVista. Benchmarking Physical Intelligence in VLMs via a Perception-Reasoning-Assessment Loop. By Xinge Peng and six others, from the University of Science and Technology of China, ByteDance, and one other institution. Published on the thirtieth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Vision-language models show strong multimodal reasoning, but whether they truly capture the physical consistency behind real-world dynamics is still unclear. Existing benchmarks tend to evaluate in fragments. They focus on isolated cognitive stages, testing one step at a time, and overlook how perception, reasoning and physical judgment work together as a whole. That missing holistic view has a practical cost. It makes it hard to diagnose whether these models can reliably judge the physical authenticity of what emerging generative models produce.

PhysVista evaluates physical intelligence through a closed cognitive loop, inspired by the way a person sees, then reasons, then assesses. It restores that loop by evaluating three stages jointly rather than one at a time. The first is perception of physical state: what is there, and in what condition. The second is reasoning about physical dynamics: how things will move and change. The third is assessment of physical plausibility: whether what was shown could really happen.

Reasoning is divided further, into event-level reasoning and scale-level reasoning, which allows a finer-grained analysis of where physical understanding holds and where it breaks. The videos come from two sources: the real world, and AI-generated content. That mix lets the benchmark span diverse domains as well as the newer scenarios that generative models create, which matters for a benchmark partly meant to test whether a model can judge generated video.

Extensive experiments across a diverse set of vision-language models reveal substantial limitations, in physical reasoning and in plausibility assessment alike. The authors read the results as a persistent gap between visual recognition and genuine physical understanding.

A model can recognise what is in a clip and still miss that the clip breaks physics. PhysVista is built to pull those two abilities apart and score them separately, and the experiments reported find the first running ahead of the second. From there the authors point toward more principled designs for multimodal intelligence that is physically grounded.