Physicalrealismbench Attributable Physical Realism Evaluation For Video World Models
Source
reka.ai
Author
Reka AI
Date
Context
PhysicalRealismBench-U asks a video-understanding model to locate a physical mistake, rather than merely say that something looks wrong. Reka AI’sbenchmark description uses simulated scenes with deliberately introduced violations and precisely known answers. To receive credit for detecting a violation under its stricter metric, the model must also identify the affected object and the relevant frames. Comparing this score with simple detection performance reveals whether a model can connect its judgment to visible evidence. That is useful when assessing models intended to judge the physical plausibility of generated video.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
