How AI is applied across API Evangelist and APIs.io. Read my AI disclosure →
API Evangelist API Evangelist
Discovery
Learnings
Guidance
Toolbox
Alignment
API Evangelist LLC

PhysicalRealismBench: Attributable Physical Realism Evaluation for Video World Models

calendar_today June 9, 2026 person domain reka-ai

PhysicalRealismBench-U is a new evaluation framework using synthetic video data to assess how well vision-language models understand physics. The benchmark shows even state-of-the-art models fail basic physical reasoning tasks, scoring as low as 0.14-0.60 on realism, and identifies failure modes like frame-skipping and spatial reasoning shortcuts.

open_in_new Read original post