WorldExam benchmarks whether video models generate worlds that actually react

Yuxue Yang · hf · 2026-08-04

WorldExam is a benchmark for treating controllable video generation systems as world models, going beyond visual quality to test whether the generated world reacts plausibly.
It defines four diagnostic levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity.
The benchmark spans 1,474 cases across 8 tasks and supports camera-, action-, and language-driven paradigms.
Evaluating 20 representative models reveals a split: camera-driven models handle camera control well but lack dynamic interaction, action-driven models control subjects more precisely but often leave the world unresponsive, and language-driven models interact better but follow complex controls less faithfully.
The takeaway is that good-looking videos and explicit instruction following are not enough to make a strong world model.

More from Multimodal