Controllable video generation models are increasingly positioned as world models, but evaluating them in that role requires more than checking visual fidelity or explicit instruction following. A new benchmark, WorldExam, introduced by researchers at Hugging Face, aims to assess whether these models can infer how a world should react from a given scene state and generate plausible consequences beyond what is explicitly described.
WorldExam is a hierarchical diagnostic benchmark spanning four levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. It comprises 1,474 cases across eight dedicated tasks and supports unified evaluation of camera-, action-, and language-driven model paradigms. The World Reactivity level specifically evaluates scene-conditioned reactions and goal-directed behaviors that go beyond explicit input specifications.
Evaluation of 20 representative models reveals a clear capability split. Camera-driven models excel at camera control but lack interfaces for dynamic interaction. Action-driven models control subjects more precisely but often leave the world unresponsive. Language-driven models perform better on interaction but follow complex controls less faithfully. Notably, no model combines broad task coverage with consistently strong performance, indicating that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity.
The benchmark provides a unified evaluation protocol across different model paradigms, making it a valuable tool for the research community. The project website is available at worldexam.github.io.