WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity

WorldExam introduces a hierarchical diagnostic benchmark for controllable video generation models viewed as world models. It evaluates four levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. The benchmark contains 1,474 cases across eight tasks and evaluates 20 representative models spanning camera-, action-, and language-driven paradigms. The reported results show a capability split: camera-driven models handle camera control well but lack interaction interfaces; action-driven models control subjects more precisely while often leaving the surrounding world unresponsive; language-driven models support interaction better but follow complex controls less faithfully. No evaluated model combines broad task coverage with consistently strong performance.