Jinpeng Lu, Dexu Zhu, Haoyuan Shi, Linghan Cai, Guo Tang, Yinda Chen, Jie Cao, Duyu Tang et al.
Current world models fail to maintain a persistent internal state that evolves independently of observation; WRBench is introduced to diagnose this blind spot.
World models must have an internal state that evolves over time, decoupled from observation, so that objects and events persist even when unobserved. Existing benchmarks only evaluate surface properties like fidelity and controllability, ignoring whether the world keeps evolving when unobserved.
WRBench treats camera motion as an intervention on observability and evaluates three aspects: (1) whether the camera executes the requested interaction, (2) whether the scene remains continuous and identifiable while in view, and (3) whether a returning target is consistent with the event that was set in motion. Experiments cover 23 models, 4 control paradigms, and 9600 videos.
All models exhibit a 'tracking shot' failure: they pause event progression when unobserved and resume from the abandoned state. This failure is consistent across paradigms, model families, and scales, indicating that robust world-state evolution does not follow from cleaner imagery or larger models. The work argues for making state kernel stability and worldline consistency first-class objectives.