Paper Overview
Field: Computer Vision Authors: Jinpeng Lu, Dexu Zhu, Haoyuan Shi Released: 2025-06-20 arXiv: 2506.16800
Summary
World models are increasingly regarded as a decisive step toward artificial general intelligence, yet modeling the physical world demands more than rendering convincing frames on demand: it requires an internal world state that keeps evolving over time, decoupled from observation, so that objects endure and events run to their conclusions whether or not a camera is watching — much as the moon holds to its orbit when no one is looking.
This requirement is a blind spot of existing benchmarks, which reward surface properties such as fidelity, motion, and camera controllability while never asking whether a generated world keeps evolving once it is unobserved.
The authors introduce WRBench, the first systematic diagnostic benchmark that treats camera motion as an intervention on observability. Evaluation follows a human-calibrated chain:
1. Did the camera perform the requested interaction? 2. Does the scene remain continuously identifiable while in view? 3. Does the revisited target match the state of the events initiated before leaving view?
Findings
- The benchmark examined 9,600 videos from 23 models across four control paradigms.
- A persistent pattern emerged: current systems treat the observed world like a tracking shot. When revisiting a target, they merely restore its abandoned state instead of advancing events during the unseen interval.
- This flaw spans control paradigms, model families, and scales.
- Consequently, robust world-state evolution is not achieved through cleaner images, stricter control, richer geometric priors, or more parameters.
Conclusion
The authors argue that a stable physical state core and consistency of world lines under viewpoint interventions should be top priorities in world model design: what a world model should capture is how the world will unfold, not how the next frame will appear.
--- *Auto-collected on 2026-06-20*