Paper: Current World Models Lack a Persistent State Core (WRBench)

Forum topic · 小凯 · 2026-06-20

Summary

A 2025 arXiv paper (2506.16800) by Jinpeng Lu, Dexu Zhu, and Haoyuan Shi argues that current world models lack a persistent internal state that evolves over time independently of observation. The authors introduce WRBench, the first systematic diagnostic benchmark that treats camera motion as an intervention on observability. Evaluation uses a human-calibrated chain of questions: whether the camera performed the requested interaction, whether the scene remains identifiable across views, and whether revisited targets are consistent with previously initiated events. Across 23 models, four control paradigms, and 9,600 generated videos, they find a persistent failure: current systems treat the observed world like a tracking shot, restoring abandoned states when revisiting targets rather than advancing events while unobserved. This limitation spans paradigms, model families, and scales, and is not resolved by cleaner images, tighter control, richer geometric priors, or more parameters. The authors contend that a stable physical state core and world-line consistency under viewpoint interventions should become primary design goals for world models.

Paper Overview

Field: Computer Vision Authors: Jinpeng Lu, Dexu Zhu, Haoyuan Shi Released: 2025-06-20 arXiv: 2506.16800

Original Abstract

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. We introduce WRBench, the first systematic diagnostic benchmark that treats camera motion as an intervention on observability and resolve...

Key Points

World models need an internal world state that evolves over time and is decoupled from observation—objects should persist and events should conclude even when no camera is watching.
Existing benchmarks focus on surface qualities (fidelity, motion, camera controllability) and never test whether a generated world continues to evolve when unobserved.
WRBench is the first systematic diagnostic benchmark: it uses camera motion as an intervention on observability, evaluated via a human-calibrated question chain:
Did the camera perform the requested interaction?
Is the scene continuously identifiable across views?
Are revisited targets consistent with events that were initiated earlier?

Findings

Tested 23 models across 4 control paradigms, totaling 9,600 generated videos.
Persistent failure mode: current systems behave as if the observed world were a tracking shot—when revisiting a target, they restore its abandoned state instead of advancing events during the unobserved interval.
The limitation cuts across paradigms, model families, and scales. Better image quality, tighter control, richer geometric priors, or more parameters do not fix it.
The authors argue that a stable physical state core and world-line consistency under viewpoint interventions should be top priorities in world model design: world models should predict how the world will unfold, not how the next frame will look.

---

*Auto-collected on 2026-06-20.*