Current World Models Lack a Persistent State Core — When AI Sees, But Forgets the World Still Runs Unobserved
> Paper: *Current World Models Lack a Persistent State Core* > Authors: Jinpeng Lu, Dexu Zhu, Haoyuan Shi, Linghan Cai, Guo Tang, Yinda Chen, Jie Cao, Duyu Tang, Yi Zhang, Yong Dai, Xiaozhu Ju > arXiv: 2606.20545
---
The Problem: An Old Philosophical Puzzle, Reborn in AI
In 1710, George Berkeley asked: *if a tree falls in a forest and no one hears it, does it make a sound?* Three centuries later, the question returns in a new form:
> If an AI "sees" a ball rolling toward a table edge, then looks away — does it "know" the ball keeps rolling and falls?
The unsettling answer for current world models is: no, it doesn't. They can predict the next frame, but they don't grasp that the world continues to evolve — objects keep moving, causes keep producing effects — even when the camera looks away.
Key points
- WRBench (World-state Robustness Benchmark) is the first systematic benchmark for testing world models' "persistent state core," treating camera motion as an intervention on observability.
- It evaluates 23 models across 9,600 videos, covering four control paradigms: text-to-video, image-to-video, video-to-video, and interactive control, on everyday physical events (rolling balls, pouring water, collisions, doors opening).
- It tests three levels: (1) does the camera perform the requested interaction; (2) is the scene consistent while in view; (3) — the core test — does the target state match the previously set event when the camera returns?
- Cross-paradigm: present in text-, image-, and video-controlled generation.
- Cross-family: affects diffusion, autoregressive, and other architectures.
- Scale-resistant: larger models, more data, and better geometric priors do not eliminate it.
- Generative model vs. physics engine: a true world model should have explicit state variables, an observation-independent state-evolution function, and a rendering function — the camera merely selects which subset is observed, without affecting state evolution (an idea extending NeRF / 3DGS with temporal dynamics).
- Intervention, not conditioning: WRBench draws on causal inference — models that truly understand the world should answer counterfactuals ("where is the ball 10 seconds after the camera leaves?"), not just conditionals.
- Human-calibrated evaluation: human judgment calibrates which state inconsistencies matter, aligning assessment with human physical intuition rather than pixel distance.
- Lu, J., et al. (2026). *Current World Models Lack a Persistent State Core*. arXiv:2606.20545.
- Ha, D., & Schmidhuber, J. (2018). *World Models*. NeurIPS.
- Pilato, G., et al. (2024). *Sora: A Review on Background, Technology, Limitations, and Opportunities of Large Vision Models*. arXiv:2402.17177.
- OpenAI. (2024). *Video Generation Models as World Simulators*. OpenAI Technical Report.
Core finding: the world freezes when unobserved
The paper identifies a "stubborn" failure mode:
> When the camera moves away and returns, current systems restore targets to their state at the moment of abandonment, rather than the state they should have evolved to during the unobserved interval.
A ball left rolling at a table edge reappears still at the edge — as if time stood still — instead of having fallen. This failure is:
Why it happens
1. No explicit world-state buffer. Current pipelines (history frames → encoder → latent → decoder → next frame) lack anything like "ball position: table edge, velocity: 0.5 m/s." Models guess frames from frames instead of predicting states from physics.
2. Training rewards "looks right," not "is consistent." Losses (MSE, LPIPS, FID, CLIP score) optimize visual plausibility, never physical consistency across unobserved intervals.
3. "Schrödinger's ball." Training data rarely includes leave-and-return camera sequences, so the safest learned strategy is to restore the last-seen state.
Deeper implications
Paths forward
The authors' central recommendation:
> "Physical state stability and worldline consistency under viewpoint interventions should be a first-class objective in world model design."
Concretely: explicit state variables, a dedicated observation-independent state-evolution module, new training data (leave-and-return sequences, multi-view + temporal captures, physical-interaction data), and possibly curriculum-style multi-stage training — echoing how infants acquire object permanence (Piaget).
Conclusion
A persistent state core is more than an architecture — it is a worldview: the world is not a sequence of frames but an evolution of states; observation is selection, not creation; time is causal flow, not frame rate. When AI internalizes this, it stops being a frame generator and becomes a genuine world simulator — one that can run the universe in its mind even with its eyes closed. That would be a real step toward AGI.