Overview
A recent paper, "Current World Models Lack a Persistent State Core" (arXiv:2606.20545), by Jinpeng Lu, Dexu Zhu, Haoyuan Shi, Linghan Cai, Guo Tang, Yinda Chen, Jie Cao, Duyu Tang, Yi Zhang, Yong Dai, and Xiaozhu Ju, delivers a striking finding: today's world models can predict the next frame, but they do not understand that the world keeps existing—and keeps evolving—when no one is watching.
The article frames this with two analogies:
- Berkeley's 1710 question: "If a tree falls in a forest and no one hears it, does it make a sound?" — rephrased for AI: if a model sees a ball rolling toward a table edge, then looks away, does it know the ball will keep rolling and fall?
- Piaget's object permanence: human infants develop this around 8–12 months; without it, the world is just a sequence of unrelated images.
- 9,600 videos, 23 models
- Four control paradigms: text-to-video, image-to-video, video-to-video, interactive control
- Everyday physical events: rolling balls, pouring water, collisions, doors opening
- Across all control paradigms (text, image, video)
- Across model families (diffusion, autoregressive, others)
- It does not disappear with larger models, more data, or better geometric priors
- Time does not pause for the camera: real-world state evolution is continuous, autonomous, and observer-independent. Current models exhibit *observation-driven* state evolution.
- Causal chains break: models can map A→B and B→C when visible, but cannot infer C from A across an unobserved B.
- No uncertainty management: belief about hidden state should spread like a probability distribution over time; current models either deterministically restore the last state or generate plausible-but-inconsistent ones.
- Interventions vs. conditioning: WRBench reframes evaluation causally—a real world model should answer counterfactual questions ("where will the ball be if I look away for 10 seconds?"), not merely conditional rendering questions.
- Generation vs. simulation: current models map conditions to frames; a true world model should behave like a physics engine—explicit state variables, an observation-independent evolution function, and a rendering function where the camera merely selects what to observe.
- 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.
The Diagnosis: WRBench
The paper introduces WRBench (World-state Robustness Benchmark), the first systematic benchmark that treats camera motion as an intervention on observability. It evaluates three levels:
1. Controllability: does the model execute the requested camera interaction? 2. Consistency in view: does the scene remain coherent while the camera moves? 3. Persistent state (core test): when the camera leaves and returns, does the object's state match how it should have evolved while unseen?
Experimental setup
Key finding
> When the camera moves away and returns, current systems restore the object to its state at the moment of abandonment, not the state it should have evolved to.
Example: a model shows a ball at the table edge; the camera moves away; when it comes back, the ball is still perched at the edge—as if time froze. In reality, it fell long ago.
This failure is systematic:
Why World Models "Forget"
Architecture: the typical pipeline—history frames → encoder → latent representation → decoder → next frame—has no explicit world-state buffer. Nothing stores "ball position: table edge; velocity: 0.5 m/s." The model guesses frame from frame instead of predicting via physical state.
Training objectives: losses like MSE, LPIPS, FID, and CLIP score reward outputs that *look* right, never requiring outputs that are *physically consistent* over unobserved time.
"Schrödinger's ball": training data rarely includes camera-leave-and-return sequences, and no objective penalizes physical inconsistency across occlusion—so the model learns the "safest" strategy: restore the last-seen state.
Deeper Implications
Path Forward
The paper's central recommendation:
> "Physical state stability and worldline consistency under viewpoint interventions should become first-order objectives in world model design."
Concretely:
1. Explicit state representations (position, velocity, material, state) rather than purely implicit weights 2. A dedicated state-evolution module independent of observation 3. Rendering decoupled from simulation, trainable and evaluable separately 4. New training data: occlusion-return sequences, multi-view + temporal footage, and physical-interaction data 5. Possibly staged/curriculum training mirroring how infants acquire object permanence
Conclusion
A persistent state core is more than an architecture component—it represents a worldview: the world is a sequence of *state evolutions*, observation is *selection* rather than creation, and time is the flow of causation. When AI grasps this, it stops being a frame generator and becomes a true world simulator—one that can run the universe in its mind even with its eyes closed. That, the author argues, would be a real step toward AGI.