CVPR 2026

Cloning Deterministic Worlds: The Critical Role of Latent Geometry in Long-Horizon World Models

Zaishuo Xia, Yukuan Lu, Xinyi Li, Yifan Xu, Yubei Chen
University of California, Davis ยท Open Path AI Foundation
Paper HTML GitHub
Figure 1
Figure 1. Representation quality is the primary bottleneck for world model fidelity.

Abstract

We study long-horizon prediction in deterministic environments and show that the main bottleneck is often not the dynamics model itself, but the geometry of the learned latent representation. Many world models can generate plausible futures, but here we focus on a stricter setting where a given initial condition and action sequence should lead to one correct future. In this setting, rollout fidelity becomes a direct test of whether the model has actually captured the environment. We find that even strong predictive models fail when operating on poorly structured latent spaces. In contrast, when the same predictive pipeline is given true underlying states, long-horizon prediction becomes much easier. This isolates representation geometry as a central issue.

Main Contributions

  1. We identify representation geometry as a primary bottleneck for long-horizon fidelity in deterministic world modeling.
  2. We propose GRWM, a simple plug-in regularization method that improves latent structure without changing the dynamics backbone.
  3. We show that better latent geometry leads to more faithful rollouts across multiple world-modeling backbones.

Results

Our results suggest that representation quality is a first-order issue for long-horizon prediction in deterministic worlds. By improving the geometry of the latent space, we obtain more faithful rollouts and reduce the gap to an oracle model with access to true states.

Figure 7
Figure 7. Visualization of latent space structure through clustering analysis. We perform k-means clustering (k = 20) on the latent representations of frames.
Figure 10
Figure 10. Visualization of generated frames at multiple time points from a single starting state. We show frames near steps 100, 200, 300, 400, sampled randomly.
Figure 11
Figure 11. Visualization of generated frames from the MC-DET sequence.

Citation

@article{xia2025cloning,
  title   = {Cloning Deterministic Worlds: The Critical Role of Latent Geometry in Long-Horizon World Models},
  author  = {Zaishuo Xia and Yukuan Lu and Xinyi Li and Yifan Xu and Yubei Chen},
  journal = {arXiv preprint arXiv:2510.26782},
  year    = {2025}
}