Correcting a learned physical invariant improves world-model rollouts
On the map: World models → Latent predictive world models
Abstract (arXiv)
World models can predict video without learning dynamics that they reliably preserve. We test whether a frozen DreamerV3 trained only on pendulum video learns a scalar that its own latent transition treats as approximately conserved. A label-free search recovers the same energy-like invariant across independently trained conservative models, while the same procedure finds no comparable invariant in matched damped models. During autonomous rollouts, this quantity drifts. Projecting the latent state back toward its initial level set reduces rollout error in all three conservative models, whereas matched random constraints usually increase it. These results distinguish a dynamically meaningful invariant from a merely decodable correlate and reveal a concrete failure mode: a world model can learn a physical constraint from pixels yet violate that constraint when it imagines forward.
Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Latent predictive world models, one of 11 directions in World models. Builds on 10 of its 15 references that are in the map. First posted on arXiv 24 Aug 2026.
Builds on
10 of 15 references.
References · 10
- 2016Understanding intermediate layers using linear classifier probescited by 1,353
- 2019Mastering Atari, Go, Chess and Shogi by Planning with a Learned Modelcited by 1,353
- 2023Mastering Diverse Domains through World Modelscited by 1,253
- 2018Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representationscited by 941
- 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
- 2018Recurrent World Models Facilitate Policy Evolutioncited by 785
- 2019Hamiltonian Neural Networkscited by 712
- 2022Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Taskcited by 431
- 2020Lagrangian Neural Networkscited by 398
- 2023Emergent Linear Representations in World Models of Self-Supervised Sequence Modelscited by 319
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026LaWM: Least Action World Models for Long-Horizon Physical Consistency from Visual Observations
- 2026Flow Equivariant World Models: Memory for Partially Observed Dynamic Environments
- 2025Zero-shot World Models via Search in Memory
- 2026Unlocking Temporal Generalization in Hamiltonian Video Dynamics Models
- 2026World Models Need a Physical Registry: Learning Internal Predictive States for Physically Plausible World Models
- 2026Off-Manifold Refinement: Guiding Video Generators with a Frozen World Model
- 2026Frozen Flows Forget: Diagnosing and Restoring Lost Motion in a Latent-flow World Model
- 2026DreamSAC: Learning Hamiltonian World Models via Symmetry Exploration