Sub-JEPA boosts world-model planning with orthogonal subspace priors
Researchers propose Sub-JEPA, a drop-in modification to LeCun's LeWorldModel that replaces the global Gaussian latent prior with multiple orthogonal subspace priors, improving performance across all four world-model benchmarks.
Sub-JEPA is a world-model architecture that addresses a structural flaw in LeWorldModel, the JEPA-based planner from Yann LeCun's group at NYU. World models learn compact latent representations for planning without pixel reconstruction — a longstanding goal in reinforcement learning and robotics. LeWorldModel achieved stable end-to-end JEPA training by enforcing an isotropic Gaussian prior over the full latent space, preventing representation collapse.
The problem: real environment dynamics live on low-dimensional manifolds, so a global high-dimensional Gaussian is geometrically mismatched to the task structure. LeWorldModel's worst results appear on low-intrinsic-dimension tasks like Two-Room, where the mismatch between prior and actual state geometry is most pronounced. Sub-JEPA keeps the same two-term objective but applies Gaussian regularization inside multiple frozen random orthogonal subspaces instead of globally. The change requires no new hyperparameters and preserves the anti-collapse benefit while relaxing the overly rigid global constraint.
Results are consistent across all four benchmarks. Sub-JEPA outperforms LeWorldModel in every case, with gains up to 10.7 percentage points on Two-Room — the task where the original model struggled most. The authors also report straighter latent trajectories and better physical-state decodability as emergent side effects, suggesting the subspace prior aligns more naturally with underlying dynamics. The preprint is available on arXiv (2605.09241) and code is open on GitHub.