Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models
Researchers have introduced Sub-JEPA, a novel method designed to enhance the stability and performance of Joint-Embedding Predictive Architectures (JEPAs) in learning world models. While JEPAs effectively predict future latent representations, they often suffer from a bias-variance tradeoff that can lead to model collapse without sufficient structural constraints. Previous solutions like LeWorldModel (LeWM) applied isotropic Gaussian priors but introduced excessive bias by ignoring the low-dimensional manifold structure of latent representations. Sub-JEPA addresses this by applying Gaussian constraints within multiple random subspaces rather than the entire ambient space. This approach relaxes global constraints while maintaining anti-collapse effects, achieving a superior balance between training stability and representation flexibility. Extensive experiments across four continuous-control environments demonstrate that Sub-JEPA consistently outperforms LeWM with significant margins. The method is noted for its simplicity and effectiveness, establishing a strong baseline for future research in JEPA-based world models. The associated code has been made publicly available to facilitate further development and verification within the machine learning community.
Wire timeline
Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models
Researchers have introduced Sub-JEPA, a novel method designed to enhance the stability and performance of Joint-Embedding Predictive Architectures (JEPAs) in learning world models. While JEPAs effectively predict future latent representations, they often suffer from a bias-variance tradeoff that can lead to model collapse without sufficient structural constraints. Previous solutions like LeWorldModel (LeWM) applied isotropic Gaussian priors but introduced excessive bias by ignoring the low-dimensional manifold structure of latent representations. Sub-JEPA addresses this by applying Gaussian constraints within multiple random subspaces rather than the entire ambient space. This approach relaxes global constraints while maintaining anti-collapse effects, achieving a superior balance between training stability and representation flexibility. Extensive experiments across four continuous-control environments demonstrate that Sub-JEPA consistently outperforms LeWM with significant margins. The method is noted for its simplicity and effectiveness, establishing a strong baseline for future research in JEPA-based world models. The associated code has been made publicly available to facilitate further development and verification within the machine learning community.
cs.AI updates on arXiv.org