Subspace-Decomposed JEPAs: Disentangling Progression and Content in Latent World Models
Researchers have developed a new method called Subspace-Decomposed JEPAs (SD-JEPA) to improve latent world models. This approach disentangles task progression from content within the model's latent space, using separate subspaces for each. SD-JEPA demonstrates improved performance on control benchmarks and a specific task called Push-T compared to existing JEPA baselines. The model's progression coordinate also acts as a scene-aware compass, accurately tracking task progress and semantic events. AI
IMPACT Introduces a novel architecture for latent world models that improves control and task progression tracking.