What it is
A domain-agnostic JEPA variant that factorizes latent targets into orthogonal predictive components, then recombines them for representation learning and world modeling across seven domains. Instead of one monolithic JEPA target and predictor, it splits the latent target into orthogonal factors with dedicated predictors, then synthesizes a full state. This allocates predictive capacity across complementary modes and supports reuse for dynamics, interventions, and analysis.
Key results
- Improves reported metrics over matched JEPA on all 10 dynamics tasks
- Cuts single-intervention prediction error on Interventional Pong by 34.8%
- Lowest 1-step and 100-step molecular errors on all 4 compared systems
Shows that explicitly factorizing predictive latent state can transfer as a common world-modeling principle across very different domains.