A recent report describes JEPA-Anything, a method that aims to replace domain-specific world models with a single, general-purpose recipe. The approach modifies the standard JEPA architecture by decomposing its single latent target into four orthogonal factors, each paired with its own predictor. This design is intended to capture diverse dynamics without requiring per-domain tuning.
The method was evaluated across seven different fields, covering ten dynamics tasks. According to the report, JEPA-Anything outperformed matched JEPA baselines on every one of those tasks. In a specific benchmark, Interventional Pong, the approach cut intervention error by 34.8% compared to the baseline.
Because the report is the only source, these results have not been independently verified. Still, the consistent gains across multiple domains suggest that a unified world-model recipe may be feasible, potentially simplifying how such models are built and applied.