Researchers introduce JEPA-Anything, a domain-agnostic world-modeling framework built on orthogonal predictive factorization (OPF), which extends joint-embedding predictive architectures by decomposing latent targets into complementary factors learned through dedicated pathways. The framework is evaluated across seven domains — vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather — including over 1,000 clinical event forecasts and 100-step molecular dynamics rollouts. Against matched JEPA baselines, it improves results on all 10 tested dynamics tasks, cuts single-intervention prediction error on Interventional Pong by 34.8%, and achieves the lowest molecular prediction errors across four systems.
