Researchers have proposed JEPA-Anything, a domain-agnostic framework for world modeling that aims to extend predictive learning beyond any single field. The work, shared on Hugging Face, builds on joint-embedding predictive architectures (JEPA) and introduces a principle the authors call orthogonal predictive factorization (OPF).
Under OPF, latent targets are decomposed into complementary factors that are learned through dedicated pathways and then recombined within a shared predictive design. The authors argue this allows a common learning principle to support world modeling across radically different systems, rather than treating each domain separately.
The framework was evaluated across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. Experiments covered representation learning, intervention prediction, out-of-distribution generalization, and long-horizon dynamics, including 10 matched dynamics tasks, forecasting of more than 1,000 clinical events, and 100-step molecular rollouts across four systems.
Against matched JEPA baselines, JEPA-Anything improved reported metrics on all 10 dynamics tasks and reduced single-intervention prediction error on Interventional Pong by 34.8%. It also achieved the lowest one-step and 100-step molecular errors among compared methods in all four systems.
Beyond prediction, the authors report that a factor-nominated biological intervention received experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and mice. Separately, latent orbital modes recovered the Keplerian scaling exponent with a fitted slope of -1.4991. The team frames these results as evidence for a common factorized predictive principle that links world modeling with intervention and experimentally grounded scientific discovery.
Code is available at github.com/Gen-Verse/JEPA-Anything.