JEPA-Anything: Learning Predictive Models across Different Worlds

JEPA-Anything. Learning Predictive Models across Different Worlds. By Taoyong Cui and twelve others. Published on the seventeenth of September. Its code repository is public and real: fifteen source files.

World modeling lets an intelligent system anticipate consequences, guide interventions, and learn from interaction. Yet predictive models are still built domain by domain. The question posed is whether one common learning principle can support world modeling across radically different systems. A weather model is built for weather and a molecular model for molecules. Each is specific to its own world.

JEPA-Anything is a domain-agnostic framework built on orthogonal predictive factorization, or OPF. It extends joint-embedding predictive architectures. OPF breaks latent targets into complementary factors, learns each through its own dedicated pathway, and recombines them within a shared predictive design. Put simply, what is to be predicted is split into complementary parts, each part gets its own route through the model, and the prediction is reassembled at the end. The same recipe is applied in every domain.

The framework is evaluated across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields and weather. The experiments cover representation learning, intervention prediction, out-of-distribution generalization, and long-horizon dynamics. They include ten matched dynamics tasks, forecasting of more than one thousand clinical events, and hundred-step molecular rollouts across four systems. The breadth is deliberate. A principle claimed to be common has to be tested on systems that share nothing else.

Against matched JEPA baselines, JEPA-Anything improves the reported metrics on all ten dynamics tasks, and reduces single-intervention prediction error on Interventional Pong by thirty-four point eight percent. It achieves the lowest one-step and hundred-step molecular errors among the methods compared, in all four systems.

Beyond prediction, a biological intervention nominated by one of the learned factors receives experimental support in cell co-cultures, patient-derived organoids, tumor fragments and mice. And latent orbital modes recover the Keplerian scaling exponent, with a fitted slope of minus one point four nine nine one. Both are claims about what the learned factors mean. One pointed to an intervention that was then tested experimentally, from cell cultures to mice; the other recovered the scaling of orbital motion from latent structure.

The authors read these results as support for a common, factorized predictive principle across very different worlds. The range is the point: the same design is asked to forecast weather, roll out molecules and nominate a treatment, and the claim connects world modeling to experimentally grounded discovery.