JEPA-Anything: cross-domain world model validated in mice

The paper reports gains on all 10 matched dynamics tasks it was tested against, a 34.8% cut in single-intervention prediction error on Interventional Pong, and a computationally nominated biological intervention that held up in mice.

The framework, called JEPA-Anything, extends joint-embedding predictive architectures with what the authors call orthogonal predictive factorization. From the preprint on arXiv: "OPF decomposes latent targets into complementary factors, learns them through dedicated pathways, and recombines them within a shared predictive design."

The evaluation runs across seven domains: vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. It includes forecasting of over 1,000 clinical events and 100-step molecular rollouts across four systems. Latent orbital modes "recover the Keplerian scaling exponent with a fitted slope of -1.4991."

The wet-lab piece is the unusual part. The abstract says a "factor-nominated biological intervention receives experimental support in cell co-cultures, patient-derived organoids, tumor fragments, and mice."

The specific intervention, the disease context, and the size of the mouse cohort are not named in the abstract, and this is a preprint that has not been through peer review.