Researchers introduced JEPA-Anything, a framework for predictive models across different domains. It uses orthogonal predictive factorization to decompose latent targets into factors. The method learns these factors through dedicated pathways. It recombines them within a shared predictive design. Tests covered seven domains including vision, biology and weather. It improved metrics on all 10 matched dynamics tasks. It cut single-intervention prediction error on Interventional Pong by 34.8%. It achieved the lowest one-step and 100-step molecular errors in all four systems. A factor-nominated biological intervention received experimental support in cell co-cultures, organoids, tumor fragments and mice. Latent orbital modes recovered the Keplerian scaling exponent. Code is available on GitHub.
What it unlocks
Building predictive models that work across vision, biology, clinical, control, molecular, physical and weather domains.
- 34.8% lower single-intervention prediction error on Interventional Pong
- 1,000+ clinical events forecasted
- 100-step molecular rollouts across 4 systems
- fitted slope of -1.4991 for Keplerian scaling
Sources