JEPA-Anything tests one predictive framework on liver cancer models and planetary orbits
PhAI Labs and researchers from Oxford, Stanford, Princeton, and the Chinese University of Hong Kong have introduced JEPA-Anything, a cross-domain framework built to learn predictive models across very different kinds of systems. The paper asks whether AI can share a deeper learning principle when modeling how things change, even when the underlying mechanisms range from cells and patients to weather systems and planetary motion.
The architecture keeps each domain’s own observation format, encoder, and context-target setup, but shares a common predictive core and a unified latent world-state interface. Its key mechanism, Orthogonal Predictive Factorization, splits target states into complementary subspaces handled by separate predictive branches, with orthogonality and activity constraints meant to reduce redundancy and collapse.
The framework was evaluated across seven classes of systems: vision, biology, clinical data, control, molecules, physical fields, and weather. In a controlled intervention benchmark, it reduced prediction error by about 11.7% on in-distribution single interventions and cut MSE by about 3.5% on unseen multi-factor combinations. In a matched benchmark of 10 tasks, it improved results in 9, including MSE reductions of 39.7% on the Burgers equation, 39.3% on shallow water equations, and 10.5% on WeatherBench 2.
The paper also highlights two analysis cases. In liver cancer experiments, internal factors led researchers to a candidate intervention, IL-18 plus NT5E/CD73 blockade, later tested in co-culture, organoids, tumor fragments, and immunocompetent mice. In simulated orbital data, latent modes recovered a frequency-semi-major-axis scaling slope of -1.4991, very close to the -1.5 value implied by Kepler’s third law, with R²=0.99999999.
