PhAI Labs, Oxford, Stanford, Princeton and CUHK Unveil Cross-Domain AI Prediction Framework JEPA-Anything
PhAI Labs, together with research teams from Oxford, Stanford, Princeton and Chinese University of Hong Kong, has released JEPA-Anything, a cross-domain framework for learning predictive models across different "worlds". It was validated on seven systems spanning vision, biology, clinical, control, molecular, physics fields and weather.
The core mechanism, Orthogonal Predictive Factorization (OPF), decomposes a target state into multiple complementary subspaces that are processed independently by separate prediction branches before being recombined. OPF enforces orthogonality to reduce redundancy, while factor activity and encoder variance constraints help prevent factor inactivity and representation collapse. The learned factors are not pre-assigned meanings; their roles emerge from the predictable structure in data.
In experiments on CITRIS Interventional Pong, JEPA-Anything reduced prediction error by about 11.7% for in-distribution single interventions and lowered MSE for unseen multi-factor combinations by about 3.5% compared to standard JEPA. On a matched dynamics benchmark, it improved performance on nine of ten tasks, including a 39.7% MSE reduction on the Burgers equation, 39.3% on shallow-water equations, and 10.5% on WeatherBench 2. It also achieved the lowest errors in long-horizon molecular dynamics for four systems.
The paper also highlights two scientific exploration cases. In liver cancer research, internal factors from the model led to a candidate intervention combining IL-18 with CD73 blockade, validated in organoids and mice. In a planetary orbit task, latent modes extracted from the model produced a fitted slope of -1.4991 for the frequency-semi-major axis relationship, close to Kepler's third law theoretical value of -1.5. The authors note the results support cross-domain applicability of a shared prediction principle, but do not yet constitute a unified world model for all domains.
