JEPA-Anything extends LeCun's architecture into one world model for seven fields, from physics to medicine

A team led by PhAI Labs, working with collaborators from Stanford, Oxford and Princeton, has introduced JEPA-Anything, The Decoder reports. It is a world model built on the Joint-Embedding Predictive Architecture that Yann LeCun pioneered. World models predict how a system will change over time, whether that system is a robot, a molecule or a patient's health. Until now, each domain has usually needed its own model. The researchers aim to show that a single shared principle can work across seven very different fields, including physics, robotics and medicine. JEPA models do not reconstruct raw data such as pixels. Instead, they predict an abstract summary of future states. The new method splits a future state into several partial predictions rather than funneling everything into one, and the authors say this lets it pick up patterns that standard models miss. The work also produced a liver cancer treatment candidate. In tests on lab samples and mice, the combination killed more tumor cells than either of its components did alone. The results come from the researchers' own experiments and are early-stage. If they hold up, they would support the idea that one general predictive architecture can replace many domain-specific models in scientific work.