Researchers stretch LeCun's JEPA AI into a universal world model that works from physics to biology

Researchers led by PhAI Labs, with collaborators from Stanford, Oxford and Princeton, have expanded the JEPA architecture pioneered by Yann LeCun into a model called JEPA-Anything that works across seven fields, including physics, robotics and medicine. Their paper introduces the method, which is built on Joint-Embedding Predictive Architectures. Rather than reconstructing raw data such as pixels, these models predict an abstract summary of a missing or future state.

The authors identify a weakness in the standard approach: predictions are funneled into a single output, so easy patterns can drown out harder ones. JEPA-Anything instead splits the predicted state into four orthogonal factors, each handled by its own prediction module, with an added constraint pushing the modules to capture different aspects. The model reassembles the partial predictions into a complete world state. The researchers do not assign meanings to the parts, letting roles emerge during training, and change only how data is prepared for each field.

Compared against a standard JEPA with the same architecture, trained on the same data under identical conditions, dynamic systems showed the clearest gains. In a simplified Pong environment, prediction error dropped by 35 percent with targeted interventions and by 13 percent for intervention combinations not seen during training. In a separate evaluation, error on the Burgers equation fell by nearly half, though over 50 prediction steps the advantage shrank to about three percent. The model assigned cell types more reliably in single-cell data and predicted more than 1,000 possible disease events slightly better on clinical data. In locomotion planning for simulated walking robots, it won in two of three environments, with the standard model ahead in the third. Differences on image tasks were small.

The top liver cancer candidate paired IL-18 with blockade of the enzyme CD73. Tested on liver cancer cells co-cultured with immune cells, organoids and tumor tissue from three patients each, and in mice, the combination killed more tumor cells than either component alone. The study does not establish whether this could become an actual therapy. In another case, the model trained on simulated orbits landed on a value of minus 1.4991 for Kepler's third law's minus 1.5 exponent; the team evaluated one training run.

The authors caution that clean separation of learned parts does not mean they capture real cause-and-effect relationships, and it remains an open question when such a model becomes reliable enough to guide experiment design. Code and models are publicly available.