Researchers have expanded Yann LeCun's JEPA architecture into a universal world model called JEPA-Anything, which works across seven different fields including physics, robotics, and medicine. The model breaks future states into multiple partial predictions, allowing it to identify patterns standard models miss.
The JEPA-Anything model, developed by PhAI Labs with collaborators from Stanford, Oxford, and Princeton, predicts an abstract summary of a missing or future state rather than reconstructing raw data like pixels. This approach allows the model to focus on relevant details while filtering out noise.
In dynamic systems tests, JEPA-Anything showed significant improvements, with prediction error dropping by 35 percent in a simplified Pong environment. The model also outperformed standard JEPA on ten test tasks, including fluid dynamics and weather forecasting, with error reductions of nearly 50 percent in some cases.
"The model's ability to split the predicted state into four orthogonal factors and reassemble them into a complete picture is a key innovation," said Jonathan Kemper, a researcher at PhAI Labs. This method allows the model to capture different aspects of a system without learning the same thing repeatedly.
The team also identified a liver cancer treatment candidate that killed more tumor cells in lab samples and mice than either component alone. While the study does not confirm clinical viability, the model's ability to propose drug combinations highlights its potential in biomedical research.
The researchers caution that while the model shows promise, it remains an open question when such a model becomes reliable enough to guide experiment design. The team's long-term goal is to use JEPA-Anything to propose and rank experiments, then feed results back into the model.
Source: thedecoder
