A new AI architecture called JEPA-Anything, built on Joint-Embedding Predictive Architectures, demonstrates performance across seven fields, including physics, robotics, and cancer research.
Key facts
- •JEPA-Anything uses multiple prediction modules to capture different aspects of a system's state.
- •The model reduced prediction error by 35 percent in a simplified Pong environment.
- •In liver cancer tests, a model-proposed combination of IL-18 and CD73 blockade increased tumor cell death in organoids and tissue samples.
- •The model learned patterns nearly identical to Kepler's third law when trained on simulated orbital data.
- •The research team has made the model code and weights publicly available.
Researchers led by PhAI Labs, with collaborators from Stanford, Oxford, and Princeton, have introduced JEPA-Anything, an expansion of Yann LeCun's JEPA architecture. The model is designed to function as a universal world model capable of predicting system evolution across diverse domains, including physics, robotics, and biology. The team also utilized the model to identify a potential liver cancer treatment candidate that was subsequently tested in laboratory settings.
By the numbers
Architecture and Performance
JEPA-Anything improves upon standard JEPA models by splitting predicted states into four orthogonal factors, each managed by a dedicated prediction module. This approach prevents easy patterns from obscuring more complex ones. In testing, the model outperformed standard JEPA across ten tasks, including fluid dynamics and simulations of water and chemical compounds. In a simplified Pong environment, the model reduced prediction errors by 35 percent during targeted interventions.
Biological and Scientific Applications
The researchers applied the model to biological data, including gene activity and CRISPR screens, to identify a liver cancer treatment candidate. The model proposed combining IL-18, a signaling molecule, with the blockade of the enzyme CD73. Laboratory tests on organoids and tumor tissue from three patients showed this combination killed more tumor cells and increased immune cell activation compared to individual treatments. Additionally, the model independently learned patterns matching Kepler's third law when trained on simulated orbits.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder.
