Researchers stretch LeCun's JEPA AI into a universal world model that works from physics to biology
decoder ·
Researchers have extended Yann LeCun’s JEPA architecture into a universal model called JEPA-Anything that operates across diverse fields such as physics, biology, and robotics. Unlike standard versions that funnel predictions into a single output, this new approach breaks future states into multiple partial predictions handled by separate modules to capture distinct patterns. The team demonstrated significant improvements in dynamic systems, reducing prediction error by 35 percent in simplified environments and nearly halving it on fluid dynamics benchmarks. In biological applications, the model analyzed gene activity and protein levels to identify a promising liver cancer treatment candidate combining IL-18 with CD73 blockade. Laboratory tests on organoids and mice confirmed that this combination killed more tumor cells and activated immune responses more effectively than either component alone. Additionally, the model successfully rediscovered Kepler's third law of planetary motion from simulated orbital data without being given physical constants. While the results are promising, the authors caution that the separation of learned factors does not necessarily imply true causal relationships. The study highlights the potential for AI agents to propose and rank experiments, with code and models now publicly available for further research.
decoder