Yann LeCun of New York University and researchers from Brown University have proposed a method that enables robots to predict how objects will behave in response to actions never observed during training. The approach measures the discrepancy between a model’s predictions and actual events, then subtracts that gap to expose hidden causal factors. Simulation tests achieved success rates of 35% to 69% on the first attempt, compared with 14% to 23% for baseline methods, and rose to 51% to 97% after multiple attempts. Published as a pre-print on November 9, 2026, the method can be integrated with minimal changes into JEPA architectures and readily available encoders such as DINOv2 and MAE, but its results have not yet been validated in real-world conditions or through peer review.
Source: Read the original article
