LeCun coauthors prove standard world models fail at counterfactuals, fix needs no retraining

alex_verem · x · 2026-11-22

Yann LeCun, Randall Balestriero and Kerda Varaku (Brown/NYU, AMI Labs) prove that world models trained on factual transitions cannot answer counterfactual questions — what would have happened under a different action given the same hidden disturbance — regardless of data or capacity. Their one-line fix: counterfactual outcome ≈ observed outcome plus the model's predicted difference between the two actions, exploiting that action–disturbance interactions are second-order in time. No counterfactual data, extra parameters, or retraining needed. They prove an O(h²) error bound and a matching minimax lower bound, characterize failure cases (discontinuities, coordinate folds, command-scaling disturbances), and validate on JEPA world models over frozen DINOv2/MAE encoders across four rendered simulators (robot hand, arm, ball-in-cup, liquid carrying), lifting counterfactual R² from 38–46% to 86–96%.

Original post →

More from Embodied

Embodied channel →