LeCun and AMI Labs propose Counterfactual World Models: JEPA reasoning about 'what if' without retraining

udmrzn · x · 2026-11-20

A new paper by Yann LeCun (NYU/AMI Labs), Randall Balestriero (Brown) and Kerda Varaku introduces Counterfactual World Models, addressing the core blind spot of standard world models: they predict average futures but cannot answer 'what would have happened had I acted differently.' In a boat-drifting simulation with hidden currents, a standard JEPA only averages over possible futures, while their model predicts each alternative as the observed outcome plus the model's predicted difference between headings—no retraining, extra parameters, or new data needed. Code is on GitHub with a Hugging Face demo.

Related event: LeCun and Colleagues Release Counterfactual World Models Paper(12 posts)→

Original post →

More from Models

Models channel →