LeCun's Counterfactual World Models paper boosts JEPA prediction R² from 38% to 96%
ylecun · x · 2026-11-20
Yann LeCun, Randall Balestriero, and Kerda Varaku (Brown/NYU/AMI Labs) released "Counterfactual World Models: JEPAs That Reason About 'What If'", enabling world models to answer what would have happened under a different action given the same hidden disturbance.
- They prove world models trained only on factual transitions cannot answer counterfactual questions on their own.
- Exploiting "natura non facit saltus," the counterfactual outcome is approximated to first order as the observed outcome plus the difference between the world model's predictions under alternative and factual actions — no counterfactual data, extra parameters, or retraining.
- They prove an O(h²) counterfactual error bound and a matching minimax lower bound, plus failure cases at discontinuities.
- With frozen DINOv2/MAE encoders on four rendered simulators, counterfactual R² improves from 38–46% to 86–96%; an underwater vehicle env reaches 90–94% with 16× more factual data.
Related event: LeCun and Colleagues Release Counterfactual World Models Paper(12 posts)→
More from Embodied
- Tesla FSD 15's real-time view isn't 3D rendering — it's the AI model's own synthetic perception — mark_k · 2026-11-22
- China Installed 59% of New Industrial Robots in 2026 Report, ~354,000 Units — ingliguori · 2026-11-22
- Apple Watch bricks after one shower while a $15 Keep band survives many — oran_ge · 2026-11-22
- Tesla FSD V15 visualizations go viral with detailed 3D night-time rendering — DimaZeniuk · 2026-11-22
- Tesla FSD visualization: from dots on a screen to photorealistic 3D night views in six years — XFreeze · 2026-11-22
- YQ Bionics' Y-Hand M2 packs 38 degrees of freedom as AI meets robotic manipulation — CurieuxExplorer · 2026-11-22