LeCun and Balestriero show JEPAs can reason counterfactually from offline data alone
randall_balestr · x · 2026-11-20
A new arXiv paper by Randall Balestriero, Yann LeCun and Kerda Varaku tackles whether world models can answer counterfactual questions — what would have happened under a different action given the same hidden disturbance.
- The authors prove that world models trained only on factual transitions cannot answer such questions regardless of data or capacity.
- Their fix leans on natura non facit saltus: when actions and disturbances act as separate forces, the counterfactual outcome is, to first order, the observed outcome plus the difference between the model's predictions under the alternative and factual actions — a simple but necessary correction term.
- This yields a counterfactual world model with no counterfactual data, extra parameters, or retraining, with an O(h²) error bound and a matching minimax lower bound.
- Validated with JEPA world models on frozen DINOv2 and MAE encoders across four rendered simulators, plus hindsight planning experiments; the paper also catalogs where each assumption holds or fails (currents, gusts, contact, long horizons).
Related event: LeCun and Colleagues Release Counterfactual World Models Paper(12 posts)→
More from Research
- Riemann hypothesis 'proven by Turkish mathematician using AI' claim mocked as the 'Organic Turk' — CatAstro_Piyush · 2026-11-22
- Andriy Burkov releases The Hundred-Page Language Models Book, hands-on with PyTorch — burkov · 2026-11-22
- Generative AI meets cardiology: Dagstuhl seminar bridges ML and clinical practice — StephanMandt · 2026-11-22
- SubRoformer hits SOTA music source separation with DSP subband filters — affige_yang · 2026-11-22
- Emily Riehl releases Lean game teaching dependent type theory at intro-to-proofs level — abhi9u · 2026-11-22
- Curated reading list: evaluating mobile-use agents on Android, from AppAgent to AndroidReality — East-Muffin-6472 · 2026-11-22