Two predictor calls fix world-model counterfactuals: R² jumps from 38–46% to 86–96%
TheTuringPost · x · 2026-11-21
A world-model research thread: standard JEPA-style models trained on factual data only return the average outcome when asked 'what if I'd acted differently', discarding information about hidden disturbances. The fix exploits that action–disturbance interaction is second-order in time: estimate counterfactual outcome as observed + prediction(other action) − prediction(taken action).
Key results:
- Counterfactual R² rises from 38–46% to 86–96% on four simulators, same world models
- Underwater vehicle (BlueROV2): 90–94%, with 16x more factual data
- Works across 4 frozen encoders (DINOv3, SigLIP 2, CLIP, EVA-02): 81–95%
- Hindsight planning finds the action that would have reached the goal on 35–69% of events vs 14–23% for an oracle planner that knows the simulator but not the outcome
The authors prove an O(h²) error bound and that no method using only unpaired factual data can achieve a better rate. They also show where it breaks: contacts, folded coordinates, and faults that scale the command.
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