Counterfactuals with two predictor calls and one addition, paper proves O(h²) optimal

TheTuringPost · x · 2026-11-21

TheTuringPost breaks down a new paper on counterfactual reasoning in world models. Standard JEPA-style world models, when asked "what if I had acted differently?", can only return the average outcome, discarding the disturbance information contained in the actually observed result.

The fix is surprisingly small: treating action and disturbance as separate forces makes their interaction second-order in time, so what-if outcome ≈ observed outcome + prediction(alternative action) − prediction(action taken). Two calls to the same predictor and one addition — no counterfactual data, no extra parameters, no retraining. The authors prove an O(h²) error bound and show no method using only unpaired factual data can achieve a better rate.

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

Original post →

More from Models

Models channel →