Standard JEPAs provably can't answer counterfactuals, however much data they get

TheTuringPost · x · 2026-11-21

From TheTuringPost's thread, explaining the blind spot the new paper targets: picture a boat pushed by a hidden current—before acting, a world model can only predict the average outcome, and only the real result reveals what the current did. Ask a standard JEPA "what if I'd steered differently?" and it returns the average again. The paper proves a model trained only on factual transitions cannot answer this on its own, whatever its data or capacity—the gap counterfactual world models are built to close. The attached figure shows the model's estimated counterfactual trajectories against the true counterfactual in the boat experiment.

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

Original post →

More from Research

Research channel →