Event

A New Method for Conformal Calibration of World Models under Policy Interventions

1 reports · 1 sources

Ayush Mittal and Dhruv Gupta propose a weighted conformal calibration method to address the failure of fixed world model uncertainty after deployment policy changes: it transfers marginal coverage to the target policy distribution and bounds the weight error using the log-loss regret of a domain classifier. In experiments based on a published DINO-feature world model and real Franka trajectories, under the strongest controlled state-action shift, the coverage of an ordinary 90% conformal set was 55.0%, while oracle calibration and learned ratio calibration reached 91.0% and 90.9%, respectively; the total variation distance between the learned alternative distribution and the target law was within 0.9%. The authors state that they have machine-verified in Lean 4 the finite-support algebra and causal identities used in the classifier regret argument, including the coverage combination step.