Directional Conformal Uncertainty Quantification from Learned Model Discrepancy

What the math gives to ML

The paper provides a transferable split-conformal mechanism for turning a nominal dynamical predictor plus a learned state-dependent discrepancy into asymmetric, finite-sample-valid uncertainty sets. The key asset is directional nonconformity: residuals aligned with the learned discrepancy are penalized less than equally large residuals in the opposite direction, while calibration remains distribution-free under exchangeability. For neural state-space models and world models, freeze a discrepancy network on a training split, calibrate its directional residual score on a held-out split, and use the resulting sets for uncertainty-aware rollout, robust MPC, or confidence-weighted training.