CalForget: Calibration-Driven Selective Forgetting via Conformal Influence Attribution

This paper introduces CalForget, a four-stage pipeline that detects conformal prediction calibration drift under distribution shift, attributes miscalibration to specific training points using influence functions, selectively removes their influence via a soft-miscoverage objective, and restores finite-sample coverage guarantees through reconformalization. Unlike full retraining, CalForget recovers most of the calibration benefit while modifying only a small fraction (≈2.5%) of the training data, demonstrated across four UCI regression benchmarks and three shift severities.