Conformal Non-Coverage Risk Control (CNCRC): Risk-Centric Guarantees for Predictive Safety in High-Stakes Settings

ICLR 2026 Conference Submission10730 Authors

Keywords: Conformal Prediction, Cost-Sensitive Learning, Risk Control, Uncertainty Quantification

TL;DR: This paper introduces Conformal Non-Coverage Risk Control (CNCRC), a new framework that adapts Conformal Prediction for high-stakes applications by directly guaranteeing a cap on the risk from costly errors.

Abstract: Standard Conformal Prediction (CP) guarantees that prediction sets contain the true label with high probability, but it is *cost-blind*, treating all errors as equally important---a critical limitation in high-stakes domains. We introduce **Conformal Non-Coverage Risk Control (CNCRC)**, a framework that replaces coverage frequency with direct risk control. CNCRC guarantees an upper bound on catastrophic **non-coverage risk** while actively reducing **ambiguity risk**, providing prediction sets that are both safe and usable. This is achieved through a principled decomposition of decision risk and the design of risk-weighted nonconformity scores that balance robustness with efficiency. Experiments show that CNCRC reliably satisfies strict risk constraints in adversarial settings and outperforms all baselines on a large-scale clinical benchmark. By offering practitioners a choice between maximum robustness and maximum efficiency, CNCRC provides a practical and theoretically grounded toolkit for deploying genuinely risk-aware machine learning systems in safety-critical applications.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 10730