<ul><li>Selective classification enhances predictive model reliability by allowing abstaining from uncertain predictions.</li><li>The study focuses on optimal selection functions using the Neyman--Pearson lemma, which characterizes the optimal rejection rule as a likelihood ratio test.</li><li>New approaches to selective classification are proposed based on the Neyman--Pearson lemma, unifying post-hoc selection baselines' behaviors.</li><li>The study evaluates the proposed methods in covariate shift scenarios across various vision and language tasks, showing consistent outperformance of existing baselines.</li></ul>

Know When to Abstain: Optimal Selective Classification with Likelihood Ratios

Discover more