Learning by Exclusion: An Evidential Contrapositive Framework for Zero-Shot OOD Detection

ICLR 2026 Conference Submission16542 Authors

Keywords: zero shot learning, out-of-distribution, contrapositive learning, image classification

TL;DR: We propose Evidential Contrapositive Framework, a zero-shot OOD detection approach that combines contrapositive reasoning with evidential uncertainty, enabling principled rejection of unseen samples without relying on labeled OOD data.

Abstract: Out-of-Distribution (OOD) detection is critical for deploying vision-language models in safety-sensitive settings. While recent approaches such as CLIPN rely solely on margin-based cosine similarity for separating in- and out-of-distribution samples, this reliance provides limited safeguards against overconfident representations. We introduce the Evidential Contrapositive Framework (ECF), a principled approach that integrates logical negation into vision-language alignment to explicitly model what a class is not. Unlike prior work that enforces separation through heuristic margins, ECF specifically introduces (i) exclusion via contrapositive loss, (ii) decoupling via negative prompt alignment, and (iii) logical consistency via contrapositive similarity regularization. To further enhance reliability, we couple this framework with evidential uncertainty modeling using a Dirichlet distribution, enabling simultaneous estimation of aleatoric and epistemic uncertainty. This combination yields interpretable uncertainty-aware decision boundaries and robust rejection of OOD inputs without requiring access to OOD samples during training. Extensive experiments on large-scale benchmarks demonstrate that ECF significantly outperforms state-of-the-art zero-shot OOD methods, both in detection accuracy and in uncertainty calibration, validating the advantage of principled contrapositive reasoning over margin-based objectives.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 16542