A Geometric Analysis of Logit Embeddings for Out-of-Distribution Detection

ICLR 2026 Conference Submission18685 Authors

Keywords: Out-of-distribution Detection; Logit Geometry; Representation Learning

TL;DR: We show that, in DL classifiers, OoD inputs concentrate near the logit-space origin while InD inputs form orthogonal, class-wise clusters—yielding a simple, geometry-driven criterion for reliable OOD detection.

Abstract:

Out-of-distribution (OOD) data pose a significant challenge to deep learning (DL) classifiers, prompting extensive research into their effective detection methods.
Current state-of-the-art OOD detection methods usually employ a scoring technique designed to assign lower scores to OOD samples compared to in-distribution (ID) ones.
Nevertheless, these approaches lack foresight into the configuration of OOD and ID data within the latent space.
Instead, they make an implicit assumption about their inherent separation or force a separation post-training by utilizing selected OOD data.
As a result, most OOD detection methods result in complicated and hard-to-validate scoring techniques.
This study conducts a thorough analysis of the logit embedding landscape, revealing that the ID and OOD data exhibit a distinct spatial configuration.
Specifically, we empirically observe that the OOD data are drawn to the center of the logit space.
In contrast, ID data are repelled from the center, dispersing outward into distinct, class-wise clusters aligned along the orthogonal axes that span the logit space.
This study highlights the critical role of the DL vision-based classifier in differentiating between ID and OOD logits.

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

Submission Number: 18685