Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection

Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection

Wonseok Shin and Songkuk Kim argue that a score suited to far-OOD detection can miss near-OOD inputs: these inputs activate class-characteristic feature directions much like in-distribution samples. They propose PLUS, a signal designed to complement existing scores rather than replace them. For each class, PLUS selects the most class-active neurons from a sparse autoencoder, forms a low-dimensional subspace from their decoder directions, and scores an input using its per-class whitened squared Mahalanobis distance within the subspace of its predicted class. The authors add this signal to a calibrated base score at one fixed global weight, with calibration statistics estimated from ID validation data alone. They report improved near-OOD AUROC across 126 detector–backbone combinations using 18 existing detectors, while far-OOD performance was broadly maintained or improved, with marginal declines where the base score was near-optimal. The proposed approach matters because it aims to add sensitivity to near-OOD inputs without discarding detectors already performing well on far-OOD. The abstract reports these evaluation results but does not identify the datasets, so their scope cannot be assessed from this evidence alone.

Publication: 884cc872-9cfc-49fe-8662-d5bf022061a4

attribution — pass: The reading correctly attributes the argument, method and reported results to authors Wonseok Shin and Songkuk Kim. It frames the results as reported by the authors, and the limitations note that they are not an independent assessment.

Authors:Wonseok Shin, Songkuk Kim

summary — uncertain: The summary faithfully tracks the abstract, but central terms are never explained for a general reader. It never expands OOD (out-of-distribution) or says what near- versus far-OOD inputs are, such as inputs unlike a model's training data. AUROC, sparse autoencoder, Mahalanobis distance, backbone and calibrated base score also go unexplained. Without these, a non-specialist cannot grasp what problem PLUS addresses or why it matters. Smallest repair: in the opening sentence, say plainly that OOD detection flags inputs unlike what the model was trained on, and that near-OOD inputs are deceptively similar ones; add tooltips for the remaining terms.

Post-hoc out-of-distribution (OOD) detection is mature on far-OOD yet remains weak on near-OOD.

date — pass: The displayed original publication date of 2026-11-04 matches the arXiv v1 submission date. The feed date of 2026-11-05 was correctly not used.

[Submitted on 4 Nov 2026]

support — pass: The claims match the abstract: the activation-direction argument, the PLUS construction, the single global fused weight using ID-only calibration, the 126 detector-backbone combinations across 18 detectors, and the far-OOD caveat. The stated significance of adding near-OOD sensitivity without discarding far-OOD detectors follows from the authors' complement framing. The limitations accurately note that the abstract does not identify the datasets.

Designed as a dedicated complement rather than a standalone detector, the signal $z$ is fused additively with any calibrated base score at a single, fixed weight.

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.