Near-OOD detection gets a complement, not a replacement: PLUS lifts all 126 detector-backbone pairs across 18 detectors
Synopsis
Near-out-of-distribution (near-OOD) inputs resemble the training classes closely, and existing post-hoc detectors that excel on far-OOD inputs lose that edge here. PLUS does not replace a detector: it uses a sparse autoencoder to pick each class's most active neurons, spans a low-dimensional subspace, and scores an input by that class's whitened Mahalanobis distance, adding the signal to any calibrated base score at one fixed weight. On the ImageNet-1K OpenOOD v1.5 benchmark, near-OOD AUROC improves in all 126 detector-backbone combinations across 18 detectors and seven backbones, while far-OOD performance is broadly maintained, with declines of at most one point where the base score is already near-saturated.
Interpretation
Near-OOD detection gains an independent signal that stacks onto existing detectors instead of being another detector that replaces them. Post-hoc detectors had relied on global feature summaries such as density, residual energy or classifier confidence; those axes separate far-OOD inputs but leave near-OOD samples, which sit on the training manifold, almost indistinguishable from in-distribution data. On the ImageNet-1K OpenOOD v1.5 benchmark, near-OOD AUROC improves in all 126 detector-backbone combinations across 18 detectors and seven backbones; for MSP on ViT-B/16, near-OOD AUROC rises from 73.52 to 78.13 and FPR95 falls from 81.85 to 68.82.
The complement comes from each class's own low-dimensional coordinates, measuring how far an input deviates from that class's training distribution inside its predicted class's subspace. A TopK sparse autoencoder is fit on in-distribution training features to give an overcomplete dictionary; for each class the 16 most active neurons are fixed, and their decoder directions are orthonormalized by sign-fixed QR to form a stable projection basis. The subspace supports and statistics come from a random subset of in-distribution training images (at most 300 per class), while score standardization uses in-distribution validation negatives; the fusion weight is fixed at a 2:1 base-favored ratio and is not adjusted per backbone, dataset or deployment.
The gain requires both the subspace and the per-class metric; neither suffices alone. With the same per-class metric inside a random 16-dimensional orthonormal subspace, there is no near-OOD gain over MDS++ alone; inside the chosen subspace, a shared or isotropic metric recovers only part of the improvement, and the shared metric gives almost nothing on ResNet-50. With MDS++ as the base, near-OOD AUROC reaches 81.23, 78.22 and 80.90 on ViT-B/16, ResNet-50 and Swin-T, versus 79.87, 74.71 and 78.92 for MDS++ alone.
Perspective
Evaluation is confined to the OpenOOD v1.5 benchmark with ImageNet-1K as in-distribution data, using SSB-hard and NINCO for near-OOD and iNaturalist, Textures and OpenImage-O for far-OOD, across ViT-B/16, ResNet-50, Swin-T and four further ImageNet backbones. PLUS keeps the base classifier frozen and needs only a one-time sparse-autoencoder fit on in-distribution training features, with one projection and one whitening operation per image at inference.
The fixed 2:1 fusion weight is a deployment default, not a claim of optimality; the sensitivity analysis shows near-OOD gains saturating around weight 1 on some backbones while far-OOD typically peaks near 1 and then declines. Textures concentrates the trade-off, with 9 of 18 detectors improving and 9 dropping on ViT-B/16 and Swin-T, because pure-texture crops deviate little within the predicted class's subspace. Far-OOD AUROC declines never exceed one point and occur on already near-saturated Mahalanobis-family bases, with ASH on ResNet-50 also showing a decline.