The paper introduces PLUS, a detector‑agnostic subspace signal that enhances near‑out‑of‑distribution (near‑OOD) detection. PLUS selects the most class‑active neurons via a sparse autoencoder, orthonormalizes their decoder directions to form a low‑dimensional subspace, and scores inputs using a per‑class whitened squared Mahalanobis distance within that subspace. The resulting signal is fused additively with any calibrated base score using a single global weight, improving near‑OOD AUROC across 126 detector–backbone combinations while largely preserving far‑OOD performance.
Post-hoc out-of-distribution (OOD) detection is mature on far-OOD yet remains weak on near-OOD. These two settings differ in their use of class-characteristic feature directions: far-OOD inputs largel...
arXiv:2607. 19393v1 Announce Type: cross Abstract: While auditing a perturbation-based OOD detector on a document benchmark, we recorded an AUROC of 0.
MahaVar is a post‑hoc out‑of‑distribution detector that augments the Mahalanobis distance with a class‑wise distance variance term. The authors observe that in‑distribution samples exhibit a sharp minimum in class‑wise Mahalanobis distances, leading to high variance across classes, whereas OOD samples show a flatter structure and lower variance. Theoretical analysis based on Neural Collapse supports this observation, and empirical results on CIFAR‑100 and ImageNet demonstrate state‑of‑the‑art performance compared to existing Mahalanobis‑based methods.
arXiv:2503. 05169v2 Announce Type: replace Abstract: Applying machine learning to increasingly high-dimensional problems with sparse or biased training data increases the risk that a model is used on inputs outside its training domain.
The paper investigates why post‑hoc out‑of‑distribution (OOD) detectors often report inflated gains. It shows that many detectors learn to recognize the specific dataset they were trained on rather than true novelty, a phenomenon called the inflation gap. By holding out entire OOD datasets during evaluation, the authors demonstrate that most reported gains vanish, and they provide a closed‑form method to predict which fits will inflate.
arXiv:2607. 23856v1 Announce Type: new Abstract: A Last-Layer Ensemble (LLE), $K$ linear units on one shared frozen feature map, is an efficient single-pass approach to the disagreement-based epistemic uncertainty for out-of-distribution (OOD) detection.
The paper introduces a new active learning signal for object detection that relies on a supervised contrastive term added to the training objective. This term shapes an embedding space where distance reflects class membership, allowing an unlabeled detection to be scored by its distance from the predicted category’s region weighted by confidence—all from a single forward pass of one network. Experiments on PASCAL VOC and MS‑COCO show that this criterion outperforms the standard posterior and remains competitive with ensemble‑based methods while incurring only a modest 8.3% increase in parameters.
arXiv:2610.09709v1 Announce Type: new Abstract: Can we determine whether a foundation model will generalize out-of-distribution (OOD) before any target data is available? Existing diagnostics require...
arXiv:2610.33063v1 Announce Type: cross Abstract: Deployed vision models auto-accept confident predictions, so high-confidence errors propagate silently. Maximum Softmax Probability (MSP), the standa...
arXiv:2609.01027v1 Announce Type: new Abstract: Out-of-distribution (OOD) detection predicts whether a test image belongs to none of the predefined classes. To evaluate this task, benchmarks need ima...
arXiv:2606. 24178v1 Announce Type: cross Abstract: Pretrained vision models often misclassify inputs that are rotated, scaled, or sheared, even though these affine transformations leave the object class unchanged.
PairCLIP‑SWA is a single‑backbone deepfake detector that fine‑tunes an OpenCLIP ViT‑L/14 visual tower using a class‑balanced corpus created by CLIP‑similarity‑guided subsampling and then applies stochastic weight averaging over the fine‑tuning trajectory. Trained on the IEEE SP Cup 2025 (DF‑Wild‑Cup) benchmark, the weight‑averaged model achieves 97.04 % accuracy, 98.64 % AU‑ROC and a 4.07 % equal error rate—outperforming the best published three‑backbone ensemble. The study shows that averaging stabilizes validation performance, prevents misranking on disjoint test sets, and that CLIP pre‑training, rather than ensembling, is the key driver of accuracy.