Research map / Label-efficient and robust learning / OOD and anomaly detection

A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Spurious-aware OOD detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Systematic OOD/misclassification-detection benchmark showing that the best confidence scoring function depends on the learned representation, with NC-geometry-based predictor recommending detector shortlists across architectures.

Problem

Rankings of OOD confidence scoring functions are unstable across datasets, architectures and training paradigms; it is unclear when a detector family works and whether it can be selected from the trained model's geometry without OOD data.

Approach

Frames detector competitiveness as a function of last-layer representation geometry: uses Neural Collapse metrics to correlate with CSF rankings (Mantel tests) and to predict a competitive detector shortlist for unseen models, plus a PCA-based projection filtering procedure (global and class-predicted variants).

Key results

• NC-based predictor reduces mean set-regret over Always-CTM by 51% (near), 67% (mid), and 84% (far) under VGG-13 to ResNet-18 transfer.
• Mantel tests show NC geometry and CSF rankings are significantly correlated across training paradigms (r = 0.31-0.68, p <= 0.004).
• Holm-corrected paired Wilcoxon tests show the predictor beats the baseline in 49 of 54 comparisons at alpha = 0.05.

Surprising findings

• Deep Gamblers training promotes Mahalanobis and Residual into far cliques otherwise dominated by CTM and fDBD, even when aggregate NC metrics remain similar.
• Some CSFs with large AUGRC gains from projection filtering (e.g., Maha on CNNs) still fail to enter any top clique.

Limitations

• Restricted to image classification on small/medium datasets with VGG-13, fine-tuned ViT, and three training paradigms.
• Cross-architecture transfer tested only for CNN to CNN (VGG-13 to ResNet-18) with a single run per held-out cell.
• Mantel association is correlational; hyperparameters tuned for AUGRC may not transfer to other ranking metrics.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 SuperCIFAR-100 TinyImageNet iSUN LSUN-cropped LSUN-resize SVHN Places365 Textures (DTD)

Benchmarks: CIFAR-10 CIFAR-100 SuperCIFAR-100 TinyImageNet iSUN LSUN SVHN Places365 DTD

Metrics: AURC AUGRC set-regret

Baselines: MSR MLS Energy Mahalanobis NNGuide fDBD CTM Always-CTM

Builds on: FD-Shifts OpenOOD Neural Collapse ViM ReAct ASH

Closest papers by meaning

• Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.92)
• Rethinking the Evaluation of Out-of-Distribution Detection: A Sorites Paradox (NeurIPS 2024 · similarity 0.91)
• Is Complex Training Necessary for Long-Tailed OOD Detection? A Re-think from Feature Geometry (NeurIPS 2026 · similarity 0.90)
• ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.90)
• Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.90)
• Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.89)