Research map / Label-efficient and robust learning / OOD and anomaly detection
Exploring the Limits of Out-of-Distribution Detection
Stanislav Fort, Jie Ren, Balaji Lakshminarayanan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune ImageNet-21k pre-trained ViT instead of prior SOTA | near-OOD detection · CIFAR-100 vs CIFAR-10 | AUROC: 85% -> 96% | Improved |
| transformer (ViT) instead of ResNet (BiT) backbone with same pre-training data (ablation) | fine-tuned OOD detector · CIFAR-100 vs CIFAR-10 | AUROC (Mahalanobis): 90.10 -> 95.53 | Improved |
| Mahalanobis score instead of MSP (ablation) | fine-tuned ViT-B_16 · CIFAR-100 vs CIFAR-10 | AUROC: 91.89 -> 95.53 | Improved |
| supervised instead of self-supervised (DINO) pre-training at matched accuracy (ablation) | ViT-B_16 · CIFAR-100 vs CIFAR-10 | AUROC (Mahalanobis): 88.78 -> 93.05 | Improved |
| scale up to ViT-L_16 (ablation) | ViT · CIFAR-100 vs CIFAR-10 | AUROC (Mahalanobis): 96.23 -> 97.98 | Improved |
| add few-shot outlier exposure (1 example per OOD class) | fine-tuned R50+ViT · CIFAR-100 vs CIFAR-10 | AUROC: 96.23 -> 98.70 | Improved |
| fine-tune on in-distribution before outlier exposure (ablation) | R50+ViT, 1-shot OE · CIFAR-100 vs CIFAR-10 | AUROC: 88.73 -> 98.70 | Improved |
| unsupervised BERT pre-training + fine-tuning instead of 1D CNN | genomics classifier · Genomics OOD | AUROC (Mahalanobis): 64.75 -> 77.49 | Improved |
| add few-shot outlier exposure | BERT genomics · Genomics OOD | AUROC: 76.73 -> 88.48 | Improved |
| use outlier class labels instead of collapsing to single OOD class (ablation) | fine-tuned BERT OE · Genomics OOD | AUROC: 84.36 -> 82.21 (1-shot); 87.94 -> 88.49 (100-shot) | Mixed |
| add OOD class names as candidate labels | CLIP zero-shot · CIFAR-100 vs CIFAR-10 | AUROC: 69.49 -> 94.68 | Improved |
| use 'the opposite of' + in-dist names as OOD labels (ablation) | CLIP zero-shot · CIFAR-100 vs CIFAR-10 | AUROC: 69.49 -> 61.37 | Hurt |
Closest papers by meaning
- How Out-of-Distribution Detection Learning Theory Enhances Transformer: Learnability and Reliability (ICML 2026 · similarity 0.91)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.91)
- Leveraging Perturbation Robustness to Enhance Out-of-Distribution Detection (CVPR 2025 · similarity 0.90)
- A Systematic Analysis of Out-of-Distribution Detection Under Representation and Training Paradigm Shifts (NeurIPS 2026 · similarity 0.90)
- Out-of-Distribution Detection with Prototypical Outlier Proxy (AAAI 2025 · similarity 0.90)
- Mysteries of the Deep: Role of Intermediate Representations in Out of Distribution Detection (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
