Research map / Label-efficient and robust learning / OOD and anomaly detection
Learning Transferable Negative Prompts for Out-of-Distribution Detection
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.
Learns a few CLIP text 'negative prompts' per ID class from ID data only (on top of CoOp positive prompts) so OOD images match negative prompts better than ID images, with transfer to unseen class names for open-vocabulary OOD detection.
Problem
Prompt-learning-based CLIP methods for out-of-distribution detection lack OOD knowledge, causing ID/OOD confusion and high false positive rates; existing methods also assume access to all ID classes during training.
Approach
Learn class-conditional negative text prompts (negative connotations of each ID label) with only ID images, using a negative-image separation loss (uniform distribution over negative prompts, OE-inspired), a negative-positive distance loss, and a negative-negative diversity loss; the negative prompts transfer to unseen class names, and scoring uses MCM-style max softmax over positive and negative
Key results
- On conventional ImageNet-1K OOD detection, NegPrompt reaches 94.81% average AUC and 23.01% FPR95, beating CLIPN (93.10 / 31.10) and LoCoOp (93.52 / 28.66).
- On hard OOD splits, NegPrompt averages 97.96% AUC and 8.18% FPR95 versus 97.06 / 11.55 for MaxLogit.
- With only 10% of ID classes for training, NegPrompt achieves 93.76% average AUC on conventional OOD detection, and keeps ID top-1 accuracy at 71.9% (vs 72.1% for CoOp).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Learn negative prompts alongside positive class prompts | NegPrompt with CLIP-B/16 · Texture | AUC: 90.93 -> 91.60 | Helped |
| Learn negative prompts alongside positive class prompts | NegPrompt with CLIP-B/16 · iNaturalist | FPR95: 16.05 -> 6.32 | Helped |
| Learn negative prompts alongside positive class prompts | NegPrompt with CLIP-B/16 · Places | AUC: 92.28 -> 93.34 | Helped |
| Learn negative prompts alongside positive class prompts | NegPrompt with CLIP-B/16 · SUN | FPR95: 23.44 -> 22.89 | Helped |
| Use ResNet-50 instead of ViT-B-16 as the backbone (ablation) | NegPrompt · ImageNet OOD detection, average AUC | AUC: 93.37 -> 85.70 | Hurt |
| Use ResNet-101 instead of ViT-B-16 as the backbone (ablation) | NegPrompt · ImageNet OOD detection, average AUC | AUC: 93.37 -> 88.04 | Hurt |
| Use ViT-B-32 instead of ViT-B-16 as the backbone (ablation) | NegPrompt · ImageNet OOD detection, average AUC | AUC: 93.76 -> 93.37 | Hurt |
| Use 1 negative prompt instead of 2 (ablation) | NegPrompt · ImageNet OOD detection, average AUC | AUC: 93.76 -> 92.73 | Hurt |
| Train positive and negative prompts together in one stage instead of two stages (ablation) | NegPrompt · ImageNet OOD detection, average AUC | AUC: 93.76 -> 80.23 | Hurt |
Surprising findings
- Training positive and negative prompts jointly in one stage collapses performance (average AUC 80.23 vs 93.76 for two-stage), showing the need to anchor positive prompts first.
Limitations
- Computational cost of negative prompts grows with the number of negative prompts per class.
- Relies on CoOp-learned positive prompts as an anchor, making training two-stage.
- Some ablations show performance varies unevenly across CLIP backbones (e.g., ViT-B-32 vs ViT-B-16).
Experimental setup
Datasets: ImageNet-1K Texture iNaturalist Places SUN TinyImageNet
Benchmarks: ImageNet-1K OOD benchmarks (Texture, iNaturalist, Places, SUN) ImageNet-1K hard OOD splits (Split-1 to Split-4)
Metrics: AUROC FPR95 Top-1 accuracy
Baselines: MCM CLIPN MSP MaxLogit Energy ReAct ODIN CoOp LoCoOp
Builds on: CoOp MCM OE (Outlier Exposure) LoCoOp
Closest papers by meaning
- Negative Label Guided OOD Detection with Pretrained Vision-Language Models (ICLR 2024 · similarity 0.88)
- ID-like Prompt Learning for Few-Shot Out-of-Distribution Detection (CVPR 2024 · similarity 0.88)
- A Noisy Elephant in the Room: Is Your Out-of-Distribution Detector Robust to Label Noise? (CVPR 2024 · similarity 0.87)
- Adaptive Multi-prompt Contrastive Network for Few-shot Out-of-distribution Detection (ICML 2025 · similarity 0.87)
- Splitting & Integrating: Out-of-Distribution Detection via Adversarial Gradient Attribution (ICML 2025 · similarity 0.87)
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.87)
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.