Research map / Label-efficient and robust learning / OOD and anomaly detection
NegRefine: Refining Negative Label-Based Zero-Shot OOD Detection
Amirhossein Ansari, Ke Wang, Pulei Xiong
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 |
|---|---|---|---|
| add negative label filtering (NegFilter) (ablation) | CSP · ImageNet-1K ID, average over 4 OOD sets | FPR95: 34.47 -> 31.88 | Improved |
| add multi-matching score S_MM (ablation) | CSP · ImageNet-1K ID, average over 4 OOD sets | FPR95: 34.47 -> 31.70 | Improved |
| add both NegFilter and S_MM (ablation) | CSP · ImageNet-1K ID, average over 4 OOD sets | FPR95: 34.47 -> 30.12 | Improved |
| add proper nouns check C1 to NegFilter (ablation) | NegRefine (S_MM on) · ImageNet-1K ID, average over 4 OOD sets | FPR95: 31.70 -> 31.09 | Improved |
| add subcategory check C2 to NegFilter (ablation) | NegRefine (S_MM on) · ImageNet-1K ID, average over 4 OOD sets | FPR95: 31.70 -> 30.82 | Improved |
| combine both filter checks C1 and C2 (ablation) | NegRefine (S_MM on) · ImageNet-1K ID, average over 4 OOD sets | FPR95: 31.70 -> 30.12 | Improved |
| swap the LLM used for filtering (Qwen2.5-14B vs Mistral, Llama-3-8B, Qwen2.5-7B) (ablation) | NegFilter · ImageNet-1K ID, average over 4 OOD sets | AUROC: 91.79 vs 91.74-91.99 | No clear effect |
| increase k from 1 to 2 in multi-matching texts (ablation) | S_MM · ImageNet-1K ID, average over 4 OOD sets | FPR95: 36.86 -> 31.94 | Improved |
| use S_MM instead of local-feature score S_L (ablation) | NegLabel score · ImageNet-1K ID, average over 4 OOD sets | AUROC | Improved |
| replace sigmoid-style exp ratio max with sum over matching texts (ablation) | S_MM · ImageNet-1K ID, average over 4 OOD sets | FPR95: 30.12 -> 49.00 | Hurt |
| increase initial negative label percentage p from 15 to 100 (ablation) | NegRefine · ImageNet-1K ID, average over 4 OOD sets | FPR95: 30.12 -> 32.16 | Hurt |
| use NegRefine instead of CSP with ImageNet-10 as ID | CSP · ImageNet-10 ID vs ImageNet-20 OOD | FPR95: 3.30 -> 3.89 | Hurt |
Closest papers by meaning
- Negative Label Guided OOD Detection with Pretrained Vision-Language Models (ICLR 2024 · similarity 0.91)
- A Close Look at Negative Label Guided Out-of-distribution Detection in Pre-trained Vision-Language Models (ICML 2026 · similarity 0.90)
- Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Learning Transferable Negative Prompts for Out-of-Distribution Detection (CVPR 2024 · similarity 0.89)
- AdaNeg: Adaptive Negative Proxy Guided OOD Detection with Vision-Language Models (NeurIPS 2024 · similarity 0.89)
- Can We Ignore Labels in Out of Distribution Detection? (ICLR 2025 · similarity 0.85)
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.
