Research map / Detection and segmentation / Open-set and few-shot detection
Beyond Common Sense: Grounding Logical Anomaly Detection in Inspection Criteria
Yuanze Li, Shihao Yuan, Zimeng Zhu, Ming LIU, Wangmeng Zuo, Guangming Shi
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 explicit inspection criteria to the prompt (ablation) | Qwen2.5-VL-72B · MVTec-LOCO (scoring) | precision: 33.28 -> 78.38 | Improved |
| add explicit inspection criteria to the prompt (ablation) | Qwen3-VL-8B · MVTec-LOCO (scoring) | recall: 21.07 -> 24.57 | Improved |
| add explicit inspection criteria to the prompt (ablation) | GPT-5.2-chat · MVTec-LOCO (scoring) | recall: 7.92 -> 37.34 | Improved |
| add explicit inspection criteria to the prompt (ablation) | SCAN (8B) · MVTec-LOCO (scoring) | accuracy: 51.48 -> 64.58 | Improved |
| train SCAN (SFT + RL) on Qwen3-VL-8B | Qwen3-VL-8B · MVTec-LOCO (scoring) | F1: 35.55 -> 53.50 | Improved |
| train SCAN on Qwen2.5-VL-7B | Qwen2.5-VL-7B · MVTec-LOCO (scoring) | precision: 67.98 -> 69.50 | Mixed |
| add SFT stage on FLAW data (ablation) | Qwen3-VL-8B · MVTec-LOCO (scoring) | recall: 24.57 -> 34.09 | Improved |
| add RL with sequence-level reward only (ablation) | SCAN SFT model · MVTec-LOCO (scoring) | precision: 80.14 -> 76.73 | Hurt |
| add rule-level reward over sequence-level reward (ablation) | SCAN RL · MVTec-LOCO (scoring) | precision: 76.73 -> 86.91 | Improved |
| add rule-level credit assignment on top of rule-level reward (ablation) | SCAN RL · MVTec-LOCO (scoring) | precision: 86.91 -> 90.05 | Improved |
| add synthetic and collection data pipelines together (FLAW) (ablation) | Qwen2.5-VL-7B SFT · MVTec-LOCO (scoring) | F1: 22.39 -> 33.98 | Improved |
| add FLAW data during SFT (ablation) | Qwen2.5-VL-7B SFT · MVTec-LOCO (scoring) | precision: 67.98 -> 61.83 | Hurt |
Closest papers by meaning
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.90)
- MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.90)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.89)
- LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction (AAAI 2025 · similarity 0.89)
- Few Shot Part Segmentation Reveals Compositional Logic for Industrial Anomaly Detection (AAAI 2024 · similarity 0.88)
- InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection (NeurIPS 2026 · 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.
