Research map / Vision-language and multimodal / Human and embodied modeling
LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction
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.
Uses GPT-4o/LLaVA via guided chain-of-thought, ROI cropping and embedding-based outlier filtering to extract text features for one-shot logical anomaly detection, plus a Prover9-based logic reasoner for explanations.
Problem
One-shot logical anomaly detection in industrial images (e.g., missing, extra, or misplaced objects), where CLIP-based and visual memory-bank methods perform poorly and lack explanations.
Approach
Replace visual features with text features extracted from autovisual-language models (AVLMs) via guided chain-of-thought prompts, GroundingDINO region crops, and LOF-based filtering of multiple text samples; compare normal and query text embeddings for anomaly scores, and optionally use an LLM-to-formal-logic pipeline with Prover9 theorem proving for explainable detection.
Key results
- On MVTec LOCO AD (one-shot), LogicAD reaches 86.0% AUROC and 83.7% F1-max, an 18.1% AUROC gain over AnomalyMoE.
- The logic reasoner variant reaches 83.3% F1-max, only 0.4 points below the format-embedding variant.
- On MVTec AD object categories containing logical anomalies (capsule, pill, transistor, toothbrush), LogicAD reports 86.5% AUROC versus 80.9% for WinCLIP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Guided CoT (ablation) | LogicAD · MVTec LOCO AD | AUROC: 23.4 -> 48.5 | Helped |
| Add ROI extraction to Guided CoT (ablation) | LogicAD · MVTec LOCO AD | AUROC: 48.5 -> 60.4 | Helped |
| Add format embedding to Guided CoT + ROI (ablation) | LogicAD · MVTec LOCO AD | AUROC: 60.4 -> 86.0 | Helped |
| Use the logic reasoner instead of format embedding (ablation) | LogicAD with Guided CoT + ROI · MVTec LOCO AD | F1-max: 83.7 -> 83.3 | Hurt |
| Use LLaVA 1.5 instead of GPT-4o | LogicAD · MVTec LOCO AD | AUROC: 86.0 -> 73.3 | Hurt |
| Use LLaVA 1.5 instead of GPT-4o | LogicAD · MVTec LOCO AD | F1-max: 83.2 -> 71.0 | Hurt |
| Use LLaVA 1.6 instead of GPT-4o | LogicAD · MVTec LOCO AD | AUROC: 86.0 -> 76.2 | Hurt |
| Use LLaVA 1.6 instead of GPT-4o | LogicAD · MVTec LOCO AD | F1-max: 83.2 -> 78.1 | Hurt |
| Use LogicAD instead of AnomalyMoE | One-shot logical anomaly detection · MVTec LOCO AD (only LA) | AUROC: 67.9 -> 86.0 | Helped |
| Use LogicAD instead of AnomalyMoE | One-shot logical anomaly detection · MVTec LOCO AD (only LA) | F1-max: 79.1 -> 83.7 | Helped |
| Use LogicAD instead of WinCLIP | Anomaly detection on LA-related MVTec AD objects · MVTec AD Object* | AUROC: 80.9 -> 86.5 | Helped |
| Use LogicAD instead of PatchCore | Anomaly detection · MVTec LOCO AD (only SA) | AUROC: 89.3 -> 81.5 | Hurt |
Surprising findings
- WinCLIP performs substantially worse on categories dominated by naive logical anomalies, and GPT-4o counting accuracy drops sharply on homogeneous-object images.
Limitations
- Results vary across different AVLM backbones.
- Inference takes a few seconds per image.
- Prompts still require manual, task-specific guided CoT design.
Experimental setup
Datasets: MVTec LOCO AD MVTec AD CountBench
Benchmarks: MVTec LOCO AD MVTec AD
Metrics: AUROC F1-max
Baselines: AnomalyMoE WinCLIP GCAD PatchCore ComAD AST AnomalyCLIP VAND
Builds on: WinCLIP GroundingDINO Prover9 Local Outlier Factor
Closest papers by meaning
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.89)
- MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.88)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.88)
- VERA: Explainable Video Anomaly Detection via Verbalized Learning of Vision-Language Models (CVPR 2025 · similarity 0.87)
- Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought (NeurIPS 2025 · similarity 0.86)
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.