Research map / Label-efficient and robust learning / OOD and anomaly detection
Global Logic and Local Search: Dual-Stream Multimodal In-Context Learning for Verifiable Industrial Anomaly Detection
Runzhi Deng, Yundi Hu, Yiming Zhong, Zhao Wang, Xixi Liu, Hongsong Wang, Caifeng Shan, Fang Zhao
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 online dual-stream verification to offline PVLA (ablation) | GLLS (Qwen3-VL-8B) · MMAD-QA MVTec (avg accuracy) | accuracy: 88.98 -> 91.18 | Improved |
| add online dual-stream verification to offline PVLA (ablation) | GLLS (Qwen2.5-VL-7B) · MMAD-QA MVTec (avg accuracy) | accuracy: 77.04 -> 85.46 | Improved |
| add offline PVLA to online-only dual-stream (ablation) | GLLS (Qwen3-VL-8B) · MMAD-QA VisA (avg accuracy) | accuracy: 79.96 -> 84.72 | Improved |
| add offline PVLA to base model (Qwen2.5-VL-7B) (ablation) | Qwen2.5-VL-7B · MMAD-QA MVTec (avg accuracy) | accuracy: 79.0 -> 77.0 | Hurt |
| add PVLA with Qwen3-VL-4B (ablation) | Qwen3-VL-4B · MMAD-QA MVTec+VisA avg | accuracy: 71.8 -> 81.2 | Improved |
| PVLA versus text-guided only atlas (ablation) | Qwen3-VL-8B · MMAD-QA MVTec+VisA avg | accuracy: 81.5 -> 84.4 | Improved |
| add Fine-Grained & Actions Stream to offline base (ablation) | GLLS (Qwen3-VL-4B) · MMAD-QA MVTec+VisA avg | accuracy: 81.2 -> 82.7 | Improved |
| add Global & Logic stream on top of fine-grained stream (ablation) | GLLS (Qwen2.5-VL-7B) · MMAD-QA MVTec+VisA avg | accuracy: 77.7 -> 81.5 | Improved |
| use MCTS search instead of sliding window (ablation) | GLLS (Qwen3-VL-8B) · MMAD-QA (Qwen3-VL-8B 1-shot search ablation) MVTec | accuracy: 89.4 -> 91.2 | Improved |
| use MCTS search instead of center crop (ablation) | GLLS (Qwen3-VL-8B) · MMAD-QA (Qwen3-VL-8B 1-shot search ablation) MVTec | accuracy: 84.2 -> 91.2 | Improved |
| raise evidence budget from 3 to 7 crops (ablation) | GLLS (Qwen3-VL-8B) · MMAD-QA (Qwen3-VL-8B 1-shot search ablation) MVTec | accuracy: 91.2 -> 90.1 | Hurt |
| raise evidence budget from 1 to 3 crops (ablation) | GLLS (Qwen3-VL-8B) · MMAD-QA (Qwen3-VL-8B 1-shot search ablation) MVTec | accuracy: 88.5 -> 91.2 | Improved |
Closest papers by meaning
- AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization (AAAI 2026 · similarity 0.90)
- MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.89)
- Judo: A Juxtaposed Domain-oriented Multimodal Reasoner for Industrial Anomaly QA (ICLR 2026 · similarity 0.89)
- Multimodal Industrial Anomaly Detection by Crossmodal Feature Mapping (CVPR 2024 · similarity 0.89)
- InCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection (NeurIPS 2026 · similarity 0.88)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.88)
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.
