Research map / Label-efficient and robust learning / OOD and anomaly detection
ADSeeker: A Knowledge-Grounded Reasoning Framework for Industry Anomaly Detection and Reasoning
Kai Zhang, Zekai Zhang, Xihe Sun, Anpeng Wang, Jingmeng Nie, Qinghui Chen, Han Hao, Jianyuan Guo, Jinglin Zhang
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 Q2K RAG (ablation) | ADSeeker baseline · MMAD two subtasks | accuracy (%): 76.6 -> 78.6 | Improved |
| add AD Expert module (ablation) | ADSeeker baseline · MMAD two subtasks | accuracy (%): 76.6 -> 77.9 | Improved |
| combine Q2K RAG and AD Expert (ablation) | ADSeeker baseline · MMAD two subtasks | accuracy (%): 76.6 -> 82.8 | Improved |
| LoRA fine-tune for 5 epochs on 43K instruction data (ablation) | baseline MLLM · MMAD two subtasks | accuracy (%): 76.6 -> 80.0 | Improved |
| extend LoRA fine-tuning from 5 to 20 epochs (ablation) | baseline MLLM · MMAD two subtasks | accuracy (%): 80.0 -> 40.9 | Hurt |
| use Q2K RAG framework instead of LoRA fine-tuning | baseline MLLM · MMAD two subtasks | accuracy (%): 80.0 -> 82.8 | Improved |
| apply Seek-Setting (Q2K RAG + AD Expert) to Qwen2.5-VL 7B | Qwen2.5-VL 7B · MMAD 1-shot | average accuracy (%): 66.62 -> 69.90 | Improved |
| apply Seek-Setting to LLaVA-NeXT 7B | LLaVA-NeXT 7B · MMAD 1-shot | average accuracy (%): 59.32 -> 65.96 | Improved |
| apply Seek-Setting to LLaVA-OV 7B | LLaVA-OV 7B · MMAD 1-shot defect description | accuracy (%): 62.19 -> 57.17 | Hurt |
| apply Seek-Setting to Qwen2.5-VL 7B | Qwen2.5-VL 7B · MMAD 1-shot object classification | accuracy (%): 92.07 -> 89.91 | Hurt |
| apply Seek-Setting to LLaVA-NeXT 7B | LLaVA-NeXT 7B · MMAD 1-shot defect classification | accuracy (%): 33.79 -> 53.29 | Improved |
| apply Seek-Setting without extra training to LLaVA-OV vs trained Anomaly-OV | LLaVA-OV · MMAD subset average | accuracy (%): 69.80 -> 73.54 | Improved |
Closest papers by meaning
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.90)
- LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction (AAAI 2025 · similarity 0.89)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.89)
- Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
- MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.88)
- AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization (AAAI 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.
