New distillation framework boosts MLLM anomaly detection accuracy

Researchers have developed ADOPD, a novel reference-privileged on-policy distillation framework designed to enhance industrial anomaly detection using multimodal large language models (MLLMs). This method internalizes the benefits of reference comparison into model parameters during training, addressing limitations where teachers might favor language priors over visual information. ADOPD achieves a 77.31% average accuracy on the MMAD benchmark in a zero-shot setting, significantly improving the Qwen3-VL-4B backbone by 6.14 points and outperforming its one-shot performance. AI

IMPACT Enhances MLLM capabilities for industrial anomaly detection, potentially improving accuracy and efficiency in visual inspection tasks.

RANK_REASON Research paper detailing a new method for MLLM-based anomaly detection. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• ADOPD
• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• MMAD benchmark
• Qwen3-VL 4B
• ScienceCast

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →