ADOPD: Reference-Privileged On-Policy Distillation for MLLM-Based Industrial Anomaly Detection
2026-08-10 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors study how to detect small visual anomalies in industrial images by using models that compare new images to known reference images. Instead of needing to look at references every time, they teach a model to internalize the comparison knowledge during training. Their method, ADOPD, uses a teacher model with access to references to guide a student model that only sees the query images, improving accuracy by making the student focus on meaningful visual differences. This approach improves performance on a standard benchmark without needing extra reference images at test time.
industrial anomaly detectionmultimodal large language modelsreference-aware teacherknowledge distillationzero-shot inferencevisual anomaly inspectiontoken-level learningsequence-level calibrationon-policy distillationMMAD benchmark
Authors
Jingtai He, Shiyuan Meng, Wenchao Meng, Qinmin Yang
Abstract
Industrial anomaly detection (IAD) requires identifying fine-grained deviations from normal visual patterns. Multimodal large language models (MLLMs) can improve recognition accuracy by comparing query images with references at inference time, but these benefits rely on additional retrieval and processing. We investigate whether the benefits of reference comparison can instead be internalized in the model parameters. Access to references during training allows a reference-aware teacher to supervise a query-only student. However, the teacher may favor plausible responses based on query cues or language priors rather than valid visual information. We propose ADOPD, a reference-privileged on-policy distillation framework. The teacher evaluates student-generated rollouts under matched and mismatched references. The matched-reference teacher-to-student log-ratio defines the token-level learning direction, specifying what the student should learn. The likelihood gap between the two reference views estimates reference-specific support and calibrates the sequence-level weight. ADOPD achieves 77.31% average accuracy on the MMAD benchmark under zero-shot inference, improving the Qwen3-VL-4B backbone by 6.14 points and outperforming its one-shot setting by 2.64 points. Experiments show that ADOPD learns a fine-grained anomaly inspection strategy from reference comparison. The project will be available at https://github.com/withTai/ADOPD.