ADOPD: Reference-Privileged On-Policy Distillation for MLLM-Based Industrial Anomaly Detection

2026-08-10

AI Summary

This work addresses the limitations of existing multimodal large language models in industrial anomaly detection, which rely on reference-image comparison during inference, struggle to internalize reference knowledge, and are susceptible to linguistic priors. The authors propose ADOPD, a novel framework that, during training, leverages a teacher model guided by reference images to supervise a student model that uses only query images. ADOPD introduces a reference-privilege mechanism based on policy distillation: by contrasting generation trajectories under matched and mismatched references, and combining token-level log-ratio guidance with sequence-level likelihood-gap weighting, the student learns to replicate comparison-based reasoning without requiring reference images at inference time. On the MMAD benchmark, ADOPD achieves a zero-shot accuracy of 77.31%, surpassing the Qwen3-VL-4B baseline by 6.14 percentage points and even outperforming its one-shot setting by 2.64 points.

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.

Problem

• industrial anomaly detection
• multimodal large language models
• reference comparison
• on-policy distillation
• zero-shot inference

Innovation

• on-policy distillation
• reference-privileged learning
• multimodal large language models
• industrial anomaly detection
• zero-shot inference