JUDO outperforms GPT-4o in industrial anomaly QA with domain knowledge injection

Industrial anomaly detection has long been bottlenecked by the lack of domain-specific knowledge in large multimodal models (LMMs). A new framework called JUDO (Juxtaposed Domain-Oriented Multimodal Reasoner) tackles this head-on. Developed by researchers including Hyunju Kang and Hogun Park, JUDO augments LMMs with two key innovations: visual reasoning via juxtaposition and domain-oriented reinforcement learning. First, the model compares the query image against a normal reference image to precisely segment defect regions — a form of fine-grained comparative inspection that mirrors how human experts work. Then, domain knowledge is injected through supervised fine-tuning (SFT) to improve context understanding, followed by GRPO (Group Relative Policy Optimization) reinforcement learning that guides the model toward domain-appropriate reasoning paths.

On the MMAD benchmark, JUDO achieves superior performance, surpassing both Qwen2.5-VL-7B and GPT-4o. This demonstrates that explicit domain knowledge injection—rather than just scaling models—is critical for accurate anomaly understanding in manufacturing, electronics, and other industrial settings. The work was accepted at ICLR 2026. For professionals, JUDO represents a practical path to deploy AI for quality control without relying on general-purpose models that lack specialized defect knowledge. The code and data are available via the arXiv preprint.