Research map / Vision-language and multimodal / Multimodal LLMs

Judo: A Juxtaposed Domain-oriented Multimodal Reasoner for Industrial Anomaly QA

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes JUDO, a three-stage Qwen2.5-VL-7B training pipeline (juxtaposed patch segmentation SFT, domain-knowledge QA SFT, and GRPO with domain-reasoning, segmentation, and choice/format rewards) that reaches 81.20% average accuracy on MMAD.

Problem

Industrial anomaly understanding with LMMs is limited by missing domain-specific knowledge and by reliance on normal reference images only at inference time, which can mislead responses on MMAD-style multiple-choice QA.

Approach

Internalize domain knowledge and normal-image context during training rather than at inference: juxtaposed patch-level segmentation between query and normal images, SFT on domain QA generated from MMAD domain snippets, then GRPO with a semantic-similarity reward to a pseudo-domain rationale plus segmentation F1 and choice/format rewards.

Key results

• JUDO reaches 81.20% average accuracy on MMAD, the highest among compared models, versus 77.29% for the re-implemented AnomalyR1 (GRPO) and 72.56% for base Qwen2.5-VL-7B.
• Stage 2 domain injection raises accuracy to 79.82% from 77.29% with GRPO alone, while RAG with the same model lowers it to 76.29%.
• Anomaly discrimination is not improved: JUDO scores 65.04%, below Kimi-VL-A3B (72.93%) and Qwen2.5-VL-7B (71.39%).

Surprising findings

• Adding RAG domain snippets at inference hurt the GRPO-trained model (76.29% vs 77.29%).
• Switching Qwen2.5-VL-7B from direct answering to chain-of-thought reasoning sharply lowered anomaly discrimination (71.39% to 61.90%).

Limitations

• Lower anomaly discrimination accuracy than some general LMMs, attributed to reasoning-mode degradation and limited vision encoder capacity.
• OmniAD was not compared because its code was unavailable; AnomalyR1 was re-implemented by the authors.
• Stage 3 uses only one training instance per category from MMAD, and the pseudo-domain rationale is generated by GPT-4o.

Experimental setup

Datasets: MMAD REAL-IAD MVTec AD MVTec LOCO VisA GoodsAD

Benchmarks: MMAD

Metrics: average accuracy

Baselines: Qwen2.5-VL-7B AnomalyR1 GPT-4o Gemini-2.5-pro InternVL3.5 Kimi-VL-A3B RAG (GRPO + RAG)

Builds on: GRPO Text4Seg AnomalyR1 MMAD

Closest papers by meaning

• AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization (AAAI 2026 · similarity 0.88)
• MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection (ICLR 2025 · similarity 0.86)
• LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction (AAAI 2025 · similarity 0.85)
• AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.85)
• Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.85)
• Vad-R1: Towards Video Anomaly Reasoning via Perception-to-Cognition Chain-of-Thought (NeurIPS 2025 · similarity 0.84)