Global Logic and Local Search: Dual-Stream Multimodal In-Context Learning for Verifiable Industrial Anomaly Detection

2026-07-04

AI Summary

This work addresses key challenges in industrial anomaly detection, including subtle defects, limited input resolution, misalignment between textual specifications and visual evidence, and scarcity of defect samples during early deployment. The authors propose a training-free, dual-stream multimodal in-context learning framework that, for the first time, integrates Monte Carlo Tree Search (MCTS) to optimize local image patch selection under a fixed computational budget. By leveraging SAM³ to extract verifiable visual facts and combining structured specifications with normal samples, the method constructs a part-aware vision-logic graph that unifies global logical reasoning with localized evidence search. Evaluated on benchmarks such as MMAD-QA, the approach significantly outperforms existing baselines while providing fully traceable diagnostic pathways grounded in explicit visual evidence.

Abstract

Large Multimodal Models (LMMs) show strong few-shot generalization, but industrial anomaly detection remains difficult because defects are small, input resolution is limited, and textual standards are not always grounded in visual evidence. Recent optimization-based methods improve alignment through fine-tuning, but they often require many defective samples, which are unavailable in early deployment. We present Global Logic and Local Search (GLLS), a training-free framework for reference-guided multimodal in-context verification. GLLS uses a Part-Aware Visual-Logical Atlas to organize normal references and structured specifications in the inference context. It combines a Global & Logic Stream, where SAM 3 extracts partially checkable visual facts, with a Fine-Grained & Actions Stream, where MCTS selects local evidence crops under a fixed budget. Experiments on MMAD-QA and additional anomaly detection datasets show consistent gains over matched and general-purpose baselines, while keeping the final diagnostic decision traceable to explicit visual evidence throughout the inspection trace.

Problem

• industrial anomaly detection
• multimodal in-context learning
• visual-textual grounding
• few-shot generalization
• defect detection

Innovation

• in-context learning
• multimodal reasoning
• anomaly detection
• visual-logical atlas
• Monte Carlo Tree Search