GRPO-Anomaly: Reinforcement Fine-Tuning Vision--Language Model for Industrial Defects Detection and Reasoning

Abstract

Industrial anomaly detection (IAD) requires reliable identification and precise localization of subtle defects, yet most existing methods depend on manually tuned decision thresholds and large collections of defect-free samples, limiting scalability in real-world production. To address these constraints, we present group relative policy optimization (GRPO)-Anomaly, a threshold-free vision–language framework tailored for industrial inspection. Here, “threshold-free” denotes the removal of manual score-threshold calibration at deployment rather than the absence of any implicit decision boundary; the final decision is produced directly by the language model, which still embodies a learned boundary. The system integrates a lightweight detector that generates pixel-level anomaly maps with a large vision–language model (LVLM) capable of joint reasoning and localization. The anomaly maps are encoded as spatial prompts and fed back into the model, forming a closed-loop mechanism that aligns low-level visual evidence with high-level semantic judgment. Furthermore, we introduce a reinforcement alignment strategy based on GRPO, which enforces structured output formats and improves decision reliability without extensive parameter updates. GRPO-Anomaly enables interpretable inspection, supports interactive refinement, and adapts to novel product categories using only a few normal exemplars. Extensive evaluation on standard benchmarks demonstrates competitive detection and localization performance, strong cross-dataset generalization, and substantial reduction of operational sensitivity to threshold selection. These results highlight the potential of GRPO-Anomaly to advance fully automated and scalable industrial quality control.