Self-Improvement Anomaly Detection via Large Language Model for Unsupervised Zero-shot Anomaly Detection

ICLR 2026 Conference Submission15608 Authors

Keywords: Unsupervised zero-shot learning, anomaly detection, LLM

Abstract: Zero-shot anomaly detection has emerged to overcome the limitations of conventional methods, which depend on learning the distribution of normal data and struggle to generalize to unseen class. However, existing zero-shot methods often rely on anomalous data during training, which is impractical in real-world settings where such data are scarce or unavailable. To address these limitations, we propose a novel unsupervised zero-shot anomaly detection framework, self-improvement anomaly detection with large language model that requires no anomalous data during training. It leverages self-improvement large language model-based architecture that refines textual responses grounded in input images. To support semantic interpretation, we design stage prompts that guide the large language model using visual features spanning from local patterns to global semantics. Our approach not only produces interpretable anomaly maps but also enhances semantic understanding of normality, offering a new direction for zero-shot anomaly detection under realistic anomaly-free constraints. Extensive experiments on nine real-world datasets from both industrial and medical domains demonstrate the effectiveness of our approach. Our self-improvement anomaly detection with large language model outperforms state-of-the-art methods across various unsupervised zero-shot anomaly detection benchmarks, validating its robustness and generalizability across diverse datasets.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 15608