LLMs Detect Tabular Anomalies via In-Context Learning

Anomaly detection in tabular data presents significant challenges, particularly because anomalies often manifest as violations of complex cross-feature dependencies rather than simple deviations. Current anomaly detectors typically rely on geometric or reconstruction-based signals. While some prior approaches have used LLMs, they often involve fine-tuning the models with normal samples or generating synthetic anomalies. This research introduces LLM-Detector, a novel framework that harnesses the in-context learning abilities of Large Language Models for tabular anomaly detection. Instead of fine-tuning, LLM-Detector transforms normal training data into structured knowledge, including statistical summaries, causal dependencies, and distilled prototypes. This structured knowledge is then organized into a prompt, which guides the LLM to synthesize a scoring engine. The generated scoring engine evaluates statistical deviation, structural inconsistency, and density-based abnormality to compute an anomaly score for each test sample. Evaluated across 24 tabular datasets and compared against 15 state-of-the-art baselines, LLM-Detector consistently shows improved performance in both mixed-type and continuous-only settings. A significant advantage of this design is the elimination of LLM fine-tuning or neural network training, which substantially reduces computational costs and makes practical anomaly detection more feasible in real-world tabular systems.