Multi-Agent LLM Framework SAGE for Time Series Anomaly Detection

A novel multi-agent system named SAGE (Specialized Analyzer Group for Expert-like Detection) has been introduced for identifying anomalies in univariate time series data. In contrast to current methods that utilize a single, general-purpose large language model to deduce anomaly indices or intervals, SAGE breaks down the analysis into four distinct Analyzers: point, structural, seasonal, and pattern anomalies. Each Analyzer employs specific numerical tools and diagnostic visualizations to produce evidence. An evidence-based Detector compiles this information into confidence-scored anomaly records, complete with intervals and candidate types. Subsequently, a Supervisor transforms these structured records into diagnostic reports for analysts. Additionally, the framework creates synthetic in-context exams to boost performance, aiming to enhance controllability, interpretability, and reliability for complex anomaly patterns.