2026 workshop program
Anomaly Detection in ASAS-SN Using Visual Embeddings and Agent-Driven Active Learning
The work
Abstract
Unusual light-curve morphologies can point to rare physical configurations, extreme evolutionary stages, or even entirely new astrophysical phenomena. However, automated searches for anomalies are often dominated by artifacts, and distinguishing genuine anomalies from false positives requires manual inspection that does not scale to large-scale photometric surveys. We present an automated framework that outsources expert vetting to large language model (LLM) agents. The framework follows an active learning workflow, in which LLM agents iteratively review and classify top-ranked anomaly candidates. The initial ranking is produced by an ensemble of isolation forests trained on DINOv2-giant visual embeddings of phase-folded light curve images, eliminating the need for hand-engineered features. After each iteration, the agents’ labels are propagated through the embedding space, prioritizing most promising anomaly candidates for subsequent rounds. Applied to ~400,000 light curves of variable stars from the ASAS-SN Sky Patrol V2 survey and following 50 iterations of anomaly vetting by Gemini 3 Flash LLM agents, our pipeline identified 24 high-confidence anomalies, including several eclipsing binaries with extremely deep primary minima and an unusually high-amplitude type II Cepheid. Our results demonstrate that visual foundation models paired with agent-driven active learning offer an efficient and scalable approach to systematic anomaly discovery in upcoming photometric surveys.