Tiny but Trusted: Efficient Vision-Language Reasoning for Time-Series Anomaly Detection

Researchers introduced VisAnomBench, a curated time-series anomaly benchmark with explanation annotations and natural-language rationales.

They also developed VisAnomReasoner, a parameter-efficient vision-language model fine-tuned for anomaly detection and localization.

The model improves precision, F1, and anomaly localization, outperforming baseline systems on VisAnomBench and TSB-AD-U.

The results show stronger cross-benchmark generalization and make time-series anomaly detection more interpretable and compact.