[arXiv]score: 0.25

LLM-Generated Feature Pools Outperform Neural Models in Time Series Anomaly Detection

September 21, 2026

A statistical pipeline using LLM-generated feature pools achieves 0.529 VUS-PR on the TSB-AD-U benchmark, surpassing the best neural models (0.45). The method relies on prompting a multimodal LLM to create domain-specific candidate pools, which serves as the primary performance driver.

HOW THIS AFFECTS YOU

●

builderYou can improve time series anomaly detection by using LLMs to engineer feature pools rather than relying solely on deep learning architectures.

●

researcherThis highlights the importance of candidate pool selection over search optimization in statistical pipelines.

read original ↗arxiv.org

DAILY DIGEST

you don't check 9 sources — we do. one email every morning, read in 2 min. free. unsubscribe anytime. privacy

← back to feed