LLMs enhanced for time-series anomaly detection with frequency-domain evidence

Researchers have developed a new framework for time-series anomaly detection using Large Language Models (LLMs) that incorporates frequency-domain evidence alongside traditional time-domain data. This approach, called AnomLLM, utilizes the Fast Fourier Transform (FFT) to extract both global and local spectral information, which can reveal changes in temporal structure like shifted periodicity or localized fluctuations. Experiments demonstrated that this evidence-augmented method improves the performance of LLM-based time-series anomaly detection when tested with models such as InternVL2-LLaMA3-76B, Qwen2.5-VL-72B-Instruct, Gemini-2.5-Flash, and GPT-4o. AI

影响 This research could lead to more robust anomaly detection systems by enabling LLMs to better understand and identify complex temporal patterns.

排序理由 The cluster contains a research paper detailing a new methodology for time-series anomaly detection using LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

• AnomLLM
• fast Fourier transform
• Gemini 2.5-Flash
• GPT-4o
• InternVL2-LLaMA3-76B
• TSB-AD-U

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →