Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ChatAD: Reasoning-Enhanced Time-Series Anomaly Detection with Multi-Turn Instruction Evolution

Hui Sun, Chang Xu, Haonan Xie, Hao Li, Yuhao Huang, Chuheng Zhang, Ming Jin, Xiaoguang Liu, Gang Wang, Jiang Bian

TL;DR

ChatAD introduces a reasoning-enhanced approach to time-series anomaly detection by leveraging a multi-agent Time Series Evolution framework (TSEvol) to synthesize rich, multi-turn reasoning data (TSEData-20K) for training a family of ChatAD models. It couples this data with Time Series Kahneman-Tversky Optimization (TKTO) to improve cross-task generalization across classification, forecasting, and imputation, and validates performance with LLADBench across seven datasets. The results show substantial gains in single-turn AD accuracy and F1, dramatic reductions in false positives, and strong multi-turn dialogue capabilities, underscoring the value of explicit reasoning and interactive data evolution for TS tasks. Despite strong performance, the work notes limitations of current transformer-based models in fully capturing numeric patterns and points to future directions in TS-tailored architectures and multi-modal extensions to broaden applicability.

Abstract

LLM-driven Anomaly Detection (AD) helps enhance the understanding and explanatory abilities of anomalous behaviors in Time Series (TS). Existing methods face challenges of inadequate reasoning ability, deficient multi-turn dialogue capability, and narrow generalization. To this end, we 1) propose a multi-agent-based TS Evolution algorithm named TSEvol. On top of it, we 2) introduce the AD reasoning and multi-turn dialogue Dataset TSEData-20K and contribute the Chatbot family for AD, including ChatAD-Llama3-8B, Qwen2.5-7B, and Mistral-7B. Furthermore, 3) we propose the TS Kahneman-Tversky Optimization (TKTO) to enhance ChatAD's cross-task generalization capability. Lastly, 4) we propose a LLM-driven Learning-based AD Benchmark LLADBench to evaluate the performance of ChatAD and nine baselines across seven datasets and tasks. Our three ChatAD models achieve substantial gains, up to 34.50% in accuracy, 34.71% in F1, and a 37.42% reduction in false positives. Besides, via KTKO, our optimized ChatAD achieves competitive performance in reasoning and cross-task generalization on classification, forecasting, and imputation.

ChatAD: Reasoning-Enhanced Time-Series Anomaly Detection with Multi-Turn Instruction Evolution

TL;DR

ChatAD introduces a reasoning-enhanced approach to time-series anomaly detection by leveraging a multi-agent Time Series Evolution framework (TSEvol) to synthesize rich, multi-turn reasoning data (TSEData-20K) for training a family of ChatAD models. It couples this data with Time Series Kahneman-Tversky Optimization (TKTO) to improve cross-task generalization across classification, forecasting, and imputation, and validates performance with LLADBench across seven datasets. The results show substantial gains in single-turn AD accuracy and F1, dramatic reductions in false positives, and strong multi-turn dialogue capabilities, underscoring the value of explicit reasoning and interactive data evolution for TS tasks. Despite strong performance, the work notes limitations of current transformer-based models in fully capturing numeric patterns and points to future directions in TS-tailored architectures and multi-modal extensions to broaden applicability.

Abstract

LLM-driven Anomaly Detection (AD) helps enhance the understanding and explanatory abilities of anomalous behaviors in Time Series (TS). Existing methods face challenges of inadequate reasoning ability, deficient multi-turn dialogue capability, and narrow generalization. To this end, we 1) propose a multi-agent-based TS Evolution algorithm named TSEvol. On top of it, we 2) introduce the AD reasoning and multi-turn dialogue Dataset TSEData-20K and contribute the Chatbot family for AD, including ChatAD-Llama3-8B, Qwen2.5-7B, and Mistral-7B. Furthermore, 3) we propose the TS Kahneman-Tversky Optimization (TKTO) to enhance ChatAD's cross-task generalization capability. Lastly, 4) we propose a LLM-driven Learning-based AD Benchmark LLADBench to evaluate the performance of ChatAD and nine baselines across seven datasets and tasks. Our three ChatAD models achieve substantial gains, up to 34.50% in accuracy, 34.71% in F1, and a 37.42% reduction in false positives. Besides, via KTKO, our optimized ChatAD achieves competitive performance in reasoning and cross-task generalization on classification, forecasting, and imputation.
Paper Structure (62 sections, 15 equations, 5 figures, 18 tables, 1 algorithm)

This paper contains 62 sections, 15 equations, 5 figures, 18 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem & Tasks Definition
  3. Related Work
  4. LLM-driven Anomaly Detection
  5. Instruction Data Evolution (IDE)
  6. Reinforcement Learning for LLM
  7. Methods
  8. Time Series Instruction Data Evolution
  9. Instruction Evolution Pool (IEP)
  10. Cognitive Reasoning Layer (CRL)
  11. Interactive Feedback-driven Layer
  12. ChatAD Model Training
  13. SFT for Anomaly Detection Task
  14. TKTO for Enhanced Generalization
  15. LLM-driven LAD Benchmark
  16. ...and 47 more sections

Figures (5)

  • Figure 1: Comparison of modeling paradigms for time series. (a) Traditional NN-based methods. (b) Enhanced paradigm using MTSLLM as an universal generator.
  • Figure 2: Pipeline of TSEvol (Only the evolution of $A_1$ and $Q_2$ are shown, with respect to the Interactive Feedback-driven Layer). Here, $H$ and $B$ denote the inputted history TS data, and short description of the TS backgrounds, respectively. The arrow $\rightarrow$ denotes information flow, with only the solid elements retained as the final result.
  • Figure 3: The model training and benchmark pipeline of time series data anomaly detection (ChatAD) model.
  • Figure 4: A real-world AD demo via using ChatAD.
  • Figure 5: Real-world example from the open-ended response question answering (OERQA) test dataset