TimeART: An Agentic Time Series Reasoning Framework via Tool Augmentation
ORIGINAL / TimeART: Towards Agentic Time Series Reasoning via Tool-Augmentation
TimeART proposes an agentic framework that integrates powerful tool-based analysis with LLM reasoning for time series question answering (TSQA). It also introduces a 100k expert trajectory corpus, TimeToolBench, and a four-stage training strategy to enhance tool-use and generalization of LLM-based time series reasoning models (TSRMs).
01 ABSTRACT
This work introduces TimeART, a framework combining strong out-of-the-box tools with LLM reasoning to serve as a fully agentic data scientist for time series question answering. The authors collect a 100k expert trajectory corpus, TimeToolBench, and devise a four-stage training strategy that enables TSRMs to learn from their own early experiences and self-reflections. Experiments show that an 8B TSRM trained on TimeToolBench and equipped with TimeART achieves consistent state-of-the-art performance across multiple TSQA tasks.
02 KEY FINDINGS
- TimeART fuses tool-based analysis with LLM reasoning to build a fully agentic framework for time series question answering.
- The authors construct TimeToolBench, a 100k expert trajectory corpus, to train LLM-based TSRMs for strategic tool use.
- A four-stage training strategy enhances TSRM generalization via early experiences and self-reflection; the 8B model achieves state-of-the-art results on multiple tasks.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG