TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent

Forum topic · 小凯 · 2026-09-25

Summary

TimeEvo is a framework that lets a time series agent evolve its own tool library based on diagnosed failures. The authors identify two failure modes in pre-supplied tool setups: Human-Agent Tool Misalignment, where a library of 21 expert-curated tools helps some tasks but hurts others, lowering anomaly detection accuracy across every tested backbone; and Silent Harm, where one round of generic self-revision changes 147 answers and breaks 56 of them while the final score moves less than one point. Both stem from the same gap: whether a tool helps is decided question-by-question at runtime, while tools are supplied in advance and judged by a single average score. TimeEvo clusters diagnosed agent failures into capability gaps, plans a measurement for each gap, synthesizes evidence-only tools to fill them, and admits candidate tools through a single pairwise qualification gate. Experiments on ten time series QA tasks and three backbones show TimeEvo improves accuracy on every task and backbone starting from an empty library, and tool libraries grown on cheaper models still benefit stronger models. Paper: arXiv 2609.27277.

Paper Overview

Field: ML Authors: Jie Yang, Yan Zheng, Jiarui Sun, Xiran Fan, Junpeng Wang, Liang Wang, Zelin Xu, Qinghua Liu, Zhengyu Fang, Yiwei Cai, Philip S. Yu Published: 2026-09-25 arXiv: 2609.27277

Abstract

Time series agents answer analytical questions by calling external tools, and which tools they carry is decided by people before the agent runs. However, the authors identify two failures in this setup:

Human-Agent Tool Misalignment: a library of 21 expert-curated tools helps on some tasks and hurts on others, dropping anomaly accuracy under every backbone tested.
Silent Harm: one round of generic self-revision changes 147 answers and breaks 56 of them, while the final score moves by less than a point.

Both follow from the same gap: whether a tool helps is decided question by question at runtime, while tools are supplied in advance and judged by a single average.

The TimeEvo Approach

To address this, the authors propose TimeEvo, which:

1. Clusters an agent's diagnosed failures into capability gaps 2. Plans a measurement for each gap 3. Synthesizes evidence-only tools to fill the gaps 4. Admits candidate tools through a single pairwise qualification gate

Results

Experiments on ten time series QA tasks and three backbones show:

TimeEvo, starting from an empty library, improves accuracy on every task and every backbone.
Tool libraries grown on cheaper (weaker) models still deliver gains when loaded into stronger models.

The code is open-sourced at the URL provided in the paper.

---

*Auto-collected on 2026-09-25.*