TimeEvo: Failure-Driven Self-Evolution of a Time Series Agent
TimeEvo builds a time-series agent tool library from diagnosed failures and raises accuracy on ten QA tasks.
Time series agents usually receive a fixed tool library chosen before runtime. A library of 21 expert tools hurt anomaly accuracy on every backbone tested, and one generic self-revision changed 147 answers and broke 56 while the overall score moved by less than a point. TimeEvo clusters diagnosed failures into capability gaps, synthesizes evidence-only tools, and admits them only through a paired gate. Starting from an empty library, it improved accuracy on all ten time-series QA tasks and three backbones, and a library grown on a cheap model still helped stronger ones.
- A fixed 21-tool library lowered anomaly accuracy on every backbone tested.
- One generic self-revision changed 147 answers and broke 56 of them.
- TimeEvo clusters failures, synthesizes tools, and admits them via a paired gate.
- Starting empty, it improved all ten tasks and three backbones, and transferred upward.
From the original article
1 min read at huggingface.co
Time series agents answer analytical questions by calling external tools, and which tools they carry is decided by people before the agent runs. However, we 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 we test.
By Jie Yang, Yan Zheng, Jiarui Sun · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.