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

TimeEvo. Failure-Driven Self-Evolution of a Time Series Agent. By Jie Yang and ten others, from the University of Illinois at Chicago, Visa Research, and three other institutions. Published on the twenty-third of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Time series agents answer analytical questions by calling external tools, and which tools an agent carries is decided by people before it ever runs. The authors identify two failures in that arrangement. The first they call Human-Agent Tool Misalignment. A library of twenty-one expert-curated tools helps on some tasks and hurts on others, and it drops anomaly accuracy under every backbone they test. The second they call Silent Harm. One round of generic self-revision changes one hundred and forty-seven answers and breaks fifty-six of them, while the final score moves by less than a point. The damage is real, and the headline number hides it.

Both failures, the paper argues, come from the same gap. Whether a tool helps is decided question by question, at runtime. But tools are supplied in advance, and judged by a single average.

TimeEvo closes that gap by letting the agent grow its own tool library out of its own mistakes. Rather than starting from what people expect to be useful, it starts from where the agent actually failed, and admits a new tool only when there is evidence that the tool helps.

The loop has four parts. First, the agent's diagnosed failures are clustered into capability gaps. Second, for each gap, TimeEvo plans a measurement. Third, it synthesizes evidence-only tools that fill those gaps. Fourth, the candidate library is admitted only by passing a paired admission gate. Nothing enters the library simply because someone expected it to help.

Experiments cover ten time series question answering tasks and three backbones. Starting from an empty library, TimeEvo improves accuracy on every task and on every backbone. There is also a transfer result. A library grown on a cheap model still gains when it is installed into stronger ones.

The paper's diagnosis and its remedy line up closely. Tools chosen in advance and judged by an average were the problem, so the remedy grows tools from observed failures and gates the candidate library before it is let in. And because a library grown on a cheap model still helps stronger ones, the growing can be done where it costs least.