TimeEvo: Failure-Driven Self-Evolution of Time Series Agents Outperforms Human-Curated Toolkits

Background and the Problem

/ Background & Pain Points The prevailing paradigm for constructing analytical agents involves human developers pre-populating an extensive static toolkit. However, empirical audits by researchers at UIC expose two severe systemic liabilities: Human-Agent Tool Misalignment, where 21 expert-curated analytical tools consistently degrade task performance on anomaly detection across all tested backbones; and Silent Harm, where standard self-reflection unhelpfully alters 147 answers, breaking 56 previously correct outputs with near-zero net benefit.

Architecture and How It Works

/ Architectural Highlights To resolve this structural mismatch, researchers introduced TimeEvo, a failure-driven self-evolving architecture:

1. Diagnostic Failure Clustering: Analyzes execution traces upon task failure, grouping underlying causes into concrete capability gaps rather than relying on superficial error messages;
2. Evidence-Only Tool Synthesis: Formulates specific measurement requirements for each capability void, synthesizing deterministic Python scripts that return objective numerical evidence without subjective prompt hallucinations;
3. Paired Admission Gate: Subjects candidate tools to counterfactual evaluation across historical error and success distributions, admitting tools only when they demonstrate verified net positive utility.

Benchmarks and Measured Results

/ Benchmark & Evaluation Extensive experiments across ten time-series QA benchmarks and three foundation model backbones demonstrate remarkable robustness:

• Outperforming Expert Libraries from Scratch: Starting from an entirely empty tool library, TimeEvo evolves specialized tools that surpass human-curated expert libraries across all ten benchmarks;
• Cross-Model Evolution Transfer: Tool libraries synthesized by a compact, inexpensive model transfer directly to frontier foundation models, delivering immediate accuracy gains without re-synthesis overhead;
• Eliminating Silent Harm: The paired admission gate cuts counterproductive answer modifications by 82.4% during reflection rounds.

Getting Started for Developers

/ Developer Practical Guide

• GitHub Repository: Accessible at https://github.com/Muyiiiii/TimeEvo;
• Theoretical Framework: Formally analyzed in arXiv preprint 2609.27277;
• Agent Design Takeaway: Rather than handcrafting extensive static tool inventories, developers should integrate failure-clustering and gated on-the-fly tool synthesis into production agent loops.