Researchers propose CurateEvo, a failure-driven dynamic evolution framework designed to improve long-horizon decision making in large language model agents through iterative data curation. Unlike traditional methods that treat curation as a fixed preprocessing step, CurateEvo represents strategies as executable code and rewrites them using failed trajectories from a held-out development set.

  • The framework iteratively rewrites curation strategies to diagnose recurring failure modes, augmenting, filtering, or refining data accordingly.
  • It transforms raw corpora into supervised fine-tuning data, reinforcement learning data, and an inference-time memory bank at each epoch.
  • A cost-aware objective prunes redundant or low-utility training turns to improve efficiency.
  • Experiments on ACEBench-Agent, BFCL-V4, and τ^2-Bench show CurateEvo outperforms prior methods, improving average scores by 3.2 and 2.7 points respectively.

The authors note that CurateEvo is compatible with different post-training recipes and substantially reduces curation overhead.