Study finds self-evolving AI agents improve rarely—only 55 of 388 skill candidates

Researchers from HKUST and collaborating labs, led by Yuxuan Liu, present a controlled evaluation of self-evolving agent skill systems—frameworks where an AI agent observes its own execution feedback and writes persistent skill updates without retraining the underlying model. The study, posted on arXiv (2608.02636), systematically varied only the feedback shown to the optimizer across 42 feedback runs on 14 supported model-benchmark settings. They tested three feedback conditions: successes and failures (Normal), failures only, and successes only. The goal was to determine when further evolution actually helps.

Key findings show that evolution is strikingly sparse: only 55 of 388 candidate skills produced byte-distinct validation bests. Validation-based selection chose an evolved skill in 11 of 14 settings, and nine of those improved released-test performance. Crucially, all 11 selections came from conditions that included failed trajectories, though the relative ranking of Normal versus Fail-only varied by setting. Test-time scaling controls using GPT-5.5 showed that oracle Parallel Sampling could nearly match the evolved SearchQA skill (within 0.43 points) but fell 30.96 points behind on SpreadsheetBench, while Sequential Refinement recovered neither. The authors conclude that persistent skill self-evolution is better understood as sparse, validation-filtered search with model- and benchmark-dependent returns, not steady improvement from more rounds.