Intel / article

Rethinking Self-Evolving Agent Skills: Feedback Dynamics over Multiple Rounds

Source: arxiv.org
Author: Yuxuan Liu, Zhaochen Su, Yuhao Zhang, Jiahe Guo, Zhongwei Xie, Huihao Jing, Lingyun Xie, Qing Zong, Yauwai Yim, Zhixiong Zhang, Haoran Li, Yangqiu Song
Date: Published Aug 5, 2026

Context

Repeatedly revising an 's saved skills produced occasional improvements rather than steady progress in this study. The researchers kept the underlying model unchanged and varied whether revisions received successful examples, failed examples or both. Only 55 of 388 candidate revisions established distinct new best results on the validation tests used for selection. Study design and findings.

Selection favored revised skills in eleven of fourteen model-and- settings, and nine of those improved performance on the separate test evaluation. Every selected revision came from feedback that included failures, although failures alone did not consistently beat mixed feedback. The results favor testing and selecting revisions carefully over assuming that another revision round will improve an agent. Reported comparison.

Key takeaways · AI-distilled

Self-evolving skills are sparse search, not steady improvement: across 42 feedback runs only 55 of 388 candidate revisions established byte-distinct validation bests.
Failed trajectories are the fuel: all 11 settings where validation selected an evolved included failure feedback - success-only feedback never won a selection.
Validation and downstream test/robustness/transfer evaluations sometimes prefer different feedback views, so a validation win does not guarantee the deployed gain.
is not a substitute everywhere: oracle parallel sampling comes within 0.43 points of the evolved skill on SearchQA but stays 30.96 points behind on SpreadsheetBench; sequential refinement recovers neither.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
test-time compute — Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.

Key quotes

“Evolution is sparse: only 55 of 388 candidates establish byte-distinct validation bests.”

“All 11 selections come from feedback conditions that include failed trajectories, although the relative ranking of Normal and Fail-only varies across settings.”

“Overall, persistent skill self-evolution is better understood as sparse, validation-filtered search with model- and benchmark-dependent returns, rather than steady improvement from additional rounds.”

Read the source arxiv.org