Reproduce “Rethinking Self-Evolving Agent Skills: Feedback Dynamics over Multiple Rounds” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuxuan Liu, Zhaochen Su, Yuhao Zhang, Jiahe Guo, Zhongwei Xie, Huihao Jing, Lingyun Xie, Qing Zong and 4 more
The authors study how feedback composition affects evolution of reusable agent skills, holding the executor, revision procedure, validation rule, and round budget fixed. Across 14 primary model–benchmark settings, validation selected an evolved skill in 11, and nine improved released-test performance. Parent-skill parallel attempts nearly matched evolution on SearchQA but lagged on SpreadsheetBench; the task-dependent result establishes no general feedback rule.
Key insights
- Skill evolution is characterized as a sparse, validation-filtered search process rather than a steady, monotonic improvement, with validated byte-distinct improvements occurring infrequently.
- Feedback conditions that include failed trajectories (Normal or Fail-only) are critical for successful skill evolution, whereas Success-only feedback was never selected as leading to an improved skill in the primary study.
- Persistent skill evolution can yield distinct and generalizable gains that, for complex procedural tasks, cannot be matched by merely increasing per-instance computational resources during inference (e.g., more attempts or sequential refinement).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.