SkillOpt: A Systematic Text-Space Optimizer for Self-Evolving Agent Skills

Forum topic · 小凯 · 2026-05-26

Summary

SkillOpt is a systematic, controllable text-space optimizer for agent skills, treating skills as the external state of a frozen agent and training them with the same reproducibility discipline as weight-space optimization. A separate optimizer model converts scored rollouts into bounded add/delete/replace edits on a single skill document, and an edit is accepted only when it strictly improves a held-out validation score. A textual learning-rate budget, a rejected-edit buffer, and epoch-wise slow/meta updates stabilize skill training while adding no extra inference-time model calls at deployment. Across six benchmarks, seven target models, and three execution frameworks (direct chat, Codex, Claude Code), SkillOpt is best or tied-best in all 52 evaluation units, outperforming human-written skills, one-shot LLM generation, Trace2Skill, TextGrad, GEPA, and EvoSkill. On GPT-5.5 it raises no-skill average accuracy by +23.5 points (direct chat), +24.8 (Codex agent loop), and +19.1 (Claude Code). Transfer experiments show optimized skills retain value across model scales, between Codex and Claude Code environments, and on adjacent math benchmarks without further optimization. Paper: arXiv 2505.21451.

Overview

Research area: NLP
Authors: Yifan Yang, Ziyang Gong, Weiquan Huang
Published: 2026-05-26
arXiv: 2505.21451

Abstract

Agent skills today are hand-crafted, generated one-shot, or evolved through loosely controlled self-revision. None of these approaches behaves like a deep-learning optimizer for the skill, and none reliably improves over its starting point under feedback. The authors argue that a skill should instead be trained as the external state of a frozen agent, following the same discipline that makes weight-space optimization reproducible.

SkillOpt is, to the authors' knowledge, the first systematic controllable text-space optimizer for agent skills:

A separate optimizer model turns scored rollouts into bounded add/delete/replace edits on a single skill document.
An edit is accepted only when it strictly improves a held-out validation score.
A textual learning-rate budget, a rejected-edit buffer, and epoch-wise slow/meta updates keep skill training stable.
Deployment adds no extra inference-time model calls.

Results

Evaluated on 6 benchmarks, 7 target models, and 3 execution frameworks (direct chat, Codex, Claude Code).
Best or tied-best in all 52 evaluation units, beating human-written skills, one-shot LLM generation, Trace2Skill, TextGrad, GEPA, and EvoSkill.
On GPT-5.5, SkillOpt improves no-skill average accuracy by +23.5 points (direct chat), +24.8 (Codex agent loop), and +19.1 (Claude Code).
Transfer experiments show optimized skills remain valuable across model scales, across Codex/Claude Code execution environments, and on adjacent math benchmarks without further optimization.