This page may contain stale information. Last updated: 2026-06-12
Definition
Text-space optimization applies deep-learning-style training discipline to natural language documents — treating skill/prompt text as trainable state with bounded edits, validation gates, and rejected-edit memory — without modifying neural network weights.
Key Points
- skillopt pioneer: separate optimizer model proposes add/delete/replace edits
- Edits accepted only when held-out validation strictly improves
- Textual learning-rate budget controls per-step change magnitude
- Contrasts with TextGrad, GEPA, EvoSkill, and ungated self-revision
Results vs Baselines
SkillOpt beats human-crafted, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills by +5.4 points average across 52 evaluation cells.