Microsoft SkillOpt: Text-Space Optimizer Trains Markdown Skills for Frozen LLMs — 52/52 Benchmark Wins

arXiv / Microsoft Research·medium signal

Microsoft Research released SkillOpt (arXiv 2605.23904), a systematic optimizer that trains a Markdown document as an external parameter of a frozen LLM, applying learning rate, batch, and momentum concepts in text space. Across 6 benchmarks, 7 models, and 3 execution harnesses (direct chat, Codex, Claude Code), SkillOpt wins or ties on all 52 evaluated cells. On GPT-5.5: +23.5 points in direct chat, +24.8 in Codex, +19.1 in Claude Code. MIT licensed, open source. For builders: if you maintain agent skill files (like CLAUDE.md), this is a systematic way to optimize them.