Microsoft Research’s SkillOpt tunes skills in text and hits 52/52 on benchmarks

eyishazyer · x · 2026-07-24

A long post breaks down Microsoft Research’s SkillOpt, a method that optimizes a natural-language “skill document” instead of model weights.

The model stays frozen; the text instruction itself is what gets iteratively improved.
The loop has four main parts: rollout on tasks, reflection by a separate optimizer model, bounded text edits as a kind of learning rate, and validation on held-out tasks.
Failed edits are stored in a rejection buffer, and slower epoch-level updates help long-horizon skills evolve without blowing up context size.
The author says SkillOpt went 52/52 across six benchmarks and seven target models, with especially large gains on tool-heavy workloads like SpreadsheetBench and OfficeQA.
The practical appeal is that the final artifact is just a short Markdown file you can drop into the system prompt, with no extra inference-time calls.

More from coding & agent