Microsoft SkillOpt Transfers Agent Skills Across Models
Microsoft researchers have introduced SkillOpt, a text-space optimizer that allows AI agents to transfer learned skills across different model scales and software harnesses without retraining.
Microsoft, alongside researchers from Shanghai Jiao Tong, Tongji, and Fudan Universities, has developed SkillOpt. This text-space optimizer refines a single natural-language skill document, saved as a best_skill.md file, while keeping the target model frozen. An optimizer model reviews scored rollouts and suggests edits, which are accepted only if they strictly improve performance on a held-out selection split. The final portable files range from 379 to 1,995 tokens, with a median of 920 tokens, built from just one to four accepted edits.
The system demonstrates that optimized skills can transfer across different model sizes within the same family. When training on GPT-5.4 and deploying on smaller variants, the SpreadsheetBench benchmark on GPT-5.4-mini retained 82 percent of its in-domain gain, moving from a 36.1 baseline to a 45.5 transferred score, compared to a 47.5 direct score. On GPT-5.4-nano, SpreadsheetBench retained 16 percent, moving from a 23.5 baseline to 26.5 transferred. For LiveMath, GPT-5.4-mini went from a 14.7 baseline to 19.2 transferred, while GPT-5.4-nano actually outperformed its direct score of 27.2 by scoring 28.8 against a 23.2 baseline.
The most significant results occurred in cross-harness transfers using GPT-5.5. On SpreadsheetBench, transferring a skill from Codex to Claude Code raised the score from a 22.1 baseline to 81.8, surpassing the 80.4 score achieved by training Claude Code directly. Conversely, transferring from Claude Code to Codex yielded a 71.1 score against an 85.0 direct score and a 27.5 baseline. For LiveMath, Claude Code to Codex retained 30 percent of the gain, scoring 48.0 against a 35.2 baseline, while Codex to Claude Code retained only 10 percent, scoring 42.4 against a 40.8 baseline.
In cross-benchmark tests from OlympiadBench to Omni-MATH, GPT-5.4 scored 60.3, GPT-5.4-mini scored 36.6, and GPT-5.4-nano scored 40.1, all exceeding their respective baselines of 56.6, 34.8, and 38.8. For practitioners, this means training costs, which range from 0.6 million tokens per point on SpreadsheetBench to 46.4 million on DocVQA, are paid only once offline. Because the optimizer adds zero inference-time overhead, developers can optimize skills where computing is cheapest and deploy them elsewhere, while auditing changes easily via traceable JSON files.
This is our own summary of reporting by MarkTechPost
