SkillLens

Collection Notes

arXiv paper associated with Microsoft's SkillLens project page. It studies model-generated agent skills across the lifecycle from experience generation to skill extraction and skill consumption.

Key Source Claims To Verify During Ingest

• Language agents improve by reusing skills, described as structured procedural artifacts distilled from past experience.
• The paper studies whether model-generated domain-level skills work, what drives skill utility, and whether extraction can be improved.
• It reports that model-generated skills are beneficial on average but show non-trivial negative transfer.
• Skill utility is not reliably predicted by model scale or baseline task strength.

Relevance

• Fills missing coverage: empirical evaluation of model-generated skills and skill lifecycle.
• Security angle: negative transfer, unsafe skill extraction, cross-agent transfer, evaluation of learned procedural artifacts.

Ingest Priority

High. Pair with SkillOpt.