arXiv 2610.17860
Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills
AI32 upvotes
Authors: Letian Zhu, Han Li, Yifan Yao, Xinping Lei, Xinyu Che, Jiahao Wang, Dianyi Wang, Yucheng Fu +4
Overview
The authors propose Socratic Skill Learning (SSL), in which an Educator guides investigation while a Student practices, tests, and authors the Skill. They evaluate on SkillTransferBench, a benchmark of 35 task families and 408 executable cases with held-out evaluation in fresh solver sessions. A variant, Issue-Guided SSL, tracks unresolved questions across Skill revisions to focus review.
Introduction
Successful task attempts can leave a procedure's conditions of use untested, so the resulting Skills may not transfer to new tasks. Learning transferable Skills therefore needs evidence about both the procedure and when it applies.
Results
Fixed-Window SSL reaches 0.801 utility on the main 155-case cohort, versus 0.723 without a Skill, 0.761 for adapted Trace2Skill, and 0.751 for adapted SKILL-KD. Issue-Guided SSL matches this with 0.805 utility while using 61.7% fewer Educator tokens and 16.9% fewer total acquisition tokens.
Conclusion
Interleaving expert guidance with student practice, and focusing review on open questions, yields more transferable agent Skills at lower teaching cost.