Socratic Skill Learning interleaves teaching with agent practice
Original title: Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills
Fixed-Window SSL scored 0.801 utility on 155 cases, above no-Skill, Trace2Skill, and SKILL-KD baselines.
The paper introduces Socratic Skill Learning, which interleaves an Educator's guidance with a Student that remains the sole environment operator and Skill author. SkillTransferBench has 35 task families and 408 executable cases for reuse beyond the learning task. On the main 22-family, 155-case cohort, Fixed-Window SSL reached 0.801 utility. Issue-Guided SSL reached 0.805 utility with 61.7% fewer Educator tokens than Fixed-Window SSL.
- Without a Skill, utility was 0.723 on the main cohort.
- Adapted Trace2Skill scored 0.761 and SKILL-KD 0.751.
- Issue-Guided SSL used 16.9% fewer total acquisition tokens.
- Interleaved guidance beat expert authoring across three model pairs.
From the original article
1 min read at huggingface.co
Agent Skills improve LLM agents by making task knowledge reusable across sessions. Learning transferable Skills requires evidence about a procedure and its conditions of use, which a successful task attempt can leave untested. We introduce Socratic Skill Learning (SSL), which interleaves expert teaching with Student practice.
By Letian Zhu, Han Li, Yifan Yao · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.