Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills
Targeted expert review cut teaching-token use by 61.7% while slightly improving utility over fixed-window review.
Socratic Skill Learning alternates expert guidance with practice, leaving the student agent in control of environment actions and skill writing. The authors test reuse in fresh solver sessions using SkillTransferBench, which contains 35 task families and 408 executable cases. On the main 155-case cohort, fixed-window teaching reached 0.801 utility, ahead of no-skill and two adapted distillation baselines. Focusing review on unresolved questions reached 0.805 utility and used 16.9% fewer total acquisition tokens.
HF Daily Papers' note
Socratic Skill Learning alternates expert guidance with practice, leaving the student agent in control of environment actions and skill writing. The authors test reuse in fresh solver sessions using SkillTransferBench, which contains 35 task families and 408 executable cases. On the main 155-case cohort, fixed-window teaching reached 0.801 utility, ahead of no-skill and two adapted distillation baselines. Focusing review on unresolved questions reached 0.805 utility and used 16.9% fewer total acquisition tokens.
HF Daily Papers' note
score 5