D-Skill: Distilling teacher model knowledge into natural-language skills
What happened
Shuo Yuan and colleagues propose D-Skill, a framework that distills the behavioral differences in paired trajectories of a teacher model and a student model into interpretable natural-language skill documents. At inference time, the student model uses these documents directly, with no weight updates required. The paper reports that the method closes 75%–95% of the performance gap between teacher and student on SpreadsheetBench-Verified, ALFWorld, and chess puzzles, and that within the same model family the student's per-task inference cost is 5.7–8.3× lower than the teacher's. These are the paper's self-reported results. According to the paper, skills are generated entirely offline, using only fixed trajectories, with no labels, no additional environment interaction, and no weight updates after rollout collection.
AI-written from the coverage · updated 2026-10-22
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Board & Classic Game AIPickD-Skill: Distilling teacher knowledge through agent skills
D-Skill distills the paired trajectory differences between a teacher and a student model into interpretable natural-language skill documents. The student uses these documents at inference time without updating its weights. The paper reports that the method closes 75%–95% of the performance gap between teacher and student on SpreadsheetBench-Verified, ALFWorld, and chess puzzles. Within the same model family, the student's per-task inference cost is 5.7x–8.3x lower than the teacher's.