SKILL-KD: Contrastive Skill Distillation for Agentic Trajectories
July 31, 2026
SKILL-KD is a framework that distills actionable discrepancies between teacher and student agent trajectories into textual skill patches. Instead of summarizing entire demonstrations, it focuses on correcting specific operational failures to improve student agent performance.
HOW THIS AFFECTS YOU
●
builderYou can use this method to improve the performance of smaller agent models by distilling specific corrective behaviors.