Skill Distillation

Compress agentic workflows and decision-making capabilities from large, expensive models (or human demonstrations) into smaller, faster models for efficient deployment. Skill distillation transfers agentic capabilities from a capable but expensive teacher (large LM or human) to a smaller student model. The student learns to imitate the teacher action sequences, tool-use decisions, and planning trajectories. The training data is generated by the teacher executing tasks and logging its full decision process. In addition to action imitation, the student learns the teacher intent — why it chose specific actions — through intermediate reasoning distillation. This enables deployable agents at a fraction of the cost.