Tag
The paper analyzes on-policy distillation, revealing it primarily suppresses low-probability tokens rather than relying on teacher guidance, and introduces OPSA, a supervision-free method that significantly enhances reasoning performance.