Weak-to-Strong Generalization via Direct On-Policy Distillation

This method explores weak-to-strong generalization using direct on-policy distillation to improve model performance. It addresses the challenge of supervising stronger models with outputs from weaker supervisory models.