Do As I Do, Not As I Say: On-Policy Distillation Inherits the Teacher in Vision-Language Models
This work compares on-policy distillation with filtered supervised fine-tuning in vision-language models. It examines how teacher supervision on student-generated trajectories affects both answer accuracy and the ability to complete a response.
The results show that on-policy supervision can transfer persistent teacher failure modes as well as knowledge, making teacher reliability on student-visited states central to successful distillation.