Teaching the Way, Not the Answer: Privileged Tutoring Distillation for Multimodal Policy Optimization

In three linesPTD-PO, a policy distillation framework for optimizing vision-language models (2B-8B parameters) via reinforcement learning. Provides dense token-level supervision without exposing answers, using structured hints (spatial attention, reasoning steps) and Top-K Jensen-Shannon divergence to stabilize training.

Summary generated by Claude — human-verified