Teaching the Way, Not the Answer: Privileged Tutoring Distillation for Multimodal Policy Optimization
The PTD-PO framework uses answer-free spatial and textual hints to provide dense, token-level guidance for failed rollouts in Reinforcement Learning with Verifiable Rewards. It conditions a self-teacher on these hints while keeping the student under the original context. This method outperforms existing methods by focusing supervision only on incorrect responses, leading to better reasoning recovery and higher policy entropy.