跳到正文
arXiv cs.AI· Shizhe Xiang, Ke An, Wenlong Yu, Yue Liu, Tingting Wu, Jian Luan, Pei Fu, Qilong Wang·· 1 天前AI 评分36

PTD-PO:面向多模态策略优化的特权提示蒸馏框架

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

AI 导读

论文提出 PTD-PO(Privileged Tutoring Distillation Policy Optimization),一种面向 LVLM 的 RLVR 训练框架,在不向学生策略暴露答案的前提下提供密集监督。

来源:arXiv cs.AI · arxiv.org