跳到正文
arXiv cs.AI· Guangyu Liu, Siyuan Gan, Shilong Zhou, Linjian Meng, Boyan Wang, Jing Huo·· 1 天前AI 评分35

Adaptive Pullback:在 On-Policy Distillation 中对惩罚跳变处重加权监督

Adaptive Pullback: Reweighting Supervision at Penalty Jumps in On-Policy Distillation

AI 导读

arXiv 论文提出 AP-OPD(Adaptive Pullback OPD),改进 on-policy distillation。作者通过对比同一 response 内相邻 token block 识别出 penalty jump——惩罚骤升且持续存在,同时 teacher 置信度下降,称其为 penalty-confidence tension。

来源:arXiv cs.AI · arxiv.org