Papers09:15Chinese content诊断与缓解On-Policy Self-Distillation中的Thinking Collapse这篇论文把OPSD训练时模型思维崩溃的原因和解决方法都讲清楚了,AD-OPSD在数学题上能提4个点,做推理优化的值得看看。#OPSD#AD-OPSD#Thinking Collapse#自蒸馏aarXiv cs.LG@Keqin Peng et al. (5)SourceRead laterReadWorth followingUsefulFollow OPSD