arXiv · NLP· Runsheng Chen (Technical University of Munich, Germany), Yihong Liu (LMU Munich, Germany, Munich Center for Machine Learning, Germany), Raoyuan Zhao (LMU Munich, Germany, Munich Center for Machine Learning, Germany), Michael A. Hedderich (LMU Munich, Germany, Munich Center for Machine Learning, Germany), Hinrich Sch\"utze (LMU Munich, Germany, Munich Center for Machine Learning, Germany), Matthias Grabmair (Technical University of Munich, Germany)·· 4 小时前AI 评分32
Bridged Self-Distillation (BSD) 统一实现跨场景策略蒸馏
Bridged Self-Distillation for On-Policy Learning
AI 导读
提出Bridged Self-Distillation (BSD)框架，通过外部教师生成推理轨迹作为特权上下文，统一解决强教师模型向小模型蒸馏时的不匹配问题。 外部教师仅提供文本生成推理轨迹，学生模型在无该上下文的rollout中利用自教师在其自身token空间内打分进行指导，结合验证器信号更新不成功rollout，适用于共享、跨tokenizer及闭源API场景，优于各场景基线。
来源：arXiv · NLP · arxiv.org