arXiv · cs.AI · Harrison Zhao, Lei Huang, Zhiyuan Zhao, Fei Liu, Wen Shi, Zijian Jin · 5 小时前模型评分42
并非每条轨迹都需要蒸馏：正确性过滤的在线策略自蒸馏 C-OPSD
Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation
导读
研究团队提出正确性过滤的在线策略自蒸馏（C-OPSD，Correctness-Filtered On-Policy Self-Distillation），对每个问题只生成一条回答，并仅蒸馏部分错误回答。
分类置信度
中 · 待人工复核
- 只有单一媒体来源，需要人工复核。
来源级别：媒体与个人
事件类型：研究发布
证据摘录
- “Motivated by these measurements, we introduce correctness-filtered OPSD (C-OPSD), which generates one response per problem and distills only a subset of incorrect responses.”
- “In a measured micro-batch-size-four implementation, executed rows decrease by approximately 68%, and with retained rows packed into full micro-batches, training takes about 40% less time while rollout generation remains unchanged.”
来源：arXiv · cs.AI · arxiv.org