跳到正文
HuggingFace Daily Papers·· 5 天前AI 评分34

DiffGate:用难度门控的教师指导改进在线策略蒸馏

DiffGate: Difficulty-Gated Teacher Guidance for On-Policy Distillation

AI 导读

DiffGate 提出一种结果门控目标,将 GRPO 与选择性、有界的教师指导结合:教师监督只施加于失败轨迹,并按组难度缩放、平滑限幅,避免极端师生差异主导优化。

来源:HuggingFace Daily Papers · arxiv.org