arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2605.30833cs.CLcs.AI
收藏

你的老师在这里帮不了你:对抗在线策略蒸馏中的监督保真度衰减

Your Teacher Can't Help You Here: Combating Supervision Fidelity Decay in On-Policy Distillation

University of Chinese Academy of Sciences(中国科学院大学);Chinese Information Processing Laboratory(中文信息处理实验室);Institute of Software, Chinese Academy of Sciences(中国科学院软件研究所);University of Chinese Academy of Sciences, Beijing, China(中国科学院大学,北京,中国)

Yanjiang Liu, Jie Lou, Xinyan Guan, Yuqiu Ji, Hongyu Lin, Ben He, Xianpei Han, Le Sun, Xing Yu, Yaojie Lu

AI总结:

针对在线策略蒸馏中监督保真度衰减问题,提出前瞻组奖励方法,通过评估学生候选词在后续步骤中诱导的教师置信度并分配组归一化奖励,结合熵触发树注意力机制,显著提升长链推理性能。

摘要:

在线策略蒸馏通过使用来自教师的 token 级反馈,在学生模型自身生成的轨迹上训练学生模型来传递推理能力。然而,我们识别出一个关键瓶颈,即 extbf{监督保真度衰减(SFD)}:随着学生生成的前缀变长,教师的下一个 token 分布变得不那么自信和更具区分性。因此,反向 KL 蒸馏中依赖教师的纠正信号减弱,导致学生漂移在长推理链中累积。为了缓解 SFD,我们引入了 extbf{前瞻组奖励(\ours{})}。基于下一步教师置信度反映了未来反向 KL 监督的区分强度这一见解,\ours{} 通过学生在后续步骤中诱导的教师置信度来评估学生的 top-K 候选 token,并分配组归一化奖励。为了保持计算效率,我们进一步设计了一种熵触发的树注意力机制。在六个数学和代码基准测试中,\ours{} 在 7B 学生模型上比 OPD 提高了 mean@8 达 extbf{2.57} 个点,在长生成任务中增益更大,在 AIME-26 上达到 + extbf{4.92} 个点(39k token)。

英文摘要:

On-policy distillation transfers reasoning capabilities by training a student model on its own generated trajectories using token-level feedback from a teacher. However, we identify a critical bottleneck, \textbf{Supervision Fidelity Decay (SFD)}: as student-generated prefixes lengthen, the teacher's next-token distribution becomes less confident and less discriminative. Consequently, the teacher-dependent corrective signal in reverse-KL distillation weakens, causing student drift to compound across long reasoning chains. To mitigate SFD, we introduce \textbf{Lookahead Group Reward (\ours{})}. Building on the insight that next-step teacher confidence reflects the discriminative strength of future reverse-KL supervision, \ours{} evaluates the student's top-K candidate tokens by the teacher confidence they induce at the subsequent step and assigns a group-normalized reward. To maintain computational efficiency, we further design an entropy-triggered tree-attention mechanism. Across six math and code benchmarks, \ours{} improves mean@8 by \textbf{2.57} points over OPD for a 7B student, with gains increasing in longer-generation and reaching +\textbf{4.92} points on AIME-26 at 39k tokens.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑