Purified OPSD: On-Policy Self-Distillation Without Losing How to Think
作者: Zhanming Shen, Jintao Tong, Shaotian Yan, Chen Shen, Hao Chen, Wentao Ye, Xiaomeng Hu, Rui Miao, Haobo Wang, Junbo Zhao, Gang Chen, Jieping Ye
分类: cs.AI, cs.LG
发布日期: 2026-07-05
💡 一句话要点
提出Purified OPSD以解决长链推理模型的自蒸馏问题
🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture) 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 长链推理 自蒸馏 点对点互信息 监督信号 模型性能提升
📋 核心要点
- 现有的在政策自蒸馏方法在长链推理模型上表现不佳,导致反思推理能力的不稳定。
- 论文提出通过构建仅依赖参考的教师模型和使用PMI机制来改善监督信号的质量。
- 实验结果显示,该方法在多个长链推理模型上均实现了显著的性能提升,超越了基线模型和标准OPSD。
📝 摘要(中文)
在现有的在政策自蒸馏(OPSD)方法中,教师模型通过参考解为学生模型提供逐步监督。然而,研究发现该方法在长链推理模型上效果不佳,导致反思推理能力的不稳定。通过对教师监督信号的分解,识别出问题根源在于参考引导的成分主导了监督信号,导致学生模型学习到特定参考的记忆性捷径。为此,论文提出了一种两步解决方案:首先构建一个仅依赖参考的教师模型,以隔离不可转移的监督成分;其次使用点对点互信息(PMI)将残差转化为学生可以直接蒸馏的目标分布。实验表明,该方法在四个长链推理模型上均取得了显著提升。
🔬 方法详解
问题定义:论文旨在解决在政策自蒸馏(OPSD)方法在长链推理模型上的效果不佳问题,现有方法导致模型学习到特定参考的捷径,影响反思推理能力。
核心思路:通过分解教师模型的监督信号,识别出参考引导成分的主导作用,提出构建仅依赖参考的教师模型来隔离不可转移的监督成分,并利用PMI机制提取有效的监督信号。
技术框架:整体流程包括两个主要阶段:首先构建参考教师模型以获取非转移成分,其次通过PMI将残差转化为目标分布供学生模型蒸馏。
关键创新:最重要的创新在于通过分解教师的监督信号,识别并隔离出影响模型学习的非转移成分,从而有效提升了长链推理模型的性能。
关键设计:在设计中,采用了参考教师模型和PMI机制,确保学生模型能够从有效的监督信号中学习,避免了参考引导的捷径影响。具体的损失函数和网络结构细节在实验部分进行了详细描述。
🖼️ 关键图片
📊 实验亮点
实验结果表明,使用Purified OPSD方法的长链推理模型在多个数据集上均实现了显著提升,相较于基线模型和标准OPSD,性能提升幅度达到X%(具体数据待补充),有效保持了模型的自然认知行为。
🎯 应用场景
该研究的潜在应用领域包括自然语言处理中的推理任务、对话系统和智能问答等。通过提升长链推理模型的性能,能够在实际应用中提供更准确的推理和决策支持,具有重要的实际价值和未来影响。
📄 摘要(原文)
On-policy self-distillation (OPSD) has emerged as a promising paradigm for improving LLM reasoning, where a privileged teacher with access to reference solutions provides token-level supervision on the student's own generated trajectories. However, we find that OPSD consistently fails on long chain-of-thought (long-CoT) reasoning models, yielding at best marginal gains while destabilizing the reflective reasoning capability these models depend on. Through a novel decomposition of the teacher's supervision signal, we identify the root cause: the teacher's supervision is dominated by a reference-induced component that drives rote memorization of reference-specific shortcuts, while the question-conditioned, inference-transferable component is ignored or actively opposed. Based on this diagnosis, we propose a two-step solution. First, we construct a reference-only teacher (the same model conditioned on the reference without the question) to isolate the non-transferable component of the supervision signal; the residual after subtracting this component captures the question-conditioned, inference-transferable correction. Second, we use pointwise mutual information (PMI) as the mechanism to transform this residual into a well-formed PMI target distribution that the student can directly distill from, filtering out the reference-induced shortcut. Experiments on four long-CoT models across two datasets demonstrate consistent improvements over both the base model and standard OPSD, while preserving the models' natural epistemic behavior throughout training.