提出新方法发现推理链中被忽略的关键证据，提升模型推理能力。
Finding the Evidence: Discovering Decision-Supporting Tokens for On-Policy Reasoning Distillation
- 通过学生不确定度识别决策点，再用隐状态相似度找支持证据
- 在数学和代码任务上最高提升5.7个百分点，显著优于传统方法
- 适合想提升模型推理可解释性和泛化能力的研究者
基于策略的推理蒸馏通过密集的词元级监督转移推理能力，但可迁移信号的本质仍不明确。我们发现推理链包含两类知识：决策（何处分支）可通过学生不确定性显现，而证据（支撑决策的中间步骤）则隐藏于学生自信却出错的位置。现有方法仅捕捉决策，证据中的实质性知识未被传递。我们提出DEAR（决策-证据感知推理蒸馏），先通过学生熵值识别决策，再利用隐状态与决策锚点的余弦相似度发现其支持证据，并借助教师-学生差异强化最大知识差距。在数学与代码基准上的三种师生配置中，DEAR持续优于标准OPD，在竞赛数学上最高提升2.5个百分点，在代码生成上最高提升5.7个百分点。
原文摘要 · Abstract (English)
On-policy distillation transfers reasoning ability through dense token-level supervision, yet the nature of the transferable signal remains unclear. We discover that reasoning chains contain two types of knowledge that require different discovery mechanisms: decisions (where to branch), which surface through student uncertainty, and evidence (intermediate steps that justify decisions), which hides in positions where the student is confident yet wrong. Current methods capture only decisions; the substantive knowledge in evidence tokens remains untransferred. We propose DEAR(Decision-Evidence Aware Reasoning Distillation), which first identifies decisions via student entropy, then discovers their supporting evidence through hidden-state cosine similarity to decision anchors, boosted by teacher-student divergence to prioritize the largest knowledge gaps. Across three student-teacher configurations on math and code benchmarks, DEAR consistently outperforms standard OPD, with up to +2.5pp on competition math and +5.7pp on code generation.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。