PulseAugur
中
实时 16:38:45
English(EN) Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why

新研究改进 AI 推理模型的 On-Policy Distillation

作者 PulseAugur 编辑部 · [5 个来源] ·

研究人员正在探索用于训练推理模型的 On-Policy Distillation (OPD) 技术，该技术使用教师模型提供每 token 的监督。然而，OPD 的有效性和最佳配置仍不清楚，其中“崩溃”（推理路径变窄）等问题是一个重大挑战。新的方法，如 RouteOPD 和 TV-Regulated OPD，分别旨在通过显式建模概率传输和稳定训练动态来改进 OPD。一篇关键性评论还综合了现有研究，确定了 token 加权、特权信息和引导动态是影响崩溃的关键因素。 AI

影响 On-Policy Distillation 的这些进展可能导致 AI 推理模型的训练更有效、更健壮，从而可能提高其在复杂任务上的性能。

排序理由 该集群包含多篇学术论文，详细介绍了特定 AI 研究领域（On-Policy Distillation）内的新方法和分析。

在 arXiv cs.AI 阅读 →

AI 生成摘要 · Google Gemini · 来自 5 个来源。 我们如何撰写摘要 →

新研究改进 AI 推理模型的 On-Policy Distillation

本文如何被排名

Signal score
0 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Research
该集群包含多篇学术论文，详细介绍了特定 AI 研究领域（On-Policy Distillation）内的新方法和分析。
Source corroboration
5 independent sources
Strong cross-source corroboration — multiple independent publishers covered this within the clustering window.
Topics
paper, model release
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
28 days old
Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.
Coverage growth since scoring
+1 source(s) since last score
New sources have picked up this story since our last re-score. Score will update on the next scoring pass.

完整方法见我们的编辑标准。

报道来源 [5]

  1. arXiv cs.CL TIER_1 English(EN) · Suwan Wu, Yumeng Lin, Pengcheng Yuan, Xiaolong Jiang ·

    面向在线策略蒸馏的统一每Token门控家族：带有多通道和偏置系数的FKL/RKL混合

    arXiv:2609.11768v1 Announce Type: cross Abstract: Per-token gating of forward/reverse KL losses has become a standard technique for on-policy knowledge distillation (OPD), but existing methods such as EOPD (Jin et al., 2026) and ToDi (Jung et al., 2025) each fix a single gating s…

  2. arXiv cs.AI TIER_1 English(EN) · Mohammadreza Armandpour, Fatih Ilhan, David Harrison, Ajay Jaiswal, Duc N. M Hoang, Fartash Faghri, Yizhe Zhang, Minsik Cho, Mehrdad Farajtabar ·

    揭秘On-Policy蒸馏：它何时有效、何时无效以及原因

    arXiv:2605.10889v2 Announce Type: replace-cross Abstract: On-policy distillation offers dense, per-token supervision for training reasoning models; however, it remains unclear under which conditions this signal is beneficial and under which it is detrimental. Which teacher model …

  3. arXiv cs.LG TIER_1 English(EN) · Tianle Xia, Lingxiang Hu, Yiding Sun, Linfang Shang, Ming Xu, Lan Xu, Ning Zheng, Wei Xu, Jie Jiang ·

    蒸馏作为概率传输：路由策略内蒸馏

    arXiv:2609.08337v1 Announce Type: new Abstract: On-policy distillation (OPD) transfers teacher knowledge on student-generated trajectories, but efficient sampled objectives reduce the teacher distribution to scalar credit on individual tokens. Such credit indicates whether a toke…

  4. arXiv cs.LG TIER_1 English(EN) · Han Xiao, Yifan Niu, Dongyi Liu, Chang Luo, Jia Li ·

    电视监管的OPD：策略蒸馏中的方向至关重要

    arXiv:2609.08341v1 Announce Type: new Abstract: On-Policy Distillation (OPD) facilitates the transfer of knowledge from domain expert to student in the post-training phase of Large Language Models (LLMs). However, the supervision signals in mainstream OPD methods suffer from high…

  5. Hugging Face Daily Papers TIER_1 English(EN) ·

    一个症状，三个杠杆：对在线策略自蒸馏的批判性回顾

    On-policy self-distillation for mathematical reasoning uses privileged information to guide a model without a larger teacher, but suffers from reasoning collapse governed by token weighting, privileged context, and guidance dynamics.

相关实体

相关话题