---
格式版本: 2
标题: "Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation"
原文链接: "https://arxiv.org/abs/2608.19098"
发布日期: "2026-08-19"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Wed, 19 Aug 2026 16:50:39 UTC (340 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-20T14:41:01+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-20T14:38:42+08:00"
入库时间: "2026-08-20T06:41:01.824Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=Oracle&searchtype=all"
匹配关键词:
  - "AI"
相关厂家:
  - "Oracle"
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 5
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "该资料为arXiv机器学习蒸馏论文，内容涉及多教师强化学习优化，与超节点、AI Rack、机柜级基础设施、供电散热互连等主题完全无关，仅因搜索词Oracle误入，判定为无关。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-20T14:42:15+08:00"
AI主题相关性: 0
AI来源权威性: 5
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 0
AI摘要: "Open-MOPD研究多教师on-policy蒸馏中能力集成不平衡问题，发现标准方法仅恢复35.6%的基线头空间，并将原因归为token级优化预算错配。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:18:31.551Z"
采集批次: "2026年8月20日14点19分32秒"
采集批次ID: "20260820-141932-079"
去重键: "https://arxiv.org/abs/2608.19098"
---

## Computer Science > Machine Learning

## Title:Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation

[View PDF](https://arxiv.org/pdf/2608.19098) [HTML (experimental)](https://arxiv.org/html/2608.19098v1)

> Abstract:Multi-teacher on-policy distillation (M-OPD) has emerged as a promising paradigm for consolidating domain-specialized reinforcement learning (RL) experts into a single generalist student via dense, token-level reward supervision. Despite its practical success, the optimization dynamics governing multi-teacher capability integration remain poorly understood, and open, rigorously reproducible recipes are conspicuously lacking. In this work, we establish a controlled M-OPD benchmark on SmolLM3-3B-Base with oracle routing, isolating capability integration from routing ambiguity. Our investigation reveals a pronounced capability integration gap: standard M-OPD captures only 35.6% of the available headroom relative to a domain-routed oracle ensemble, with concise tasks such as instruction following suffering severe degradation and premature stagnation. Crucially, we show that this failure stems not from gradient conflict, but from a severe misallocation of the token-level optimization budget. This pathology is driven by three orthogonal factors: structural sequence-length disparities across domains, dynamic convergence drift due to non-uniform learning rates, and multi-step reward staleness from asynchronous policy updates. To resolve these imbalances, we introduce Open-MOPD, a principled framework incorporating token-share balancing, gap-aware dynamic budget allocation, and student reward refresh. Together, these mechanisms systematically restore cross-domain balance, elevating headroom recovery from 35.6% to 83.4% in a single deployable student. We fully open-source our end-to-end post-training recipe, training trajectories, and evaluation suites on an academically accessible hardware budget.

| Comments: |  |
| --- | --- |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | [arXiv:2608.19098](https://arxiv.org/abs/2608.19098) \[cs.LG\] |
|  | (or [arXiv:2608.19098v1](https://arxiv.org/abs/2608.19098v1) \[cs.LG\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.19098](https://doi.org/10.48550/arXiv.2608.19098) |

## Submission history

From: Huan-Ang Gao \[[view email](https://arxiv.org/show-email/f8ee4177/2608.19098)\]  
**\[v1\]** Wed, 19 Aug 2026 16:50:39 UTC (340 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.19098) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))