---
格式版本: 2
标题: "Diffract: Spectral View of LLM Domain Adaptation"
原文链接: "https://arxiv.org/abs/2608.10850"
发布日期: "2026-08-11"
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\\]** Tue, 11 Aug 2026 12:23:28 UTC (5,295 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: "2026-08-12T11:29:47+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: "2026-08-12T11:26:18+08:00"
入库时间: "2026-08-12T03:29:47.974Z"
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=AI&searchtype=all"
匹配关键词:
  - "AI"
相关厂家:
  []
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: 0
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "主题完全无关，论文讨论LLM领域自适应，未涉及超节点/AI Rack/机柜级AI基础设施或任何相关技术、商业信号。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-12T11:30:20+08:00"
AI主题相关性: 0
AI来源权威性: 0
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 0
采集批次: "2026年8月12日2点04分20秒"
采集批次ID: "20260812-020420-289"
去重键: "https://arxiv.org/abs/2608.10850"
---

## Computer Science > Machine Learning

## Title:Diffract: Spectral View of LLM Domain Adaptation

[View PDF](https://arxiv.org/pdf/2608.10850) [HTML (experimental)](https://arxiv.org/html/2608.10850v1)

> Abstract:We study continual pre-training (CPT) as a mechanism for adapting general-purpose large language models to specialized domains: mathematics, instruction, code, and natural text. Using singular value decomposition of weight matrices, we find that CPT leaves singular value spectra largely invariant, with adaptation driven mainly by changes in singular vectors. An analysis of attention-head projection matrices reveals strong, domain-dependent head heterogeneity, which we exploit to define a head importance criterion: up to 60% of head updates can be removed without measurable quality loss. Selectively rewinding low-importance heads to their pre-trained state improves benchmark accuracy by up to 4% versus the fully trained baseline. Finally, we identify domain connectivity - linear interpolation between CPT checkpoints yields smooth domain-quality interpolation without notable degradation on either domain - and release Diffract, an open-source toolkit for scalable spectral analysis of billion-parameter models.

| Comments: |  |
| --- | --- |
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | [arXiv:2608.10850](https://arxiv.org/abs/2608.10850) \[cs.LG\] |
|  | (or [arXiv:2608.10850v1](https://arxiv.org/abs/2608.10850v1) \[cs.LG\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.10850](https://doi.org/10.48550/arXiv.2608.10850) |

## Submission history

From: Nikita Borodin S. \[[view email](https://arxiv.org/show-email/dc667240/2608.10850)\]  
**\[v1\]** Tue, 11 Aug 2026 12:23:28 UTC (5,295 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.10850) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))