---
格式版本: 2
标题: "From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix"
原文链接: "https://arxiv.org/abs/2609.01572"
发布日期: "2026-09-01"
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, 1 Sep 2026 17:39:26 UTC (1,607 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: "2026-09-02T18:51:42+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: "2026-09-02T18:51:36+08:00"
入库时间: "2026-09-02T10:51:42.971Z"
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=GPU&searchtype=all"
匹配关键词:
  - "GPU"
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: 40
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "正文主线是通过GRPO专家训练与SLERP合并提升单一自托管LLM，并以116M次/月、承接50%流量衡量应用效果；GPU池仅为背景，没有机架拓扑、互连、RAS、供电或液冷等可复用基础设施机制。arXiv原始论文页来源可追溯，但当前仅提供摘要。固定知识库未发现同一研究，不能据此认定首次出现；其新增事实主要是模型训练方法及生产流量结果，并非超节点新增。命中“应用与模型效率”硬否决项，最高54分，不予准入。"
AI质检模型: "gpt-5.6-sol"
AI质检时间: "2026-09-02T18:51:53+08:00"
AI主题相关性: 2
AI来源权威性: 12
AI新颖性: 11
AI技术细节: 5
AI商业部署信号: 4
AI完整性: 6
AI评分提示词版本: "v17-精简生产版"
AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d"
AI评分知识库版本: "knowledge_base_v1-20260819+runtime.81"
AI评分知识库SHA256: "3b93d12e47749b3512f545f51c44c011bdc0931677c2cfe61e4df59a2b1a5a48"
AI评分知识库检索词: "[\"GPU\",\"LLM\",\"PDF\",\"arxiv.org/pdf/2609.01572\",\"HTML\",\"arxiv.org/html/2609.01572v1\",\"LLMs\",\"v1\",\"GRPO\",\"SLERP\",\"arxiv.org/abs/2609.01572\",\"arxiv.org/abs/2609.01572v1\"]"
AI评分知识库命中: "[{\"id\":\"july-correct-0111\",\"title\":\"StrataCL: Fabric-Native Communication Library for Production Supernodes\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"LLM\",\"PDF\",\"HTML\",\"v1\"],\"rank\":-13.739354454471368},{\"id\":\"runtime-569639751a0dbe7ef3ffccc5\",\"title\":\"[2608.17503] Predict Before Replay: Joint FEC and Flight Control for Reliable Scale-Up Links\",\"sourceType\":\"ai_excellent_article\",\"time\":\"2026-08-18\",\"matchedTerms\":[\"PDF\",\"HTML\",\"v1\"],\"rank\":-10.630171179861968},{\"id\":\"july-correct-0115\",\"title\":\"锚定 300kW 整机柜演进方向 OAII 社区三项规范联合发布，树立AIDC基础设施建设标准答案\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"LLM\",\"HTML\",\"v1\"],\"rank\":-9.525234386891393},{\"id\":\"july-correct-0010\",\"title\":\"Schneider Electric and AMD release first Helios platform reference design to accelerate AI Factory deployment\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"GPU\",\"LLM\",\"PDF\",\"v1\"],\"rank\":-9.365959513879561},{\"id\":\"runtime-bf6a5bdbe28961c50fb91e45\",\"title\":\"Amazon EC2 P6-B300 instances are now available in additional AWS Regions\",\"sourceType\":\"ai_excellent_article\",\"time\":\"2026-08-28\",\"matchedTerms\":[\"GPU\",\"LLM\",\"LLMs\"],\"rank\":-7.299574186623396}]"
AI摘要: "一项研究将 200 多个内部应用的请求流量整合到单一自托管 LLM，通过按指令遵循、函数调用等分轴训练的 GRPO 专家与两阶段 SLERP 合并来补齐质量差距；"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-02T17:14:15.221Z"
采集批次: "2026年9月2日18点28分48秒"
采集批次ID: "20260902-182848-894"
去重键: "https://arxiv.org/abs/2609.01572"
---

## Computer Science > Computation and Language

## Title:From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

[View PDF](https://arxiv.org/pdf/2609.01572) [HTML (experimental)](https://arxiv.org/html/2609.01572v1)

> Abstract:Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored by deterministic verifiers or calibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domain reward interference, we train a separate GRPO expert per axis and merge them via two-stage SLERP. Each expert's reward exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a ${\sim}7\times$ larger by total parameters baseline on the in-house Arena with 69.6 to 65.8, instruction following with 0.85 to 0.83, and function-calling with 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.

| Subjects: | Computation and Language (cs.CL) |
| --- | --- |
| Cite as: | [arXiv:2609.01572](https://arxiv.org/abs/2609.01572) \[cs.CL\] |
|  | (or [arXiv:2609.01572v1](https://arxiv.org/abs/2609.01572v1) \[cs.CL\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2609.01572](https://doi.org/10.48550/arXiv.2609.01572) |

## Submission history

From: Olga Tsymboi \[[view email](https://arxiv.org/show-email/9df985a8/2609.01572)\]  
**\[v1\]** Tue, 1 Sep 2026 17:39:26 UTC (1,607 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2609.01572) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))