<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Zhihao Lin · 博客</title>
    <link>https://mathieu0905.github.io/blog</link>
    <description>研究心得、论文解读与技术探索</description>
    <language>zh-CN</language>
    <lastBuildDate>Sat, 24 Oct 2026 00:00:00 GMT</lastBuildDate>
    <atom:link href="https://mathieu0905.github.io/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>2026 博士生青培计划入选</title>
      <link>https://mathieu0905.github.io/blog/2026-phd-youth-training-program</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/2026-phd-youth-training-program</guid>
      <pubDate>Sat, 24 Oct 2026 00:00:00 GMT</pubDate>
      <description>很高兴入选 2026 博士生青培计划。带着新的认可与期待，继续在 AI 与软件工程交叉方向做有价值的研究。</description>
      <category>notes</category>
      <category>个人动态</category>
      <category>博士生</category>
      <category>学术成长</category>
    </item>
    <item>
      <title>OpenHarmony Bench：能编译，不等于真的完成了应用需求</title>
      <link>https://mathieu0905.github.io/blog/paper-openharmony-bench</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-openharmony-bench</guid>
      <pubDate>Mon, 17 Aug 2026 00:00:00 GMT</pubDate>
      <description>我在实习期间参与的 OpenHarmony Bench 技术报告已上线 arXiv。它用可构建、可安装、可交互验证的 ArkTS 应用，评估 coding agents 是否真的完成了端到端需求。</description>
      <category>research</category>
      <category>我的论文</category>
      <category>Code Agent</category>
      <category>OpenHarmony</category>
      <category>ArkTS</category>
      <category>Benchmark</category>
    </item>
    <item>
      <title>AtomicCommitBench：代码智能体不该只交一坨 diff</title>
      <link>https://mathieu0905.github.io/blog/paper-atomiccommitbench</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-atomiccommitbench</guid>
      <pubDate>Wed, 08 Jul 2026 00:00:00 GMT</pubDate>
      <description>我们的 AtomicCommitBench 论文已上线 arXiv。它研究 coding agents 能否把一个最终 squashed patch 重新组织成可 replay、可 review、可 selective revert 的原子提交历史。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>Code Agent</category>
      <category>软件演化</category>
      <category>原子提交</category>
    </item>
    <item>
      <title>RepoRescue：旧仓库不是坏了，是世界变了</title>
      <link>https://mathieu0905.github.io/blog/paper-reporescue</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-reporescue</guid>
      <pubDate>Fri, 03 Jul 2026 00:00:00 GMT</pubDate>
      <description>我们的 RepoRescue 论文已上线 arXiv。它讲的是一个很现实的故事：旧仓库曾经能跑，现代生态却把它弄坏了；代码智能体能不能真正修源码，而不是走改测试的捷径？</description>
      <category>research</category>
      <category>论文解读</category>
      <category>Code Agent</category>
      <category>兼容性维护</category>
    </item>
    <item>
      <title>论文精读：语言模型是少样本学习者（Language Models are Few-Shot Learners）</title>
      <link>https://mathieu0905.github.io/blog/gpt3-few-shot-learners-reading</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/gpt3-few-shot-learners-reading</guid>
      <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
      <description>从通用语言模型的动机、in-context learning、175B 参数规模、数据配比到测试集污染，整理一次 GPT-3 论文精读后的知识笔记。</description>
      <category>reading</category>
      <category>论文阅读</category>
      <category>GPT-3</category>
      <category>大模型</category>
      <category>Scaling</category>
    </item>
    <item>
      <title>论文精读：注意力就是你所需要的一切（Attention Is All You Need）</title>
      <link>https://mathieu0905.github.io/blog/attention-is-all-you-need-reading</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/attention-is-all-you-need-reading</guid>
      <pubDate>Thu, 02 Jul 2026 00:00:00 GMT</pubDate>
      <description>从 seq2seq、RNN/CNN、QKV 到位置编码和长上下文，整理一次从原论文读出来的 Transformer 知识笔记。</description>
      <category>reading</category>
      <category>论文阅读</category>
      <category>Transformer</category>
      <category>大模型</category>
      <category>深度学习</category>
    </item>
    <item>
      <title>代码基模与 Agentic Training 前沿：从 GLM-5.2、DeepSeek-V4 到 SWE Agent RL</title>
      <link>https://mathieu0905.github.io/blog/agentic-code-model-training-2026</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/agentic-code-model-training-2026</guid>
      <pubDate>Sat, 27 Jun 2026 00:00:00 GMT</pubDate>
      <description>一份更像读论文现场的深度笔记：用 alphaXiv、论文图和官方技术报告，把代码基模、长上下文、verifier、SWE 环境和 agent RL 的训练闭环讲清楚。</description>
      <category>reading</category>
      <category>大模型</category>
      <category>代码模型</category>
      <category>Agentic Training</category>
      <category>SWE Agent</category>
      <category>论文阅读</category>
    </item>
    <item>
      <title>To Run or Not to Run：代码智能体到底要不要跑测试？</title>
      <link>https://mathieu0905.github.io/blog/paper-run-less</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-run-less</guid>
      <pubDate>Thu, 25 Jun 2026 00:00:00 GMT</pubDate>
      <description>ISSTA 2026：基于 7,745 条公开轨迹与 3,000 次受控修复实验，分析代码执行的边际收益、成本与适用边界。已更新预印本、正式 DOI 和复现材料。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>程序修复</category>
      <category>Code Agent</category>
    </item>
    <item>
      <title>CodeAnchor：代码智能体到底需要多少静态结构？</title>
      <link>https://mathieu0905.github.io/blog/paper-codeanchor</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-codeanchor</guid>
      <pubDate>Thu, 25 Jun 2026 00:00:00 GMT</pubDate>
      <description>我们研究轻量静态结构如何作为 deterministic anchors，帮助代码智能体更稳定、更可复现地定位代码。论文被 ISSTA 2026 正式接收。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>静态分析</category>
      <category>Code Agent</category>
    </item>
    <item>
      <title>Phantom Rendering：揪出 UI 中看不见的性能杀手</title>
      <link>https://mathieu0905.github.io/blog/paper-phantom-rendering</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-phantom-rendering</guid>
      <pubDate>Mon, 22 Jun 2026 00:00:00 GMT</pubDate>
      <description>我们系统刻画了 Phantom Rendering：移动应用在 UI 线程做了大量计算，却没有产生可见渲染输出。论文发表于 PACMSE/FSE 2026，并开源了 HapPRDetection。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>性能分析</category>
      <category>UI</category>
    </item>
    <item>
      <title>EAGER：让代码生成和执行同时进行</title>
      <link>https://mathieu0905.github.io/blog/paper-eager</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-eager</guid>
      <pubDate>Sat, 04 Apr 2026 00:00:00 GMT</pubDate>
      <description>LLM 生成代码时你在干等？EAGER 让生成和执行并行，端到端延迟降低最高 55%。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>LLM</category>
      <category>代码生成</category>
    </item>
    <item>
      <title>当 LLM 遇上小众语言：仓颉语言微调实战</title>
      <link>https://mathieu0905.github.io/blog/paper-cangjie-finetuning</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-cangjie-finetuning</guid>
      <pubDate>Wed, 01 Apr 2026 00:00:00 GMT</pubDate>
      <description>如何让 LLM 学会一门训练数据极其稀缺的编程语言？我们以华为仓颉语言为案例，探索低资源语言的有效微调策略。被 EMSE 2026 接收。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>LLM</category>
      <category>低资源语言</category>
    </item>
    <item>
      <title>MazeBreaker：用多智能体强化学习破解 LLM 安全防线</title>
      <link>https://mathieu0905.github.io/blog/paper-mazebreaker</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-mazebreaker</guid>
      <pubDate>Sun, 15 Jun 2025 00:00:00 GMT</pubDate>
      <description>我们提出了一种多智能体协作的越狱评估框架，揭示了当前 LLM 安全防御的系统性弱点。被 ICSE 2026 接收。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>LLM安全</category>
      <category>强化学习</category>
    </item>
    <item>
      <title>HapRepair：让 AI 学会修复鸿蒙应用</title>
      <link>https://mathieu0905.github.io/blog/paper-haprepair</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-haprepair</guid>
      <pubDate>Thu, 20 Mar 2025 00:00:00 GMT</pubDate>
      <description>面向 OpenHarmony 生态的自动化程序修复工具，结合领域规则与 LLM，在华为真实项目中落地验证。被 FSE Industry 2025 接收。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>程序修复</category>
      <category>OpenHarmony</category>
    </item>
    <item>
      <title>当我们给 AI 一份代码蓝图，结果出乎意料</title>
      <link>https://mathieu0905.github.io/blog/ai-blueprint</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/ai-blueprint</guid>
      <pubDate>Tue, 21 Jan 2025 00:00:00 GMT</pubDate>
      <description>如果我们不再仅仅向 AI 提供原始文本，而是给它一份结构化的代码&apos;地图&apos;，会怎样？研究发现出人意料且反直觉。</description>
      <category>notes</category>
      <category>论文解读</category>
      <category>LLM</category>
      <category>静态分析</category>
    </item>
    <item>
      <title>开源 AI 编程工具的机遇与挑战：一份全景综述</title>
      <link>https://mathieu0905.github.io/blog/paper-opensource-ai-se</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-opensource-ai-se</guid>
      <pubDate>Sat, 15 Jun 2024 00:00:00 GMT</pubDate>
      <description>第一篇被接收的论文！系统梳理开源 AI 软件工程工具的生态现状、协作模式和未来挑战。发表于 TOSEM 2024。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>综述</category>
      <category>开源</category>
    </item>
    <item>
      <title>Exploring Code Analysis: Zero-Shot Insights on Syntax and Semantics with LLMs</title>
      <link>https://mathieu0905.github.io/blog/paper-llm-code-analysis</link>
      <guid isPermaLink="true">https://mathieu0905.github.io/blog/paper-llm-code-analysis</guid>
      <pubDate>Wed, 15 Nov 2023 00:00:00 GMT</pubDate>
      <description>我的第一篇投稿：系统评估 LLM 对代码语法和语义的理解能力，发现它们像 AST 解析器一样擅长语法，但在语义理解上漏洞百出。TOSEM 已接收，正式版本题目更新为 Exploring Code Analysis: Zero-Shot Insights on Syntax and Semantics with LLMs。</description>
      <category>research</category>
      <category>论文解读</category>
      <category>LLM</category>
      <category>静态分析</category>
    </item>
  </channel>
</rss>