
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>HANS</title>
      <link>https://www.hansking.cn/blog</link>
      <description>A posts created with Next.js and Tailwind.css</description>
      <language>en-us</language>
      <managingEditor>admin@hansking.cn (HANS)</managingEditor>
      <webMaster>admin@hansking.cn (HANS)</webMaster>
      <lastBuildDate>Sun, 17 Aug 2025 10:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.hansking.cn/tags/模型评测/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.hansking.cn/blog/2026/Agent-ping-ce-ren-wu-cheng-gong-l-gui-ji-zhi-liang-yu-cheng-ben</guid>
    <title>Agent 评测：任务成功率、轨迹质量与成本</title>
    <link>https://www.hansking.cn/blog/2026/Agent-ping-ce-ren-wu-cheng-gong-l-gui-ji-zhi-liang-yu-cheng-ben</link>
    <description>[{&quot;value&quot;:&quot;Agent 评测：任务成功率、轨迹质量与成本&quot;,&quot;url&quot;:&quot;#agent-评测任务成功率轨迹质量与成本&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先定义任务成功而不是模型成功&quot;,&quot;url&quot;:&quot;#一先定义任务成功而不是模型成功&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、任务成功率只是第一层&quot;,&quot;url&quot;:&quot;#二任务成功率只是第一层&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、轨迹质量看 Agent 是怎么做的&quot;,&quot;url&quot;:&quot;#三轨迹质量看-agent-是怎么做的&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、错误预算不能只奖励冒险&quot;,&quot;url&quot;:&quot;#四错误预算不能只奖励冒险&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、成本要算到成功任务&quot;,&quot;url&quot;:&quot;#五成本要算到成功任务&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、恢复能力是长任务的分水岭&quot;,&quot;url&quot;:&quot;#六恢复能力是长任务的分水岭&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、评测集要有难度和风险分层&quot;,&quot;url&quot;:&quot;#七评测集要有难度和风险分层&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、离线评测和线上反馈要互相连接&quot;,&quot;url&quot;:&quot;#八离线评测和线上反馈要互相连接&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、评测报告要帮助做决策&quot;,&quot;url&quot;:&quot;#九评测报告要帮助做决策&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：Agent 评测是在评估一套行为&quot;,&quot;url&quot;:&quot;#总结agent-评测是在评估一套行为&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sun, 17 Aug 2025 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>Agent</category><category>模型评测</category><category>成本优化</category><category>top</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026/BrowserGym-yu-ren-wu-xing-Agent-ji-zhun-ce-shi</guid>
    <title>BrowserGym 与任务型 Agent 基准测试</title>
    <link>https://www.hansking.cn/blog/2026/BrowserGym-yu-ren-wu-xing-Agent-ji-zhun-ce-shi</link>
    <description>[{&quot;value&quot;:&quot;BrowserGym 与任务型 Agent 基准测试&quot;,&quot;url&quot;:&quot;#browsergym-与任务型-agent-基准测试&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先定义任务的成功状态&quot;,&quot;url&quot;:&quot;#一先定义任务的成功状态&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、浏览器环境要可复现&quot;,&quot;url&quot;:&quot;#二浏览器环境要可复现&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、记录完整轨迹&quot;,&quot;url&quot;:&quot;#三记录完整轨迹&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、指标至少分五类&quot;,&quot;url&quot;:&quot;#四指标至少分五类&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;1. 任务结果&quot;,&quot;url&quot;:&quot;#1-任务结果&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;2. 动作质量&quot;,&quot;url&quot;:&quot;#2-动作质量&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;3. 效率&quot;,&quot;url&quot;:&quot;#3-效率&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;4. 恢复能力&quot;,&quot;url&quot;:&quot;#4-恢复能力&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;5. 安全性&quot;,&quot;url&quot;:&quot;#5-安全性&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;五、轨迹评测和最终评测要分开&quot;,&quot;url&quot;:&quot;#五轨迹评测和最终评测要分开&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、基准任务要覆盖边界&quot;,&quot;url&quot;:&quot;#六基准任务要覆盖边界&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、不要让 Agent 记住答案&quot;,&quot;url&quot;:&quot;#七不要让-agent-记住答案&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、工具执行要有沙箱&quot;,&quot;url&quot;:&quot;#八工具执行要有沙箱&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、失败回放比排行榜更有用&quot;,&quot;url&quot;:&quot;#九失败回放比排行榜更有用&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十、BrowserGym 类基准如何服务真实产品&quot;,&quot;url&quot;:&quot;#十browsergym-类基准如何服务真实产品&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：Agent 的路迹也是答案&quot;,&quot;url&quot;:&quot;#总结agent-的路迹也是答案&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sun, 31 Aug 2025 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>Agent</category><category>BrowserGym</category><category>模型评测</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026/Claude3-xi-lie-yu-mo-xing-neng-li-ping-ce-fang-fa</guid>
    <title>Claude 3 系列与模型能力评测方法</title>
    <link>https://www.hansking.cn/blog/2026/Claude3-xi-lie-yu-mo-xing-neng-li-ping-ce-fang-fa</link>
    <description>[{&quot;value&quot;:&quot;Claude 3 系列与模型能力评测方法&quot;,&quot;url&quot;:&quot;#claude-3-系列与模型能力评测方法&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、Claude 3 系列为什么值得写进发展史&quot;,&quot;url&quot;:&quot;#一claude-3-系列为什么值得写进发展史&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、先建立能力矩阵&quot;,&quot;url&quot;:&quot;#二先建立能力矩阵&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、评测集比一句神奇 Prompt 更重要&quot;,&quot;url&quot;:&quot;#三评测集比一句神奇-prompt-更重要&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、自动指标和人工指标要分工&quot;,&quot;url&quot;:&quot;#四自动指标和人工指标要分工&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、不能把长答案当成高质量&quot;,&quot;url&quot;:&quot;#五不能把长答案当成高质量&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、长上下文要测“找得到”和“用得对”&quot;,&quot;url&quot;:&quot;#六长上下文要测找得到和用得对&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、多模态评测要看视觉细节&quot;,&quot;url&quot;:&quot;#七多模态评测要看视觉细节&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、调用参数要固定，结果才可比较&quot;,&quot;url&quot;:&quot;#八调用参数要固定结果才可比较&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、模型裁判可以用，但不能独裁&quot;,&quot;url&quot;:&quot;#九模型裁判可以用但不能独裁&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十、评测结果要能指导路由&quot;,&quot;url&quot;:&quot;#十评测结果要能指导路由&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十一、上线后继续评测&quot;,&quot;url&quot;:&quot;#十一上线后继续评测&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十二、我的总结：模型评测是地图，不是冠军赛&quot;,&quot;url&quot;:&quot;#十二我的总结模型评测是地图不是冠军赛&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sun, 28 Apr 2024 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>Claude 3</category><category>模型评测</category><category>多模态</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026/LLM-as-a-Judge-zi-dong-ping-shen-de-pian-cha-yu-jiao-zhun</guid>
    <title>LLM-as-a-Judge：自动评审的偏差与校准</title>
    <link>https://www.hansking.cn/blog/2026/LLM-as-a-Judge-zi-dong-ping-shen-de-pian-cha-yu-jiao-zhun</link>
    <description>[{&quot;value&quot;:&quot;LLM-as-a-Judge：自动评审的偏差与校准&quot;,&quot;url&quot;:&quot;#llm-as-a-judge自动评审的偏差与校准&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先定义裁判到底要判断什么&quot;,&quot;url&quot;:&quot;#一先定义裁判到底要判断什么&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、给裁判一份明确的评分标准&quot;,&quot;url&quot;:&quot;#二给裁判一份明确的评分标准&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、裁判常见的几种偏差&quot;,&quot;url&quot;:&quot;#三裁判常见的几种偏差&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;1. 长度偏差&quot;,&quot;url&quot;:&quot;#1-长度偏差&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;2. 位置偏差&quot;,&quot;url&quot;:&quot;#2-位置偏差&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;3. 自我偏好&quot;,&quot;url&quot;:&quot;#3-自我偏好&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;4. 迎合语气&quot;,&quot;url&quot;:&quot;#4-迎合语气&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;5. 参考答案锚定&quot;,&quot;url&quot;:&quot;#5-参考答案锚定&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;四、先做盲测和一致性检查&quot;,&quot;url&quot;:&quot;#四先做盲测和一致性检查&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、人工标注用来校准，不是只用来盖章&quot;,&quot;url&quot;:&quot;#五人工标注用来校准不是只用来盖章&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、让裁判输出证据和解释&quot;,&quot;url&quot;:&quot;#六让裁判输出证据和解释&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、评审 Prompt 也要版本化&quot;,&quot;url&quot;:&quot;#七评审-prompt-也要版本化&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、什么时候不要使用 LLM-as-a-Judge&quot;,&quot;url&quot;:&quot;#八什么时候不要使用-llm-as-a-judge&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、把分数连接到行动&quot;,&quot;url&quot;:&quot;#九把分数连接到行动&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：自动裁判要先证明自己值得被信任&quot;,&quot;url&quot;:&quot;#总结自动裁判要先证明自己值得被信任&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sat, 31 Jan 2026 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>LLM-as-a-Judge</category><category>模型评测</category><category>数据质量</category>
  </item>

    </channel>
  </rss>
