https://beanhsiang.github.io/tags/Evaluation/ Recent content in Evaluation on Innovation with tech Hugo zh-Hans Mon, 05 Oct 2026 12:00:00 +0800 https://beanhsiang.github.io/post/2026-10-05-evaluate-more-spend-less-batching-microsoft-foundry-evaluators/ Mon, 05 Oct 2026 12:00:00 +0800 https://beanhsiang.github.io/post/2026-10-05-evaluate-more-spend-less-batching-microsoft-foundry-evaluators/ <p> <a data-fancybox="gallery" href="https://beanhsiang.github.io/uploads/20261007113027-foundry-composite-eval-title.png"> <img class="mx-auto" alt="Evaluate More, Spend Less: Batching Microsoft Foundry Evaluators" src="https://beanhsiang.github.io/uploads/20261007113027-foundry-composite-eval-title.png" /> </a> </p> <p><em>题图来自 Microsoft Foundry 博客的 <a href="https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/evaluate-more-spend-less-batching-microsoft-foundry-evaluators-for-efficient-eva/4560841?WT.mc_id=AI-MVP-5003172">Evaluate More, Spend Less: Batching Microsoft Foundry Evaluators for Efficient Evaluation</a>，作者是 Salma Elshafey、Ali Mahmoudzadeh、Kayla Ames、Ahmad Qardahji、Vivek Bhadauria、Morteza Ziyadi 和 April Kwong。</em></p> <p>给 Agent 做评估时，同一段对话往往要分别发给好几个评估器。这篇文章讲的是微软 Foundry 团队的一项实验：用一次 Judge 调用同时跑五六个 Microsoft Foundry 评估器，总输入 token 少了 61.25% 到 71.07%，实测运行耗时少了 35.74% 到 46.10%，而且在几个价值较高的评估维度上，前沿 Judge 模型的质量没有明显波动。我把它读完后觉得，数字好看是一回事，更有用的是它顺带给出了哪些评估器适合合并、哪些得留着单独跑。</p>