跳到主要内容Skip to main content

评估Evaluation, Evals

Evaluation, Evals

  • 选型与成本Selection & cost
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-03refs checked 2026-11-03
解决什么问题What problem it solves

判断一个方案到底好不好,而不是「看起来不错」。Decide whether an approach is actually good, rather than merely plausible-sounding.

这是什么What it is

用一批固定题目和评分标准,量化比较不同模型、不同提示词、不同方案的效果。

评 agent 时还要看轨迹:工具调用是否正确、参数是否对、步数与成本是否可接受。

A fixed set of questions plus scoring criteria, to compare models, prompts or approaches quantitatively.

为什么需要知道Why it matters

没有评估就没有改进依据。这是区分「认真做的产品」和「说得好听的产品」的分水岭;agent 评测还要看过程,不只看看最终答案。Without evaluation there is no basis for improvement. It is the dividing line between a product built seriously and one that merely sounds good.

出处Sources
  • OpenAI Evals 官方指南OpenAI Working with Evals ↗
  • HELM 论文Holistic Evaluation of Language Models (HELM) ↗