榜单怎么读How to Read a Leaderboard

先查三件事：谁维护、题目从哪来、分数怎么算——三者不明的榜单只当广告看。

看题目发布时间：榜单晚于模型训练截止才有区分度；疑似泄漏的公开题集分数普遍虚高。

总分拆开看：同一模型在不同任务子集上排名可以完全颠倒，按你要的任务子集比。

对比要同口径：不同榜单的分数不能互比，同一榜单里不同评测条件（上下文长度、工具开关）也不能混比。

看分数差有没有意义：1-2 分常在评测噪声范围内，排名相邻基本等于打平。

自建口径补位：公共榜单回答「普遍水平」，你自己的 20-50 道题回答「在我这活儿上行不行」。

Check three things first: who maintains it, where the questions come from, and how scores are computed — a leaderboard missing all three is an ad.

Look at when the questions were published: a benchmark only discriminates if it postdates the model's training cutoff; suspected-contaminated public sets inflate scores across the board.

Break the aggregate apart: the same model can rank completely differently across task subsets — compare on the subset you care about.

Compare only like with like: scores from different leaderboards are not comparable, and neither are different eval settings (context length, tools on/off) within one.

Ask whether the gap is meaningful: 1–2 points is usually within eval noise; adjacent ranks are effectively a tie.

Fill the gap with your own eval: public leaderboards answer "general level"; your own 20–50 questions answer "is it good at my job".