LMArena (Chatbot Arena)

What it tests: Millions of blind side-by-side human votes: users compare two anonymous models and pick the better answer. Scores are Bradley-Terry ratings (Elo-like).

考什么:数百万次真人盲测投票:用户并排看两个匿名模型的回答,选更好的那个。分数是 Bradley-Terry 评分(类 Elo)。

Why it counts: The only large-scale measure of what humans actually prefer — but 2025's 'Leaderboard Illusion' paper showed big labs can game it, so we cap its weight at 5%.

为什么算数:唯一大规模反映'真人真实偏好'的数据——但 2025 年《排行榜幻觉》论文实锤了大厂能刷榜,所以权重压到 5% 并保留争议说明。