PenaltyBench
Can an AI agent referee?AI 智能体能当裁判吗？
We tested nine AI agents on 2026 penalty clips cut before the referee's call. The best two tie the referee with VAR.我们让九个 AI 智能体看 2026 年的点球争议片段，片段剪在裁判判罚之前。最好的两个与借助 VAR 的裁判打平。
About简介
An AI agent watches a silent clip of a 2026 penalty incident and makes the call. Referee experts are the answer key.AI 智能体看一段无声的 2026 年点球争议片段，然后做出判罚。裁判专家的结论是标准答案。
You are the VAR你来当 VAR
Leaderboard排行榜
Accuracy on 101 clips.101 个片段上的准确率。
Data数据
Each item is a disputed 2026 penalty-area incident from a top competition. We cut the broadcast clip so that it ends before the referee's call, remove the sound and black out on-screen graphics. We keep only incidents that Gemini 4 Carbon cannot recall from a text description. The agent gets the clip in a sandbox with no network and answers penalty or no penalty. A call is correct when it matches the expert verdict; where no expert ruled on the incident (6 of 101), the referee's final decision stands.每道题是 2026 年顶级赛事里禁区内的一次争议判罚。我们把转播片段剪在裁判判罚之前，去掉声音，遮住屏幕上的图文。只保留 Gemini 4 Carbon 凭文字描述想不起来的事件。智能体在断网的沙箱里拿到片段，回答是不是点球。与专家结论相同即算判对；没有专家结论的 6 个事件维持原判。
Traces轨迹
Where each agent looked, and why they disagreed.每个智能体看了哪里，又为什么判得不一样。