Repovive 数学竞赛 AI 判卷评测：GPT 与 Claude 模型评判数学证明的能力对比
RT by @karinanguyen: 1/4 How well can models judge mathematical reasoning? After our first math contest on Repovive, we worked with a team of IOI, IMO, and ICPC finalists to evaluate our AI judge. We selected 534 suspicious or borderline submissions for review, established a shared grading standard, and created expert reference judgments. We then compared models under different grading instructions against those judgments. The main question: can models reliably understand and check challenging mathematical proofs? Accuracy is higher across the full set of contest submissions.
Repovive 联合 IOI、IMO、ICPC 决赛选手评测 AI 判卷能力，从数学竞赛中选出 534 份可疑或边缘提交，建立统一评分标准与专家参考判断。模型整体偏严，更常拒绝有效提交；清晰指令让 GPT 模型平均提升约 2 个百分点，加入评分示例后 Claude 平均提升约 10 个百分点、GPT 约 5 个百分点。Grok 几乎拒绝一切，仅接受 1–5% 提交，而金标准接受 64.4%。
来源：X：Karina Nguyen（@karinanguyen） · x.lingyaoai.com