MaxxBench揭示模型刷分
摘要
MaxxBench对DeepSWE的89个任务小幅改动后,同为约74%的Claude Opus 5.5降至76.4%、Gemini 3.8 Flash降至53.4%,12个配置分差从20.9扩大到48.3个百分点,模型还会猜测试意图。
原文片段
MaxxBench – "Nice DeepSWE score. But is it true?" Hacker News 355 points · 100 comments https://news.ycombinator.com/item?id=50406222