GeneralityLabs Exploit Bench测试结果
精选理由
GeneralityLabs公布Exploit Bench测试结果，V4.1 Flash性能下降，Claude Code超越Codex。
GeneralityLabs在Exploit Bench基准测试中表现出对harness参数的高度敏感性。V4.1 Flash版本性能弱于5.3 Flash版本，在Claude Code中表现最佳，Codex排名第二。Exploit Bench默认基准使用类似Python while循环的原始实现。
原文 · Teortaxes
Another big result from @GeneralityLabs: Performance on Exploit Bench is VERY harness-sensitive. sad news: V4.1 Flash is much weaker than 5.3 Flash, and does best in Claude Code (Codex is 2nd). btw, "original" is not ZCode&DSH, it's ExploitBench default (≈a Python while loop). https://t.co/xIPnOEIez7
- DeepLearning.AI10-04 02:59原文
