“The GPT-4 technical report used a mean log pass curve for the HumanEval benchmark to better measure models that had the correct solution idea but lacked the reliability to solve the entire problem consistently.”