Happiest Minds：2025年大语言模型时代质量保障：生成式AI系统评估与验证方法白皮书（英文版）

Evaluating large language models (LLMs) poses a challenge due to their black-box characteristics. With traditional software, it is easy to predict the outcome and thereby errors can be traced to specific segments of code. With LLMs, the scope of inputs and generated outputs is unimaginably large. This makes error identification and correction nearly impossible. The evaluation consists of verifying how well the LLM comprehends the intended message of the text and produces it in human language,