“如履薄冰”
用英语怎么说？

In theory, standard machine learning protocol requires a strict holdout method, meaning practitioners should lock away test data and evaluate systems only once. Ignoring this puts developers on thin ice, as repeated exposure enables algorithms to game leaderboards without genuine comprehension. Furthermore, absolute accuracy scores fluctuate wildly when test distributions undergo even trivial modifications. Hardt cautions that marketing claims equating a high mathematical benchmark score with human-level reasoning are inherently misleading. Treating these empirical scores as absolute measurements leads to flawed takeaways about whether a system truly possesses general intelligence.

理论上，标准的机器学习规范要求严格执行留出法，这意味着从业者应当封存测试数据，并仅对系统进行一次评估。忽视这一点会让开发者如履薄冰，因为重复接触能让算法在缺乏真正理解的情况下操纵排行榜。此外，即便测试集分布发生微小变动，绝对准确率分值也会剧烈波动。哈特警告称，将数学基准上的高分等同于人类水平推理能力的营销宣传本质上具有误导性。将这些经验得分视为绝对衡量标准，会导致在评估系统是否真正具备通用智能时得出有缺陷的结论。