“评估危机”
用英语怎么说？

Historically, machine learning avoided a debilitating replicability crisis because relative rankings across competing architectures remained surprisingly robust across diverse datasets. Even without understanding the underlying theoretical mechanisms, the higher-ranked model in a tournament reliably proved superior in practical deployments. However, this accidental reliability has collapsed in the frontier model era. Because proprietary training corpora remain undisclosed, engineers can deliberately curate pre-training mixes to match evaluation criteria. When developers intentionally optimize systems on prior datasets-a practice known as benchmaxxing-a weaker model can eclipse a superior architecture merely through targeted preparation, creating an acute evaluative crisis.

在历史上，机器学习避免了一场毁灭性的可复现性危机，因为竞争架构之间的相对排名在不同数据集上表现出惊人的稳健性。即使不理解其底层的理论机制，锦标赛中排名较高的模型在实际部署中也确实表现更佳。然而，这种偶然的可靠性在前沿大模型时代已然瓦解。由于专有训练语料保持不公开，工程师可以蓄意挑选预训练混合数据以契合评估标准。当开发人员针对先前的测试集刻意优化系统--这种做法被称为刷榜--较弱的模型仅凭针对性准备就能超越更优异的架构，从而引发了严峻的评估危机。