一项学术研究对人工智能基准测试的饱和现象进行了系统分析。[1]研究团队在60个语言模型基准测试上进行了调查，采用14个与饱和相关的属性进行评估。[1]研究发现，近半数基准测试已经出现饱和现象，且这一比例随着年份推移而持续上升。[1]
研究表明，基准测试的设计方式对其抗饱和能力有重要影响。[1]专家策划的测试数据相比公开发布的测试数据，更能够延缓饱和现象的出现。[1]该研究指出，通过优化设计选择，可以有效延长基准测试的有效使用寿命。[1]
Researchers have conducted a comprehensive analysis of saturation phenomena across artificial intelligence benchmarks, examining 60 language model evaluation datasets [1]. The study employed 14 attributes related to saturation to assess benchmark performance, revealing that nearly half of the benchmarks examined exhibit saturation characteristics [1]. The saturation rate increases with time, indicating a growing challenge in maintaining benchmark relevance [1].
The findings suggest that design choices significantly influence a benchmark's resistance to saturation [1]. Notably, benchmarks curated by experts demonstrate greater resilience to saturation compared to those based on publicly available test data [1]. These insights point toward strategies for extending the operational lifespan of evaluation frameworks in AI research [1]. The research was submitted on February 18, 2026, with final revisions completed on June 29, 2026 [1].