When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation
ETHZ - ETH Zurich · Stanford University · Alinia AI · Northeastern University · Independent / Eleuther AI · University of Maryland, College Park · Department of Computer Science, ETHZ - ETH Zurich · SFU · Meta · EleutherAI · MIT, IBM · Comenius University in Bratislava · Weizenbaum Institut · Queen Mary University of London · Harvard University · StickFlux Labs & Queer in AI · University College London · Scale AI · Cohere · Congruity360 · ETH Zurich; The University of Hong Kong · University of Illinois Chicago · CMU, Carnegie Mellon University · Hebrew University, Hebrew University of Jerusalem · Iowa State University · IIT Bombay · Universität Leipzig · University of California, Los Angeles · Swiss Federal Institute of Technology · Stanford University // Virtue AI · University of Edinburgh, University of Edinburgh · Hugging Face
PDF 由论文原始站点提供，PaperCompass 不保存论文文件。
摘要
Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.