NeurIPS 2026 ED Oral: Alibaba Data Research Team evaluates the effectiveness of 158 evaluation benchmarks

The Ali Data Research Team evaluated the validity of 158 AI evaluation benchmarks and found that some benchmarks have approached saturation due to improved model capabilities, with narrowing score gaps leading to reduced reference value. The research team assessed the current performance of the benchmarks through a scoring system, pointing out which benchmarks can still effectively measure model performance.