Benchmark^2: Systematic Evaluation of LLM Benchmarks
Qi Qian, Chengsong Huang, Jingwen Xu, Changze Lv, Muling Wu, Wenhao Liu, Xiaohua Wang, Zhenghua Wang, Zisu Huang, Muzhao Tian, Jianhan Xu, Kun Hu, He-Da Wang, Yao Hu, Xuanjing Huang, Xiaoqing Zheng
2026-01-08
Summary
This paper addresses the issue of how we can trust the tests used to evaluate large language models, like ChatGPT. With so many different benchmarks popping up, it's hard to know which ones are actually good at measuring a model's true abilities.
What's the problem?
Currently, there are tons of different ways to test these AI models, but no real way to check if those tests themselves are reliable. A benchmark might say one model is better, while another says the opposite. This makes it difficult to compare models fairly and understand which ones are genuinely improving. The core issue is a lack of quality control *for* the benchmarks themselves.
What's the solution?
The researchers created a system called Benchmark^2, which is essentially a set of tools to evaluate benchmarks. It looks at three main things: first, if a benchmark ranks models in a similar way to other benchmarks; second, how well a benchmark can actually tell the difference between strong and weak models; and third, if there are weird cases where a weaker model does better than a stronger one on the same test. They tested this system on 15 different benchmarks and 11 different AI models to see how well it worked, and found they could even use fewer test questions while still getting good results.
Why it matters?
This work is important because it helps us build more confidence in the evaluation of AI models. If we can identify and improve the quality of benchmarks, we can more accurately track progress in the field and ensure that we're developing AI systems that are truly capable and reliable. It means we can focus on the *right* tests, and not waste time on ones that don't give us useful information.
Abstract
The rapid proliferation of benchmarks for evaluating large language models (LLMs) has created an urgent need for systematic methods to assess benchmark quality itself. We propose Benchmark^2, a comprehensive framework comprising three complementary metrics: (1) Cross-Benchmark Ranking Consistency, measuring whether a benchmark produces model rankings aligned with peer benchmarks; (2) Discriminability Score, quantifying a benchmark's ability to differentiate between models; and (3) Capability Alignment Deviation, identifying problematic instances where stronger models fail but weaker models succeed within the same model family. We conduct extensive experiments across 15 benchmarks spanning mathematics, reasoning, and knowledge domains, evaluating 11 LLMs across four model families. Our analysis reveals significant quality variations among existing benchmarks and demonstrates that selective benchmark construction based on our metrics can achieve comparable evaluation performance with substantially reduced test sets.