Benchmark^2: Systematic Evaluation of LLM Benchmarks
Benchmark² evaluates LLM benchmarks via ranking consistency, discriminability, and capability alignment deviation, revealing quality variations and enabling smaller effective test sets.
Published Jan 7, 2026▲ 34 on Hugging FacearXiv ↗
Only vote on papers you've read. Sign in with GitHub to vote.
Benchmark² offers a rigorous, teachable framework measuring benchmark consistency, discriminability, and family-level failure ratios across 15 benchmarks to expose quality gaps and shrink test sets, though its ranking consistency risks circularity, family failure examples remain unnamed…
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.