metabench -- A Sparse Benchmark of Reasoning and Knowledge in Large Language Models

TL;DR

This work tackles benchmark redundancy in evaluating large language models by introducing metabench, a sparse, psychometrics-guided benchmark distilled from six prominent tasks. Using data from over $5000$ LLMs, the authors apply cross-validated subsampling, one-dimensional IRT, and GAMs to construct latent-ability estimators that can reconstruct original benchmark scores with RMSEs around $1\%$–$1.5\%$ and a grand mean with RMSE near $0.6\%$, while revealing a single general factor that accounts for most variance. The framework achieves substantial efficiency gains (reducing $d=28{,}632$ items to $<3\%$ of the original) and demonstrates strong cross-benchmark synergy, enabling accurate score reconstruction with a compact item bank and even enabling adaptive testing simulations. By providing two disjoint versions for repeated evaluation and an open-source reproducibility plan, metabench offers a practical, interpretable, and scalable path for faster, more fair LLM evaluation across domains.

Abstract

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n > 5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d = 28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.24% root mean square error (RMSE), (2) reconstruct the original total score with 0.58% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.94.

Figures (12)

• Figure 1: Score Reconstruction. (A) After distilling all six benchmarks to in total $858$ items, the latent abilities estimated by the reduced set can reconstruct each normalized benchmark score with, on average, $1.24\%$ root mean squared error (RMSE) on the test split. (B) Empirical test RMSE distributions for $10,000$ randomly subsampled benchmarks of matching sizes. Colors match benchmarks in the other panels, and the asterisks are their corresponding test RMSEs. Our method outperforms every best randomly sampled subset by a significant margin. (C) The test RMSE is further reduced to $0.58$% when the mean normalized score across all benchmarks is predicted.
• Figure 2: Processing pipeline. Explained in \ref{['sec:two']}.
• Figure 3: Conditional score distributions. Kernel density estimations of the normalized scores for each benchmark - all resp. last uses scores from all LLMs resp. the last listed LLM per unique username.
• Figure 4: Results of cross-validated subsampling. The best error from the validation set (solid line) is compared to the corresponding error from the test set (dashed line) for varying subset sizes.
• Figure 5: Maximum a posteriori score reconstructions from logistic regression models that use scores as predictors. Results are shown for the respective training sets with $350$ items.