You Don't Need to Run Every Eval

A new study finds that a frontier model’s scores across 133 benchmarks can be largely determined by just two numbers, suggesting that many costly evaluations may be redundant [1]. The researchers compiled a public score matrix of 84 models and designed a method called BenchPress to recover held-out scores to within 4.6 points [2]. The study, posted to arXiv on June 22, 2026, assembled a matrix of 2,604 cells, of which only 23.3 percent were filled, tracking 84 frontier models across 133 benchmarks [2]. The authors found the matrix is approximately rank-2, meaning a model’s performance across all benchmarks is largely determined by two underlying factors. Two factors already explain over 90 percent of the variation among models on the benchmarks they share [2]. Building on this finding, the team designed BenchPress, a logit-space rank-2 matrix completion method that recovers held-out scores to within 4.6 points and includes a confidence layer that indicates when each prediction can be trusted [2]. Using BenchPress, the researchers identified a subset of five benchmarks—GPQA-D, HLE, Codeforces, MMLU-Pro, and ARC-AGI-1—that can recover the rest of a model’s public scorecard to within 3.93 points. For a tighter inference budget, a cheaper set—GPQA-D, MMLU-Pro, Aider Polyglot, MATH-500, and AIME 2026—can predict a model’s evals to within 4.55 points [2]. The work arrives as the machine-learning community increasingly seeks ways to make research more reproducible and accessible. Since October 2021, Hugging Face Spaces has been used to build and share over 12,000 open-source machine learning demos, enabling anyone with a browser to try models without writing code [9]. In 2022, Hugging Face collaborated with arXiv to integrate these demos directly onto paper abstract pages through a Demo tab, allowing readers to explore results interactively [10]. The BenchPress authors have released the score matrix, the BenchPress code, and an interactive tool that predicts any model’s score on any benchmark [2].