GPQA

GPQA (Graduate-Level Google-Proof Q&A) is a benchmark of expert-written multiple-choice science questions, with its hardest subset known as GPQA Diamond, designed to be resistant to simple web lookup and to require genuine graduate-level reasoning. It is widely used to track frontier model reasoning ability, and cases where a model's performance on GPQA diverges sharply from its performance on other reasoning tests are treated as important signals about what these benchmarks actually measure.