GPQA: A Graduate-Level Google-Proof Q&A Benchmark

TL;DR

GPQA is a challenging, expert-created multiple-choice question dataset in science, designed to be difficult for both humans and AI, to aid in developing scalable oversight methods for AI systems.

Contribution

The paper introduces GPQA, a high-quality, difficult science question dataset that is resistant to Google searches and challenging for state-of-the-art AI, facilitating research on AI oversight.

Findings

Experts achieve 65% accuracy on GPQA questions.

Non-experts reach only 34% accuracy despite extensive web access.

GPT-4 achieves 39% accuracy on GPQA questions.

Abstract

We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult: experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while highly skilled non-expert validators only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web (i.e., the questions are "Google-proof"). The questions are also difficult for state-of-the-art AI systems, with our strongest GPT-4 based baseline achieving 39% accuracy. If we are to use future AI systems to help us answer very hard questions, for example, when developing new scientific knowledge, we need to develop scalable oversight methods that enable humans to supervise their outputs, which may be difficult even if the supervisors are themselves skilled and knowledgeable. The difficulty of GPQA both for skilled non-experts and frontier AI systems should enable realistic scalable oversight experiments, which we hope can help devise ways for human experts to reliably get truthful information from AI systems that surpass human capabilities.