GPQA

GPQA (Google-Proof Q&A) is a graduate-level science question set created by domain experts. Questions are written so that simply searching the web does not help: 'google-proof' means that even human experts with full search access only answer about a third of them correctly in the non-expert setting.

GPQA measures general reasoning and comprehension. In detail, it checks:

• deep science knowledge beyond common exam questions
• step-by-step scientific reasoning
• robustness against attractive wrong answers
• 'closed-book' problem solving without external help

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

What is the GPQA benchmark?

GPQA (Google-Proof Q&A) is a graduate-level science question set created by domain experts. Questions are written so that simply searching the web does not help: 'google-proof' means that even human experts with full search access only answer about a third of them correctly in the non-expert setting.

What does GPQA test?

GPQA mainly measures general reasoning and comprehension. Concretely, it checks: deep science knowledge beyond common exam questions; step-by-step scientific reasoning; robustness against attractive wrong answers.

How is the GPQA score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get GPQA data and scores?

The official source is https://arxiv.org/abs/2311.12022. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/gpqa/scores) or pull it directly from the Open LLM Radar API with any HTTP client.