GPQA Diamond
Diamond split
公表値の高い順・評価条件は異なります · %
評価条件と原典を確認する
専門家が作成した科学の難問で、知識と推論を測る。
Diamond分割を使用。ツール、推論設定、集計方法を区別します。
評価条件 1
Diamond / No tools / xhigh / Accuracy · trial count not specified / OpenAI research environment · API
評価条件 2
Diamond / No tools / xhigh / Accuracy · trial count not specified / OpenAI research environment · API
評価条件 3
Diamond / No tools / high / Accuracy · trial count not specified / OpenAI research environment · API
評価条件 4
Diamond / Not specified in source / Chain of thought / 0-shot · accuracy / Meta instruction-tuned evaluation · English
評価条件 5
Diamond / Not specified in source / Chain of thought / 0-shot · accuracy / Meta instruction-tuned evaluation · English
評価条件 6
Diamond / Not specified in source / Chain of thought / 5-shot · trial count not specified / Mistral provider evaluation · text
評価条件 7
Diamond / Not specified in source / Thinking / pass@1 · trial count not specified / DeepSeek research evaluation · temperature 1.0 · 128k context
評価条件 8
GPQA Diamond; dataset revision notreported / notreported / Maximum score across effort settings; exact effort notreported / notreported / OpenAI research environment or API; exact harness notreported
評価条件 9
GPQA Diamond; dataset revision notreported / No tools / High thinking / Single attempt (pass@1); default sampling; no majority voting or parallel test-time compute; exact trial count notreported / Gemini API; exact harness version notreported
Gemini 3.1 Flash Lite Preview · 86.9%
Google · google:gemini-3.1-flash-lite-preview
開発元の公表値
資料公表 2026-03-03 · 収集・原典確認 2026-09-08 · 測定日 未公表
Evaluation > Results (as of March 2026) > GPQA Diamond Scientific knowledge > Gemini 3.1 Flash-Lite High column; No tools row note
一部の評価条件は未公表です。
評価手法PDFがmodel-id gemini-3.1-flash-lite-previewを明記。3月のPreview評価であり、現行stable版への転記は行わない。
評価条件 10
Diamond · 198 questions / Tool setting not specified in §2.9 / Adaptive thinking · max effort / Mean over 10 trials / Anthropic evaluation · default temperature/top_p
評価条件 11
Diamond / No tools / Thinking (High) / pass@1 · mean over trials / Gemini API · default sampling
