SimpleQA

SimpleQA: Measuring short-form factuality in large language models

4,326 个简短的事实型问题,每题有唯一无可争议的答案,针对 GPT-4o 对抗式收集并经两名独立标注者验证,涵盖科学、历史、艺术、地理等。评分模型将每个回答标记为正确、错误或未尝试;主指标为正确率,同时报告 F 分数和校准。它衡量模型是否知道自己知道什么:易产生幻觉的模型即便其他能力出色,得分也会很低。

62.5%
gpt-4.5-preview-2025-02-27
厂商自报
2024-11-07 → 2026-08-13: 38.2% → 57.9%
发布
2024-11
维护者
OpenAI (Wei et al.)
状态
active
污染风险
high
指标
accuracy (correct) (percent, ↑)
题量
4,326
领域
factuality knowledge
人类
无实测基线

备注. Answers are static facts, so retrieval or web search makes the test trivial; only no-tools runs are comparable. Google's SimpleQA Verified (1,000 relabeled questions, 2025) is a separate, stricter variant and its scores should not be mixed with the original. OpenAI stopped reporting SimpleQA in simple-evals in July 2025.

完整账本

系统开发者分数日期来源条件
gpt-4.5-preview-2025-02-27OpenAI62.5%厂商自报tools: no
simple-evals README benchmark table (repo deprecated July 2025).
DeepSeek V4 Pro 0813DeepSeek57.9%聚合站
benchlm.ai lists this under SimpleQA, but DeepSeek's own model card reports 57.9 as SimpleQA-Verified (Pass@1), a different 1,000-question variant; treat as not comparable with the rows above.
o3OpenAI49.4%厂商自报tools: no
simple-evals README; o3-high 48.6, o4-mini 20.2.
GPT-4oOpenAI38.2%论文tools: no
Table 3 of the SimpleQA paper: percent correct; o1-preview scored 42.7.