benchgap
Calibration

SimpleQA → MMLU

MMLU is estimated from SimpleQA with a Michaelis–Menten curve fitted on 6 models measured on both: y = 0.9250·x / (0.00914 + x), R² = 0.64, cross-validated error 4.1 pp. It is used for 1 estimate.

Estimated modelSimpleQAMMLUSource
Gemini 2.0 Flash-Lite16.5%87.6%estimated ± 4.1 pp, medium confidence

Fitted on

ModelSimpleQAMMLU
DeepSeek-R130.1%90.8%
GPT-4.562.5%90.8%
Kimi K231.0%89.5%
Mistral Small 3.110.4%80.6%
o1-mini7.6%85.2%
o4-mini20.2%90.0%