benchgap
Calibration

HLE w/o tools → Vals MMLU-Pro

Vals MMLU-Pro is estimated from HLE w/o tools with a Michaelis–Menten curve fitted on 19 models measured on both: y = 0.9970·x / (0.05462 + x), R² = 0.76, cross-validated error 1.9 pp. It is used for 3 estimates.

Estimated modelHLE w/o toolsVals MMLU-ProSource
Claude Haiku 5.545.9%89.1%estimated ± 1.9 pp, high confidence
Gemma 4 26B A4B8.7%61.2%estimated ± 1.9 pp, medium confidence
Pareto 26.949.0%89.7%estimated ± 1.9 pp, high confidence