Calibration
HLE w/o tools → GPQA
GPQA is estimated from HLE w/o tools with a linear curve fitted on 25 models measured on both: y = 0.3619·x + 0.7702, R² = 0.86, cross-validated error 2.1 pp. It is used for 3 estimates.
| Estimated model | HLE w/o tools | GPQA | Source |
|---|---|---|---|
| Claude Haiku 5.5 | 45.9% | 93.6% | estimated ± 2.1 pp, high confidence |
| Gemma 4 26B A4B | 8.7% | 80.2% | estimated ± 2.1 pp, high confidence |
| Pareto 26.9 | 49.0% | 94.8% | estimated ± 2.1 pp, high confidence |