Calibration
HLE w/o tools → HealthBench Professional
HealthBench Professional is estimated from HLE w/o tools with a Michaelis–Menten + offset curve fitted on 5 models measured on both: y = 0.0000 + 1.7850·x / (1.04560 + x), R² = 0.79, cross-validated error 4.8 pp. It is used for 3 estimates.
| Estimated model | HLE w/o tools | HealthBench Professional | Source |
|---|---|---|---|
| Claude Haiku 5.5 | 45.9% | 54.5% | estimated ± 4.8 pp, medium confidence |
| Gemma 4 26B A4B | 8.7% | 13.7% | estimated ± 4.8 pp, low confidence |
| Pareto 26.9 | 49.0% | 57.0% | estimated ± 4.8 pp, medium confidence |