Calibration
GPQA Diamond → HealthBench Professional
HealthBench Professional is estimated from GPQA Diamond with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.95106^6.00 + x^6.00), R² = 0.53, cross-validated error 4.8 pp. It is used for 14 estimates.
| Estimated model | GPQA Diamond | HealthBench Professional | Source |
|---|---|---|---|
| A.X K2 | 85.6% | 41.7% | estimated ± 4.8 pp, low confidence |
| K-EXAONE 2.0 | 82.2% | 35.3% | estimated ± 4.8 pp, low confidence |
| LLaDA2.2-mini | 44.4% | 1.2% | estimated ± 4.8 pp, low confidence |
| LongCat-Flash-Lite-Sparse | 69.5% | 15.8% | estimated ± 4.8 pp, low confidence |
| Mercury 2.5 | 79.0% | 29.7% | estimated ± 4.8 pp, low confidence |
| MiniCPM5-1B | 26.3% | 0.1% | estimated ± 4.8 pp, low confidence |
| MiniCPM5-2B | 70.2% | 16.7% | estimated ± 4.8 pp, low confidence |
| Pareto 26.10 Preview | 92.4% | 54.8% | estimated ± 4.8 pp, medium confidence |
| Beam | 90.5% | 51.1% | estimated ± 4.8 pp, low confidence |
| Solar Open 2 | 86.3% | 43.0% | estimated ± 4.8 pp, low confidence |
| Solar Pro 4 | 89.0% | 48.2% | estimated ± 4.8 pp, low confidence |
| Step 5 Preview | 93.5% | 56.9% | estimated ± 4.8 pp, medium confidence |
| Trinity-Large-Preview | 63.3% | 9.6% | estimated ± 4.8 pp, low confidence |
| Trinity-Large-Thinking | 76.3% | 25.3% | estimated ± 4.8 pp, low confidence |