Calibration

AA-Omniscience Accuracy → GPQA

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

GPQA is estimated from AA-Omniscience Accuracy with a Hill curve fitted on 60 models measured on both: y = 0.0000 + (0.9452 − 0.0000)·x^1.70 / (0.07005^1.70 + x^1.70), R² = 0.65, cross-validated error 5.9 pp. It is used for 0 estimates.