benchgap
Calibration

GPQA → AA-HLE

AA-HLE is estimated from GPQA with a inverse Michaelis–Menten curve fitted on 61 models measured on both: y = 0.10129·(x − 0.3649) / (0.3649 + 0.7010 − x), R² = 0.85, cross-validated error 6.5 pp. It is used for 3 estimates.

Estimated modelGPQAAA-HLESource
Claude 3.5 Sonnet59.4%4.9%estimated ± 6.5 pp, medium confidence
LFM2.5-VL-450M25.7%0.0%estimated ± 6.5 pp, low confidence
Qwen3 235B 250777.5%14.3%estimated ± 6.5 pp, medium confidence