Calibration

RefCOCO (avg) → MMMU

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

MMMU is estimated from RefCOCO (avg) with a offset logistic curve fitted on 5 models measured on both: y = 0.4599 + (0.8456 − 0.4599) / (1 + exp(−127.16·(x − 0.9004))), R² = 1.00, cross-validated error 0.4 pp. It is used for 1 estimate.

Estimated model	RefCOCO (avg)	MMMU	Source
Interfaze Beta	82.1%	46.0%	estimated ± 0.4 pp, low confidence