Calibration

VideoMMMU → MMMU-Pro

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

MMMU-Pro is estimated from VideoMMMU with a inverse Michaelis–Menten curve fitted on 11 models measured on both: y = 1.04252·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.45, cross-validated error 2.4 pp. It is used for 0 estimates.