benchgap
Calibration

Video-MME (with subtitle) → AI2D_TEST

AI2D_TEST is estimated from Video-MME (with subtitle) with a Michaelis–Menten + offset curve fitted on 6 models measured on both: y = 0.2534 + 2.0000·x / (1.71462 + x), R² = 0.80, cross-validated error 1.3 pp. It is used for 2 estimates.

Estimated modelVideo-MME (with subtitle)AI2D_TESTSource
MiniMax M385.4%91.8%estimated ± 1.3 pp, medium confidence
LFM3-VL-2B69.0%82.7%estimated ± 1.3 pp, low confidence

Fitted on

ModelVideo-MME (with subtitle)AI2D_TEST
Claude Opus 4.577.6%87.7%
Gemini 3 Pro88.4%94.1%
GPT-5.286.0%92.2%
Kimi K2.587.4%90.8%
Qwen3.5 397B87.5%93.9%
Qwen3.6-35B-A3B86.6%92.7%