benchgap
Calibration

MathVision → SimpleVQA

SimpleVQA is estimated from MathVision with a linear curve fitted on 9 models measured on both: y = 0.9269·x + -0.0987, R² = 0.80, cross-validated error 7.3 pp. It is used for 10 estimates.

Estimated modelMathVisionSimpleVQASource
Kimi K394.3%77.5%estimated ± 7.3 pp, medium confidence
Inkling 2 Small92.8%76.2%estimated ± 7.3 pp, medium confidence
Seed 2.1 Pro92.6%76.0%estimated ± 7.3 pp, medium confidence
Qwen3.8-Omni-Flash91.8%75.2%estimated ± 7.3 pp, medium confidence
Qwen3.8-Flash-Next90.6%74.1%estimated ± 7.3 pp, medium confidence
Seed 2.1 Turbo90.1%73.6%estimated ± 7.3 pp, medium confidence
Kimi K2.687.4%71.1%estimated ± 7.3 pp, medium confidence
Gemma 4 12B79.7%64.0%estimated ± 7.3 pp, medium confidence
Nemotron 3.5 Super VL 120B A12B78.2%62.6%estimated ± 7.3 pp, medium confidence
Nemotron 3.5 Super VL 120B A12B NVFP477.8%62.2%estimated ± 7.3 pp, medium confidence

Fitted on

ModelMathVisionSimpleVQA
Claude Opus 4.574.3%65.7%
dots3-note Preview87.7%72.5%
Gemini 3 Pro86.6%73.2%
GPT-5.283.0%55.8%
Kimi K2.584.2%71.2%
LFM3-VL-2B51.6%36.3%
Qwen3.5 397B88.6%67.1%
Qwen3.7 Plus90.3%81.7%
Qwen3.8 Max95.2%75.0%