benchgap
Calibration

SimpleVQA → We-Math

We-Math is estimated from SimpleVQA with a linear curve fitted on 6 models measured on both: y = 0.4451·x + 0.5212, R² = 0.52, cross-validated error 6.7 pp. It is used for 2 estimates.

Estimated modelSimpleVQAWe-MathSource
Step 3.7 Flash79.2%87.4%estimated ± 6.7 pp, low confidence
Muse Spark Open73.6%84.9%estimated ± 6.7 pp, low confidence

Fitted on

ModelSimpleVQAWe-Math
Claude Opus 4.565.7%70.0%
Gemini 3 Pro73.2%86.9%
GPT-5.255.8%79.0%
Kimi K2.571.2%84.7%
LFM3-VL-2B36.3%68.6%
Qwen3.5 397B67.1%87.9%