benchgap
Calibration

MathVision → PerceptionBench

PerceptionBench is estimated from MathVision with a inverse Michaelis–Menten curve fitted on 5 models measured on both: y = 0.70017·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.58, cross-validated error 3.6 pp. It is used for 12 estimates.

Estimated modelMathVisionPerceptionBenchSource
Qwen3.8-Omni-Flash91.8%59.4%estimated ± 3.6 pp, medium confidence
Qwen3.8-Flash-Next90.6%58.0%estimated ± 3.6 pp, medium confidence
Qwen3.7 Plus90.3%57.6%estimated ± 3.6 pp, medium confidence
Qwen3.5 397B88.6%55.7%estimated ± 3.6 pp, medium confidence
Qwen3.6 Plus88.0%55.0%estimated ± 3.6 pp, medium confidence
Kimi K2.687.4%54.3%estimated ± 3.6 pp, low confidence
Gemini 3 Pro86.6%53.5%estimated ± 3.6 pp, low confidence
Qwen3.5-122B-A10B86.2%53.0%estimated ± 3.6 pp, low confidence
Qwen3.5-27B86.0%52.8%estimated ± 3.6 pp, low confidence
Kimi K2.584.2%50.9%estimated ± 3.6 pp, low confidence
Qwen3.5-35B-A3B83.9%50.6%estimated ± 3.6 pp, low confidence
GPT-5.283.0%49.7%estimated ± 3.6 pp, low confidence

Fitted on

ModelMathVisionPerceptionBench
dots3-note Preview87.7%53.4%
Inkling 2 Mini89.0%57.5%
Inkling 2 Small92.8%64.7%
Kimi K394.3%58.5%
Qwen3.8 Max95.2%63.5%