Calibration
CharXiv → MathVision
MathVision is estimated from CharXiv with a Michaelis–Menten curve fitted on 15 models measured on both: y = 2.0000·x / (1.05549 + x), R² = 0.79, cross-validated error 2.6 pp. It is used for 8 estimates.
| Estimated model | CharXiv | MathVision | Source |
|---|---|---|---|
| Claude Sonnet 4.6 | 77.4% | 84.6% | estimated ± 2.6 pp, high confidence |
| Command A+ | 52.7% | 66.6% | estimated ± 2.6 pp, medium confidence |
| Gemini 3.1 Flash-Lite | 73.2% | 81.9% | estimated ± 2.6 pp, high confidence |
| Gemini 3.5 Flash | 84.2% | 88.7% | estimated ± 2.6 pp, high confidence |
| MiMo-V2.5 | 81.0% | 86.8% | estimated ± 2.6 pp, high confidence |
| Qwen3.6-35B-A3B | 78.0% | 85.0% | estimated ± 2.6 pp, high confidence |
| Sakana Fugu | 85.1% | 89.3% | estimated ± 2.6 pp, high confidence |
| Sakana Fugu-Ultra | 86.6% | 90.1% | estimated ± 2.6 pp, high confidence |