Calibration
MathVision → ERQA
ERQA is estimated from MathVision with a inverse Michaelis–Menten curve fitted on 7 models measured on both: y = 0.70342·(x − 0.0000) / (0.0000 + 1.8176 − x), R² = 0.69, cross-validated error 2.4 pp. It is used for 12 estimates.
| Estimated model | MathVision | ERQA | Source |
|---|---|---|---|
| Claude Opus 4.5 | 74.3% | 48.6% | estimated ± 2.4 pp, low confidence |
| dots3-note Preview | 87.7% | 65.6% | estimated ± 2.4 pp, low confidence |
| Gemini 3 Pro | 86.6% | 64.0% | estimated ± 2.4 pp, low confidence |
| Gemma 4 12B | 79.7% | 54.9% | estimated ± 2.4 pp, low confidence |
| GPT-5.2 | 83.0% | 59.1% | estimated ± 2.4 pp, low confidence |
| Kimi K2.6 | 87.4% | 65.2% | estimated ± 2.4 pp, low confidence |
| Kimi K3 | 94.3% | 75.8% | estimated ± 2.4 pp, medium confidence |
| Qwen3.5-122B-A10B | 86.2% | 63.5% | estimated ± 2.4 pp, low confidence |
| Qwen3.5-27B | 86.0% | 63.2% | estimated ± 2.4 pp, low confidence |
| Qwen3.5-35B-A3B | 83.9% | 60.3% | estimated ± 2.4 pp, low confidence |
| Qwen3.5 397B | 88.6% | 66.9% | estimated ± 2.4 pp, low confidence |
| Qwen3.6 Plus | 88.0% | 66.0% | estimated ± 2.4 pp, low confidence |