Calibration
MathVision → RealWorldQA
RealWorldQA is estimated from MathVision with a offset logistic curve fitted on 5 models measured on both: y = 0.0000 + (0.8818 − 0.0000) / (1 + exp(−200.00·(x − 0.8814))), R² = 0.72, cross-validated error 0.8 pp. It is used for 14 estimates.
| Estimated model | MathVision | RealWorldQA | Source |
|---|---|---|---|
| Claude Opus 4.5 | 74.3% | 0.0% | estimated ± 0.8 pp, low confidence |
| dots3-note Preview | 87.7% | 26.0% | estimated ± 0.8 pp, low confidence |
| Gemini 3 Pro | 86.6% | 3.9% | estimated ± 0.8 pp, low confidence |
| Gemma 4 12B | 79.7% | 0.0% | estimated ± 0.8 pp, low confidence |
| GPT-5.2 | 83.0% | 0.0% | estimated ± 0.8 pp, low confidence |
| Kimi K2.6 | 87.4% | 16.4% | estimated ± 0.8 pp, low confidence |
| Kimi K3 | 94.3% | 88.2% | estimated ± 0.8 pp, medium confidence |
| Qwen3.5-122B-A10B | 86.2% | 1.8% | estimated ± 0.8 pp, low confidence |
| Qwen3.5-27B | 86.0% | 1.2% | estimated ± 0.8 pp, low confidence |
| Qwen3.5-35B-A3B | 83.9% | 0.0% | estimated ± 0.8 pp, low confidence |
| Qwen3.5 397B | 88.6% | 63.2% | estimated ± 0.8 pp, low confidence |
| Qwen3.6 Plus | 88.0% | 38.1% | estimated ± 0.8 pp, low confidence |
| Seed 2.1 Pro | 92.6% | 88.2% | estimated ± 0.8 pp, medium confidence |
| Seed 2.1 Turbo | 90.1% | 86.5% | estimated ± 0.8 pp, medium confidence |