Calibration
RealWorldQA → AI2D_TEST
AI2D_TEST is estimated from RealWorldQA with a offset logistic curve fitted on 7 models measured on both: y = 0.8760 + (0.9325 − 0.8760) / (1 + exp(−200.00·(x − 0.8087))), R² = 0.94, cross-validated error 1.4 pp. It is used for 4 estimates.
| Estimated model | RealWorldQA | AI2D_TEST | Source |
|---|---|---|---|
| Ternary Bonsai 2 27B | 80.1% | 88.7% | estimated ± 1.4 pp, medium confidence |
| LFM2.5-VL-3B | 73.1% | 87.6% | estimated ± 1.4 pp, medium confidence |
| LFM2.5-VL-450M | 58.4% | 87.6% | estimated ± 1.4 pp, low confidence |
| North Micro Vision Instruct | 62.2% | 87.6% | estimated ± 1.4 pp, low confidence |
Fitted on
| Model | RealWorldQA | AI2D_TEST |
|---|---|---|
| Claude Opus 4.5 | 77.0% | 87.7% |
| Gemini 3 Pro | 83.3% | 94.1% |
| GPT-5.2 | 83.3% | 92.2% |
| Kimi K2.5 | 81.0% | 90.8% |
| Qwen3.5 397B | 83.9% | 93.9% |
| Qwen3.6-35B-A3B | 85.3% | 92.7% |
| ZAYA1-VL-8B | 65.0% | 87.5% |