Calibration
RealWorldQA → CharXiv
CharXiv is estimated from RealWorldQA with a linear curve fitted on 7 models measured on both: y = 3.3633·x + -2.0450, R² = 0.72, cross-validated error 4.1 pp. It is used for 5 estimates.
| Estimated model | RealWorldQA | CharXiv | Source |
|---|---|---|---|
| LFM2.5-VL-3B | 73.1% | 41.4% | estimated ± 4.1 pp, low confidence |
| LFM2.5-VL-450M | 58.4% | 0.0% | estimated ± 4.1 pp, low confidence |
| North Micro Vision Instruct | 62.2% | 4.7% | estimated ± 4.1 pp, low confidence |
| Ternary Bonsai 2 27B | 80.1% | 65.0% | estimated ± 4.1 pp, low confidence |
| ZAYA1-VL-8B | 65.0% | 14.1% | estimated ± 4.1 pp, low confidence |