Calibration
SimpleVQA → RealWorldQA
RealWorldQA is estimated from SimpleVQA with a Michaelis–Menten curve fitted on 11 models measured on both: y = 0.9379·x / (0.08093 + x), R² = 0.49, cross-validated error 3.8 pp. It is used for 0 estimates.
Fitted on
| Model | SimpleVQA | RealWorldQA |
|---|---|---|
| Claude Opus 4.5 | 65.7% | 77.0% |
| Gemini 3 Pro | 73.2% | 83.3% |
| GPT-5.2 | 55.8% | 83.3% |
| Kimi K2.5 | 71.2% | 81.0% |
| LFM2.5-VL-3B | 35.4% | 73.1% |
| LFM3-VL-2B | 36.3% | 79.3% |
| Qwen3.5 397B | 67.1% | 83.9% |
| Qwen3.6-27B | 56.1% | 84.1% |
| Qwen3.6-35B-A3B | 58.9% | 85.3% |
| Qwen3.7 Plus | 81.7% | 86.9% |
| Qwen3.8 Max | 75.0% | 88.0% |