Calibration
SimpleVQA → MLVU (M-Avg)
MLVU (M-Avg) is estimated from SimpleVQA with a Michaelis–Menten curve fitted on 10 models measured on both: y = 1.0506·x / (0.15104 + x), R² = 0.65, cross-validated error 5.0 pp. It is used for 2 estimates.
| Estimated model | SimpleVQA | MLVU (M-Avg) | Source |
|---|---|---|---|
| Step 3.7 Flash | 79.2% | 88.2% | estimated ± 5.0 pp, high confidence |
| Muse Spark Open | 73.6% | 87.2% | estimated ± 5.0 pp, high confidence |
Fitted on
| Model | SimpleVQA | MLVU (M-Avg) |
|---|---|---|
| Claude Opus 4.5 | 65.7% | 81.7% |
| Gemini 3 Pro | 73.2% | 83.0% |
| GPT-5.2 | 55.8% | 85.6% |
| Kimi K2.5 | 71.2% | 85.0% |
| LFM3-VL-2B | 36.3% | 71.3% |
| Qwen3.5 397B | 67.1% | 86.7% |
| Qwen3.6-27B | 56.1% | 86.6% |
| Qwen3.6-35B-A3B | 58.9% | 86.2% |
| Qwen3.7 Plus | 81.7% | 87.4% |
| Qwen3.8 Max | 75.0% | 90.8% |