Calibration
MMMU → RefCOCO (avg)
RefCOCO (avg) is estimated from MMMU with a Michaelis–Menten curve fitted on 5 models measured on both: y = 1.0216·x / (0.08874 + x), R² = 0.90, cross-validated error 1.8 pp. It is used for 6 estimates.
| Estimated model | MMMU | RefCOCO (avg) | Source |
|---|---|---|---|
| Command A+ | 75.1% | 91.4% | estimated ± 1.8 pp, medium confidence |
| LFM2.5-VL-450M | 32.7% | 80.3% | estimated ± 1.8 pp, low confidence |
| Qwen3.5-122B-A10B | 83.9% | 92.4% | estimated ± 1.8 pp, low confidence |
| Qwen3.5-27B | 82.3% | 92.2% | estimated ± 1.8 pp, medium confidence |
| Qwen3.5-35B-A3B | 81.4% | 92.1% | estimated ± 1.8 pp, medium confidence |
| Qwen3.6 Plus | 86.0% | 92.6% | estimated ± 1.8 pp, low confidence |