Calibration
Video-MME (with subtitle) → We-Math
We-Math is estimated from Video-MME (with subtitle) with a offset logistic curve fitted on 6 models measured on both: y = 0.6930 + (0.8734 − 0.6930) / (1 + exp(−175.74·(x − 0.8592))), R² = 0.98, cross-validated error 3.5 pp. It is used for 3 estimates.
| Estimated model | Video-MME (with subtitle) | We-Math | Source |
|---|---|---|---|
| MiMo-V2.5 | 87.7% | 86.6% | estimated ± 3.5 pp, medium confidence |
| Qwen3.6-35B-A3B | 86.6% | 83.2% | estimated ± 3.5 pp, medium confidence |
| MiniMax M3 | 85.4% | 74.5% | estimated ± 3.5 pp, medium confidence |
Fitted on
| Model | Video-MME (with subtitle) | We-Math |
|---|---|---|
| Claude Opus 4.5 | 77.6% | 70.0% |
| Gemini 3 Pro | 88.4% | 86.9% |
| GPT-5.2 | 86.0% | 79.0% |
| Kimi K2.5 | 87.4% | 84.7% |
| LFM3-VL-2B | 69.0% | 68.6% |
| Qwen3.5 397B | 87.5% | 87.9% |