Calibration
We-Math → ERQA
ERQA is estimated from We-Math with a inverse Michaelis–Menten curve fitted on 5 models measured on both: y = 0.89565·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.97, cross-validated error 2.7 pp. It is used for 1 estimate.
| Estimated model | We-Math | ERQA | Source |
|---|---|---|---|
| Kimi K2.5 | 84.7% | 65.8% | estimated ± 2.7 pp, medium confidence |
Fitted on
| Model | We-Math | ERQA |
|---|---|---|
| Claude Opus 4.5 | 70.0% | 46.8% |
| Gemini 3 Pro | 86.9% | 70.5% |
| GPT-5.2 | 79.0% | 59.8% |
| LFM3-VL-2B | 68.6% | 48.2% |
| Qwen3.5 397B | 87.9% | 67.5% |