Calibration
We-Math → CountBench
CountBench is estimated from We-Math with a offset logistic curve fitted on 6 models measured on both: y = 0.9183 + (0.9730 − 0.9183) / (1 + exp(−200.00·(x − 0.8487))), R² = 0.92, cross-validated error 1.8 pp. It is used for 0 estimates.
Fitted on
| Model | We-Math | CountBench |
|---|---|---|
| Claude Opus 4.5 | 70.0% | 90.6% |
| Gemini 3 Pro | 86.9% | 97.3% |
| GPT-5.2 | 79.0% | 91.9% |
| Kimi K2.5 | 84.7% | 94.1% |
| LFM3-VL-2B | 68.6% | 93.0% |
| Qwen3.5 397B | 87.9% | 97.2% |