Calibration
MMLU-ProX → PolyMath
PolyMath is estimated from MMLU-ProX with a offset logistic curve fitted on 5 models measured on both: y = 0.3824 + (0.8666 − 0.3824) / (1 + exp(−133.95·(x − 0.8394))), R² = 0.98, cross-validated error 11.4 pp. It is used for 7 estimates.
| Estimated model | MMLU-ProX | PolyMath | Source |
|---|---|---|---|
| Qwen3.6 Plus | 84.7% | 73.8% | estimated ± 11.4 pp, low confidence |
| GLM-5 | 83.1% | 50.1% | estimated ± 11.4 pp, low confidence |
| Nemotron 3 Ultra | 83.0% | 49.0% | estimated ± 11.4 pp, low confidence |
| Qwen3.5-122B-A10B | 82.2% | 42.5% | estimated ± 11.4 pp, low confidence |
| Qwen3.5-27B | 82.2% | 42.5% | estimated ± 11.4 pp, low confidence |
| Qwen3.5-35B-A3B | 81.0% | 39.2% | estimated ± 11.4 pp, low confidence |
| Qwen3 235B 2507 | 79.4% | 38.4% | estimated ± 11.4 pp, low confidence |
Fitted on
| Model | MMLU-ProX | PolyMath |
|---|---|---|
| Claude Opus 4.5 | 85.7% | 79.0% |
| Kimi K2.5 | 82.3% | 43.1% |
| Qwen3.5 397B | 84.7% | 73.3% |
| Qwen3.7 Max | 87.0% | 86.5% |
| Qwen3.7 Plus | 85.4% | 84.0% |