Calibration
V* → CharXiv
CharXiv is estimated from V* with a offset logistic curve fitted on 12 models measured on both: y = 0.0000 + (0.7995 − 0.0000) / (1 + exp(−160.93·(x − 0.6589))), R² = 0.82, cross-validated error 3.6 pp. It is used for 1 estimate.
| Estimated model | V* | CharXiv | Source |
|---|---|---|---|
| Step 3.7 Flash | 95.3% | 80.0% | estimated ± 3.6 pp, high confidence |
Fitted on
| Model | V* | CharXiv |
|---|---|---|
| Claude Opus 4.5 | 67.0% | 68.5% |
| Gemini 3 Pro | 88.0% | 81.4% |
| GPT-5.2 | 75.9% | 82.1% |
| Kimi K2.6 | 96.9% | 80.4% |
| Nemotron 3.5 Super VL 120B A12B | 85.6% | 80.8% |
| Nemotron 3.5 Super VL 120B A12B NVFP4 | 86.0% | 79.9% |
| Qwen3.5-122B-A10B | 93.2% | 77.2% |
| Qwen3.5-27B | 93.7% | 79.5% |
| Qwen3.5-35B-A3B | 92.7% | 77.5% |
| Qwen3.5 397B | 95.8% | 80.8% |
| Qwen3.6-27B | 94.7% | 78.4% |
| Qwen3.6 Plus | 96.9% | 81.5% |