Calibration
VITA-Bench → DeepPlanning
DeepPlanning is estimated from VITA-Bench with a offset logistic curve fitted on 5 models measured on both: y = 0.2622 + (1.2000 − 0.2622) / (1 + exp(−83.89·(x − 0.4617))), R² = 1.00, cross-validated error 9.1 pp. It is used for 7 estimates.
| Estimated model | VITA-Bench | DeepPlanning | Source |
|---|---|---|---|
| Agents-A1 | 38.8% | 26.4% | estimated ± 9.1 pp, low confidence |
| Agents-A1-4B | 40.3% | 26.9% | estimated ± 9.1 pp, low confidence |
| Claude Sonnet 4.5 | 17.0% | 26.2% | estimated ± 9.1 pp, low confidence |
| DeepSeek V3.2 | 18.5% | 26.2% | estimated ± 9.1 pp, low confidence |
| GLM-4.7 | 15.5% | 26.2% | estimated ± 9.1 pp, low confidence |
| LongCat-Flash-Lite-Sparse | 21.7% | 26.2% | estimated ± 9.1 pp, low confidence |
| Qwen3.7 Max | 47.9% | 100.0% | estimated ± 9.1 pp, low confidence |