Calibration
LongBench v2 → AA-LCR
AA-LCR is estimated from LongBench v2 with a offset logistic curve fitted on 10 models measured on both: y = 0.5648 + (0.7333 − 0.5648) / (1 + exp(−98.57·(x − 0.4546))), R² = 0.47, cross-validated error 7.0 pp. It is used for 4 estimates.
| Estimated model | LongBench v2 | AA-LCR | Source |
|---|---|---|---|
| Agents-A1 | 60.2% | 73.3% | estimated ± 7.0 pp, low confidence |
| Agents-A1-4B | 52.1% | 73.3% | estimated ± 7.0 pp, low confidence |
| LLaDA2.2-mini | 35.0% | 56.5% | estimated ± 7.0 pp, low confidence |
| Beam | 65.5% | 73.3% | estimated ± 7.0 pp, low confidence |