Calibration
AA-HLE → LABBench2
LABBench2 is estimated from AA-HLE with a linear curve fitted on 7 models measured on both: y = 0.4257·x + 0.6276, R² = 0.64, cross-validated error 2.4 pp. It is used for 11 estimates.
| Estimated model | AA-HLE | LABBench2 | Source |
|---|---|---|---|
| Claude Haiku 5.5 | 44.4% | 81.7% | estimated ± 2.4 pp, medium confidence |
| Claude Sonnet 5.5 | 55.0% | 86.2% | estimated ± 2.4 pp, medium confidence |
| DeepSeek V4.1 Flash | 39.2% | 79.5% | estimated ± 2.4 pp, low confidence |
| GPT-4 Turbo | 3.1% | 64.1% | estimated ± 2.4 pp, low confidence |
| Grok 4.7 | 43.1% | 81.1% | estimated ± 2.4 pp, medium confidence |
| Ling 3.1 Flash | 39.4% | 79.5% | estimated ± 2.4 pp, low confidence |
| Mercury 2.5 | 11.8% | 67.8% | estimated ± 2.4 pp, low confidence |
| MiMo-V2.6-Flash | 35.1% | 77.7% | estimated ± 2.4 pp, low confidence |
| MiMo-V2.6-Pro | 49.4% | 83.8% | estimated ± 2.4 pp, medium confidence |
| Mistral Large 4 | 35.0% | 77.7% | estimated ± 2.4 pp, low confidence |
| Step 5 Preview | 46.5% | 82.6% | estimated ± 2.4 pp, medium confidence |