Positioning spread
| Benchmark | Task acc | Confidence | F₁ | Leans |
|---|---|---|---|---|
| SQuAD (factual recall) | 0.47 | 0.30 | 0.57 | -0.18 cautious |
| MMLU-Pro (knowledge) | 0.41 | 0.69 | 0.57 | +0.28 overconfident |
| LegalBench (legal reasoning) | 0.84 | 0.51 | 0.63 | -0.33 cautious |
| MathBench (competition math) | 0.83 | 0.91 | 0.91 | +0.08 overconfident |
| OmniMath (advanced math) | 0.34 | 0.43 | 0.57 | +0.10 overconfident |
| SciCode (scientific code) | 0.43 | 0.42 | 0.56 | -0.01 calibrated |
In the full cloud
Qwen 2.5 72B conditions all other model/condition points equal relative confidence and pass rate
Pairwise footprint
Match accuracy controls for the performance base-rate gap
Qwen 2.5 72B pairs
18/ 171
Qwen 2.5 72B mean tau
+0.052
All-pairs mean
+0.037
Qwen 2.5 72B p<0.05
11(61.1%)
all model pairs (observed) base-rate-matched null calibration-preserving null Qwen 2.5 72B pair (filled = p<0.05) Qwen 2.5 72B mean all-pairs mean