Question Answering on MMLU-Pro n=8312 (test) (Act, Acc, WA)
97.6Activation (Act)CRC
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CRCbeta=0.15, T=42026.06 | 97.6 | 84.8 | 14.8 | |
| Selective Predictionbeta=0.15, T=42026.06 | 97.3 | 85.1 | 14.5 | |
| Isotonic Confidencebeta=0.15, T=42026.06 | 97.3 | 85 | 14.6 | |
| Learned Stopperbeta=0.15, T=42026.06 | 96.6 | 83.6 | 16.4 | |
| Calibrated Learnedbeta=0.15, T=42026.06 | 95.9 | 85.4 | 14 | |
| Consensusbeta=0.15, T=42026.06 | 94.5 | 84.5 | 15.5 | |
| Oraclebeta=0.15, T=42026.06 | 86 | 100 | 0 | |
| Confidence Thresholdbeta=0.15, T=42026.06 | 69.8 | 84.5 | 15.5 | |
| kNN no biasbeta=0.15, T=4, ablated_component=no modulus2026.06 | 37.9 | 97.7 | 0.9 | |
| Adaptive Local k-NN Certification (Ours)beta=0.15, T=42026.06 | 29.8 | 98 | 0.6 | |
| Final-roundbeta=0.15, T=4, bound_at=t=T2026.06 | 29.3 | 97.9 | 0.6 |