Reasoning Fidelity on AIME 2025 (ref6)
0.917Top-1 AccuracyCASK
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| CASKBudget=5122026.04 | 0.917 | 0.996 | 0.254 | 17.5 | 0.372 | |
| CASKBudget=3842026.04 | 0.903 | 0.991 | 0.313 | 18.2 | 0.633 | |
| TriAttentionBudget=5122026.04 | 0.899 | 0.99 | 0.321 | 17.5 | 0.448 | |
| TriAttentionBudget=3842026.04 | 0.891 | 0.988 | 0.357 | 18.2 | 0.595 | |
| CASKBudget=2562026.04 | 0.868 | 0.979 | 0.504 | 17.5 | 0.559 | |
| TriAttentionBudget=2562026.04 | 0.857 | 0.978 | 0.5 | 17.5 | 0.634 |