Mathematical Reasoning on MATH500 (Acc, CI, Token Usage, Flips)
76.3AccuracySelective active verify
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Selective active verifyPolicy strategy=Selective active verify2026.06 | 76.3 | 72.8 | 48.2 | 2,791 | 7,104 | 1 | |
| Long baseGeneration limit=8,1922026.06 | 76 | 72.2 | 0 | 0 | 5,124 | 0 | |
| Always active verifyPolicy strategy=Always active verify2026.06 | 75.5 | 71.9 | 100 | 3,812 | 8,125 | 2.2 | |
| Selective continuePolicy strategy=Selective continue2026.06 | 73.6 | 69.9 | 48.2 | 2,751 | 7,064 | 1.7 | |
| Always continuePolicy strategy=Always continue2026.06 | 72 | 68.4 | 100 | 3,694 | 8,007 | 3.6 | |
| BaseGeneration limit=4,0962026.06 | 59 | 54.8 | 0 | 0 | 4,313 | 0 |