Mathematical Problem Solving on ORZ-57K, AIME24/25, Minerva, Math500, AMC, OlympiadBench (test)
31.66Accuracy (Minerva)ACPO
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| ACPOBackbone=Qwen2.5-7B, Policy Regime=OffP.2026.06 | 31.66 | 80.48 | 51.2 | 16.35 | 10.1 | 42.17 | 38.66 | |
| ACPOBackbone=Qwen2.5-7B, Policy Regime=N-OnP.2026.06 | 31.02 | 79.78 | 54.37 | 16.77 | 14.37 | 42.83 | 39.86 | |
| Base ModelBackbone=Qwen2.5-7B, Policy Regime=N-OnP.2026.06 | 13.32 | 41.48 | 26.2 | 0.83 | 0.62 | 18.02 | 16.75 |