Mathematical Reasoning on AIME24, AIME25, MATH500, AMC23, GSM8K, Minerva
15.42AIME24 AccuracyAsymPO
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| AsymPOBackbone=LLaMA-3.2-3B-Instruct, Evaluation Rollouts=82026.06 | 15.42 | 3.33 | 49.95 | 23.91 | 79.48 | 18.08 | 31.7 | |
| GRPOBackbone=Qwen3-4B-Base, Evaluation Rollouts=82026.06 | 13.33 | 11.25 | 78.72 | 44.84 | 90.63 | 34.33 | 45.52 | |
| AsymPOBackbone=Qwen3-4B-Base, Evaluation Rollouts=82026.06 | 12.5 | 9.16 | 77.53 | 44.57 | 91.15 | 35.48 | 45.07 | |
| SPOBackbone=Qwen3-4B-Base, Evaluation Rollouts=82026.06 | 12.08 | 8.75 | 77.45 | 48.37 | 91.09 | 34.93 | 45.45 | |
| GRPOBackbone=LLaMA-3.2-3B-Instruct, Evaluation Rollouts=82026.06 | 10 | 3.33 | 49.37 | 23.37 | 80.88 | 18.08 | 30.84 | |
| SPOBackbone=LLaMA-3.2-3B-Instruct, Evaluation Rollouts=82026.06 | 8.33 | 3.33 | 48.53 | 20.92 | 79.56 | 16.96 | 29.61 | |
| AsymPOBackbone=Qwen3-1.7B-Base, Evaluation Rollouts=82026.06 | 7.5 | 4.16 | 66.2 | 33.7 | 82.87 | 27.44 | 36.98 | |
| GRPOBackbone=Qwen3-1.7B-Base, Evaluation Rollouts=82026.06 | 6.67 | 4.58 | 65.87 | 37.5 | 81.79 | 28.26 | 37.45 | |
| SPOBackbone=Qwen3-1.7B-Base, Evaluation Rollouts=82026.06 | 6.25 | 4.16 | 64.85 | 33.7 | 83.14 | 26.33 | 36.41 |