Math Reasoning on AMC23, AMC24, AIME24, AIME25, MATH500
93.1AMC23 Scoreµ-GRPO
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| µ-GRPOBackbone Model=DEEPSEEK-7B, µ (Staleness level)=10242026.05 | 93.1 | 68.9 | 50.2 | 37.9 | 92.8 | 68.6 | 21.4 | 51.3 | |
| GRPOBackbone Model=DEEPSEEK-7B, µ (Staleness level)=42026.05 | 91.4 | 66.7 | 46.8 | 33.2 | 92.4 | 66.1 | 26 | 61.9 | |
| M2POBackbone Model=DEEPSEEK-7B, µ (Staleness level)=10242026.05 | 90.6 | 69.4 | 50.8 | 35 | 93 | 67.8 | 27.9 | 68.3 | |
| GRPOBackbone Model=DEEPSEEK-7B, µ (Staleness level)=10242026.05 | 82.6 | 63.5 | 39.5 | 26.8 | 88.7 | 60.2 | 22.3 | 54.5 | |
| µ-GRPOBackbone Model=QWEN3-BASE-8B, µ (Staleness level)=10242026.05 | 69.4 | 51.1 | 21.5 | 17.4 | 81.8 | 48.2 | 7.9 | 25.9 | |
| GRPOBackbone Model=QWEN3-BASE-8B, µ (Staleness level)=42026.05 | 68.8 | 53.3 | 23.5 | 19.4 | 86.3 | 50.3 | 21.3 | 58.2 | |
| GRPOBackbone Model=QWEN2.5-MATH-7B, µ (Staleness level)=42026.05 | 68.1 | 42.8 | 29 | 15.8 | 81 | 47.3 | 8.8 | 23.7 | |
| M2POBackbone Model=QWEN2.5-MATH-7B, µ (Staleness level)=10242026.05 | 66.9 | 40.6 | 29.6 | 14.6 | 82.5 | 46.8 | 6.7 | 20.2 | |
| M2POBackbone Model=QWEN3-BASE-8B, µ (Staleness level)=10242026.05 | 65.6 | 53.9 | 24.2 | 19.6 | 87.1 | 50.1 | 23.2 | 49.7 | |
| µ-GRPOBackbone Model=QWEN2.5-MATH-7B, µ (Staleness level)=10242026.05 | 63.1 | 47.2 | 31.5 | 16.5 | 81.7 | 48 | 4.7 | 18.1 | |
| GRPOBackbone Model=QWEN2.5-MATH-7B, µ (Staleness level)=10242026.05 | 61.3 | 36.2 | 23.3 | 9.8 | 78.2 | 41.8 | 3.8 | 17.9 | |
| GRPOBackbone Model=QWEN3-BASE-8B, µ (Staleness level)=10242026.05 | 59.6 | 43.2 | 17.6 | 11.2 | 74.6 | 41.2 | 7.4 | 23.6 | |
| GRPOBackbone Model=QWEN3-BASE-1.7B, µ (Staleness level)=42026.05 | 42.5 | 27.2 | 7.5 | 3.8 | 66.6 | 29.5 | 11.2 | 20.5 | |
| µ-GRPOBackbone Model=QWEN3-BASE-1.7B, µ (Staleness level)=10242026.05 | 40.8 | 24.9 | 7.5 | 6.5 | 68.7 | 29.7 | 5.2 | 12.4 | |
| M2POBackbone Model=QWEN3-BASE-1.7B, µ (Staleness level)=10242026.05 | 35 | 24.4 | 10 | 4.4 | 66.6 | 28.1 | 12.9 | 24.2 | |
| GRPOBackbone Model=QWEN3-BASE-1.7B, µ (Staleness level)=10242026.05 | 32.5 | 19.3 | 7.7 | 4.5 | 63.5 | 25.5 | 6.5 | 13.3 | |
| M2POBackbone Model=LLAMA-3.2-3B-INSTRUCT, µ (Staleness level)=10242026.05 | 30.6 | 15 | 10.6 | 0.2 | 50.7 | 21.4 | 15.3 | 21.3 | |
| µ-GRPOBackbone Model=LLAMA-3.2-3B-INSTRUCT, µ (Staleness level)=10242026.05 | 29.4 | 17.7 | 14 | 0.5 | 52.8 | 22.9 | 12.7 | 18.6 | |
| GRPOBackbone Model=LLAMA-3.2-3B-INSTRUCT, µ (Staleness level)=42026.05 | 25 | 20.6 | 12.3 | 0.2 | 55.1 | 22.6 | 15.1 | 22.7 | |
| GRPOBackbone Model=LLAMA-3.2-3B-INSTRUCT, µ (Staleness level)=10242026.05 | 21.9 | 12.2 | 9.8 | 0.6 | 46.8 | 18.3 | 13.2 | 19.3 |