Mathematical Reasoning on Minerva Math (avg@8, pass@8)
54.3Average Score (avg@8)ASPO-Math-30B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ASPO-Math-30BBackbone=Qwen3-30B-A3B2025.10 | 54.3 | 64.3 | |
| GRPOBackbone=Qwen3-30B-A3B2025.10 | 51.9 | 61.8 | |
| ASPO-Math-8BBackbone=Qwen3-8B2025.10 | 50.6 | 60.7 | |
| GRPOBackbone=Qwen3-8B2025.10 | 50.2 | 61.4 | |
| ASPO-Math-4BBackbone=Qwen3-4B2025.10 | 49 | 60.7 | |
| Qwen3-30B-A3BBackbone=Qwen3-30B-A3B2025.10 | 47.7 | 59.6 | |
| GRPOBackbone=Qwen3-4B2025.10 | 47.2 | 58.5 | |
| Qwen3-8BBackbone=Qwen3-8B2025.10 | 43.9 | 58.1 | |
| Qwen3-4BBackbone=Qwen3-4B2025.10 | 41.5 | 56.3 | |
| Nemotron-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.10 | 35.3 | 47.8 | |
| ASPO-Math-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.10 | 35.1 | 50.4 | |
| DeepScaleR-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.10 | 30.3 | 51.1 | |
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.10 | 29.2 | 46.3 | |
| DeepSeek-R1-1.5BBackbone=DeepSeek-R1-Distill-Qwen-1.5B2025.10 | 27.6 | 48.2 |