Mathematical Reasoning on MATH 500 (Avg@32, Pass@32)
86Avg@32MASPO
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MASPOBackbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 86 | 94.7 | |
| SAPOBackbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 85.5 | 92.9 | |
| Clip HigherBackbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 85.3 | 95.4 | |
| DACBackbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 85.1 | 94.8 | |
| BAPOBackbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 85 | 94.5 | |
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 84.8 | 92.4 | |
| Entropy Adv.Backbone=DeepSeek-R1-Distill-Qwen-7B2026.02 | 84.8 | 92.5 | |
| SAPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 79.1 | 91.4 | |
| Entropy Adv.Backbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 78.5 | 89.7 | |
| DACBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 78.1 | 92.5 | |
| MASPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 78 | 89.7 | |
| GRPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 77.6 | 90.8 | |
| BAPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 74.9 | 91 | |
| Clip HigherBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.02 | 71.7 | 88.6 |