Mathematical Reasoning on Minerva (Mean@32 accuracy)
36.5Mean@32 AccuracyBAPO
Evaluation Results
| Method | Links | |
|---|---|---|
| BAPOBackbone=Qwen3-8B-Instruct2026.05 | 36.5 | |
| HTPOBackbone=Qwen3-8B-Instruct2026.05 | 36.5 | |
| GSPOBackbone=Qwen3-8B-Instruct2026.05 | 35.9 | |
| GRPO†Backbone=Qwen3-8B-Instruct2026.05 | 34.8 | |
| 80/20-RuleBackbone=Qwen3-8B-Instruct2026.05 | 34.3 | |
| DAPOBackbone=Qwen3-8B-Instruct2026.05 | 33.8 | |
| SAPOBackbone=Qwen3-8B-Instruct2026.05 | 33.1 | |
| GSPOBackbone=Qwen3-8B-Base2026.05 | 30.3 | |
| HTPOBackbone=Qwen3-8B-Base2026.05 | 30.3 | |
| SAPOBackbone=Qwen3-8B-Base2026.05 | 30.1 | |
| DAPOBackbone=Qwen3-8B-Base2026.05 | 29.4 | |
| GRPO†Backbone=Qwen3-8B-Base2026.05 | 28.5 | |
| BAPOBackbone=Qwen3-8B-Base2026.05 | 28.3 | |
| 80/20-RuleBackbone=Qwen3-8B-Base2026.05 | 27.6 |