Mathematical Reasoning on AIME 2025 (Mean@32 accuracy)
66Mean@32 AccuracyHTPO
Evaluation Results
| Method | Links | |
|---|---|---|
| HTPOBackbone=Qwen3-8B-Instruct2026.05 | 66 | |
| DAPOBackbone=Qwen3-8B-Instruct2026.05 | 64.7 | |
| GSPOBackbone=Qwen3-8B-Instruct2026.05 | 63.7 | |
| SAPOBackbone=Qwen3-8B-Instruct2026.05 | 63.6 | |
| 80/20-RuleBackbone=Qwen3-8B-Instruct2026.05 | 63.4 | |
| BAPOBackbone=Qwen3-8B-Instruct2026.05 | 59.2 | |
| GRPO†Backbone=Qwen3-8B-Instruct2026.05 | 58.6 | |
| HTPOBackbone=Qwen3-8B-Base2026.05 | 30.4 | |
| GSPOBackbone=Qwen3-8B-Base2026.05 | 30.3 | |
| SAPOBackbone=Qwen3-8B-Base2026.05 | 28.4 | |
| 80/20-RuleBackbone=Qwen3-8B-Base2026.05 | 25.9 | |
| BAPOBackbone=Qwen3-8B-Base2026.05 | 24.7 | |
| DAPOBackbone=Qwen3-8B-Base2026.05 | 23.7 | |
| GRPO†Backbone=Qwen3-8B-Base2026.05 | 22.9 |