Mathematical Reasoning on OlympiadBench (Mean@32 accuracy)
60.9Mean@32 AccuracyHTPO
Evaluation results
| Method | Links | |
|---|---|---|
| HTPOBackbone=Qwen3-8B-Instruct2026.05 | 60.9 | |
| 80/20-RuleBackbone=Qwen3-8B-Instruct2026.05 | 60 | |
| SAPOBackbone=Qwen3-8B-Instruct2026.05 | 59.9 | |
| DAPOBackbone=Qwen3-8B-Instruct2026.05 | 59.3 | |
| GSPOBackbone=Qwen3-8B-Instruct2026.05 | 58.4 | |
| BAPOBackbone=Qwen3-8B-Instruct2026.05 | 56.9 | |
| GRPO†Backbone=Qwen3-8B-Instruct2026.05 | 53.5 | |
| HTPOBackbone=Qwen3-8B-Base2026.05 | 47.6 | |
| DAPOBackbone=Qwen3-8B-Base2026.05 | 45.7 | |
| 80/20-RuleBackbone=Qwen3-8B-Base2026.05 | 45.1 | |
| SAPOBackbone=Qwen3-8B-Base2026.05 | 44.8 | |
| GSPOBackbone=Qwen3-8B-Base2026.05 | 44.6 | |
| BAPOBackbone=Qwen3-8B-Base2026.05 | 42.3 | |
| GRPO†Backbone=Qwen3-8B-Base2026.05 | 39.8 |