Mathematical Reasoning on MATH 500 (Avg@4)
83.1Avg@4GSPO + CBS
Evaluation results
| Method | Links | |
|---|---|---|
| GSPO + CBSBackbone=Qwen3-8B-Base2026.02 | 83.1 | |
| GSPO + CBS*Backbone=Qwen3-8B-Base2026.02 | 79.65 | |
| GSPOBackbone=Qwen3-8B-Base2026.02 | 79.2 | |
| DAPO + CBSBackbone=Qwen3-8B-Base2026.02 | 78.25 | |
| GSPO + CBS*Backbone=Qwen3-4B-Base2026.02 | 77.85 | |
| DAPOBackbone=Qwen3-8B-Base2026.02 | 77.65 | |
| GSPOBackbone=Qwen3-4B-Base2026.02 | 77.55 | |
| GRPO + CBSBackbone=Qwen3-8B-Base2026.02 | 77.45 | |
| GRPO + CBS*Backbone=Qwen3-8B-Base2026.02 | 77.45 | |
| DAPO + CBS*Backbone=Qwen3-8B-Base2026.02 | 76.85 | |
| DAPO + CBS*Backbone=Qwen3-4B-Base2026.02 | 76.3 | |
| GRPOBackbone=Qwen3-8B-Base2026.02 | 76.15 | |
| GSPO + CBSBackbone=Qwen3-4B-Base2026.02 | 76.05 | |
| DAPO + CBSBackbone=Qwen3-4B-Base2026.02 | 75.85 | |
| GRPO + CBS*Backbone=Qwen3-4B-Base2026.02 | 75.55 | |
| DAPOBackbone=Qwen3-4B-Base2026.02 | 75.05 | |
| GRPO + CBSBackbone=Qwen3-4B-Base2026.02 | 74.45 | |
| GRPOBackbone=Qwen3-4B-Base2026.02 | 73.7 | |
| GSPO + CBS*Backbone=Qwen3-1.7B-Base2026.02 | 67.5 | |
| GSPOBackbone=Qwen3-1.7B-Base2026.02 | 65.95 | |
| GSPO + CBSBackbone=Qwen3-1.7B-Base2026.02 | 65.7 | |
| GRPO + CBS*Backbone=Qwen3-1.7B-Base2026.02 | 64.8 | |
| GRPO + CBSBackbone=Qwen3-1.7B-Base2026.02 | 63.9 | |
| DAPO + CBSBackbone=Qwen3-1.7B-Base2026.02 | 63.65 | |
| DAPO + CBS*Backbone=Qwen3-1.7B-Base2026.02 | 62.45 | |
| GRPOBackbone=Qwen3-1.7B-Base2026.02 | 60.95 | |
| DAPOBackbone=Qwen3-1.7B-Base2026.02 | 58.7 | |
| Base ModelBackbone=Qwen3-8B-Base2026.02 | 53.25 | |
| Base ModelBackbone=Qwen3-4B-Base2026.02 | 31.9 | |
| Base ModelBackbone=Qwen3-1.7B-Base2026.02 | 29.1 |