Mathematical Reasoning on AMC23 (pass@1, pass@256)
46.5Pass@1 ScoreGRPO + Branch
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRPO + BranchBackbone=Qwen2.5-Math-7B-Base2026.05 | 46.5 | 96.5 | |
| GRPOBackbone=Qwen2.5-Math-7B-Base2026.05 | 43 | 96 | |
| GRPO + Forward KLBackbone=Qwen2.5-Math-7B-Base2026.05 | 42 | 97.5 | |
| GRPO w/o KLBackbone=Qwen2.5-Math-7B-Base2026.05 | 41 | 93 | |
| GRPO + RandomBackbone=Qwen2.5-Math-7B-Base2026.05 | 41 | 96.5 | |
| GRPO + TokenBackbone=Qwen2.5-Math-7B-Base2026.05 | 39.5 | 96 | |
| Base ModelBackbone=Qwen2.5-Math-7B-Base2026.05 | 35.5 | 95 |