Math Reasoning on AIME 2024 (Acc@16)
40Accuracy @16Ctrl-R
Evaluation Results
| Method | Links | |
|---|---|---|
| Ctrl-RBackbone=Qwen3-8B-Base, beta=0.22026.03 | 40 | |
| DAPOBackbone=Qwen3-8B-Base2026.03 | 36.04 | |
| Reward ShapingBackbone=Qwen3-8B-Base2026.03 | 34.79 | |
| NL GuidanceBackbone=Qwen3-8B-Base2026.03 | 34.58 | |
| Ctrl-RBackbone=Qwen3-1.7B-Base, beta=0.22026.03 | 15.21 | |
| Reward ShapingBackbone=Qwen3-1.7B-Base2026.03 | 14.37 | |
| NL GuidanceBackbone=Qwen3-1.7B-Base2026.03 | 14.16 | |
| DAPOBackbone=Qwen3-1.7B-Base2026.03 | 13.12 | |
| Ctrl-RBackbone=Qwen3-1.7B-Base, beta=1, Importance Sampling (IS)=true2026.03 | 12.29 | |
| Ctrl-RBackbone=Qwen3-1.7B-Base, beta=0, Importance Sampling (IS)=false2026.03 | 12.08 |