Math Reasoning on Olympiad (Acc@1)
59.05Accuracy@1Reward Shaping
Evaluation results
| Method | Links | |
|---|---|---|
| Reward ShapingBackbone=Qwen3-8B-Base2026.03 | 59.05 | |
| Ctrl-RBackbone=Qwen3-8B-Base, beta=0.22026.03 | 58.6 | |
| NL GuidanceBackbone=Qwen3-8B-Base2026.03 | 57.27 | |
| DAPOBackbone=Qwen3-8B-Base2026.03 | 57.12 | |
| NL GuidanceBackbone=Qwen3-1.7B-Base2026.03 | 33.82 | |
| Ctrl-RBackbone=Qwen3-1.7B-Base, beta=0.22026.03 | 33.23 | |
| DAPOBackbone=Qwen3-1.7B-Base2026.03 | 32.78 | |
| Reward ShapingBackbone=Qwen3-1.7B-Base2026.03 | 32.78 | |
| Ctrl-RBackbone=Qwen3-1.7B-Base, beta=1, Importance Sampling (IS)=true2026.03 | 32.78 | |
| Ctrl-RBackbone=Qwen3-1.7B-Base, beta=0, Importance Sampling (IS)=false2026.03 | 32.04 |