Mathematical Reasoning on OlympiadBench (pass@4)
67.6Pass@4ThinkTwice
Evaluation results
| Method | Links | |
|---|---|---|
| ThinkTwiceBackbone=Qwen3-4B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 67.6 | |
| DAPOBackbone=Qwen3-4B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 67.5 | |
| DrGRPOBackbone=Qwen3-4B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 66.51 | |
| ThinkTwiceBackbone=OLMo3-7B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 65.81 | |
| DrGRPOBackbone=OLMo3-7B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 65.09 | |
| GRPOBackbone=Qwen3-4B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 63.8 | |
| DAPOBackbone=OLMo3-7B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 63.8 | |
| GRPOBackbone=OLMo3-7B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 62.42 | |
| Base ModelBackbone=OLMo3-7B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 61.36 | |
| Base ModelBackbone=Qwen3-4B, Evaluation Protocol=Direct prompting, Sampling count (n)=322026.04 | 57.9 |