Mathematical Reasoning on GSM8K (Accuracy, Token Usage, Compression Ratio)
96.3AccuracyChainofDraft
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ChainofDraftBase Model=Qwen3-14B2025.10 | 96.3 | 698 | 44.5 | |
| BeConciseBase Model=Qwen3-14B2025.10 | 96.1 | 1,004 | 64 | |
| WHISPERBase Model=Qwen3-14B2025.10 | 96.1 | 440 | 28.1 | |
| OriginalBase Model=Qwen3-14B2025.10 | 95.9 | 1,568 | 100 | |
| OriginalBase Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 95.8 | 1,540 | 100 | |
| GEPABase Model=Qwen3-14B2025.10 | 95.8 | 751 | 47.9 | |
| DEER*Base Model=Qwen3-14B2025.10 | 95.5 | 934 | 59.6 | |
| BeConciseBase Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 95.2 | 1,265 | 82.1 | |
| NoThinkingBase Model=Qwen3-14B2025.10 | 94.8 | 289 | 18.4 | |
| WHISPERBase Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 94.4 | 655 | 42.5 | |
| OriginalBase Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 92 | 1,715 | 100 | |
| DEER*Base Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 92 | 758 | 49.2 | |
| GEPABase Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 91.9 | 836 | 54.3 | |
| WHISPERBase Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 91.4 | 624 | 36.4 | |
| BeConciseBase Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 91.3 | 1,482 | 86.4 | |
| NoThinkingBase Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 90.7 | 250 | 16.2 | |
| ChainofDraftBase Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 90.2 | 1,101 | 64.2 | |
| GEPABase Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 89.8 | 1,045 | 60.9 | |
| ChainofDraftBase Model=DeepSeek-R1-Distill-Qwen-14B2025.10 | 89.1 | 501 | 30 | |
| DEER*Base Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 88.7 | 909 | 53 | |
| NoThinkingBase Model=DeepSeek-R1-Distill-LLaMA-8B2025.10 | 80 | 266 | 15.5 |