Mathematical Reasoning on GSM8K (Acc, Tokens, Latency, ActRatio)
92.17AccuracyOriginal
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| OriginalBackbone=Qwen2.5-7B Series, Ratio=-2025.10 | 92.17 | 298.63 | 3.831 | — | |
| UCoTBackbone=Qwen2.5-7B Series, Ratio=0.92025.10 | 91.72 | 244.88 | 3.321 | 0.85 | |
| PromptBackbone=Qwen2.5-7B Series, Ratio=0.92025.10 | 91.47 | 281.65 | 3.741 | 0.97 | |
| TokenskipBackbone=Qwen2.5-7B Series, Ratio=0.92025.10 | 91.131 | 277.21 | 3.621 | 0.93 | |
| CoDBackbone=Qwen2.5-7B Series, Ratio=0.92025.10 | 90.981 | 257.77 | 3.431 | 0.89 | |
| PromptBackbone=Qwen2.5-7B Series, Ratio=0.72025.10 | 90.441 | 287.32 | 3.771 | 0.99 | |
| PromptBackbone=Qwen2.5-7B Series, Ratio=0.52025.10 | 89.692 | 283.2 | 3.691 | 0.98 | |
| UCoTBackbone=Qwen2.5-7B Series, Ratio=0.72025.10 | 87.984 | 194.63 | 2.571 | 0.65 | |
| CoDBackbone=Qwen2.5-7B Series, Ratio=0.72025.10 | 87.544 | 223.02 | 2.971 | 0.77 | |
| PromptBackbone=Llama3.1-8B Series, Ratio=0.92025.10 | 87.32 | 207.89 | 2.401 | 0.98 | |
| OriginalBackbone=Llama3.1-8B Series, Ratio=-2025.10 | 87.26 | 212.13 | 2.441 | — | |
| PromptBackbone=Llama3.1-8B Series, Ratio=0.52025.10 | 87.16 | 189.32 | 2.151 | 0.89 | |
| PromptBackbone=Llama3.1-8B Series, Ratio=0.72025.10 | 86.79 | 205.88 | 2.331 | 0.97 | |
| UCoTBackbone=Qwen2.5-7B Series, Ratio=0.52025.10 | 86.555 | 140.36 | 1.862 | 0.48 | |
| UCoTBackbone=Llama3.1-8B Series, Ratio=0.92025.10 | 86.32 | 180.31 | 2.191 | 0.85 | |
| CoDBackbone=Llama3.1-8B Series, Ratio=0.92025.10 | 85.821 | 197.28 | 2.281 | 0.93 | |
| TokenskipBackbone=Llama3.1-8B Series, Ratio=0.92025.10 | 85.791 | 197.16 | 2.371 | 0.94 | |
| TokenskipBackbone=Qwen2.5-7B Series, Ratio=0.72025.10 | 85.696 | 223.83 | 2.871 | 0.75 | |
| CoDBackbone=Qwen2.5-7B Series, Ratio=0.52025.10 | 84.967 | 150.61 | 1.991 | 0.52 | |
| UCoTBackbone=Llama3.1-8B Series, Ratio=0.72025.10 | 84.173 | 140.01 | 1.731 | 0.66 | |
| CoDBackbone=Llama3.1-8B Series, Ratio=0.72025.10 | 83.903 | 154.85 | 1.841 | 0.73 | |
| UCoTBackbone=Llama3.1-8B Series, Ratio=0.52025.10 | 83.623 | 101.82 | 1.261 | 0.48 | |
| TokenskipBackbone=Qwen2.5-7B Series, Ratio=0.52025.10 | 83.478 | 157.58 | 2.071 | 0.53 | |
| TokenskipBackbone=Llama3.1-8B Series, Ratio=0.72025.10 | 82.574 | 161.21 | 1.861 | 0.76 | |
| TokenskipBackbone=Llama3.1-8B Series, Ratio=0.52025.10 | 81.805 | 135.76 | 1.561 | 0.64 | |
| TruncationBackbone=Qwen2.5-7B Series, Ratio=0.92025.10 | 79.6812 | 272.32 | 3.471 | 0.91 | |
| CoDBackbone=Llama3.1-8B Series, Ratio=0.52025.10 | 73.4613 | 110.31 | 1.271 | 0.52 | |
| TruncationBackbone=Llama3.1-8B Series, Ratio=0.92025.10 | 62.5524 | 193.91 | 2.221 | 0.91 | |
| TruncationBackbone=Qwen2.5-7B Series, Ratio=0.72025.10 | 52.4239 | 210.55 | 2.681 | 0.7 | |
| TruncationBackbone=Llama3.1-8B Series, Ratio=0.72025.10 | 27.4859 | 148.49 | 1.721 | 0.7 | |
| TruncationBackbone=Qwen2.5-7B Series, Ratio=0.52025.10 | 12.3079 | 154.68 | 2.011 | 0.52 | |
| TruncationBackbone=Llama3.1-8B Series, Ratio=0.52025.10 | 7.3279 | 108.19 | 1.251 | 0.51 |