Mathematical Reasoning on ASDiv (Acc, Tokens, Latency, ActRatio)
96.58AccuracyUCoT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| UCoTBackbone=Deepseek-R1-Distill-Qwen-7B, Threshold=0.52025.10 | 96.58 | 257.93 | 3.01 | 55 | |
| UCoTBackbone=Deepseek-R1-Distill-Qwen-7B, Threshold=0.72025.10 | 96.01 | 300.12 | 3.52 | 64 | |
| OriginalBackbone=Deepseek-R1-Distill-Qwen-7B2025.10 | 95.47 | 468.94 | 5.48 | — | |
| UCoTBackbone=Deepseek-R1-Distill-Qwen-7B, Threshold=0.92025.10 | 95.43 | 384.61 | 4.52 | 82 | |
| TokenskipBackbone=Deepseek-R1-Distill-Qwen-7B, Threshold=0.92025.10 | 95.18 | 431.55 | 5.07 | 92 | |
| TokenskipBackbone=Deepseek-R1-Distill-Qwen-7B, Threshold=0.72025.10 | 94.76 | 337.63 | 3.95 | 72 | |
| TokenskipBackbone=Deepseek-R1-Distill-Qwen-7B, Threshold=0.52025.10 | 94.33 | 301.32 | 3.52 | 64 | |
| OriginalBackbone=Qwen3-8B2025.10 | 93.74 | 475.49 | 4.72 | — | |
| UCoTBackbone=Qwen3-8B, Threshold=0.92025.10 | 93.53 | 413.67 | 4.12 | 87 | |
| TokenskipBackbone=Qwen3-8B, Threshold=0.72025.10 | 93.48 | 361.37 | 3.59 | 76 | |
| UCoTBackbone=Qwen3-8B, Threshold=0.72025.10 | 93.32 | 304.31 | 3.03 | 64 | |
| UCoTBackbone=Qwen3-8B, Threshold=0.52025.10 | 92.77 | 257.76 | 2.57 | 54 | |
| TokenskipBackbone=Qwen3-8B, Threshold=0.92025.10 | 92.55 | 446.96 | 4.44 | 94 | |
| TokenskipBackbone=Qwen3-8B, Threshold=0.52025.10 | 91.57 | 294.83 | 2.93 | 62 |