Mathematical Reasoning on Math500 (ACC, Tokens)
92.35AccuracyQwQ-32B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| QwQ-32Bstrategy=Full COT2025.08 | 92.35 | 5,955.39 | |
| QwQ-32Bstrategy=step-entropy pruning, pruning_ratio=80%2025.08 | 91.75 | 3,978.06 | |
| Qwen3-8Bstrategy=Full COT2025.08 | 91.37 | 7,138.49 | |
| Qwen3-8Bstrategy=step-entropy pruning, pruning_ratio=80%2025.08 | 91.13 | 5,209.24 | |
| DeepSeek-R1-7Bstrategy=Full COT2025.08 | 88.17 | 3,703.83 | |
| DeepSeek-R1-7Bstrategy=step-entropy pruning, pruning_ratio=80%2025.08 | 88.17 | 2,604.23 | |
| DeepSeek-R1-14Bstrategy=Full COT2025.08 | 84.37 | 2,853.73 | |
| DeepSeek-R1-14Bstrategy=step-entropy pruning, pruning_ratio=80%2025.08 | 82.16 | 1,980.97 | |
| LC-PromptCompression Type=Chain-of-Thought compression2025.08 | -1.6 | -9.8 | |
| R1-CompressCompression Type=Chain-of-Thought compression2025.08 | -3.2 | -20.3 | |
| SFT + RLCompression Type=Chain-of-Thought compression2025.08 | -3.2 | -35 | |
| TokenSkipCompression Type=Chain-of-Thought compression2025.08 | -5.2 | -11.1 | |
| Token-efficient PromptsCompression Type=Chain-of-Thought compression2025.08 | -6.4 | -11.1 | |
| CoT-ValveCompression Type=Chain-of-Thought compression2025.08 | -10.6 | -48.4 |