Mathematical Reasoning on GSM8K (Accuracy, Tokens, CR, TE)
0.9303AccuracyOriginal
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| OriginalModel=Qwen2.5-14B, Compression Setting=None2026.01 | 0.9303 | 313.94 | 1 | 29.63 | |
| OriginalModel=Qwen2.5-7B, Compression Setting=None2026.01 | 0.9158 | 299.22 | 1 | 3.27 | |
| CtrlCoTModel=Qwen2.5-14B, Compression Setting=High+2026.01 | 0.9014 | 120.92 | 0.39 | 74.55 | |
| TokenSkipModel=Qwen2.5-14B, Compression Setting=High+2026.01 | 0.884 | 135.74 | 0.43 | 65.12 | |
| CtrlCoTModel=Qwen2.5-14B, Compression Setting=High++2026.01 | 0.8749 | 100.41 | 0.32 | 87.13 | |
| TokenSkipModel=Qwen2.5-14B, Compression Setting=High++2026.01 | 0.8446 | 106.01 | 0.34 | 79.67 | |
| OriginalModel=Qwen2.5-3B, Compression Setting=None2026.01 | 0.8324 | 316.94 | 1 | 3.81 | |
| CtrlCoTModel=Qwen2.5-7B, Compression Setting=High+2026.01 | 0.8317 | 109.8 | 0.37 | 75.74 | |
| CtrlCoTModel=Qwen2.5-14B, Compression Setting=High+++2026.01 | 0.7983 | 67.97 | 0.22 | 117.44 | |
| TokenSkipModel=Qwen2.5-7B, Compression Setting=High+2026.01 | 0.793 | 130.45 | 0.44 | 60.79 | |
| CtrlCoTModel=Qwen2.5-7B, Compression Setting=High++2026.01 | 0.7908 | 83.36 | 0.28 | 94.86 | |
| TokenSkipModel=Qwen2.5-14B, Compression Setting=High+++2026.01 | 0.7718 | 75.95 | 0.24 | 101.62 | |
| CtrlCoTModel=Qwen2.5-3B, Compression Setting=High+2026.01 | 0.771 | 125.44 | 0.4 | 61.47 | |
| CtrlCoTModel=Qwen2.5-3B, Compression Setting=High++2026.01 | 0.743 | 106.24 | 0.34 | 69.93 | |
| CtrlCoTModel=Qwen2.5-7B, Compression Setting=High+++2026.01 | 0.7369 | 58.83 | 0.2 | 125.27 | |
| TokenSkipModel=Qwen2.5-7B, Compression Setting=High++2026.01 | 0.724 | 93.01 | 0.31 | 77.85 | |
| TokenSkipModel=Qwen2.5-3B, Compression Setting=High+2026.01 | 0.6801 | 134.17 | 0.42 | 50.69 | |
| CtrlCoTModel=Qwen2.5-3B, Compression Setting=High+++2026.01 | 0.6338 | 74.4 | 0.23 | 85.19 | |
| TokenSkipModel=Qwen2.5-7B, Compression Setting=High+++2026.01 | 0.6118 | 66.95 | 0.22 | 91.38 | |
| TokenSkipModel=Qwen2.5-3B, Compression Setting=High++2026.01 | 0.5868 | 108.07 | 0.34 | 54.3 | |
| TokenSkipModel=Qwen2.5-3B, Compression Setting=High+++2026.01 | 0.4807 | 80.7 | 0.25 | 59.56 | |
| TruncationModel=Qwen2.5-7B, Compression Setting=High+2026.01 | 0.2328 | 201.79 | 0.67 | 11.54 | |
| TruncationModel=Qwen2.5-3B, Compression Setting=High+2026.01 | 0.1835 | 202.69 | 0.64 | 9.05 | |
| TruncationModel=Qwen2.5-14B, Compression Setting=High+2026.01 | 0.1835 | 203 | 0.65 | 9.04 | |
| TruncationModel=Qwen2.5-7B, Compression Setting=High++2026.01 | 0.0675 | 153.81 | 0.51 | 4.39 | |
| TruncationModel=Qwen2.5-3B, Compression Setting=High++2026.01 | 0.0614 | 153.92 | 0.49 | 3.99 | |
| TruncationModel=Qwen2.5-14B, Compression Setting=High++2026.01 | 0.0569 | 153.93 | 0.49 | 3.69 | |
| TruncationModel=Qwen2.5-3B, Compression Setting=High+++2026.01 | 0.0288 | 102 | 0.32 | 2.82 | |
| TruncationModel=Qwen2.5-7B, Compression Setting=High+++2026.01 | 0.0258 | 102 | 0.34 | 2.53 | |
| TruncationModel=Qwen2.5-14B, Compression Setting=High+++2026.01 | 0.0205 | 102 | 0.32 | 2.01 |