Mathematical Reasoning on GSM8K (Accuracy, Tokens, ActRatio)
103.69TokensTruncation
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TruncationBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.52025.08 | 103.69 | 0.49 | 7 | |
| ASCoTBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.52025.08 | 118.57 | 0.55 | 79.5 | |
| ASCoTBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.62025.08 | 128.38 | 0.6 | 81.6 | |
| TruncationBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.72025.08 | 149.97 | 0.7 | 25.9 | |
| ASCoTBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.72025.08 | 151.32 | 0.71 | 83.7 | |
| ASCoTBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.82025.08 | 169.96 | 0.79 | 85.1 | |
| PromptBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.52025.08 | 188.82 | 0.89 | 83.7 | |
| ASCoTBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.92025.08 | 197.65 | 0.92 | 86.1 | |
| TruncationBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.92025.08 | 202.06 | 0.95 | 70.2 | |
| PromptBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.72025.08 | 209.39 | 0.98 | 84.9 | |
| Extra-CoT (CHRPO)Ratio=<POLICY>, Backbone=Qwen3-1.7B2026.02 | 210 | 0.24 | 85.8 | |
| OriginalBackbone=LLaMA-3.1-8B-Instruct, Ratio=1.02025.08 | 213.17 | 1 | 86.2 | |
| ASCoTBackbone=LLaMA-3.1-8B-Instruct, Ratio=1.02025.08 | 214.24 | 1 | 86.9 | |
| PromptBackbone=LLaMA-3.1-8B-Instruct, Ratio=0.92025.08 | 226.37 | 1.06 | 84.1 | |
| TokenSkipRatio=0.2, Backbone=Qwen3-1.7B2026.02 | 273 | 0.3 | 59.1 | |
| Extra-CoTRatio=0.2, Backbone=Qwen3-1.7B2026.02 | 303 | 0.34 | 80.2 | |
| Thinkless* (DeGRPO)Ratio=–, Backbone=Qwen3-1.7B2026.02 | 356 | 0.41 | 85.5 | |
| Extra-CoTRatio=0.4, Backbone=Qwen3-1.7B2026.02 | 469 | 0.53 | 82.3 | |
| TokenSkipRatio=<POLICY>, Backbone=Qwen3-1.7B2026.02 | 473 | 0.52 | 70 | |
| TokenSkipRatio=0.4, Backbone=Qwen3-1.7B2026.02 | 516 | 0.56 | 74.2 | |
| Extra-CoTRatio=<POLICY>, Backbone=Qwen3-1.7B2026.02 | 569 | 0.65 | 86.3 | |
| Extra-CoTRatio=0.6, Backbone=Qwen3-1.7B2026.02 | 641 | 0.73 | 85.5 | |
| TruncationRatio=0.6, Backbone=Qwen3-1.7B2026.02 | 692 | 0.8 | 81.7 | |
| LC-PromptRatio=0.6, Backbone=Qwen3-1.7B2026.02 | 763 | 0.88 | 86.2 | |
| TokenSkipRatio=0.6, Backbone=Qwen3-1.7B2026.02 | 770 | 0.84 | 78.5 | |
| TokenSkipRatio=0.8, Backbone=Qwen3-1.7B2026.02 | 794 | 0.91 | 84.3 | |
| Extra-CoTRatio=0.8, Backbone=Qwen3-1.7B2026.02 | 807 | 0.92 | 86.7 | |
| Base ModelBackbone=Qwen3-1.7B2026.02 | 873 | — | 86.8 | |
| Extra-CoTRatio=1.0, Backbone=Qwen3-1.7B2026.02 | 902 | 1.03 | 86.7 | |
| TokenSkipRatio=1.0, Backbone=Qwen3-1.7B2026.02 | 916 | 1.05 | 83.7 |