Mathematical Reasoning on GSM8K (Accuracy and Token Consumption)
97.8AccuracyMAD-M2 (S)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MAD-M2 (S)Backbone=QwQ-32B, Masking Strategy=Subjective2026.03 | 97.8 | 1.32 | |
| CoTBackbone=QwQ-32B2026.03 | 97.4 | 0.08 | |
| CoT-SCBackbone=QwQ-32B2026.03 | 97.4 | 0.44 | |
| MADBackbone=QwQ-32B2026.03 | 97.2 | 1 | |
| MAD-M2 (O)Backbone=QwQ-32B, Masking Strategy=Objective2026.03 | 96.6 | 0.56 | |
| CoT-SCBackbone=Qwen2.5-Math-7B-Instruct2026.03 | 96.4 | 0.45 | |
| MAD-M2 (O)Backbone=Qwen2.5-Math-7B-Instruct, Masking Strategy=Objective2026.03 | 95.4 | 0.6 | |
| CoTBackbone=Qwen2.5-Math-7B-Instruct2026.03 | 95.2 | 0.08 | |
| MADBackbone=Qwen2.5-Math-7B-Instruct2026.03 | 95.2 | 1 | |
| MAD-M2 (S)Backbone=Qwen2.5-Math-7B-Instruct, Masking Strategy=Subjective2026.03 | 95.2 | 1.44 | |
| MADBackbone=Qwen2.5-7B-Instruct2026.03 | 91.8 | 1 | |
| MAD-M2 (S)Backbone=Qwen2.5-7B-Instruct, Masking Strategy=Subjective2026.03 | 89 | 1.25 | |
| MAD-M2 (O)Backbone=Qwen2.5-7B-Instruct, Masking Strategy=Objective2026.03 | 89 | 0.72 | |
| CoT-SCBackbone=DeepSeek-Math-7B-Instruct2026.03 | 88.8 | 0.52 | |
| CoT-SCBackbone=Qwen2.5-7B-Instruct2026.03 | 83.6 | 0.51 | |
| MAD-M2 (O)Backbone=DeepSeek-Math-7B-Instruct, Masking Strategy=Objective2026.03 | 82.2 | 0.71 | |
| MADBackbone=DeepSeek-Math-7B-Instruct2026.03 | 81.2 | 1 | |
| MAD-M2 (S)Backbone=DeepSeek-Math-7B-Instruct, Masking Strategy=Subjective2026.03 | 80.8 | 1.33 | |
| CoTBackbone=DeepSeek-Math-7B-Instruct2026.03 | 79 | 0.09 | |
| CoTBackbone=Qwen2.5-7B-Instruct2026.03 | 64.8 | 0.09 |