Mathematical Reasoning on GSM8K (Accuracy, Tokens, CR, Think Rate)
95.5AccuracyThinking
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ThinkingModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 95.5 | 1,510 | 100 | 100 | |
| DEERModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 95.5 | 1,059 | 70 | 100 | |
| DAMI-PromptModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 95.5 | 1,028 | 68 | 59 | |
| MI-07Model Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 95.4 | 1,222 | 81 | 100 | |
| MI-05Model Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 95.1 | 1,084 | 72 | 83 | |
| TAModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 95 | 1,225 | 81 | 100 | |
| RoutingModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94.5 | 430 | 28 | 2.7 | |
| DAMI-PrefModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94.5 | 843 | 56 | 25 | |
| DAMI-ConfModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94.5 | 856 | 57 | 41 | |
| DAMI-PredModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94.4 | 365 | 24 | 0 | |
| InstructModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94.3 | 349 | 23.1 | 0 | |
| TIESModel Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94 | 363 | 24 | 0 | |
| MI-03Model Base Pair=Qwen3-4B-2507-Thinking/Qwen3-4B-2507-Instruct, Sampling Rounds=2x2026.01 | 94 | 383 | 25 | 0 | |
| DAMI-PrefModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 92.4 | 1,121 | 59 | 48 | |
| MI-05Model Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 91.9 | 617 | 33 | 71 | |
| DAMI-ConfModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 91.8 | 1,050 | 56 | 55 | |
| MI-07Model Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 91.2 | 902 | 48 | 99 | |
| TAModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 90.7 | 769 | 41 | 99 | |
| ThinkingModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 89.6 | 1,886 | 100 | 100 | |
| DEERModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 89.5 | 873 | 46 | 100 | |
| TIESModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 86.5 | 346 | 18 | 0 | |
| MI-03Model Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 85.5 | 379 | 20 | 0 | |
| DAMI-PredModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 83.5 | 410 | 22 | 0 | |
| DAMI-PromptModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 82.1 | 431 | 23 | 0 | |
| RoutingModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 64.5 | 1,045 | 55 | 0 | |
| InstructModel Base Pair=DeepSeek-R1-Distill-Qwen-7B/Qwen2.5-Math-7B, Sampling Rounds=2x2026.01 | 55.3 | 1,007 | — | 0 |