Mathematical Reasoning on MATH, AIME25, AMC, MINERVA, KAOYAN, OLYMPIAD, CN_MATH24 Suite
91.1MATH AccuracyLocal Highest
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Local HighestStudent Model=Qwen2.5-32B-Instruct, Selection Strategy=Local Average Log Probability, Decoding Strategy=non-greedy decoding, Temperature=0.6, Top-p=0.95, Samples=8, Prompt Set=LIMO2025.10 | 91.1 | 66.2 | 96.9 | 36.1 | 65.8 | 67.5 | 85 | 72.7 | |
| Global HighestStudent Model=Qwen2.5-32B-Instruct, Selection Strategy=Global Average Log Probability, Decoding Strategy=non-greedy decoding, Temperature=0.6, Top-p=0.95, Samples=8, Prompt Set=LIMO2025.10 | 86.2 | 44.2 | 90.3 | 33.8 | 62.9 | 63.4 | 72.1 | 64.7 | |
| All 3 TeachersStudent Model=Qwen2.5-32B-Instruct, Selection Strategy=All 3 Teachers, Decoding Strategy=non-greedy decoding, Temperature=0.6, Top-p=0.95, Samples=8, Prompt Set=LIMO2025.10 | 83.5 | 40 | 88.7 | 33.5 | 57.8 | 55.9 | 58.3 | 59.7 | |
| Original ModelStudent Model=Qwen2.5-32B-Instruct, Decoding Strategy=non-greedy decoding, Temperature=0.6, Top-p=0.95, Samples=8, Prompt Set=LIMO2025.10 | 82.6 | 12.1 | 71.5 | 29.5 | 41.6 | 46.1 | 23.7 | 43.9 |