Mathematical Reasoning on AIME 2022-2024 (Accuracy, Length, Speed, Modification)
65.6Accuracy32B model
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| 32B modelSpeculative Model=32B, Protocol=pass@12025.04 | 65.6 | — | 12,274.3 | — | 30 | — | — | |
| 14B modelSpeculative Model=14B, Protocol=pass@12025.04 | 60 | — | 12,600.2 | — | 49.9 | — | — | |
| Speculative ThinkingSpeculative Model=7B, Target Model=32B, Protocol=pass@12025.04 | 53.3 | 4.4 | 13,213.6 | -0.3 | 41 | 36.8 | 18 | |
| 7B modelSpeculative Model=7B, Protocol=pass@12025.04 | 48.9 | — | 13,250.4 | — | 56.4 | — | — | |
| Speculative ThinkingSpeculative Model=1.5B, Target Model=14B, Protocol=pass@12025.04 | 33.3 | 7.7 | 16,691.2 | -6.2 | 110.3 | 121.1 | 18 | |
| Speculative ThinkingSpeculative Model=1.5B, Target Model=32B, Protocol=pass@12025.04 | 32.2 | 6.6 | 15,706.1 | -11.7 | 85.8 | 185.9 | 19 | |
| Deepseek-Distilled Qwen-2.5-1.5BSpeculative Model=1.5B, Protocol=pass@12025.04 | 25.6 | — | 17,800 | — | 198.9 | — | — | |
| RoPE theta × 1Operation=RoPE theta × 12025.05 | 8.44 | — | 8,257 | — | — | — | — | |
| 0.7 × RoPE theta × 16 + 0.3 × 1M-ModelOperation=0.7 × RoPE theta × 16 + 0.3 × 1M-Model, recipe=LCTMerge2025.05 | 7.11 | — | 6,859 | — | — | — | — | |
| RoPE theta × 16Operation=RoPE theta × 162025.05 | 3.78 | — | 13,411 | — | — | — | — |