Speculative Decoding on Math
7.69Match RateHedgeSpec
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| HedgeSpecBackbone=LLaMA-3.1-8B-IT, Inference Setting=Greedy2025.10 | 7.69 | 98.63 | |
| HedgeSpecTarget Model=LLaMA-3.1-8B-IT, Inference Mode=Non-greedy2025.10 | 7.32 | 88.58 | |
| HedgeSpecBackbone=Qwen-3-8B, Inference Setting=Greedy2025.10 | 7.27 | 79.55 | |
| HedgeSpecTarget Model=Qwen-3-8B, Inference Mode=Non-greedy2025.10 | 7.09 | 76.9 | |
| HedgeSpecBackbone=Qwen-3-32B, Inference Setting=Greedy2025.10 | 6.96 | 45.14 | |
| HedgeSpecTarget Model=Qwen-3-32B, Inference Mode=Non-greedy2025.10 | 6.71 | 47.37 | |
| EagleBackbone=LLaMA-3.1-8B-IT, Inference Setting=Greedy2025.10 | 5.88 | 76.35 | |
| Vanilla EagleTarget Model=LLaMA-3.1-8B-IT, Inference Mode=Non-greedy2025.10 | 5.78 | 74.52 | |
| UCBSpecBackbone=LLaMA-3.1-8B-IT, Inference Setting=Greedy2025.10 | 5.75 | 77.57 | |
| EXP3SpecBackbone=LLaMA-3.1-8B-IT, Inference Setting=Greedy2025.10 | 5.59 | 74.21 | |
| UCBSpecTarget Model=LLaMA-3.1-8B-IT, Inference Mode=Non-greedy2025.10 | 5.5 | 72.61 | |
| EXP3SpecTarget Model=LLaMA-3.1-8B-IT, Inference Mode=Non-greedy2025.10 | 5.33 | 69.6 | |
| UCBSpecBackbone=Qwen-3-8B, Inference Setting=Greedy2025.10 | 5.3 | 61.28 | |
| UCBSpecTarget Model=Qwen-3-8B, Inference Mode=Non-greedy2025.10 | 5.25 | 56.77 | |
| UCBSpecBackbone=Qwen-3-32B, Inference Setting=Greedy2025.10 | 5.24 | 38.69 | |
| UCBSpecTarget Model=Qwen-3-32B, Inference Mode=Non-greedy2025.10 | 5.22 | 38.04 | |
| EXP3SpecTarget Model=Qwen-3-8B, Inference Mode=Non-greedy2025.10 | 5.2 | 55.19 | |
| EXP3SpecBackbone=Qwen-3-32B, Inference Setting=Greedy2025.10 | 5.13 | 37.42 | |
| EXP3SpecBackbone=Qwen-3-8B, Inference Setting=Greedy2025.10 | 5.1 | 59.06 | |
| EXP3SpecTarget Model=Qwen-3-32B, Inference Mode=Non-greedy2025.10 | 4.98 | 36.22 | |
| Vanilla EagleTarget Model=Qwen-3-8B, Inference Mode=Non-greedy2025.10 | 4.58 | 50.59 | |
| EagleBackbone=Qwen-3-8B, Inference Setting=Greedy2025.10 | 4.52 | 50.88 | |
| Vanilla EagleTarget Model=Qwen-3-32B, Inference Mode=Non-greedy2025.10 | 3.36 | 25.18 | |
| EagleBackbone=Qwen-3-32B, Inference Setting=Greedy2025.10 | 3.36 | 24.16 | |
| LLaMA-3.1-8B-ITTarget Model=LLaMA-3.1-8B-IT, Inference Mode=Non-greedy2025.10 | 1 | 17.97 | |
| Qwen-3-8BTarget Model=Qwen-3-8B, Inference Mode=Non-greedy2025.10 | 1 | 15.08 | |
| Qwen-3-32BTarget Model=Qwen-3-32B, Inference Mode=Non-greedy2025.10 | 1 | 8.57 | |
| LLaMA-3.1-8B-ITBackbone=LLaMA-3.1-8B-IT, Inference Setting=Greedy2025.10 | 1 | 18.87 | |
| Qwen-3-8BBackbone=Qwen-3-8B, Inference Setting=Greedy2025.10 | 1 | 14.58 | |
| Qwen-3-32BBackbone=Qwen-3-32B, Inference Setting=Greedy2025.10 | 1 | 7.96 |