Mathematical Reasoning on GSM8K (Speedup vs. AR)
1.93Speedup (vs. AR)HiSpec
Evaluation Results
| Method | Links | |
|---|---|---|
| HiSpecModel=Llama3-8B2025.10 | 1.93 | |
| HiSpecModel=Llama2-7B2025.10 | 1.92 | |
| LayerSkipModel=Llama3-8B2025.10 | 1.77 | |
| LayerSkipModel=Llama2-13B2025.10 | 1.76 | |
| HiSpecModel=Llama2-70B2025.10 | 1.72 | |
| LayerSkipModel=Llama2-7B2025.10 | 1.69 | |
| HiSpecModel=Llama2-13B2025.10 | 1.65 | |
| HiSpecModel=CodeLlama-34B2025.10 | 1.6 | |
| AdaDecodeModel=CodeLlama-34B2025.10 | 1.56 | |
| LayerSkipModel=CodeLlama-34B2025.10 | 1.52 | |
| LookAheadModel=Llama2-7B2025.10 | 1.5 | |
| LayerSkipModel=Llama2-70B2025.10 | 1.48 | |
| LookAheadModel=Llama2-13B2025.10 | 1.43 | |
| AdaDecodeModel=Llama3-8B2025.10 | 1.39 | |
| LookAheadModel=CodeLlama-34B2025.10 | 1.32 | |
| LookAheadModel=Llama3-8B2025.10 | 1.2 | |
| SWIFTModel=CodeLlama-34B2025.10 | 1.12 | |
| SWIFTModel=Llama2-70B2025.10 | 1.09 | |
| SWIFTModel=Llama2-13B2025.10 | 1.01 |