Code Generation on HumanEval (Speedup vs. AR)
1.91Speedup vs ARHiSpec
Evaluation Results
| Method | Links | |
|---|---|---|
| HiSpecModel=Llama3-8B2025.10 | 1.91 | |
| HiSpecModel=Llama2-7B2025.10 | 1.84 | |
| LayerSkipModel=Llama2-7B2025.10 | 1.71 | |
| HiSpecModel=Llama2-70B2025.10 | 1.69 | |
| AdaDecodeModel=CodeLlama-34B2025.10 | 1.67 | |
| HiSpecModel=CodeLlama-34B2025.10 | 1.67 | |
| LayerSkipModel=Llama3-8B2025.10 | 1.54 | |
| HiSpecModel=Llama2-13B2025.10 | 1.52 | |
| LayerSkipModel=Llama2-13B2025.10 | 1.51 | |
| LayerSkipModel=CodeLlama-34B2025.10 | 1.5 | |
| AdaDecodeModel=Llama3-8B2025.10 | 1.36 | |
| LayerSkipModel=Llama2-70B2025.10 | 1.32 | |
| LookAheadModel=Llama2-7B2025.10 | 1.24 | |
| SWIFTModel=CodeLlama-34B2025.10 | 1.24 | |
| LookAheadModel=Llama2-13B2025.10 | 1.23 | |
| LookAheadModel=Llama3-8B2025.10 | 1.19 | |
| LookAheadModel=CodeLlama-34B2025.10 | 1.18 | |
| SWIFTModel=Llama2-70B2025.10 | 1.12 | |
| SWIFTModel=Llama2-13B2025.10 | 1.05 |