Knowledge-intensive Question Answering on GPQA (Speedup)
3.11SpeedupFailFast
Evaluation Results
| Method | Links | |
|---|---|---|
| FailFastTarget Model=Qwen2.5-32B-Instruct2025.12 | 3.11 | |
| FailFastTarget Model=Qwen2.5-14B-Instruct2025.12 | 2.54 | |
| Fast-dLLMTarget Model=Qwen2.5-32B-Instruct2025.12 | 2.42 | |
| EAGLE-3 (w/ draft tree)Target Model=Qwen2.5-32B-Instruct2025.12 | 2.42 | |
| AR Draft ModelTarget Model=Qwen2.5-32B-Instruct2025.12 | 2.26 | |
| EAGLE-3 (w/ draft tree)Target Model=Qwen2.5-14B-Instruct2025.12 | 2.04 | |
| EAGLE-3 (w/ draft tree)Target Model=Qwen2.5-7B-Instruct2025.12 | 1.98 | |
| FailFastTarget Model=Qwen2.5-7B-Instruct2025.12 | 1.89 | |
| Fast-dLLMTarget Model=Qwen2.5-14B-Instruct2025.12 | 1.86 | |
| AR Draft ModelTarget Model=Qwen2.5-14B-Instruct2025.12 | 1.71 | |
| Fast-dLLMTarget Model=Qwen2.5-7B-Instruct2025.12 | 1.34 | |
| AR Draft ModelTarget Model=Qwen2.5-7B-Instruct2025.12 | 1.25 |