Multi-path Speculative Decoding on Held-out (test)
6.84Average Block EfficiencyKhisti NDE
Evaluation results
| Method | Links | |
|---|---|---|
| Khisti NDEBackbone Model=Llama2026.02 | 6.84 | |
| TraversalBackbone Model=Llama2026.02 | 6.78 | |
| SpecTr NDEBackbone Model=Llama2026.02 | 6.74 | |
| SpecInfer NDEBackbone Model=Llama2026.02 | 6.64 | |
| NaiveTree NDEBackbone Model=Llama2026.02 | 6.47 | |
| NSS NDEBackbone Model=Llama2026.02 | 5.78 | |
| TraversalBackbone Model=Qwen2026.02 | 5.33 | |
| TraversalBackbone Model=Average2026.02 | 5.31 | |
| Khisti NDEBackbone Model=Average2026.02 | 5.23 | |
| SpecTr NDEBackbone Model=Average2026.02 | 5.05 | |
| SpecInfer NDEBackbone Model=Qwen2026.02 | 5.03 | |
| SpecInfer NDEBackbone Model=Average2026.02 | 5.03 | |
| SpecTr NDEBackbone Model=Qwen2026.02 | 4.97 | |
| Khisti NDEBackbone Model=Qwen2026.02 | 4.97 | |
| NaiveTree NDEBackbone Model=Qwen2026.02 | 4.94 | |
| NaiveTree NDEBackbone Model=Average2026.02 | 4.94 | |
| NSS NDEBackbone Model=Average2026.02 | 4.48 | |
| NSS NDEBackbone Model=Qwen2026.02 | 4.31 | |
| Khisti NDEBackbone Model=Gemma2026.02 | 3.89 | |
| TraversalBackbone Model=Gemma2026.02 | 3.81 | |
| SpecTr NDEBackbone Model=Gemma2026.02 | 3.44 | |
| NaiveTree NDEBackbone Model=Gemma2026.02 | 3.42 | |
| SpecInfer NDEBackbone Model=Gemma2026.02 | 3.42 | |
| NSS NDEBackbone Model=Gemma2026.02 | 3.34 |