Text Generation Acceleration on General LLM Evaluation
2.5Max Speedup FactorEAGLE
Evaluation results
| Method | Links | ||
|---|---|---|---|
| EAGLEDraft Model=Required (feature-level), Decision Layer=Verification (exact), Decision Level=Token-level, Params (Decision)=02026.06 | 2.5 | — | |
| MedusaDraft Model=Not required, Decision Layer=Gate network (MLP), Decision Level=Token-level, Params (Decision)=∼1M2026.06 | 2.2 | — | |
| Speculative DecodingDraft Model=Required (separate), Decision Layer=Verification (exact), Decision Level=Token-level, Params (Decision)=02026.06 | 2 | — | |
| Speculative SamplingDraft Model=Required (separate), Decision Layer=Verification (approx.), Decision Level=Token-level, Params (Decision)=02026.06 | 2 | — | |
| LookaheadDraft Model=Not required, Decision Layer=Jacobi iteration, Decision Level=Token-level, Params (Decision)=02026.06 | 1.5 | — | |
| CLPDraft Model=Not required, Decision Layer=CLP (linear), Decision Level=Span-level, Params (Decision)=4.6K–7.7K2026.06 | 1.14 | — |