KL certificate distribution analysis on Decoding Steps (Evaluation set)
1.814Mean KL DivergenceTree
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| TreeBackbone=Qwen3-4B, Criterion=Tree, m=82026.06 | 1.814 | 2.126 | 0.663 | 1.52 | |
| TreeBackbone=Qwen3-1.7B, Criterion=Tree, m=82026.06 | 1.736 | 2.038 | 0.531 | 1.355 | |
| TreeBackbone=Qwen3-4B, Criterion=Tree, m=42026.06 | 1.276 | 1.541 | 0.327 | 0.995 | |
| TreeBackbone=Qwen3-1.7B, Criterion=Tree, m=42026.06 | 1.213 | 1.456 | 0.249 | 0.862 | |
| TreeBackbone=Qwen3-4B, Criterion=Tree, m=22026.06 | 0.82 | 1.036 | 0.099 | 0.538 | |
| TreeBackbone=Qwen3-1.7B, Criterion=Tree, m=22026.06 | 0.772 | 0.962 | 0.071 | 0.448 | |
| MedusaBackbone=Qwen3-4B, Criterion=Medusa, epsilon_0=0.1, delta_0=0.092026.06 | 0.606 | 0.664 | 0.378 | 0.536 | |
| MedusaBackbone=Qwen3-1.7B, Criterion=Medusa, epsilon_0=0.1, delta_0=0.092026.06 | 0.598 | 0.646 | 0.375 | 0.522 | |
| Mult.Backbone=Qwen3-4B, Criterion=Multiplicative, alpha=0.12026.06 | 0.573 | 0.643 | 0.296 | 0.465 | |
| Mult.Backbone=Qwen3-1.7B, Criterion=Multiplicative, alpha=0.12026.06 | 0.555 | 0.61 | 0.279 | 0.435 | |
| AdditiveBackbone=Qwen3-4B, Criterion=Additive, t=0.32026.06 | 0.505 | 0.625 | 0.107 | 0.315 | |
| Mult.Backbone=Qwen3-4B, Criterion=Multiplicative, alpha=0.52026.06 | 0.492 | 0.619 | 0.066 | 0.288 | |
| AdditiveBackbone=Qwen3-1.7B, Criterion=Additive, t=0.32026.06 | 0.482 | 0.573 | 0.102 | 0.272 | |
| AdditiveBackbone=Qwen3-4B, Criterion=Additive, t=0.12026.06 | 0.475 | 0.619 | 0.025 | 0.236 | |
| Strict greedyBackbone=Qwen3-4B, Criterion=Strict greedy2026.06 | 0.468 | 0.619 | 0.007 | 0.217 | |
| Mult.Backbone=Qwen3-1.7B, Criterion=Multiplicative, alpha=0.52026.06 | 0.464 | 0.56 | 0.061 | 0.235 | |
| AdditiveBackbone=Qwen3-1.7B, Criterion=Additive, t=0.12026.06 | 0.445 | 0.559 | 0.023 | 0.178 | |
| Strict greedyBackbone=Qwen3-1.7B, Criterion=Strict greedy2026.06 | 0.438 | 0.559 | 0.006 | 0.16 |