Evaluate additional self-speculative decoding methods

Evaluate self-speculative decoding methods such as LayerSkip and Kangaroo on models pretrained with layer dropout, and determine whether they provide the same or greater inference speedups as Draft-and-Verify.

Background

The paper evaluates Draft-and-Verify with several search procedures for selecting draft layers and finds improved speedups for models trained with higher layer-dropout rates. Other self-speculative approaches are not assessed.

The unresolved question is whether methods based on early-exit drafting or adapter-assisted early exit similarly exploit the structural elasticity induced by layer-dropout pretraining.

References

We leave for future work the evaluation of other self-speculative techniques like LayerSkip \citep{elhoushi-etal-2024-layerskip}, which uses early exit for drafting, and Kangaroo \citep{kangaroo}, which employs early exit with adapters trained in a similar manner to Balcony~\citep{balcony}.

— Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference  (2609.05275 - Elhoushi et al., 4 Sep 2026) in Section 6, Subsection “Self-Speculative Decoding”