PRESTO: Tree-Based Drafting for Diffusion Speculative Decoding
A new method called PRESTO (Prefix-Aligned Tree Drafting) improves speculative decoding for diffusion large language models (dLLMs). Unlike autoregressive models, dLLMs generate tokens in parallel, making them efficient draft models. However, existing diffusion-based drafting uses linear drafting, which limits acceptance length due to the combinatorial space of candidate tokens. PRESTO applies tree-based drafting to explore multiple candidate paths, but naive tree drafting is suboptimal because diffusion marginals are prefix-blind, mismatching the prefix-based autoregressive verification. PRESTO introduces a principled framework that aligns tree drafting with prefix verification, enhancing decoding efficiency. The paper is published on arXiv under ID 2607.22634.
Key facts
- PRESTO extends tree-based drafting to diffusion drafters for speculative decoding.
- Diffusion LLMs generate tokens in parallel, unlike autoregressive models.
- Existing diffusion drafting methods use linear drafting, limiting efficiency.
- Naive tree drafting is suboptimal due to prefix-blind diffusion marginals.
- PRESTO aligns tree drafting with prefix-based AR verification.
- The method aims to improve acceptance length and decoding efficiency.
- The paper is available on arXiv with ID 2607.22634.
- PRESTO stands for Prefix-Aligned Tree Drafting.
Entities
Institutions
- arXiv