PRESTO Speeds Up Diffusion LLM Decoding with Tree Drafting.

Zheng Wang, Zhifan Ye, Qi Cheng, Yonggan Fu, Ziyan Wang, Feng Zhu, Haozhe Zhao, Jan Kautz, Pavlo Molchanov, Humphrey Shi, Minjia Zhang· July 28, 2026 View original

Summary

This research introduces PRESTO, a framework that enhances the efficiency of Diffusion Large Language Models (dLLMs) by applying prefix-aligned tree-based drafting for speculative decoding. It addresses the mismatch between diffusion model confidence and autoregressive verification, achieving significant throughput speedups.

Diffusion Large Language Models (dLLMs) offer a promising alternative to traditional autoregressive LLMs by generating tokens in parallel, making them suitable for speculative decoding. However, existing diffusion-based drafting methods often use linear approaches, which limit efficiency because dLLMs produce multiple candidate tokens, creating a vast space of potential decoding paths. This new framework, PRESTO, extends tree-based drafting to diffusion models to explore these diverse candidate paths more effectively. It tackles a core challenge: diffusion models' marginal probabilities are "prefix-blind," meaning they don't align well with the prefix-based nature of autoregressive verification. PRESTO introduces prefix-aligned scoring and a priority-based tree search to resolve this mismatch. Its key principles ensure that candidate ranking aligns with verification needs and that tree construction prioritizes paths with high verification potential. Experiments show PRESTO can achieve up to 1.5x throughput speedup on state-of-the-art diffusion drafters and 1.12x on self-speculative diffusion LLMs.

Why it matters

This research offers a significant advancement in making diffusion-based LLMs faster and more practical for real-time applications, potentially reducing inference costs and latency for generative AI systems.

How to implement this in your domain

  1. 1Evaluate current LLM inference pipelines for potential bottlenecks in token generation speed.
  2. 2Investigate integrating speculative decoding techniques, specifically those leveraging diffusion models, into existing LLM deployments.
  3. 3Explore open-source implementations of PRESTO or similar tree-based drafting methods to benchmark performance improvements.
  4. 4Consider contributing to or adopting research that further refines prefix-aligned scoring for custom diffusion models.

Who benefits

AI/ML DevelopmentCloud ComputingContent GenerationSoftware Development

Key takeaways

  • PRESTO significantly accelerates diffusion LLM inference through a novel tree-based drafting approach.
  • It resolves the mismatch between diffusion model confidence and prefix-based autoregressive verification.
  • The method prioritizes candidate paths with high verification potential for maximum acceptance length.
  • Achieves up to 1.5x throughput speedup, making dLLMs more efficient for practical use.

Original post by Zheng Wang, Zhifan Ye, Qi Cheng, Yonggan Fu, Ziyan Wang, Feng Zhu, Haozhe Zhao, Jan Kautz, Pavlo Molchanov, Humphrey Shi, Minjia Zhang

"arXiv:2607.22634v1 Announce Type: new Abstract: Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to autoregressive (AR) LLMs, generating tokens in parallel. This makes them effective draft models for speculative decoding (SD), producing an entire bl…"

View on X

Originally posted by Zheng Wang, Zhifan Ye, Qi Cheng, Yonggan Fu, Ziyan Wang, Feng Zhu, Haozhe Zhao, Jan Kautz, Pavlo Molchanov, Humphrey Shi, Minjia Zhang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses