---
title: "DARTree: Speculative Diffusion Decoding with Autoregressive Draft Trees"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.13524v1"
permalink: "https://feed.astralor.com/items/itm_3928289022811be714dc61a4"
published: "2026-08-13T17:43:44.000Z"
score: 77
channel: "ai"
tags: ["DARTree","Diffusion Models","Speculative Decoding","Reasoning Acceleration"]
---

# DARTree: Speculative Diffusion Decoding with Autoregressive Draft Trees

## AI commentary

> Speculative decoding is a key path for LLM inference acceleration. DARTree cleverly combines the parallelism of diffusion drafters with the sequential correction of autoregressive heads into a tree structure without any training. This paper is worth attention for engineers working on LLM inference optimization and deployment, and it also showcases tree pruning and batched scoring techniques for researchers.

## At a glance

Speculative decoding losslessly accelerates autoregressive language models by verifying multiple draft tokens in parallel. Diffusion-based drafters reduce proposal latency by predicting an entire token block in parallel, but their position-wise distributions are marginal rather than conditioned on tokens selected along each draft path. DARTree is a training-free speculative decoding method that extends a pretrained AR correction head from chains to trees. It constructs a fixed-width candidate tree by expanding and scoring all nodes at each depth in a single batch, then applies best-first pruning to select the verification tree. Across seven math, code, and chat benchmarks, DARTree achieves the highest average acceptance length and speedup in all four model-temperature configurations, accepting up to 12.97 tokens per verification round, 98.6% more than DFlash and 27.9% more than Domino, and reaching up to 9.73x lossless speedup over autoregressive decoding.

## Key points

- DARTree is the first training-free speculative decoding method that extends a pretrained autoregressive correction head from chains to trees.
- It decouples AR-head inference from sequential heap operations via batched expansion and best-first pruning.
- It achieves the highest average acceptance length and speedup across all four model-temperature configurations on seven benchmarks.
- It accepts up to 12.97 tokens per verification round, 98.6% more than DFlash and 27.9% more than Domino, with up to 9.73x lossless speedup.

## Full text

Speculative decoding losslessly accelerates autoregressive language models by verifying multiple draft tokens in parallel. Diffusion-based drafters further reduce proposal latency by predicting an entire token block in parallel, but their position-wise distributions are marginal rather than conditioned on tokens selected along each draft path. Existing recurrent correction incorporates causal information along a single draft chain, whereas diffusion-based tree construction broadens candidate coverage without carrying this correction along individual branches. We introduce DARTree, a training-free speculative decoding method that extends a pretrained AR correction head from chains to trees. DARTree first constructs a fixed-width candidate tree by expanding and scoring all nodes at each depth in a single batch, and then only applies best-first pruning to select the verification tree, decoupling AR-head inference from sequential heap operations. Across seven math, code, and chat benchmarks, DARTree achieves the highest average acceptance length and speedup in all four model--temperature configurations, accepting up to 12.97 tokens per verification round, 98.6\% more than DFlash and 27.9\% more than Domino in the same setting, and reaching up to 9.73$\times$ lossless speedup over locally measured autoregressive decoding.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_3928289022811be714dc61a4