SAID: Accelerating Diffusion-Based Language Models via Scaffold-Aware Iterative Decoding

Na Li, Chengda Wang, Mingju Gao, Hao Tang
6/3/2026
cs.CL
diffusion-flow

Official code: TH-AI-Lab-PKU/SAID (2 stars, last commit 2026-10-19).

Abstract

Diffusion large language models (DLLMs) enable non-autoregressive generation by iteratively denoising corrupted token sequences with bidirectional context. Despite their ability to update multiple positions in parallel, inference remains costly due to the many denoising steps required for high-quality generation. We propose SAID, a Scaffold-Aware Iterative Decoding framework that accelerates DLLMs by reallocating computation across tokens. SAID first spends denoising computation on scaffold tokens to establish the coarse semantic structure, and then completes predictable detail tokens with fewer steps. We further adapt SAID to block-wise diffusion decoding and introduce Confidence-Hierarchical Layered Generation (CHLG), which assigns additional steps only to low-confidence tokens. Experiments on LLaDA-8B and LLaDA 1.5 across math, coding, and knowledge benchmarks show that SAID significantly accelerates DLLM inference with a maximum speedup of 9.1x while maintaining competitive performance. Our code is publicly available: https://github.com/TH-AI-Lab-PKU/SAID.

View on arXivView PDF

Code Implementations(1)

TH-AI-Lab-PKU/SAIDOfficial100%

The official implement for "SAID: Accelerating Diffusion-Based Language Models via Scaffold-Aware Iterative Decoding"

20CSS, HTMLJun 2, 20261 weeks ago

Similar Papers

PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Decoding

93% match2026

Dependency-Aware Revocable Decoding for Efficient Diffusion Large Language Model Inference

92% match2026

SimSD: Simple Speculative Decoding in Diffusion Language Models

92% match2026

Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs

91% match2026

BitLM: Unlocking Multi-Token Language Generation with Bitwise Continuous Diffusion

91% match2026

DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models

91% match2026

Cite this paper

@article{li2026said,
  title  = {SAID: Accelerating Diffusion-Based Language Models via Scaffold-Aware Iterative Decoding},
  author = {Na Li and Chengda Wang and Mingju Gao and Hao Tang},
  year   = {2026},
  eprint = {2606.04974},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2606.04974v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2606.04974.svg)](https://codeofpaper.com/paper/2606.04974)

Discussion