AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Step-dLLM: Adaptive Step-aware Sparse Attention for Efficient Diffusion LLM Inference

Zhichen Zeng, Xichong zhang, Junpan Wu, Yifei Zuo, Chi-Chih Chang, Jiayi Wang, Maohua Nie, Ji Liu, Ang Li, Banghua Zhu

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Step-dLLM, an adaptive step-aware sparse attention method that speeds up diffusion LLM inference by re-deciding sparsity across denoising steps rather than reusing a single first-step mask.

Problem

Diffusion LLMs use bidirectional attention, which blocks KV caching and forces full attention recomputation at every denoising step, giving quadratic per-step cost. Existing sparse attention methods for dLLMs compute a mask once at the first step and reuse it, which the abstract suggests is a limita

Approach

Makes the sparse attention pattern step-aware and adaptive across denoising iterations, instead of computing one sparse mask at the first denoising step and reusing it throughout. The abstract is truncated, so the exact allocation scheme is not stated.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
refresh sparse mask at anchor steps (f=0.2) vs one-shot mask (ablation)SparseD-style uniform budget · GSM8K, 90% sparsity, LLaDA-1.5accuracy: 29.64 -> 43.90Improved
add head-adaptive global budget (no refresh) (ablation)SparseD · GSM8K, 90% sparsity, LLaDA-1.5accuracy: 29.64 -> 35.18Improved
combine refresh and head-adaptive budget (ablation)Step-dLLM · GSM8K, 90% sparsity, LLaDA-1.5accuracy: 35.18 -> 61.79Improved
add head-adaptive budget (ablation)SparseD · RULER-VT, 95% sparsityaccuracy: 86.16 -> 92.32Improved
refresh at low sparsity (50%) (ablation)Step-dLLM vs SparseD · GSM8K, 50% sparsityaccuracy: 80.59 vs 77.48 at f=1.0; 80.97 vs 81.65 at f=0.2No clear effect
use Step-dLLM instead of SparseDLLaDA-1.5 · GSM8K, 90% sparsityaccuracy: 29.64 -> 61.79Improved
use Step-dLLM instead of SparseDDream-7B-Instruct · NIAH, 90% sparsityaccuracy: 77.06 -> 76.12Hurt
use causal-style sink-plus-window sparsity (StreamingLLM)LLaDA-1.5 · GSM8Kaccuracy: 74.07 -> 0.08 (50% to 90%)Hurt
increase attention block size 16 to 32 (ablation)Step-dLLM · GSM8K, 90% sparsityaccuracy: 61.79 -> 39.50Hurt
use Step-dLLM at 50% sparsity vs denseLLaDA-1.5 ~1K seq · latencyms/step: 50.39 vs 51.23No clear effect
use sparse kernel at 50% sparsity short contextblock-sparse kernel · 1K and 2K contextspeedup: 0.83xHurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Limitations

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy