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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| refresh sparse mask at anchor steps (f=0.2) vs one-shot mask (ablation) | SparseD-style uniform budget · GSM8K, 90% sparsity, LLaDA-1.5 | accuracy: 29.64 -> 43.90 | Improved |
| add head-adaptive global budget (no refresh) (ablation) | SparseD · GSM8K, 90% sparsity, LLaDA-1.5 | accuracy: 29.64 -> 35.18 | Improved |
| combine refresh and head-adaptive budget (ablation) | Step-dLLM · GSM8K, 90% sparsity, LLaDA-1.5 | accuracy: 35.18 -> 61.79 | Improved |
| add head-adaptive budget (ablation) | SparseD · RULER-VT, 95% sparsity | accuracy: 86.16 -> 92.32 | Improved |
| refresh at low sparsity (50%) (ablation) | Step-dLLM vs SparseD · GSM8K, 50% sparsity | accuracy: 80.59 vs 77.48 at f=1.0; 80.97 vs 81.65 at f=0.2 | No clear effect |
| use Step-dLLM instead of SparseD | LLaDA-1.5 · GSM8K, 90% sparsity | accuracy: 29.64 -> 61.79 | Improved |
| use Step-dLLM instead of SparseD | Dream-7B-Instruct · NIAH, 90% sparsity | accuracy: 77.06 -> 76.12 | Hurt |
| use causal-style sink-plus-window sparsity (StreamingLLM) | LLaDA-1.5 · GSM8K | accuracy: 74.07 -> 0.08 (50% to 90%) | Hurt |
| increase attention block size 16 to 32 (ablation) | Step-dLLM · GSM8K, 90% sparsity | accuracy: 61.79 -> 39.50 | Hurt |
| use Step-dLLM at 50% sparsity vs dense | LLaDA-1.5 ~1K seq · latency | ms/step: 50.39 vs 51.23 | No clear effect |
| use sparse kernel at 50% sparsity short context | block-sparse kernel · 1K and 2K context | speedup: 0.83x | Hurt |
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.
- SparseD — SparseD: Sparse Attention for Diffusion Language Models (ICLR 2026) · better on 9 of 9 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
- Full text unavailable; only the truncated abstract was used, so method details and results are unknown.
Closest papers by meaning
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · similarity 0.90)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- MemDLM: Memory-Enhanced DLM Training (NeurIPS 2026 · similarity 0.90)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.89)
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.
