Research map / Architectures and efficiency / Attention and transformers
SparseD: Sparse Attention for Diffusion Language Models
Zeqing Wang, Gongfan Fang, Xinyin Ma, Xingyi Yang, Xinchao Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add skipping sparse attention in early denoising steps (ablation) | SparseD (LLaDA-1.5) · RULER-4k | accuracy: 87.91 -> 90.89 | Improved |
| add skipping sparse attention in early denoising steps (ablation) | SparseD (LLaDA-1.5) · RULER 64k sample latency | latency (s): 1552 -> 1695 | Hurt |
| add sparse pattern reusing across denoising steps (ablation) | SparseD (LLaDA-1.5) · RULER 64k sample latency | latency (s): 30020 -> 1695 | Improved |
| add sparse pattern reusing across denoising steps (ablation) | SparseD (LLaDA-1.5) · RULER-4k | accuracy: 90.82 -> 90.89 | No clear effect |
| add isolated selection of prefill and generation tokens (ablation) | SparseD (LLaDA-1.5) · RULER-4k | accuracy: 90.53 -> 90.89 | Improved |
| add isolated selection of prefill and generation tokens (ablation) | SparseD (LLaDA-1.5) · RULER 64k sample latency | latency (s): 1687 -> 1695 | No clear effect |
| use SparseD instead of StreamingLLM | Dream-7B-Instruct · Average of MMLU, GSM8k, HumanEval, RULER-4k, RULER-8k | accuracy (avg): 50.17 -> 72.38 | Improved |
| use SparseD instead of dKV-Cache on long context | Dream-7B-Instruct · RULER-8k | accuracy: 55.08 -> 72.47 | Improved |
| use SparseD instead of Fast-dLLM on long context | LLaDA-1.5 · RULER-8k | accuracy: 47.76 -> 62.44 | Improved |
| apply SparseD to LLaDA-1.5 | LLaDA-1.5 · Average of MMLU, GSM8k, HumanEval, RULER-4k, RULER-8k | accuracy (avg): 67.33 -> 67.62 | No clear effect |
| apply SparseD to Dream-7B-Instruct | Dream-7B-Instruct · Average of MMLU, GSM8k, HumanEval, RULER-4k, RULER-8k | accuracy (avg): 72.42 -> 72.38 | No clear effect |
| increase diffusion steps from 128 to 1024 | SparseD (Dream-7B-Instruct) · 64k context | speedup vs FlashAttention: 1.23x -> 1.50x | Improved |
Later papers that compared with this method
2 later accepted papers report SparseD as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · better on 10 of 10 shared results)
- Step-dLLM: Adaptive Step-aware Sparse Attention for Efficient Diffusion LLM Inference (NeurIPS 2026 · better on 9 of 9 shared results)
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.
- FlashAttention — FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (NeurIPS 2022) · better on 2 of 2 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.
Closest papers by meaning
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.89)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.89)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.89)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.89)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · 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.
