Research map / Architectures and efficiency / Attention and transformers
LoSA: Locality Aware Sparse Attention in Diffusion Language Models
Haocheng Xi, Harman Singh, Yuezhou Hu, Coleman Richard Charles Hooper, Rishabh Tiwari, Aditya Tomar, Minjae Lee, Wonjun Kang, Michael W. Mahoney, Chenfeng Xu, Kurt Keutzer, Amir Gholami
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 |
|---|---|---|---|
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 128 · LongBench avg | accuracy: 31.54% -> 41.97% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 256 · LongBench avg | accuracy: 34.64% -> 42.14% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 1024 · LongBench avg | accuracy: 37.69% -> 42.88% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | SDAR-8B-Chat, budget 128 · LongBench avg | accuracy: 32.57% -> 40.95% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | SDAR-8B-Chat, budget 256 · LongBench avg | accuracy: 36.21% -> 41.20% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-4B-Instruct, block size 32, budget 128 · LongBench avg | accuracy: 28.88 -> 30.46 | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-4B-Instruct, block size 16, budget 128 · LongBench avg | accuracy: 33.16 -> 33.93 | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 128 · Commonsense avg | accuracy: 69.45 -> 69.91 | No clear effect |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 256 · Commonsense avg | accuracy: 70.60 -> 69.85 | No clear effect |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 128 · LongBench KV-cache density | density: 4.24% -> 2.58% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, budget 1024 · LongBench KV-cache density | density: 25.66% -> 18.23% | Improved |
| replace naive per-query QUEST with LoSA (reuse cached prefix attention for stable tokens) | Trado-8B-Instruct, 64K context, RTX A6000 · prefix attention latency | speedup over dense: 3.26× -> 4.14× | Improved |
Later papers that compared with this method
1 later accepted paper reports LoSA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- SALR: Sparsity-Aware Low-Rank Representation for Efficient Fine-Tuning of Large Language Models (AAAI 2026 · better on 8 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.
- QUEST — QUEST: Query-Aware Sparsity for Efficient Long-Context LLM Inference (ICML 2024) · better on 18 of 19 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
- MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM (NeurIPS 2026 · similarity 0.91)
- Attention-Discounted Adaptive Sampler for Masked Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.89)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- More Value per Key: Asymmetric Sparse Attention for Faster LLM Decoding (NeurIPS 2026 · similarity 0.89)
- UltraLLaDA: Scaling the Context Length to 128K for Diffusion Large Language Models (ICLR 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.
