Research map / Architectures and efficiency / Attention and transformers
Sparser Block-Sparse Attention via Token Permutation
Xinghao Wang, Pengyu Wang, Dong Zhang, Chenkun Tan, Shaojun Zhou, Zhaoxiang Liu, Shiguo Lian, Fangxu Liu, Kai Song, Xipeng Qiu
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 segmented key permutation to MeanPooling block selection (ablation) | MeanPooling block-sparse attention · LongBench v2 (Qwen-2.5-7B-1M) | accuracy: 26.24 -> 34.39 (31% relative) | Improved |
| add segmented key permutation to MeanPooling block selection (ablation) | MeanPooling block-sparse attention · RULER (Llama-3.1-8B) | accuracy: 82.76 -> 85.97 (+3.21 avg, +7.66 at 128K) | Improved |
| add segmented key permutation to MeanPooling block selection (ablation) | MeanPooling block-sparse attention · LongBench (Qwen-2.5-7B-1M) | score: 31.83 -> 36.37 | Improved |
| add segmented key permutation to MeanPooling block selection (ablation) | MeanPooling block-sparse attention · LongBench (Llama-3.1-8B) | score: 36.67 -> 37.37 | Improved |
| add permutation to XAttention antidiagonal block scoring (PBS-Attn+) (ablation) | XAttention · RULER | accuracy: +1.41 on Llama-3.1-8B, +1.05 on Qwen-2.5-7B-1M | Improved |
| add permutation before block selection (ablation) | MeanPooling block selection · block-level density | relative sparsity: 7% absolute improvement at 8K context | Improved |
| add permutation before block selection (ablation) | Llama-3.1-8B · selected blocks | selected-block count: reduces selected blocks by 10.7% at 8K and 14.4% at 128K | Improved |
| permute keys randomly instead of by global importance (ablation) | segmented permutation · sparsity gain at 97.5% coverage (NarrativeQA) | absolute sparsity gain: -6.11 (Random) vs 7.90 (Random-Query-Attn) | Hurt |
| sort keys by global attention importance instead of greedy local query alignment (ablation) | segmented permutation · sparsity gain at 97.5% coverage (NarrativeQA) | absolute sparsity gain: 3.00 -> 7.77 (Greedy vs Last-Block-Query-Attn) | Improved |
| also permute queries in addition to keys (ablation) | PBS-Attn · LongBench vs density | score-density trade-off | No clear effect |
| apply permutation across heads of Llama-3.1-8B at 32K (ablation) | PBS-Attn · niah_single head analysis | head sparsity: 70.8% heads helped, 24.0% unchanged, 5.2% harmed | Mixed |
| use block size B=128 rather than 64 or 256 (ablation) | PBS-Attn · LongBench vs density | LongBench score | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- XAttention — XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025) · better on 6 of 6 shared results
- FlexPrefill — FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025) · better on 6 of 6 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
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.91)
- FlexPrefill: A Context-Aware Sparse Attention Mechanism for Efficient Long-Sequence Inference (ICLR 2025 · similarity 0.91)
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.91)
- Block Sparse Flash Attention (NeurIPS 2026 · similarity 0.91)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.90)
- MInference 1.0: Accelerating Pre-filling for Long-Context LLMs via Dynamic Sparse Attention (NeurIPS 2024 · similarity 0.90)
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.
