Research map / Architectures and efficiency / Attention and transformers
Hilbert-Guided Sparse Local Attention
Yunge Li, Lanyu Xu
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 Hilbert reordering to Flex window attention | WSA (Flex) · Window attention I=128x128, W=16x16 | forward time: 5.68 ms -> 0.68 ms (4.0x over dense WSA) | Improved |
| add Hilbert reordering to Flex window attention | WSA (Flex) · Window attention I=64x64, W=8x8 | forward time: 0.40 ms -> 0.12 ms | Improved |
| add Hilbert reordering to Flex window attention | WSA (Flex) · Window attention I=96x96, W=12x12 | forward time: 2.02 ms -> 0.59 ms | Improved |
| use FlexAttention block-sparse kernel without Hilbert reordering | WSA · Window attention I=128x128, W=16x16 | forward time: 2.74 ms -> 5.68 ms | Hurt |
| add Hilbert reordering (increase empty-block ratio) | WSA (Flex) · Window attention I=128x128, W=16x16 | sparsity: 87.50% -> 98.44% | Improved |
| increase block size from 128 to 1024 (ablation) | HWA (Flex) · Window attention I=128, W=16 | forward time: 0.68 ms -> 2.78 ms | Hurt |
| increase block size from 128 to 512 (ablation) | HWA (Flex) · Window attention I=128, W=16 | forward time: 0.68 ms -> 1.43 ms | Hurt |
| increase block size from 128 to 1024 (ablation) | WSA (Flex) · Window attention I=128, W=16 | forward time: 5.68 ms -> 5.68 ms | No clear effect |
| use Hilbert reordering at low resolution | HWA (Flex) · Window attention I=56x56, W=7x7 | forward time | No clear effect |
| replace Swin-T windows with Hilbert windows (HWT-T) | Swin-T · ImageNet-1K val 224x224 | top-1 accuracy: 81.2 -> 81.0 | No clear effect |
| replace NAT-mini neighborhoods with Hilbert neighborhoods (HNT-mini) | NAT-mini · ImageNet-1K val 224x224 | top-1 accuracy: 81.8 -> 81.6 | No clear effect |
| replace Swin with HWT | SWIN · CIFAR100 | accuracy: 76.6% -> 76.4% | No clear effect |
Closest papers by meaning
- Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level (NeurIPS 2024 · similarity 0.87)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.85)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.84)
- Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction (NeurIPS 2025 · similarity 0.84)
- Sparse Attention as Compact Kernel Regression (NeurIPS 2026 · similarity 0.84)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.84)
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.
