Research map / Architectures and efficiency / Attention and transformers
VecAttention: Vector-wise Sparse Attention for Accelerating Long Context Inference
Anmin Liu, Ruixuan Yang, Huiqiang Jiang, Bin Lin, Minmin Sun, Yong Li, Chen Zhang, Tao Xie
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 |
|---|---|---|---|
| use vertical-vector sparse attention instead of coarse line/block patterns | Qwen2.5-VL-7B-Instruct · video understanding avg | avg accuracy: 47.2 -> 49.9 at 73.6 -> 78.5% sparsity (vs XAttention) | Improved |
| use VecAttention on InternVL-3.5-8B | InternVL-3.5-8B · video understanding avg | avg accuracy: 52.7 -> 51.1 | Mixed |
| VecAttention vs XAttention and SVG on video generation | Wan2.1 and HunyuanVideo · VBench | PSNR: Hunyuan 21.8 -> 22.8 PSNR; LPIPS 0.326 -> 0.330 vs SVG | Mixed |
| replace row-wise topP with sorting-free minS filter (ablation) | vector selection · 64K context, sparsity 0.9 | selection latency: 3.77x speedup | Improved |
| fuse selection into tiled GEMM (TilingSelect) (ablation) | minS selection · 64K context, sparsity 0.9 | HBM access / latency: 18.3 GB -> 1.8 GB; 2.42x speedup | Improved |
| use vertical-vector instead of horizontal-vector pattern (ablation) | attention sparsity pattern · VLM attention recall at matched sparsity | recall: 0.89 -> 0.95 (V-Vec vs H-Vec at ~69% sparsity) | Improved |
| increase vector size Pq beyond 64 (ablation) | VecAttention · VideoMME | accuracy/latency | Mixed |
| increase K-block group size Gk to enable cross-tile minS accumulation (ablation) | VecAttention · VideoMME | accuracy | Improved |
| use VecAttention for end-to-end prefill | VLM · VideoMME TTFT | TTFT speedup: 1.17x end-to-end vs 2.65x attention-only | Improved |
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.
- XAttention — XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025) · better on 1 of 1 shared result
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
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.89)
- PAROAttention: Pattern-Aware ReOrdering for Efficient Sparse and Quantized Attention in Visual Generation Models (NeurIPS 2025 · similarity 0.88)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.88)
- HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion (NeurIPS 2026 · similarity 0.88)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.88)
- XAttention: Block Sparse Attention with Antidiagonal Scoring (ICML 2025 · similarity 0.87)
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.
