Research map / Video / Video-language understanding
VideoNSA: Native Sparse Attention Scales Video Understanding
Enxin Song, Wenhao Chai, Shusheng Yang, Ethan J. Armand, Xiaojun Shan, Haiyang Xu, Jianwen Xie, Zhuowen Tu
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 learnable native sparse attention (VideoNSA) instead of dense attention | Qwen2.5-VL-7B · LongTimeScope | accuracy: 40.7 -> 44.4 | Improved |
| use VideoNSA | Qwen2.5-VL-7B · Tomato | accuracy: 22.6 -> 26.5 | Improved |
| use VideoNSA | Qwen2.5-VL-7B · VSIBench | accuracy: 29.7 -> 36.1 | Improved |
| use VideoNSA | Qwen2.5-VL-7B · LongVideoBench | accuracy: 58.7 -> 60.0 | Mixed |
| replace token compression with sparse attention | Qwen2.5-VL-7B · long video / VSIBench | accuracy: FastV 41.8 vs VideoNSA 51.8 on MLVU | Improved |
| use all three branches (CMP, SLC, SWD) instead of one (ablation) | VideoNSA · LongVideoBench | accuracy: 49.1 -> 60.0 | Improved |
| use all three branches instead of selection only (ablation) | VideoNSA · TimeScope | accuracy: 63.7 -> 83.7 | Improved |
| transfer learned sparse weights to dense attention (Dense-NSA) (ablation) | Qwen2.5-VL-7B · VSIBench | accuracy: 30.6 -> 33.1 | Improved |
| transfer sparse weights to dense attention (ablation) | Qwen2.5-VL-7B · LongVideoBench | accuracy: 57.8 -> 56.1 | Hurt |
| keep runtime sparsity and dynamic gating vs dense transfer (ablation) | VideoNSA · LongTimeScope | accuracy: 40.9 -> 44.4 | Improved |
| increase global attention (more blocks) vs local sliding window at fixed budget (ablation) | VideoNSA · video benchmarks | accuracy | Improved |
| scale context beyond training length (36K to 128K) (ablation) | VideoNSA · video benchmarks | accuracy | Improved |
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.90)
- VideoAnchor: Reinforcing Subspace-Structured Visual Cues for Coherent Visual-Spatial Reasoning (ICLR 2026 · similarity 0.89)
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.89)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.89)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.88)
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.
