Video understanding in multimodal language models remains limited by context length: models often miss key transition frames and struggle to maintain coherence across long time scales. To address this, we adapt Native Sparse Attention (NSA) to video-language models. Our method, VideoNSA, adapts Qwen2.5-VL through end-to-end training on a 216K video instruction dataset. We employ a hardware-aware hybrid approach to attention, preserving dense attention for text, while employing NSA for video. Compared to token-compression and training-free sparse baselines, VideoNSA achieves improved performance on long-video understanding, temporal reasoning, and spatial benchmarks. Further ablation analysis reveals four key findings: (1) reliable scaling to 128K tokens; (2) an optimal global-local attention allocation at a fixed budget; (3) task-dependent branch usage patterns; and (4) the learnable combined sparse attention help induce dynamic attention sinks. Project Page: https://enxinsong.com/VideoNSA-web/, Code: https://github.com/Espere-1119-Song/VideoNSA

翻译：多模态语言模型中的视频理解仍受限于上下文长度：模型常忽略关键过渡帧，且难以在长时间尺度上保持连贯性。为解决此问题，我们将原生稀疏注意力（NSA）适配至视频-语言模型。我们的方法VideoNSA通过在216K视频指令数据集上进行端到端训练，对Qwen2.5-VL模型进行适配。我们采用硬件感知的混合注意力机制：对文本保留稠密注意力，对视频则采用NSA。相较于令牌压缩与免训练的稀疏基线方法，VideoNSA在长视频理解、时序推理及空间基准测试中均实现了性能提升。进一步的消融分析揭示了四个关键发现：（1）可稳定扩展至128K令牌；（2）在固定计算预算下存在最优的全局-局部注意力分配比例；（3）任务依赖的分支使用模式；（4）可学习的组合稀疏注意力有助于诱导动态注意力汇聚点。项目页面：https://enxinsong.com/VideoNSA-web/，代码：https://github.com/Espere-1119-Song/VideoNSA