Research map / Video / Video-language understanding
H2SVC: Head-aware Heterogeneous Streaming Video Cache for Online Video Understanding
han li, Xinyi Zhang, Wenrui Dai, Yaoming Wang, Xinyu Peng, Fan He, Hang Xu, Ziyang Zheng, Chenglin Li, Junni Zou, Hongkai Xiong
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 |
|---|---|---|---|
| head-wise instead of layer-wise classification and routing (ablation) | H2SVC (LLaVA-Video-7B) · Video-MME / OVO-Bench | accuracy: +2.5 on Video-MME, +0.7 on OVO-Bench | Improved |
| mix three head types instead of only Instantaneous heads (ablation) | H2SVC (LLaVA-Video-7B) · Video-MME | accuracy: 49.1 -> 63.0 | Improved |
| mix three head types instead of only Episodic heads (ablation) | H2SVC (LLaVA-Video-7B) · OVO-Bench | accuracy: 59.5 -> 62.6 | Improved |
| scale memory budget from S to M (ablation) | H2SVC (LLaVA-Video-7B) · Video-MME / OVO-Bench | accuracy: +1.3 on Video-MME, +2.7 on OVO-Bench | Improved |
| scale memory budget from M to L (ablation) | H2SVC (LLaVA-Video-7B) · Video-MME / OVO-Bench | accuracy: +1.5 on Video-MME, degrades OVO-Bench by 1.1 | Mixed |
| curvature-based STCE instead of Uniform Eviction (ablation) | H2SVC (LLaVA-Video-7B) · OVO-Bench Avg | accuracy: 61.6 -> 62.6 | Improved |
| curvature-based STCE instead of Similarity Eviction (ablation) | H2SVC (LLaVA-Video-7B) · OVO-Bench Avg | accuracy: 61.2 -> 62.6 | Improved |
| use Value states instead of Key states for trajectory (ablation) | STCE (LLaVA-Video-7B) · OVO-Bench Avg | accuracy: 62.1 -> 62.6 | Improved |
| add Magnitude Gating term (ablation) | STCE (LLaVA-Video-7B) · OVO-Bench Avg | accuracy: 61.4 -> 62.6 | Improved |
| add Temporal Weight term (ablation) | STCE (LLaVA-Video-7B) · OVO-Bench Avg | accuracy: 62.5 -> 62.6 | Improved |
| combine TAD_V2V and TAD_T2V instead of TAD_V2V only (ablation) | H2SVC head classification · Video-MME | accuracy: 62.2 -> 63.0 | Improved |
| combine TAD_V2V and TAD_T2V instead of TAD_T2V only (ablation) | H2SVC head classification · Video-MME / OVO-Bench | accuracy: 61.2 on Video-MME -> 63.0, 62.4 on OVO-Bench -> 62.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- StreamingTOM — StreamingTOM: Streaming Token Compression for Efficient Video Understanding (CVPR 2026) · better on 2 of 2 shared results
- StreamForest — StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025) · better on 2 of 2 shared results
- InfiniPot-V — InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding (NeurIPS 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
- CoRDS: Coreset-Based Representative and Diverse Selection for Streaming Video Understanding (NeurIPS 2026 · similarity 0.90)
- VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.90)
- StreamForest: Efficient Online Video Understanding with Persistent Event Memory (NeurIPS 2025 · similarity 0.88)
- InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding (NeurIPS 2025 · similarity 0.88)
- FreshMem: Brain-Inspired Frequency-Space Hybrid Memory for Streaming Video Understanding (NeurIPS 2026 · similarity 0.88)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · 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.
