Research map / Video / Video-language understanding
MuKV: Multi-Grained KV Cache Compression for Long Streaming Video Question-Answering
Junbin Xiao, Jiajun Chen, Tianxiang Sun, Xun Yang, Angela Yao
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 patch- and frame-level KV caches to segment-level only (ablation) | MuKV (0.5B) · RVS-Ego | accuracy: 54.9 -> 56.5 | Improved |
| add DCP compression at 67% to uncompressed multi-grained cache (ablation) | MuKV (0.5B) · RVS-Ego | accuracy: 53.7 -> 57.3 | Improved |
| add frequency signal to attention-only compression (ablation) | MuKV (0.5B) · RVS-Ego | accuracy: 55.9 -> 57.3 | Improved |
| randomly drop half of KV tokens instead of informed compression (ablation) | ReKV (0.5B) · RVS-Ego | accuracy: 51.5 -> 46.7 | Hurt |
| add DCP(50%) compression (ablation) | ReKV (0.5B) · RVS-Ego | accuracy: 51.5 -> 56.1 | Improved |
| sparser frame sampling (Half Frames) (ablation) | ReKV (0.5B) · RVS-Ego / RVS-Movie | accuracy: 51.5 -> 53.0 on Ego; 42.3 -> 41.9 on Movie | Mixed |
| prune low-frequency tokens rather than high-frequency tokens (ablation) | ReKV (0.5B) · RVS-Ego | accuracy: 55.1 -> 51.4 | Hurt |
| add semi-hierarchical retrieval over parallel retrieval (ablation) | MuKV (0.5B) · RVS-Ego / RVS-Movie | accuracy: 56.5 -> 57.9 on Ego; 46.0 -> 45.2 on Movie | Mixed |
| use naive hierarchical retrieval instead of parallel (ablation) | MuKV (0.5B) · RVS-Ego | accuracy: 56.5 -> 52.9 | Hurt |
| use last-layer scores instead of penultimate layer (ablation) | MuKV (0.5B) · RVS-Ego | accuracy: 54.9 -> 57.3 | Improved |
| raise frame rate from 0.5 to 3 FPS (ablation) | MuKV (7B) · StreamingBench Real-Time (All) | accuracy: 64.4 -> 71.4 | Improved |
| raise frame rate from 0.5 to 3 FPS (ablation) | ReKV (7B) · StreamingBench Real-Time (All) | accuracy: 62.3 -> 62.5 | No clear effect |
Closest papers by meaning
- Streaming Video Question-Answering with In-context Video KV-Cache Retrieval (ICLR 2025 · similarity 0.94)
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.90)
- BIMBA: Selective-Scan Compression for Long-Range Video Question Answering (CVPR 2025 · similarity 0.89)
- Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning (ICLR 2025 · similarity 0.89)
- SVBench: A Benchmark with Temporal Multi-Turn Dialogues for Streaming Video Understanding (ICLR 2025 · similarity 0.88)
- InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video 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.
