Research map / Video / Video-language understanding
EAKV: An Entropy-Driven Adaptive KV Compression Framework for Long Video Understanding
Hengrui Hu, Jingyu Li, Juntao Liang, Guanyu Chen, Lan Zhang
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 |
|---|---|---|---|
| apply EAKV to Qwen2.5-VL-7B | Qwen2.5-VL-7B · LVBench | accuracy: 45.3 -> 50.1 | Improved |
| apply EAKV to Qwen2-VL-7B | Qwen2-VL-7B · LVBench | accuracy: 42.4 -> 48.9 | Improved |
| apply EAKV to LLaVA-Video-7B | LLaVA-Video-7B · VideoMME (All) | accuracy: 62.9 -> 66.2 | Improved |
| apply EAKV to InternVL3.5-8B | InternVL3.5-8B · LVBench | accuracy: 41.8 -> 45.2 | Improved |
| EAKV entropy-driven allocation vs attention-based and static KV compression methods | Qwen2.5-VL-3B · LVBench | accuracy: 43.9 -> 46.7 | Improved |
| add Entropy-Driven Contention (cache cut 8K to 4K) (ablation) | Qwen2.5-VL-7B · 4-benchmark average | avg accuracy gain: +1.2 | Improved |
| add Asymmetric Temporal Integration and double frames to 256 (ablation) | Qwen2.5-VL-7B · 4-benchmark average | avg accuracy gain: +1.2 | Improved |
| add Layer-wise Safety Net (ablation) | Qwen2.5-VL-7B · 4-benchmark average | avg accuracy gain: +0.7 | Improved |
| scale input to 2048 frames and 16K context with full method (ablation) | Qwen2.5-VL-7B · 4-benchmark average | avg accuracy gain: +0.8 | Improved |
| apply EAKV compression to decoding | Qwen2.5-VL-7B · decode latency, 2048 frames | TPOT: 48.5 ms/token -> 15.2 ms/token | Improved |
| entropy computation during prefill | Qwen2.5-VL-7B · prefill latency | TTFT: 3.12 s -> 3.28 s | Hurt |
| apply EAKV to reduce KV cache memory | Qwen2.5-VL-7B · peak GPU memory, 2048 frames | peak memory: OOM -> 51.6 | Improved |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.91)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.90)
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.88)
- Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models (ICLR 2026 · similarity 0.88)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.88)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (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.
