AI Paper Atlas

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

ICML 2026 · Video › Video-language understanding · approach: Long video understanding

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 changedWhereMeasuredEffect
apply EAKV to Qwen2.5-VL-7BQwen2.5-VL-7B · LVBenchaccuracy: 45.3 -> 50.1Improved
apply EAKV to Qwen2-VL-7BQwen2-VL-7B · LVBenchaccuracy: 42.4 -> 48.9Improved
apply EAKV to LLaVA-Video-7BLLaVA-Video-7B · VideoMME (All)accuracy: 62.9 -> 66.2Improved
apply EAKV to InternVL3.5-8BInternVL3.5-8B · LVBenchaccuracy: 41.8 -> 45.2Improved
EAKV entropy-driven allocation vs attention-based and static KV compression methodsQwen2.5-VL-3B · LVBenchaccuracy: 43.9 -> 46.7Improved
add Entropy-Driven Contention (cache cut 8K to 4K) (ablation)Qwen2.5-VL-7B · 4-benchmark averageavg accuracy gain: +1.2Improved
add Asymmetric Temporal Integration and double frames to 256 (ablation)Qwen2.5-VL-7B · 4-benchmark averageavg accuracy gain: +1.2Improved
add Layer-wise Safety Net (ablation)Qwen2.5-VL-7B · 4-benchmark averageavg accuracy gain: +0.7Improved
scale input to 2048 frames and 16K context with full method (ablation)Qwen2.5-VL-7B · 4-benchmark averageavg accuracy gain: +0.8Improved
apply EAKV compression to decodingQwen2.5-VL-7B · decode latency, 2048 framesTPOT: 48.5 ms/token -> 15.2 ms/tokenImproved
entropy computation during prefillQwen2.5-VL-7B · prefill latencyTTFT: 3.12 s -> 3.28 sHurt
apply EAKV to reduce KV cache memoryQwen2.5-VL-7B · peak GPU memory, 2048 framespeak memory: OOM -> 51.6Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy