AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Revisiting Multimodal KV Cache Compression: A Frequency-Domain-Guided Outlier-KV-Aware Approach

Yaoxin Yang, Peng Ye, Xudong Tan, Chongjun Tu, Maosen Zhao, Jia Hao, Tao Chen

CVPR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 2 citations (OpenAlex)

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
add dynamic budget allocation (ablation)FlashCache · ImageNeedleInAHaystackscore: 24.69 -> 29.69Improved
add dynamic budget allocation (ablation)FlashCache · CLEVR-Changescore: 35.85 -> 41.04Improved
add dynamic budget allocation (ablation)FlashCache · ALFREDscore: 34.32 -> 34.39No clear effect
add dynamic budget allocation (ablation)FlashCache · GPR1200score: 14.67 -> 15.50Improved
raise low-pass cut-off factor from 0.1 to 0.9 (ablation)FlashCache · ImageNeedleInAHaystackscore: 29.06 -> 20.08Hurt
keep outlier KV pairs instead of attention-score-based eviction (ablation)FlashCache vs H2O · MileBench Needle in a Haystackscore: 12.66 -> 26.72Improved
FlashCache vs full cache on needle task (ablation)FlashCache · MileBench Needle in a Haystackscore: 27.35 -> 26.72Hurt
compatible with FlashAttention avoids OOMFlashCache · MUIRBench (retention 0.05)accuracy: OOM -> 44.42Improved
compress KV cache with FlashCacheQwen2.5-VL-7B · HR-Bench (retention 0.05)accuracy: 70.75 -> 72.38Improved
compress KV cache at retention 0.05Qwen2.5-VL-7B · FAVOR-Bench allaccuracy: 40.91 -> 30.71Hurt
use FlashCache vs SnapKV on video at retention 0.05FlashCache · FAVOR-Bench allaccuracy: 29.95 -> 30.71Improved
FlashCache selection latency vs SnapKVFlashCache · Method overhead at 8Klatency ms: 9.57 -> 6.77Improved

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