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
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 |
|---|---|---|---|
| add dynamic budget allocation (ablation) | FlashCache · ImageNeedleInAHaystack | score: 24.69 -> 29.69 | Improved |
| add dynamic budget allocation (ablation) | FlashCache · CLEVR-Change | score: 35.85 -> 41.04 | Improved |
| add dynamic budget allocation (ablation) | FlashCache · ALFRED | score: 34.32 -> 34.39 | No clear effect |
| add dynamic budget allocation (ablation) | FlashCache · GPR1200 | score: 14.67 -> 15.50 | Improved |
| raise low-pass cut-off factor from 0.1 to 0.9 (ablation) | FlashCache · ImageNeedleInAHaystack | score: 29.06 -> 20.08 | Hurt |
| keep outlier KV pairs instead of attention-score-based eviction (ablation) | FlashCache vs H2O · MileBench Needle in a Haystack | score: 12.66 -> 26.72 | Improved |
| FlashCache vs full cache on needle task (ablation) | FlashCache · MileBench Needle in a Haystack | score: 27.35 -> 26.72 | Hurt |
| compatible with FlashAttention avoids OOM | FlashCache · MUIRBench (retention 0.05) | accuracy: OOM -> 44.42 | Improved |
| compress KV cache with FlashCache | Qwen2.5-VL-7B · HR-Bench (retention 0.05) | accuracy: 70.75 -> 72.38 | Improved |
| compress KV cache at retention 0.05 | Qwen2.5-VL-7B · FAVOR-Bench all | accuracy: 40.91 -> 30.71 | Hurt |
| use FlashCache vs SnapKV on video at retention 0.05 | FlashCache · FAVOR-Bench all | accuracy: 29.95 -> 30.71 | Improved |
| FlashCache selection latency vs SnapKV | FlashCache · Method overhead at 8K | latency ms: 9.57 -> 6.77 | Improved |
Closest papers by meaning
- EchoKV: Efficient KV Cache Compression via Similarity-Based Reconstruction (NeurIPS 2026 · similarity 0.90)
- ZipCache: Accurate and Efficient KV Cache Quantization with Salient Token Identification (NeurIPS 2024 · similarity 0.90)
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.90)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.90)
- AttentionPredictor: Temporal Patterns Matter for KV Cache Compression (NeurIPS 2025 · similarity 0.90)
- Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models (ICLR 2026 · similarity 0.90)
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.
