Research map / Vision-language and multimodal / Multimodal LLMs
DAVID: Dual-stage Adaptive Vision-text Integrated Decoupling for Multimodal KV Cache Eviction
Yifeng Gu, Jianxiu Jin, Kailing Guo, Xiangmin Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DAVID, a layer-adaptive KV cache eviction strategy for multimodal LLMs that adapts to the degree of vision-text fusion across layers.
Problem
KV cache memory overhead from long multimodal inputs makes deploying multimodal LLMs on low-resource devices difficult.
Approach
Eviction adapts to the degree of modality fusion at each layer, based on analysis of features across layers, rather than applying a uniform policy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add NCAR-based layer switching to decoupled eviction (ablation) | DAVID (LLaVA-1.5-7B) · OCRBench | final score: 298 -> 305 | Improved |
| add decoupled vision/text eviction in all layers (ablation) | SnapKV-style eviction (LLaVA-1.5-7B) · OCRBench | final score: 295 -> 298 | Improved |
| apply DAVID on top of H2O scoring (ablation) | H2O (LLaVA-1.5-7B) · OCRBench | final score: 271 -> 285 | Improved |
| DAVID vs SnapKV on Qwen-VL | DAVID · OCRBench | score: 469 -> 478 | Improved |
| DAVID vs SnapKV on Qwen-VL | DAVID · MM-Vet | score: 38.43 -> 40.31 | Improved |
| DAVID vs SnapKV at 20% sparsity | DAVID (LLaVA-1.5-13B) · OCRBench | final score: 322 -> 324 | No clear effect |
| DAVID vs full cache | DAVID (LLaVA-1.5-7B) · OCRBench | final score: 308 -> 305 | Hurt |
| modality-fixed LOOK-M-style merging (LOOK-M, MEDA) vs unimodal SnapKV | LOOK-M / MEDA · OCRBench | final score: SnapKV 295 -> LOOK-M 158, MEDA 156 | Hurt |
| reduce KV budget to 20% with DAVID | DAVID · MathVista subset | decoding latency: 26.23 -> 19.17 ms/token | Improved |
| reduce KV budget to 20% with DAVID | DAVID · MathVista subset | GPU memory: 1.26GB -> 0.34GB | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MEDA — MEDA: Medical-Oriented Activation Editing for Hallucination Mitigation in Medical Large Vision-Language Model (ICML 2026) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models (ICLR 2026 · similarity 0.88)
- Q Cache: Visual Attention Is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model (AAAI 2026 · similarity 0.87)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.86)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.86)
- TGV-KV: Text-Grounded KV Eviction for Vision-Language Models (ICML 2026 · similarity 0.86)
- Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical Findings (NeurIPS 2025 · similarity 0.86)
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.
