Research map / Vision-language and multimodal / Multimodal LLMs
Layer-wise Vision Injection with Disentangled Attention for Efficient LVLMs
Xuange Zhang, Dengjie Li, Bo Liu, Zenghao Bao, Yao Zhou, Baisong Yang, Zhongying Liu, Yujie Zhong, Tongtong Yuan
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 layer-wise disentangled vision injection (LVIDA) to the decoder | Vicuna-7B LVLM · Decoder FLOPs | GFLOPs: 10800 -> 1310 | Improved |
| add LVIDA to the decoder | Vicuna-7B LVLM · Time to first token | TTFT: 215.6 -> 54.4 | Improved |
| add LVIDA to the decoder | Vicuna-7B LVLM · Peak memory | GB: 24.3 -> 17.2 | Improved |
| add LVIDA to the decoder | Vicuna-7B LVLM · MM-Vet | accuracy: 40.0 -> 37.5 | Hurt |
| add LVIDA to the decoder | Vicuna-7B LVLM · MME | score: 1439.1 -> 1498.9 | Improved |
| add LVIDA to the decoder | TinyLlama-1.1B LVLM · TextVQA | accuracy: 49.6 -> 44.1 | Hurt |
| add LVIDA instead of FastV or VTW visual token pruning | Llama3-1B LVLM · TextVQA | accuracy: 38.7 -> 50.6 | Improved |
| add LVIDA instead of FastV (K=3,R=75%) pruning | Vicuna-7B LVLM · GQA | accuracy: 58.8 -> 62.4 | Improved |
| use a smaller TinyCLIP-39M encoder instead of SigLIP (ablation) | LVIDA TinyLlama-1.1B · VQAv2 | accuracy: 75.5 -> 67.1 | Hurt |
| increase TinyCLIP feature dimension from 512 to 640 (40M to 61M) (ablation) | LVIDA TinyLlama-1.1B · VQAv2 | accuracy: 64.2 -> 63.7 | No clear effect |
| use disentangled projection instead of vanilla projection (ablation) | LVIDA Llama3-1B · GFLOPs | GFLOPs: 2040 -> 192 | Improved |
| use disentangled projection instead of vanilla projection (ablation) | LVIDA Llama3-1B · VQAv2 | accuracy: 76.8 -> 76.2 | No clear effect |
Closest papers by meaning
- A-VL: Adaptive Attention for Large Vision-Language Models (AAAI 2025 · similarity 0.90)
- Large Vision-Language Models Get Lost in Attention (ICML 2026 · similarity 0.90)
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.89)
- ROSE: Rotate Your Large Language Model to See (CVPR 2026 · similarity 0.89)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.89)
- Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models (ICLR 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.
