AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Pruning All-Rounder: Rethinking and Improving Inference Efficiency for Large Vision Language Models

Wei Suo, Ji Ma, Mengyang Sun, Lin Yuanbo Wu, Peng Wang, Yanning Zhang

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 4 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 DPO training to the layer-skipping router (ablation)PAR layer pruning · POPEaccuracy: 81.4 -> 85.8Improved
add DPO training to the layer-skipping router (ablation)PAR layer pruning · SEED-Bench-Imageaccuracy: 61.6 -> 65.1Improved
add DPO training to token router (ablation)PAR token pruning (T=176) · POPEaccuracy: 72.8 -> 74.9Improved
add Gaussian noise perturbation on top of DPO (ablation)PAR token pruning (T=176) · POPEaccuracy: 74.9 -> 76.5Improved
add Gaussian noise perturbation on top of DPO (ablation)PAR token pruning (T=176) · SEED-Bench-Imageaccuracy: 61.2 -> 63.0Improved
add learned joint router (DPO and noise) instead of naive combination of layer skipping and token pruning (ablation)PAR joint pruning · POPEaccuracy: 64.4 -> 76.1Improved
add learned joint router (DPO and noise) instead of naive combination of layer skipping and token pruning (ablation)PAR joint pruning · SEED-Bench-Imageaccuracy: 54.5 -> 62.7Improved
naively combine rule-based layer skipping with token pruning (ablation)FastV plus layer skipping · POPEaccuracy: 72.8 -> 64.4Hurt
use PAR instead of FastV at same computeLLaVA-1.5-7B · POPEaccuracy: 65.6 -> 75.9Improved
use PAR instead of ShortGPT-style layer skippingLLaVA-1.5-7B · POPEaccuracy: 79.3 -> 82.9Improved
use specialized token-router instead of FastVLLaVA-1.5-7B · SEED-Bench-Imageaccuracy: 56.3 -> 59.0Improved
use one jointly trained router for token-only compression instead of retraining a token routerPAR · POPEaccuracy: 71.3 vs 72.7Mixed

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