CLASP: Class-Adaptive Layer Fusion and Dual-Stage Pruning for Multimodal Large Language Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Dang, Yunkai, Jiang, Yizhu, Jiang, Yifan, Fan, Qi, Shi, Yinghuan, Li, Wenbin, Gao, Yang
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908963120873472
author Dang, Yunkai
Jiang, Yizhu
Jiang, Yifan
Fan, Qi
Shi, Yinghuan
Li, Wenbin
Gao, Yang
author_facet Dang, Yunkai
Jiang, Yizhu
Jiang, Yifan
Fan, Qi
Shi, Yinghuan
Li, Wenbin
Gao, Yang
contents Multimodal Large Language Models (MLLMs) suffer from substantial computational overhead due to the high redundancy in visual token sequences. Existing approaches typically address this issue using single-layer Vision Transformer (ViT) features and static pruning strategies. However, such fixed configurations are often brittle under diverse instructions. To overcome these limitations, we propose CLASP, a plug-and-play token reduction framework based on class-adaptive layer fusion and dual-stage pruning. Specifically, CLASP first constructs category-specific visual representations through multi-layer vision feature fusion. It then performs dual-stage pruning, allocating the token budget between attention-salient pivot tokens for relevance and redundancy-aware completion tokens for coverage. Through class-adaptive pruning, CLASP enables prompt-conditioned feature fusion and budget allocation, allowing aggressive yet robust visual token reduction. Extensive experiments demonstrate that CLASP consistently outperforms existing methods across a wide range of benchmarks, pruning ratios, and MLLM architectures. Code will be available at https://github.com/Yunkaidang/CLASP.
format Preprint
id arxiv_https___arxiv_org_abs_2604_12767
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CLASP: Class-Adaptive Layer Fusion and Dual-Stage Pruning for Multimodal Large Language Models
Dang, Yunkai
Jiang, Yizhu
Jiang, Yifan
Fan, Qi
Shi, Yinghuan
Li, Wenbin
Gao, Yang
Computer Vision and Pattern Recognition
Artificial Intelligence
Multimodal Large Language Models (MLLMs) suffer from substantial computational overhead due to the high redundancy in visual token sequences. Existing approaches typically address this issue using single-layer Vision Transformer (ViT) features and static pruning strategies. However, such fixed configurations are often brittle under diverse instructions. To overcome these limitations, we propose CLASP, a plug-and-play token reduction framework based on class-adaptive layer fusion and dual-stage pruning. Specifically, CLASP first constructs category-specific visual representations through multi-layer vision feature fusion. It then performs dual-stage pruning, allocating the token budget between attention-salient pivot tokens for relevance and redundancy-aware completion tokens for coverage. Through class-adaptive pruning, CLASP enables prompt-conditioned feature fusion and budget allocation, allowing aggressive yet robust visual token reduction. Extensive experiments demonstrate that CLASP consistently outperforms existing methods across a wide range of benchmarks, pruning ratios, and MLLM architectures. Code will be available at https://github.com/Yunkaidang/CLASP.
title CLASP: Class-Adaptive Layer Fusion and Dual-Stage Pruning for Multimodal Large Language Models
topic Computer Vision and Pattern Recognition
Artificial Intelligence
url https://arxiv.org/abs/2604.12767