Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Feng, Yilin, Gulhan, Ahmed Burak, Kandemir, Mahmut Taylan
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.29535
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911726865219584
author Feng, Yilin
Gulhan, Ahmed Burak
Kandemir, Mahmut Taylan
author_facet Feng, Yilin
Gulhan, Ahmed Burak
Kandemir, Mahmut Taylan
contents Vision-Language Models (VLMs) process thousands of visual tokens per image alongside comparatively few text tokens, yet existing compression methods treat both modalities uniformly. We observe that the two modalities have fundamentally different properties: vision tokens are spatially redundant and dominate prefill, while text tokens are causally dependent and accumulate during decoding. Based on this asymmetry, we propose and empirically evaluate AsymVLM, which applies aggressive pruning to vision tokens before prefill using a learned importance scorer with per-sample adaptive budgeting, and temporal threshold-based eviction to text tokens only when they exceed a fixed budget. Our experiments indicate that AsymVLM achieves the highest FLOPs savings (up to 54%) among state-of-the-art methods while outperforming existing approaches by 2--3% on document and chart understanding tasks where visual information is spatially localized and query-specific, and maintaining competitive accuracy on holistic benchmarks. In text-dominated scenarios, our eviction strategy substantially outperforms standard LLM cache compression methods by adapting to the short-context nature of VLM.
format Preprint
id arxiv_https___arxiv_org_abs_2605_29535
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle AsymVLM: Asymmetric Token Pruning for Efficient Vision-Language Model Inference
Feng, Yilin
Gulhan, Ahmed Burak
Kandemir, Mahmut Taylan
Machine Learning
Vision-Language Models (VLMs) process thousands of visual tokens per image alongside comparatively few text tokens, yet existing compression methods treat both modalities uniformly. We observe that the two modalities have fundamentally different properties: vision tokens are spatially redundant and dominate prefill, while text tokens are causally dependent and accumulate during decoding. Based on this asymmetry, we propose and empirically evaluate AsymVLM, which applies aggressive pruning to vision tokens before prefill using a learned importance scorer with per-sample adaptive budgeting, and temporal threshold-based eviction to text tokens only when they exceed a fixed budget. Our experiments indicate that AsymVLM achieves the highest FLOPs savings (up to 54%) among state-of-the-art methods while outperforming existing approaches by 2--3% on document and chart understanding tasks where visual information is spatially localized and query-specific, and maintaining competitive accuracy on holistic benchmarks. In text-dominated scenarios, our eviction strategy substantially outperforms standard LLM cache compression methods by adapting to the short-context nature of VLM.
title AsymVLM: Asymmetric Token Pruning for Efficient Vision-Language Model Inference
topic Machine Learning
url https://arxiv.org/abs/2605.29535