Index-Preserving Lightweight Token Pruning for Efficient Document Understanding in Vision-Language Models

Jaemin Son, Sujin Choi, Inyong Yun
9/8/2025
cs.CVcs.AIcs.CL

No official code found. Best implementation: jaeminSon/index-preserving-lightweight-token-pruning (1 star, last commit 2025-12-17).

Abstract

Recent progress in vision-language models (VLMs) has led to impressive results in document understanding tasks, but their high computational demands remain a challenge. To mitigate the compute burdens, we propose a lightweight token pruning framework that filters out non-informative background regions from document images prior to VLM processing. A binary patch-level classifier removes non-text areas, and a max-pooling refinement step recovers fragmented text regions to enhance spatial coherence. Experiments on real-world document datasets demonstrate that our approach substantially lowers computational costs, while maintaining comparable accuracy.

View on arXivView PDF

Code Implementations(1)

jaeminSon/index-preserving-lightweight-token-pruningCommunity66%

Huggingface model and misc codes in paper "Index-Preserving Lightweight Token Pruning for Efficient Document Understanding in Vision-Language Models"

10PythonSep 8, 202510 months ago

Similar Papers

VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm

84% match1 repo122025

VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm

84% match1 repo122025

Focus-Scan-Refine: From Human Visual Perception to Efficient Visual Token Pruning

83% match1 repo42026

Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration

83% match2024

VDInstruct: Zero-Shot Key Information Extraction via Content-Aware Vision Tokenization

83% match2025

Efficient Vision-Language Reasoning via Adaptive Token Pruning

83% match2025

Cite this paper

@article{son2025indexpreserving,
  title  = {Index-Preserving Lightweight Token Pruning for Efficient Document Understanding in Vision-Language Models},
  author = {Jaemin Son and Sujin Choi and Inyong Yun},
  year   = {2025},
  eprint = {2509.06415},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2509.06415v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2509.06415.svg)](https://codeofpaper.com/paper/2509.06415)

Discussion