Starve to Perceive: Taming Lazy Perception in VLMs with Constrained Visual Bandwidth

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wu, Yuhuan, Wei, Cong, Lin, Fangzhen, Chen, Wenhu, Wang, Haozhe
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914578469748736
author Wu, Yuhuan
Wei, Cong
Lin, Fangzhen
Chen, Wenhu
Wang, Haozhe
author_facet Wu, Yuhuan
Wei, Cong
Lin, Fangzhen
Chen, Wenhu
Wang, Haozhe
contents Vision-Language Models (VLMs) deployed as situated agents in high-resolution visual environments require active perception -- the ability to dynamically decide where to look through operations like zooming, cropping, and panning. However, current training paradigms produce models that mimic the surface form of such operations without functionally depending on their outputs, a phenomenon we term lazy perception. We trace this to a fundamental learning asymmetry: when coarse global views combined with language priors suffice for moderate accuracy, the model has no incentive to learn harder multi-step visual search. If a model can succeed without actively looking, it will never learn to look. This motivates Starve to Perceive, a training paradigm that constrains visual bandwidth -- restricting each observation to a tight token budget so that no single view suffices for task completion, making active perception the only viable strategy. Despite requiring no auxiliary losses, reward shaping, or architectural changes -- serving as a minimal, plug-in modification to standard post-training pipelines -- models trained under perceptual starvation achieve substantial gains of 5% average relative improvement across diverse benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2605_18603
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Starve to Perceive: Taming Lazy Perception in VLMs with Constrained Visual Bandwidth
Wu, Yuhuan
Wei, Cong
Lin, Fangzhen
Chen, Wenhu
Wang, Haozhe
Computer Vision and Pattern Recognition
Vision-Language Models (VLMs) deployed as situated agents in high-resolution visual environments require active perception -- the ability to dynamically decide where to look through operations like zooming, cropping, and panning. However, current training paradigms produce models that mimic the surface form of such operations without functionally depending on their outputs, a phenomenon we term lazy perception. We trace this to a fundamental learning asymmetry: when coarse global views combined with language priors suffice for moderate accuracy, the model has no incentive to learn harder multi-step visual search. If a model can succeed without actively looking, it will never learn to look. This motivates Starve to Perceive, a training paradigm that constrains visual bandwidth -- restricting each observation to a tight token budget so that no single view suffices for task completion, making active perception the only viable strategy. Despite requiring no auxiliary losses, reward shaping, or architectural changes -- serving as a minimal, plug-in modification to standard post-training pipelines -- models trained under perceptual starvation achieve substantial gains of 5% average relative improvement across diverse benchmarks.
title Starve to Perceive: Taming Lazy Perception in VLMs with Constrained Visual Bandwidth
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2605.18603