Look Where It Matters: High-Resolution Crops Retrieval for Efficient VLMs

TL;DR

AwaRes is a framework that dynamically balances accuracy and efficiency in vision-language models by selectively retrieving high-resolution image segments only when necessary for the query.

Contribution

It introduces a novel spatial-on-demand approach that combines low-resolution global views with targeted high-resolution retrieval using tool-calling and multi-turn trajectories.

Findings

Improves accuracy-efficiency trade-off in VLMs

Reduces computational costs while maintaining high accuracy

Demonstrates effectiveness on benchmark datasets

Abstract

Vision-language models (VLMs) typically process images at a native high-resolution, forcing a trade-off between accuracy and computational efficiency: high-resolution inputs capture fine details but incur significant computational costs, while low-resolution inputs advocate for efficiency, they potentially miss critical visual information, like small text. We present AwaRes, a spatial-on-demand framework that resolves this accuracy-efficiency trade-off by operating on a low-resolution global view and using tool-calling to retrieve only high-resolution segments needed for a given query. We construct supervised data automatically: a judge compares low- vs.\ high-resolution answers to label whether cropping is needed, and an oracle grounding model localizes the evidence for the correct answer, which we map to a discrete crop set to form multi-turn tool-use trajectories. We train our framework with cold-start SFT followed by multi-turn GRPO with a composite reward that combines semantic answer correctness with explicit crop-cost penalties. Project page: https://nimrodshabtay.github.io/AwaRes