Beyond Static Cropping: Layer-Adaptive Visual Localization and Decoding Enhancement

TL;DR

This paper introduces a dynamic, layer-adaptive approach to visual grounding in large vision-language models, improving reasoning and localization by selecting task-specific layers during inference.

Contribution

It proposes VAQ, a new metric for identifying relevant layers, and LASER, a training-free method that adaptively enhances visual localization and reasoning tasks.

Findings

LASER improves VQA accuracy across diverse benchmarks

Layer-wise sensitivity analysis reveals different layers for simple and complex tasks

VAQ effectively identifies task-relevant layers for visual grounding

Abstract

Large Vision-Language Models (LVLMs) have advanced rapidly by aligning visual patches with the text embedding space, but a fixed visual-token budget forces images to be resized to a uniform pretraining resolution, often erasing fine-grained details and causing hallucinations via over-reliance on language priors. Recent attention-guided enhancement (e.g., cropping or region-focused attention allocation) alleviates this, yet it commonly hinges on a static "magic layer" empirically chosen on simple recognition benchmarks and thus may not transfer to complex reasoning tasks. In contrast to this static assumption, we propose a dynamic perspective on visual grounding. Through a layer-wise sensitivity analysis, we demonstrate that visual grounding is a dynamic process: while simple object recognition tasks rely on middle layers, complex visual search and reasoning tasks require visual information to be reactivated at deeper layers. Based on this observation, we introduce Visual Activation by Query (VAQ), a metric that identifies the layer whose attention map is most relevant to query-specific visual grounding by measuring attention sensitivity to the input query. Building on VAQ, we further propose LASER (Layer-adaptive Attention-guided Selective visual and decoding Enhancement for Reasoning), a training-free inference procedure that adaptively selects task-appropriate layers for visual localization and question answering. Experiments across diverse VQA benchmarks show that LASER significantly improves VQA accuracy across tasks with varying levels of complexity.