Perceiving Beyond Language Priors: Enhancing Visual Comprehension and Attention in Multimodal Models

TL;DR

This paper investigates how Multimodal Large Language Models (MLLMs) can better understand visual content by analyzing their internal mechanisms and introducing techniques to enhance visual comprehension and attention, leading to improved performance on visual tasks.

Contribution

The paper provides new insights into the internal visual understanding of MLLMs and proposes techniques to amplify visual comprehension and attention, improving multimodal performance.

Findings

Enhanced ability to predict visually-dependent tokens

10-point boost on visually challenging tasks

Deeper understanding of visual content in MLLMs

Abstract

Achieving deep alignment between vision and language remains a central challenge for Multimodal Large Language Models (MLLMs). These models often fail to fully leverage visual input, defaulting to strong language priors. Our approach first provides insights into how MLLMs internally build visual understanding of image regions and then introduces techniques to amplify this capability. Specifically, we explore techniques designed both to deepen the model's understanding of visual content and to ensure that these visual insights actively guide language generation. We demonstrate the superior multimodal understanding of our resultant model through a detailed upstream analysis quantifying its ability to predict visually-dependent tokens as well as 10 pt boost on visually challenging tasks.