Large Vision-Language Models Get Lost in Attention

TL;DR

This paper introduces a theoretical framework to analyze large vision-language models, revealing that attention and feedforward networks serve distinct functions and that current models may be inefficiently using attention mechanisms.

Contribution

It provides a unified information-theoretic and geometric analysis of residual modules, highlighting functional decoupling and inefficiencies in current LVLM architectures.

Findings

Attention acts as a subspace-preserving reconfiguration operator.

Feedforward networks serve as subspace-expanding semantic drivers.

Replacing learned attention with predefined values maintains or improves performance.

Abstract

Despite the rapid evolution of training paradigms, the decoder backbone of large vision--language models (LVLMs) remains fundamentally rooted in the residual-connection Transformer architecture. Therefore, deciphering the distinct roles of internal modules is critical for understanding model mechanics and guiding architectural optimization. While prior statistical approaches have provided valuable attribution-based insights, they often lack a unified theoretical basis. To bridge this gap, we propose a unified framework grounded in information theory and geometry to quantify the geometric and entropic nature of residual updates. Applying this unified framework reveals a fundamental functional decoupling: Attention acts as a subspace-preserving operator focused on reconfiguration, whereas FFNs serve as subspace-expanding operators driving semantic innovation. Strikingly, further experiments demonstrate that replacing learned attention weights with predefined values (e.g., Gaussian noise) yields comparable or even superior performance across a majority of datasets relative to vanilla models. These results expose severe misallocation and redundancy in current mechanisms, suggesting that state-of-the-art LVLMs effectively ``get lost in attention'' rather than efficiently leveraging visual context.