Predicting Middle-Layer Attention in MLLMs for Efficient Visual Token Pruning

A new arXiv paper (2608.06411) addresses efficiency bottlenecks in multimodal large language models (MLLMs) caused by processing numerous visual tokens. The authors propose a method to predict middle-layer attention, which is used to guide visual token pruning. Their analysis reveals that the optimal layer for attention varies across samples, making fixed-layer selection suboptimal. Additionally, obtaining attention from the appropriate middle layer requires processing many visual tokens through several layers, incurring significant computation. The proposed approach aims to predict the attention from the most responsive layer without full processing, thereby reducing computational cost. The paper is announced as a new submission and is available on arXiv.