Rethinking Visual Layer Selection in Multimodal LLMs

Multimodal large language models (MLLMs) have achieved impressive performance across a wide range of tasks, typically using CLIP-ViT as their visual encoder due to its strong text-image alignment capabilities. While prior studies suggest that different CLIP-ViT layers capture different types of information, with shallower layers focusing on fine visual details and deeper layers aligning more closely with textual semantics, most MLLMs still select visual features based on empirical heuristics rather than systematic analysis. In this work, we propose a Layer-wise Representation Similarity approach to group CLIP-ViT layers with similar behaviors into {shallow, middle, and deep} categories and assess their impact on MLLM performance. Building on this foundation, we revisit the visual layer selection problem in MLLMs at scale, training LLaVA-style models ranging from 1.4B to 7B parameters. Through extensive experiments across 10 datasets and 4 tasks, we find that: (1) deep layers are essential for OCR tasks; (2) shallow and middle layers substantially outperform deep layers on reasoning tasks involving counting, positioning, and object localization; (3) a lightweight fusion of features across shallow, middle, and deep layers consistently outperforms specialized fusion baselines and single-layer selections, achieving gains on 9 out of 10 datasets. Our work offers the first principled study of visual layer selection in MLLMs, laying the groundwork for deeper investigations into visual representation learning for MLLMs.