Multimodal Language Models See Better When They Look Shallower
This work addresses performance bottlenecks in MLLMs for fine-grained visual tasks, offering a principled approach to enhance model accuracy, though it is incremental in optimizing existing architectures.
The study tackled the problem of suboptimal visual feature extraction in multimodal large language models (MLLMs) by analyzing ViT layer selection, finding that shallow and middle layers outperform deep layers on fine-grained visual tasks like counting and localization, with a proposed fusion method achieving consistent improvements across 10 benchmarks.
Multimodal large language models (MLLMs) typically extract visual features from the final layers of a pretrained Vision Transformer (ViT). This widespread deep-layer bias, however, is largely driven by empirical convention rather than principled analysis. While prior studies suggest that different ViT layers capture different types of information, with shallower layers focusing on fine visual details and deeper layers aligning more closely with textual semantics, the impact of this variation on MLLM performance remains underexplored. We present the first comprehensive study of visual layer selection for MLLMs, analyzing representation similarity across ViT layers to establish shallow, middle, and deep layer groupings. Through extensive evaluation of MLLMs (1.4B-7B parameters) across 10 benchmarks encompassing 60+ tasks, we find that while deep layers excel in semantic-rich tasks like OCR, shallow and middle layers significantly outperform them on fine-grained visual tasks including counting, positioning, and object localization. Building on these insights, we propose a lightweight feature fusion method that strategically incorporates shallower layers, achieving consistent improvements over both single-layer and specialized fusion baselines. Our work offers the first principled study of visual layer selection in MLLMs, showing that MLLMs can often see better when they look shallower.