Do VLMs Need Vision Transformers? Evaluating State Space Models as Vision Encoders

TL;DR

This paper evaluates state space model (SSM) vision backbones as alternatives to transformer-based encoders in vision-language models, demonstrating competitive performance and robustness, especially at smaller scales.

Contribution

It systematically compares SSM and transformer backbones in VLMs, showing SSMs as a viable, efficient alternative with improved robustness and competitive results.

Findings

SSM backbones achieve strong performance in VQA and grounding.

Dense-task tuning enhances backbone performance.

Larger ImageNet accuracy does not guarantee better VLM results.

Abstract

Large vision--language models (VLMs) often use a frozen vision backbone, whose image features are mapped into a large language model through a lightweight connector. While transformer-based encoders are the standard visual backbone, we ask whether state space model (SSM) vision backbones can be a strong alternative. We systematically evaluate SSM vision backbones for VLMs in a controlled setting. Under matched ImageNet-1K initialization, the SSM backbone achieves the strongest overall performance across both VQA and grounding/localization. We further adapt both SSM and ViT-family backbones with detection or segmentation training and find that dense-task tuning generally improves performance across families; after this adaptation, the SSM backbone remains competitive while operating at a substantially smaller model scale. We further observe that (i) higher ImageNet accuracy or larger backbones do not reliably translate into better VLM performance, and (ii) some visual backbones are unstable in localization. Based on these findings, we propose stabilization strategies that improve robustness for both backbone families and highlight SSM backbones as a strong alternative to transformer-based vision encoders in VLMs.