Do VLMs Need Vision Transformers? Evaluating State Space Models as Vision Encoders
This work addresses the problem of efficient and robust vision encoders for VLMs, offering an incremental improvement by proposing SSMs as an alternative to transformers.
The paper investigates whether state space model (SSM) vision backbones can replace transformer-based encoders in vision-language models (VLMs), finding that SSM backbones achieve strong performance in VQA and grounding tasks, remain competitive after adaptation, and operate at a smaller scale.
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.