Stateful visual encoders improve vision-language models

Arxiv CS.CV (computer vision)

Paper proposes stateful visual encoders that process video frames with memory, enabling models to detect visual changes without relying solely on language. Outperforms existing VLMs on multi-image and video tasks by encoding temporal context directly in the vision backbone.

vision-language-models

Arxiv CS.CV (computer vision)·Zirui WangJunwei YuAdam YalaDavid M. ChanJoseph E. GonzalezTrevor Darrell

7 days ago