Adapting ViT Representations to Specific Tasks

y_m_asano · x · 2026-07-19

This post shares a research diagram arguing that **pre-trained ViT representations should not be entirely static**. The author suggests that while a standard model's output won't change no matter how many times it runs a forward pass, allowing the model to "know what to look for" can adjust its representation to better suit the attributes of the current task. The diagram contrasts standard DINOv2 attention with SteerViT across different attributes: for the same input, SteerViT can direct its attention to task-relevant areas like a "person" or a "potted plant," rather than fixating on a single salient object like a static representation would.

Original post →

More from Research

Research channel →