How Do Vision Models Work? (HOW)

How Do Vision Models Work (HOW) is a follow-up to last year’s Mechanistic Interpretability for Vision (MIV) workshop, expanding its focus from mechanistic analysis to a broader scientific study of how vision models learn, represent, and reason. As large-scale training has made learned representations increasingly opaque, developing a systematic understanding of these models has become a central scientific challenge. This workshop brings together researchers to investigate the principles and mechanisms underlying the behavior of vision models through observation, experimentation, and theory.