Mirage Probes: How Vision Models Fake Visual Understanding

Abstract

Mirage Probes uses matched questions and internal model activations to study answers that appear visually grounded even without the image. It distinguishes reliance on textual biases from false visual content in the learned representations, suggesting different interventions for each failure mode.

Ravid Shwartz-Ziv

AI Researcher

AI researcher at Meta MSL with a background in information theory and computational neuroscience, working on world models, memory, and compression. Former Assistant Professor and Faculty Fellow at NYU, collaborating on research across academia and industry.