Reproduce “Mirage Probes: How Vision Models Fake Visual Understanding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daniel Ben-Levi, Judah Goldfeder, Weiliang Zhao, Raz Lapid, Amit LeVi, Allen G. Roush, Ravid Shwartz-Ziv, Hod Lipson
Researchers at Columbia University developed Mirage Probes, a contrastive framework that decodes mirage behavior from Vision-Language Model internal representations, even when an image is present. Their findings suggest mirages arise from two distinct mechanisms – "spurious images" and "textual biases" – indicating a need for mechanism-specific mitigation strategies.
Key insights
- Mirage behavior is linearly decodable from VLM internal representations, even when an image is present, suggesting it's an accessible property within the latent space.
- The shift in VLM representations between image-present and image-absent conditions strongly correlates with mirage behavior, enabling highly accurate detection.
- Mirages arise from at least two distinct mechanisms: 'spurious images' (false internal visual content) and 'textual biases' (reliance on language priors), rather than a single phenomenon.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.