Blog

Personal perspectives on AI research, trustworthy AI, AI for science, and teaching. For research write-ups, see the research pages and the DREAM Lab blog.

  • Vision Should Be Structural

    Why my group spent a few years turning images into SVG code that language models can read, and what I think still holds.

  • A Unified View of Trustworthy AI

    Robustness, fairness, and interpretability are usually studied apart, but many of their methods answer the same question about data.

  • We Evaluate Prompt-Injection Defenses Too Simply

    A low attack success rate can mean a model handled untrusted text correctly, or that it threw the text away, and the standard metric does not say which.