My research studies what representations learn from supervision—particularly language supervision—and when such learning leads to identifiable latent structure.

I work on understanding when modern learning objectives recover latent structure beyond predictive performance, using tools from identifiability theory, latent-variable modeling, and representation geometry. In particular, I study the equivalence classes of representations induced by learning objectives, and how cross-modal supervision shapes the geometry of vision-language models. I am also interested in how these learned representations relate to human-interpretable concepts.