My research focuses on understanding how language and vision-language models
represent and process information. I am interested in white-box interpretability methods to study how knowledge is organized and transformed inside deep neural networks, analyzing the relationship between the geometry and semantics of neural representations. I also study how to combine mechanistic interpretability with geometric and unsupervised methods from manifold learning to characterize the computational principles underlying modern multimodal models.

I got my Ph.D. from the International School for Advanced Studies (SISSA, Trieste), where I developed an algorithm for estimating the intrinsic dimension of high-dimensional datasets. Using this method, I showed how geometric compression is linked to semantic abstraction in transformers trained with self-supervision. I also investigated the density structures of hidden representations in convolutional neural networks, showing how they relate to the hierarchical organization of concepts, and how the low-dimensional geometry of hidden representations can explain the generalization capabilities of overparameterized neural networks.