Diogo Sousa Soares

Recent research suggests that as neural networks scale, their internal representations converge towards a shared "platonic" structure. My PhD project aims to move this from a theoretical observation to a practical tool for building better models. While current models are powerful, they often struggle to measure how well their latent spaces actually capture real-world logic, e.g. the evolutionary distance between proteins or the semantic proximity of human concepts.

I propose developing a framework based on novel, robust similarity metrics to directly align these latent "manifolds" with known physical and biological patterns. This work will tackle three key areas: Descriptive (measuring how much "real-world" knowledge a model has actually learned), Optimization (using scientific biases to train models faster with less data), and Controllability (ensuring that the prompts we give generative models actually map correctly to their internal logic). By making latent spaces more interpretable and geometrically grounded, this research aims to create Al systems that are not just high-performing, but scientifically consistent and easier to steer.