Gabriel Franco

I am Gabriel Franco, a fifth-year Computer Science PhD candidate at Boston University, advised by Prof. Mark Crovella. My research aims to reverse-engineer the internal computations of large language models (LLMs), moving beyond correlational observations to understand how and why they truly work.

My primary focus is on mechanistic interpretability. I develop methods to uncover the causal drivers of model behavior, with a particular interest in causality within the attention mechanism. My approach is to tackle interpretability by using the model’s own computations, leveraging the low-rank structures that naturally arise in these systems.

Before focusing on interpretability, my Master’s research with Prof. Giovanni Comarela at the Federal University of Viçosa explored weakly supervised learning, specifically the problem of Learning from Label Proportions (LLP). I also have industry experience as a Data Scientist at SEEK and Localiza, where I designed, deployed, and optimized production-level machine learning models and recommender systems.