cv | Gabriel Franco

Full Name	Gabriel Franco
Email	gvfranco@bu.edu
Research Interests	Mechanistic interpretability, representation geometry, feature alignment, and causal circuit discovery. Previous work includes LLM evaluation and learning from label proportions.

2021 – now
Research Assistant

Boston University, Boston, MA
- Mechanistic interpretability of LLMs: developed spectral decomposition methods to identify causal communication circuits in attention, enabling circuit tracing and low-rank model interventions.
- Learning from Label Proportions (LLP): characterized LLP into distinct variants and proposed benchmark datasets and model-selection strategies for each.
Summer 2024
Data Scientist Intern

Microsoft, Redmond, WA
- Fine-tuned and optimized multi-modal small language models for local image tagging, reaching about 85% of a large foundation model's performance under a sub-5s inference target.

Spring 2025

Teaching Assistant — CAS CS 132 (Geometric Algorithms)

Boston University
Fall 2023

Teaching Assistant — CDS DS 701 (Tools for Data Science)

Boston University