About Me
Hi, I'm Adrian. I'm a CS PhD student at Sapienza University of Rome, co-advised by Prof. Emanuele Rodolà (PI of the GLADIA Lab) and Prof. Iacopo Masi (PI of the OmnAI Lab).
My research focuses on Large Language Models (LLMs), specifically exploring model merging and uncertainty estimation. The primary goal of my work is to develop methods that make these models more robust, efficient, and interpretable.
Background
I have always been driven by a desire to automate tasks and optimize processes, an inclination that led me to explore electronics and programming from a young age. This path began with a high school focus on Electronics, followed by a Bachelor’s degree in Computer Engineering. Towards the end of my undergraduate studies, I began exploring Machine Learning, fascinated by the prospect of machines discovering the "hidden rules" of the world autonomously rather than relying on manually defined logic.
Driven by this passion, I enrolled in a Master’s in Artificial Intelligence. As I built increasingly complex models, the emergence of LLMs like ChatGPT pushed me to investigate the field further through research projects. This eventually led me to start a PhD, under the co-supervision of Emanuele Rodolà and Iacopo Masi. However, my excitement was often tempered by frustration over the inherent inefficiency and lack of interpretability in deep learning models. This directed my research toward novel methods in model merging and uncertainty estimation to push the state of the art, studying LLMs through the dual perspectives of data modeling and parameter optimization. I keep coming back to a few questions I can't quite shake: How do LLMs encode knowledge into their parameters? Can we define and quantify what a model knows, and what it does not?
In 2025, I joined the Adaptive Bayesian Intelligence team at RIKEN AIP in Tokyo as a research intern under the supervision of Thomas Möllenhoff and Emtiyaz Khan. While I had touched upon the elegance of Bayesian Inference during my Master's, I had previously set it aside due to its perceived intractability for deep learning. Well, I was very happy to find out that I'd been wrong. My time at RIKEN allowed me to deep dive into variational learning and approximate bayes, offering a rigorous theoretical framework that unified my previous work and opened up new perspectives. I now view my research through this lens, using it to design even more efficient and robust methods, and I am eager to continue this trajectory in my studies.
Contact
Email: [email protected]