Davide Baldelli

Hi, I’m Davide 😊. I am a second-year PhD student at MILA and Polytechnique Montréal, advised by Sarath Chandar and Amal Zouaq.

I am currently participating in MATS 10.0, an AI alignment fellowship, in the LawZero stream, where I am mentored by Damiano Fornasiere and Mirko Bronzi. We are working on introspection capabilities in language models.

Since the beginning of my PhD, I have worked on language models for computer-aided design, including building a large annotated CAD dataset and fine-tuning open-source code LLMs for text-driven sequential CAD generation. I then focused on memory for language-model agents, studying why standard chat-based agents cannot reliably generate and retain private secret state across interactions (for example, why they cannot properly play Hangman). More recently, I have worked on probabilistic calibration, addressing the problem that language models are often poorly calibrated random samplers through fine-tuning methods.

Previously, I was a researcher at the National Institute of Informatics (Tokyo) under the supervision of Akiko Aizawa. I received an MSc in Artificial Intelligence from the University of Bologna in October 2023, where I was supervised by Paolo Torroni.