Nicolò De Sabbata

Hello there! Welcome to my corner of the web. I'm an engineer and researcher passionate about deep learning, reverse-engineering human intelligence, and exploring practical applications of AI in the real world. Currently, I'm a ML Research Intern at EPFL in Switzerland, where I recently earned my Master's degree in Computer Science. I'm part of the EPFL NLP group, working under the supervision of Prof. Antoine Bosselut.

My research interests lie in improving our understanding of the architecture and learning dynamics of contemporary AI systems and leveraging these insights to develop the next generation of intelligent agents. Specifically, my work focuses on two primary goals: (i) understanding the capabilities and limitations of LLMs through comparisons with the human mind, and (ii) addressing these limitations to build agents that can learn, reason, and communicate with efficiency and flexibility.

Previously, I was a Visiting Student Researcher at the Computational Cognitive Science Lab at Princeton University, collaborating with Prof. Tom Griffiths. My industry experience includes internships at Amazon in Luxembourg, where I automated ETL pipelines monitoring for ATS, and at Axa in Lausanne, where I designed and implemented LLM-based agents to automate the processing of insurance claims.

Email  /  CV  /  GitHub  /  Linkedin  /  Google Scholar  /  Twitter

We developed a novel method to optimize reasoning in large language models using a novel reward function based on the Value of Computation, enabling models to selectively use intermediate reasoning steps only when necessary. We demonstrated the effectiveness of this approach across diverse datasets, reducing inference costs by up to 37% without compromising performance.

Investigated vision-language models' limitations in multi-object reasoning, linking them to cognitive science's binding problem and trade-offs between flexibility and capacity. Demonstrated that VLMs face human-like limits in multi-object scene processing in tasks like counting, localization, and visual analogy.

A GPT-2 model fine-tuned on demonstrations distilled from a 100B - parameter scale LLM, with generation quality enhanced by a reward model trained via reinforcement learning with human feedback (RLHF).

Implemented a deep convolutional encoder-decoder neural network for image denoising in PyTorch, trained without a clean reference image.

A route planner that calculates the fastest route based on desired arrival times and analyzes historical delay data for accurate predictions using Apache Spark on Swiss Federal Railways (SBB) data.

Developed a Spark-based distributed recommendation system in Scala, processing millions of reviews. Implemented an approximate and distributed K-Nearest Neighbors algorithm to enhance recommendation precision.

Processed millions of quotes using statistical and sentiment analysis, followed by topic clustering with a pre-trained BERT encoder to predict the speaker's political party affiliation. Created and presented comprehensive visualizations of the results through a website, data story, and plots.

Explored the feasibility of Neural Networks models for the quantitative discrimination between different conformational species of proteins.

Built a classifier using various ML models to predict Higgs boson decay signatures from background noise, achieving a categorical accuracy of 0.840 and an F1 score of 0.756, ranking 6th among nearly 300 teams.

Completed two data analysis projects to enhance skills in data wrangling, data visualization, regression analysis, observational studies, statistics, and supervised learning.

Developed an online multiplayer board game in Java, playable on both JavaFX GUI and CLI, with robust server features for connection stability. Managed the full development cycle using Agile, from UML modeling to implementation, ensuring scalability, reliability, and security through effective design patterns.

Developed the backend and frontend of a website that allows users to securely upload songs to a server, ensuring safe database communications.

Website template credits to Jon Barron.