Remo Sasso

Queen Mary University of London

I hold a BSc and MSc in Artificial Intelligence at the University of Groningen (NL) and am currently a PhD student at the Queen Mary University of London under the supervision of Paulo Rauber. In addition to my academic work, I have worked as a Machine Learning engineer, and am currently the Head of AI at xDNA, an AI/Cybersecurity-based start-up from the Netherlands. Here I'm leading the initiative Project Aletheia, where we develop AI-driven tools to optimize the workflow of professional fact-checkers, with the overarching goal of ensuring information integrity in the world.

Foundation World Models and Foundation Agents for Reinforcement Learning

My research focuses on developing reinforcement learning algorithms that are both scalable and sample-efficient through Bayesian methods and model-based approaches, recently with a particular emphasis on Large Language Models (LLMs). My previous research focused on principled, efficient and scalable exploration algorithms for reinforcement learning, e.g. Poster Sampling for Deep Reinforcement Learning (ICML 2023), where we developed a reinforcement learning algorithm that can be considered state-of-the-art in Atari games. 

Currently I'm particularly interested in the integration of LLMs in the reinforcement learning framework, both as decision-making agents and simulators. My current research, called "Foundation World Models and Foundation Agents for Reinforcement Learning" investigates this integration in-depth and shows that large models show significant potential in various reinforcement learning tasks, ranging from decision-making in stochastic environments to serving as world models.