Both human brains and AI models learn complex representations of visual and linguistic stimuli, yet their learning processes differ substantially. Do they nonetheless converge on similar representations? In my research, I investigate this question by measuring the alignment between the internal representations of vision-and-language AI models and human brain responses recorded while participants view visual or linguistic stimuli. This comparison can reveal both shared representational structures and key differences between biological and artificial intelligence, improving our understanding of both.

My research is part of the HAICu project, which deploys Artificial Intelligence to make digital heritage collections more accessible. Specifically, I work on using large language models to improve information access. This includes scientific text simplification and efficient retrieval-augmented generation.

Even though Large Language Models (LLMs) have become ubiquitous in modern society, these models still exhibit many undesirable behaviors, such as sycophancy, stereotypical bias and unfairness in high-stakes advice. In my research, I study whether users are equally affected by these behaviors, depending on aspects of the user's identity such as the language they speak, their sociodemographic characteristics but also what part of their identity they share in a conversation.

I work at the intersection of philosophy and logic, formulating better formalisms to understand what modality, assertion and truth are. Modality is a very important notion in logic, and modal logic is used to, for example, model how we reason about statements about the future, about our beliefs, the dynamics of information or about semantic rules. I work on the development of semantics of modal logic without possible worlds. Assertion (and denial) are about some the fundamental tools we have at our disposal to communicate as humans being, and understanding the various ways we can express our agreement and disagreement is what multilateral logic studies. I have proposed a more informative model theory for multilateral logic and its extensions. Finally, truth, which needs no introduction. Together with the team of the Project GENERALISE we have proposed a proper way to discuss the existence of both definitions of truth in higher-order logic and axiomatic theories of truth in first-order logic.

I work at the intersection of machine learning and logic. Logic helps us understand AI: by describing what a model does in logical terms, we can make it more explainable and more likely to generalize. AI helps us do logic: whether solving a crossword or writing a mathematical proof, we want models we can trust to follow the rules.

Traditionally, the meaning of a sentence is understood in terms of the conditions under which it is true. A shortcoming of this analysis is that it cannot capture the meaning of questions, because they are not true or false. The recently developed framework of inquisitive semantics identifies the meaning of a sentence with the conditions under which it is supported by a given information state. Within this framework, I study the interaction between modality and questions, from a linguistic as well as from a logical perspective.

Our reasoning, as individuals and as groups, is subject to limitations, e.g. in time, memory, attention, and prone to systematic divergence from the standard picture of rationality. However, logical models for human reasoning often neglect such findings. In this project, we argue for the importance of building bridges between logic, psychology, and epistemology, and develop formal frameworks to this end.

Neural networks are the powerful engines of modern artificial intelligence. However, we still don't really understand them: unlike classical computers, we don't have a good language or framework to explain and predict the behaviour of neural networks. Finding such a framework is a fascinating open problem at the intersection of AI, logic, math, and cognitive science.

Why are blackbox artificial neural network models so good at many natural language processing tasks? I hope that, by studying them, I can learn more about what makes natural language what it is and then use that knowledge to make computer models of language even better.

There are many things that are not the case, but that we know could have been the case. I could have taken public transport to get to work instead of taking the bike. Having knowledge of these non-actual situations is crucial for our future planning, risk assessment, and counterfactual reasoning. I use tools of logic, cognitive science, and philosophy to try to understand how it is that we know what is possible and whether we can justify this knowledge.

Making decisions collectively in a fair and efficient way is a central part of a well-functioning society. We design and analyse methods for group decision making using tools from the areas of mathematics and logic in small-scale but also in large-scale contexts, like national elections, business choices, negotiations between friends and families, etc.

Quantum computers may bring us great computational power, but how can we make that power available to users that cannot house an entire quantum machine? By analyzing the information contained in quantum states, I design algorithms for delegated quantum computation that, ideally, ensure privacy and verifiability for the user.

While all animals have a communication system, human language is unique in some of its properties. Which of our learning abilities are we using when we acquire new language? And how do these learning abilities differ from those of animals?

Most of the time, we do not think about what we do. We just respond in an adequate way to what our environment has to offer. I am interested in how brain, body and world contribute jointly to such episodes of everyday skillful behavior, but also in how unreflective and more reflective actions mesh together. I use phenomenology, theoretical neuroscience and embodied cognitive science to try to better understand these questions.

My research interests revolve around the intersection of Category Theory and Modal Logic. I am interested in relating the two at an abstract level, investigating whether techniques and results can be transferred from one to the other, and in specific applications, such as topics in Theoretical Computer Science, Quantum Theory and formal models of social phenomena.