MAIN PhD Seminars 2025

All the seminars at 13:00 in Aula Magna di Fisica.

Sandro Gabriele Tiddia: LLM Agents: Definitions and Real-World Applications

In this seminar, we explore the concept of ‘agents’ in artificial intelligence (AI), with a particular focus on the role of Large Language Models (LLMs) in powering these systems. The seminar begins by discussing a real-world application where LLMs are used to build a question-answering (QA) system, showing how LLMs can function as ‘agents’ within such a system. We then examine various definitions of ‘agent’ across AI subfields and consider how agents interact with their environment, make decisions, and pursue goals autonomously. Additionally, we revisit earlier works on agency in AI, reflecting on their original, more profound ideas, and connect them to recent developments and applications of LLM-powered agents. The seminar concludes by exploring experiments and use cases from recent literature, highlighting the capabilities and potential of LLM agents across different domains. The goal is to provide a clear introduction to the concept of LLM-powered agents, their role in AI systems, and how this concept has evolved from theoretical foundations to practical applications.

Andrea Azzarelli: Fractional Laplacian and ADMM for glyph extraction

In archaeology it is a common task to extract incisions or glyphs from a surface. This procedure is usually done manually and, therefore, it is prone to errors and it can be extremely time consuming. In this talk we present a variational model to automatically extract these incisions from a smooth surface. We provide a procedure to generate realistic synthetic data and we show the performances of the proposed method on this kind of data.

Valentino Artizzu: End-User Development for Extended Reality: Empowering Users to Create and Understand XR Environments

In this seminar we will see how to enable individuals without prior XR development experience to create and understand XR environments. It focuses on using End-User Development (EUD) techniques to allow users to design, build, and adapt XR systems. The research specifically explores methodologies and tools for non-programmers to construct XR environments. It examines how EUD can facilitate novice developers in comprehending existing VR environments for enhancement purposes. The study further investigates how EUD can empower domain experts to tailor these environments to meet diverse requirements. Additionally, it delves into how EUD can guide domain experts in configuring XR environments to support task learning and demonstration. The goal is to provide a clear overview of the topic using the experiences and the applications developed during the three years of a PhD career.

Nicola Piras: Global and local fit measures for latent class models and extensions

Latent class (LC) analysis is a powerful and flexible statistical tool for model-based clustering with categorical data. An important task in LC analysis is the choice of the number of clusters or classes. The choice of the number of classes is a selection model problem and usually Information Criteria are considered for this purpose. These are measures that weigh model fit (log-likelihood) and model complexity (based on the number of free parameters). The LC models formulation is subject to an assumption of conditional independence between the variables involved. Adherence to this assumption and the correct estimation of the parameters is central to evaluate how well the model fits to the data. While global selection and the goodness of fit is verified through Information Criteria, model conditions must also be checked. Specific statistics can be defined that allow to verify the local independence assumption. In the literature of LC analysis the statistics considered are the Bivariate residuals. The standard LC model can be modified to handle more complex data structures, and fit measures must be adapted to the new formulations. In this talk, after having briefly discussed the results in the standard formulation, I will also present the extension to the case in which data have a multilevel cross-classified structure. This structure is present when observations are simultaneously nested within two groups, for example, children nested within both schools and neighborhoods. An application is illustrated using an Italian dataset on the evaluation of students of their degree programmes, where degree programmes are nested in both universities and fields of study.

Sara Vergallo: Mathematics as a foundation for learning Machine Learning from primary school onward

The proliferation of artificial intelligence (AI) in many people’s daily lives has led the education sector to recognize the importance of teaching elements of AI—and in particular Machine Learning (ML)—from the earliest stages of schooling (Karalekas et al., 2023). Consequently, there is a need for resources, guidelines, and studies on the feasibility and methods for integrating ML into lower‐school settings, beginning in kindergarten (Sanusi et al., 2023), so that children can understand how the machines they interact with every day work (Lin & Brummelen, 2021). Teaching machine learning in primary and secondary school is very challenging, also due to students’ deficiencies in data analysis—especially in classification (sets) and data representation (trees, two‐way tables) (Grillenberger & Romeike, 2019)—an essential competency included in mathematics curricula from primary school onward and in secondary‐school computer science curricula, yet insufficiently promoted or stimulated (Grillenberger & Romeike, 2019). We investigated the level of these mathematical skills in a fifth‐grade class at a primary school and proposed a didactic pathway composed almost entirely of unplugged activities for learning the basics of ML. The activity fostered an improvement in the children’s classification and representation skills and received a high level of enthusiasm; this latter point is significant, as one of the critical issues identified in the literature is the lack of engaging activities within the school context (Grillenberger & Romeike, 2019). The results from these in‐class research activities will then be considered within a broader research overview concerning the use of non‐standard approaches (such as game‐based learning) to enhance mathematical skills, which are also necessary for a better understanding of computer‐science content.

Giorgia Nieddu: Use of GenAI to Support Learning and Teaching Mathematics

This seminar will present our recent experiments on the use of large multimodal models for learning and teaching mathematics. The first, conducted in North Macedonia, explored how students interact with GenAI in electronics problem-solving activities with mathematical content; the second, carried out in collaboration with the University of Turin, investigated how GenAI can support teachers in preparing educational materials. In this latter study, conducted within teacher training courses, we observed teachers’ behaviors and attitudes, aiming to understand whether AI could be useful in their lesson planning and how.

Lejzer Javier Castro Tapia: Mod-2 Cohomological Classification of Orbit Spaces of Free Involutions on the 2-Fold Projective Product Space

The action of a compact Lie group on a topological space describes the symmetries of our space, in that sense the properties of the orbit space of these actions has attracted many mathematicians over the world since the beginning of the twentieth century. In this talk we present in an informative way a cohomological classification via spectral sequences of orbit spaces of free involutions on the two-fold projective product space, a manifold that generalizes the usual projective spaces and wich was introduced for the first time by Donald Davis in 2010.

Matteo Mocci: Automatic Walkability Assessment using AI and multi-input image classification

Walkability is a key element of sustainable and livable cities, influencing public health, environmental impact, and social connectivity. Traditional methods for assessing walkability, such as surveys and audits, are often time-consuming and limited in scale. Recent advancements in artificial intelligence and computer vision offer new opportunities to automate and enhance these assessments using street-level and aerial imagery. This seminar explores how deep learning and multi-perspective image analysis can provide more comprehensive and scalable walkability evaluations. By integrating insights from urban planning, AI, and geospatial analysis, we will discuss the potential and challenges of these emerging technologies in shaping more pedestrian-friendly cities.

Giuseppe Zecchini: On the algebraic study of substructural logics by means of Płonka sums

Logic can be intuitively defined as the science of correct reasoning: given a certain set of premises, we want to be able to establish their consequences. Algebraic Logic can naively be defined as the study of Logic through the methods of Algebra. In the first part of this talk, we will explain in detail what a logic formally is and what it means to study it algebraically. In the second part, we will introduce and motivate substructural logics, which are traditionally defined by means of Gentzen-style sequent calculi in which one or more of the structural rules (exchange, weakening, contraction) of Gentzen’s LK calculus for classical logic are restricted or eliminated. Finally, in the third part, we will present the algebraic counterpart of substructural logics, residuated lattices, and make some remarks on the study of their structure through the construction of Płonka sums, a construction introduced in Universal Algebra in the 1960s by the eponymous Polish mathematician.

Antonio Pio Contrò: A Recent Specialization on the Notion of Formality

In topology, we can intuitively define the homotopy type of a space as “its behavior up to continuous deformations”. In the 1970s, Daniel Quillen and Dennis Sullivan developed the so-called rational homotopy theory for topological spaces, an elegant framework that makes it possible to completely understand a specific homotopy type (called the rational homotopy type) of certain topological spaces from a special algebraic structure associated with each of them: the minimal model of the space. In this context, the notion of a formal space was introduced for the first time. Roughly speaking, a formal space is a topological space for which the rational homotopy type can be reconstructed from its cohomology algebra with rational coefficients. The term “formal” comes from the fact that “the rational homotopy type and the minimal model are a formal consequence of the cohomology algebra.”. More recently, in the context of complex geometry, new notions of formality have been introduced for complex manifolds by J. Stelzig (LMU) and A. Milivojević (University of Waterloo). In this talk, we will present both notions of formality, beginning with the definition of a complex manifold and introducing the associated algebraic structure on differential forms. We will conclude with a brief, general overview of the existing relationships among the various concepts of formality.

Matteo Palmieri: A First Approach to the Bergman Kernel and Metric

The Bergman kernel and metric have been foundational tools in geometric analysis, since their invention by Stefan Bergman in 1922. The basic idea is to associate each bounded domain in the complex Euclidean space with a special Hilbert space — the square-integrable holomorphic functions on the domain — and study the geometric properties of such domain by analyzing the reproducing kernel of the space and the derived metric. From this concept, the Bergman kernel and metric evolved into powerful instruments for function theory, analysis, differential geometry, partial differential equations. The objective of this seminar is to present the Bergman kernel and metric, focusing on the well-behavior of these objects. To this end, we will review the necessary background on this topic and provide concrete examples and possible applications. Some natural questions will arise along the way, leading to a brief overview of the main directions of the current research.

Michele Faedda: Boolean Operators: Practical Problems and Real Applications

Boolean operations on 3D meshes — such as union, intersection, and difference — are essential in computer graphics, CAD, and digital fabrication. Despite their wide use, performing these operations robustly remains a challenge due to the limitations of floating-point arithmetic, which can lead to topological errors and unreliable results. This seminar explores the geometric and computational difficulties behind mesh-based Boolean operations, and presents a hybrid solution that combines floating-point and exact arithmetic to improve robustness without sacrificing efficiency. We will discuss how this approach enables more reliable modeling, and highlight potential applications, including version control for 3D assets through mesh differencing like git system.

Giuseppe Scarpi: GENIS, a simple and explainable algorithm for sentiment analysis of e-commerce reviews

GENIS is a simple algorithm that analyzes textual e-commerce reviews, calculating a rating that is more reliable than the typical “stars.” Compared to many similar algorithms, GENIS aims to be more explainable, providing elements that help users understand why a certain score was given. We will look at what GENIS does, but also the path taken to develop it—including mistakes and missteps—because sometimes, the best part of the journey is the road itself, not the destination.