9:30

Stephanie Malvicini

A Principled Hybrid Multi-Agent Architecture for Simulating Social Platforms: Applications to Misinformation and Polarization

Simulating social platforms has become increasingly important as access to real-world data grows more restricted, while controlled experimentation on existing platforms remains impractical or infeasible. Yet beyond methodological constraints, we argue that simulation serves a deeper epistemic function: making visible the social and psychological dynamics that remain opaque to direct observation. Multi-agent systems offer a natural framework for implementing such simulations, as they allow populations of autonomous agents to interact, form beliefs, and produce emergent collective behaviors within a shared environment, mirroring the structure of real online platforms. These agents can be implemented along a spectrum of cognitive fidelity, from symbolic rule-based models to fully language-driven ones. Traditional Agent-Based Models (ABM) and other classical cognitive agents’ architectures, such as Belief-Desire-Intention (BDI), offer transparency and strong theoretical foundations but fall short in environments characterized by rich natural language interaction. Language is, however, the primary medium of social media: users argue, persuade, spread narratives, and form identities through text, making linguistic realism a critical dimension of any faithful simulation. Large Language Model (LLM)-driven agentic approaches enable more realistic linguistic behavior, often at the cost of interpretability and reproducibility. We work towards the design and implementation of a Social Platform Simulator Framework built on a flexible Social Agent model that spans a spectrum from fully LLM-based agents to structured symbolic architectures, explicitly modeling core platform components while allowing access to the agents' internal state. By developing a framework that supports varying degrees of cognitive fidelity and control, we aim to investigate how architectural choices shape emergent social phenomena and to provide a foundation for the systematic study of online social dynamics. We demonstrate the potential of the framework through a series of case studies focused on social phenomena such as affective polarization, misinformation diffusion, and argumentation dynamics.

10:00

Bjoern Komander

Automated Monitoring of Online Communities

This thesis investigates the complex dynamics governing the evolution and stability of online communities. Recognizing the challenges these platforms face, we leverage advancements in temporal and dynamic Graph Neural Networks (GNNs) to analyze the large-scale, time-varying relational data inherent in community interactions. The research involves developing and validating methods, including the use of synthetic network models, to ensure GNNs effectively capture fundamental network processes like growth and temporal shifts. We direct these techniques towards understanding critical community phenomena, such as evolving user activity patterns, shifts in network structure, and the overall factors influencing community health and longevity over time. The ultimate aim is to develop robust computational tools capable of monitoring and potentially predicting key dynamic changes within online communities, contributing to a deeper understanding of their lifecycle and sustainability.

10:30

Coffee

11:00

Alba Aguilera

Agent-based Modeling for Equitable Policy-Making

Artificial intelligence (AI) has the potential to play a crucial role in informing societal decision-making. AI-based simulations, in particular, can help anticipate the impact of legal norms in diverse environments in a non-invasive way. This research focuses on the development of agent-based models (ABMs) or simulations for policy-making analysis in contexts of social inequity. The work is grounded in the Capability Approach (CA), a widely recognized framework for analyzing, promoting and assessing human well-being and development. We use the CA to ground both the decision-making and evaluation of the simulation, distinguishing between the agents' opportunities and realized outcomes by explicitly considering their contextual constraints and individual choices. By analyzing and testing different decision-making architectures, such as rule-based and MDP-based approaches, and calibrating them using likelihood-based techniques; this research aims to provide a simulation framework that serves as a basis for policy simulation tools across different social contexts.

11:30

Shuxian Pan

Automated Extraction of Online Community Norms

To develop cooperative multiagent systems effectively, we aim to create an architecture that facilitates the agents' dynamic adoption of conventions. It expands an existing agent model's action selection architecture with a component that applies Natural Language Processing techniques with a domain-specific knowledge base. This component embeds conventions into agent interaction strategies to improve the predictability of other agents' actions. At the same time, Natural Language Processing techniques allow the users to introduce the conventions or provide their domain-specific knowledge to the multi-agent system in a more user-friendly way.

12:00

Amir Nadeem

Integrating Symbolic Reasoning and Machine Learning in Boolean Optimization

Boolean Satisfiability (SAT) and its optimization variants, MaxSAT and MinSAT, are central problems in computer science and artificial intelligence, serving as a universal modeling language for a wide range of industrial and scientific applications. Although Conflict-Driven Clause Learning (CDCL) revolutionized SAT solving, current solvers face substantial challenges as industrial instances continue to grow in size and structural complexity. In MaxSAT and MinSAT, optimization introduces additional layers of search, bounding, and relaxation that exacerbate scalability issues. To address these limitations, this project aims to design, implement, and evaluate new solving techniques that integrate machine learning (ML) into symbolic reasoning systems. Specifically, we propose to develop novel hybrid MaxSAT and MinSAT solvers that incorporate ML techniques at multiple levels: heuristic guidance, encoding optimization, and end-to-end neural approximations. Ultimately, this research aims to bridge data-driven methods and formal logic to advance the state of the art in combinatorial optimization.

12:30

Jairo Lefebre Lobaina

Explainable Dialogue-Based Multi-Agent Recommender System

This thesis addresses the challenge of building trustworthy multi-agent recommender systems by integrating formal explainability, interactive dialogue, and computational trust into a unified architecture. The overarching goal is to ensure that users not only receive accurate recommendations from multiple independent agents, but also receive rigorous and understandable justifications for those recommendations, and are empowered to challenge and refine them through structured interaction, including understanding the reasons behind the ordering of recommendations. We pursued this goal through three main research directions. First, we developed a comprehensive suite of polynomial-time algorithms for computing sample-based formal explanations, enabling any black-box classifier to justify its decisions with provable correctness guarantees. Second, we designed and evaluated an interactive dialogue protocol through which users can impose constraints, request alternative explanations, and explore counterfactual scenarios in real time. Third, we extended the PAAS trust model into the EPAAS framework, which learns fine-grained trust distributions from user feedback and uses them to rank recommendations in multi-agent settings.

13:00

Laura Rodriguez Cima

Value-Driven Negotiation in Multi-Agent Systems: Embedding Social Preferences Within Utility Functions

This thesis presents a self‐contained, value‐driven negotiation framework in which each autonomous agent internalizes its own notion of fairness or other values directly within its utility function, removing dependence on external fairness benchmarks such as Nash or Kalai points. We formalize this dual‐component utility model, combining an individual‐utility term with a social‐utility term that dynamically aggregates estimates of opponents’ individual utilities. To support genuine value alignment, we introduce a lightweight feedback language embedded in the Alternating Offers Protocol, enabling agents to refine one another’s value‐weight models during negotiation. Implemented using the NegMAS platform, our approach is validated through a simple bilateral use case of two students dividing a prize. Experiments under various configurations show how embedded social commitments influence negotiation dynamics, such as the final agreement and convergence speed measured by the number of rounds to agreement. Finally, we outline a roadmap for scaling to more complex multilateral scenarios.

13:30

Lunch

14:30

Ion Mikel Liberal

Proof complexity beyond Resolution

The Boolean satisfaction problem (SAT) consists in determining whether a propositional CNF formula can be satisfied within the framework of classical logic. Although it has purely logical motivations on its own, SAT is frequently used to model a wide range of problems in computer science and industry. In the optimization variant of SAT, known as MaxSAT, the goal is not only to determine satisfiability but also to maximize the number of satisfied clauses. In this context, we propose CSat and CSimple, two novel SAT-based algorithms and the Comparator Calculus, a formal proof system that leverages a propositional proof system as an oracle and is designed to model SAT-based reasoning in MaxSAT solvers. Additionaly, cardinality constraints are one of the most widely used high-level constraints in SAT-based reasoning and, recently, it has been observed experimentally that the input ordering of cardinality encodings can significantly affect the performance of SAT solvers, sometimes even more than the choice of the encoding itself. We also study this phenomenon from a proof-complexity perspective, showing that encoding the contradictory pair of constraints $x_1+\cdots+x_n \geq k+1$ and  $x_1+\cdots+x_n \leq k$, with totalizers and different input orderings, gives rise to exponentially hard formulas for Resolution.

15:00

Ronghao Zhang 

Few-Shot Learning on Malware Detection and classification

Malware detection and classification remain challenging due to the rapid evolution of malicious software and the limited availability of labeled samples for emerging malware families. Few-shot learning (FSL) is therefore a promising direction, as it aims to recognize novel classes with limited supervision. However, existing studies often rely on image-derived representations, which may discard important structural information from executable files and have shown limited effectiveness for malware detection. This project investigates few-shot learning for malware analysis through a more systematic and realistic evaluation framework. The  work includes a literature review of recent methods, representations, datasets, and evaluation protocols, as well as the development of a benchmark for metric-based few-shot malware learning. The benchmark adapts representative metric based approaches to images, raw bytes, and feature based inputs, and evaluates how existing methods behave when the number of malware families increases, when unknown families appear at inference time, and when compared with standard supervised classifiers. Building on this benchmark, our next work will develop malware specific FSL methods by drawing on advanced structures from computer version while redesigning them for malware analysis. The focus also will be on methods that remain effective when the candidate family space becomes large and that support unseen family rejection under open-set scenarios.

15:30

Luca Pallonetto

Social Embeddings

This project targets two main research objectives: (1) how to build compact, yet semantics-preserving, embeddings to represent arbitrary social environments; how to fully characterize these embeddings, including their latent semantics; (2) precisely frame the socio-cognitive skill of social awareness enabled by social embeddings, and demonstrate it on social robots. The DC will particularly focus on social dynamics, by characterizing the trajectories of on-going social situations in the embedding space; discontinuities in the embedding space, that might represent unexpected changes of social dynamics; and social situation predictions, by extrapolating trajectories in the embedding space.

16:00

Ramon Vallés

Design and Implementation of Multi-Objective Scheduling Algorithms for the Cherenkov Telescope Array Observatory (CTAO)

The proposed research plan focuses on leveraging different optimization approaches in order to design a multi-objective scheduling algorithm for the Cherenkov Telescope Array Observatory (CTAO). The study aims to address the unique challenges of scheduling astronomical observations across multiple sub-arrays, considering objectives such as observation priority, resource utilization, and time constraints. This research is expected to contribute to the operational efficiency of CTAO, balancing these objectives efficiently and enabling optimal scientific outcomes.

9:30

Francesco Manfucci

An algebraic approach to conditional logics 

The research project investigates conditional logics through the lens of algebraic logic. From a general perspective, conditional statements are sentences of the form “if A then B”, and whose formalization goes beyond the usual material implication. Conditional logics are those formal systems that are designed to reason about conditional statements. Given their capability of expressing hypothetical situations, conditional formulas are pivotal in representing knowledge and reasoning abilities of rational agents. Conditional reasoning features in a wide range of areas spanning non- monotonic reasoning, causal inference, learning, and more generally reasoning under uncertainty. For this reason, conditional logics are at the cornerstone of many disciplines and their development is key to their advancement. 

Traditionally, conditional logics have been studied along two main methodological paths: semantic approaches based on possible worlds models, and syntactic (i.e. proof-theoretical) approaches. This project proposes a unifying algebraic framework that integrates these perspectives. Specifically, in this project we aim to: (1) develop an algebraic theory of conditional logics that bridges existing proof-theoretical methods with novel algebraic techniques; (2) connect possible worlds semantics with universal algebra through duality theory, extending methods successfully applied in modal logic to the conditional setting; (3) exploit these connections to provide a uniform framework for studying the foundations of probability theory within conditional logics. By bringing together semantic, syntactic, and algebraic methods, this research seeks to establish a cohesive foundation for conditional reasoning and to open new avenues for its application. "

10:00

Rocco Ballester

Harnessing Quantum Computing for Advancements in Federated Learning and Optimization

The rapid advancements in quantum computing are opening up new possibilities in areas like machine learning and optimization. Quantum computers, with their unique capabilities, can potentially revolutionize how we solve intricate computational problems. This could lead to significant improvements in the efficiency and effectiveness of machine learning algorithms and optimization techniques, enabling breakthroughs in various scientific and industrial applications. This industrial PhD program aims to investigate the integration of quantum computing techniques into federated learning, with a particular emphasis on discovering innovative methods to optimize federated learning processes. Furthermore, the research will explore the use of quantum annealers for optimization tasks, providing novel solutions to difficult optimization problems.

10:30

Break

11:00

Lluis Subirana

When does a belief become knowledge?

The objective of this thesis is to advance in the understanding of the formal distinction among the fundamental notions of beliefs and knowledge, being the former intuitively stronger than the latter. Subsequently our aim is also to understand up to which extent, and according to which rules, a belief can be converted in a knowledge by gaining what we can call an ‘’epistemic value’’. Methodologically speaking we will base our investigations on two pillars: the Lockean thesis (Foley) from one side, and a Bayesian/probabilistic version of revisions (Alchourron, Gärdenfors and Makinson) and updates (Katsuno Mendelzon) for the other. Intuitively, while Lockean thesis defines an agent beliefs in terms of subjective probabilistic values, belief change theories describes how a belief can be converted in knowledge by subsequent steps of revisions and updates.

11:30

Daniel Pardo 

AI-Powered cheese manufacturing: a digital future for dairy

Milk’s high nutritional value and versatility make it a key raw material for products such as yogurt, cheese, butter, milk powder, and cream, all of which extend shelf life. Its composition can vary owing to factors such as the animals’ diet, stage of lactation, genetic background, and health and welfare. This variability determines how the milk will be used and the performance and quality of the products obtained. Cheese, the most important derivative for direct consumption and as an ingredient, relies on three critical stages: coagulation, syneresis, and ripening . Coagulation sets the course for the rest of the process: the exact moment the curd is cut determines whey syneresis, solids retention, and ultimately final texture and yield. The objective of this thesis is to develop an artificial intelligence-based predictive model for the key variables in cheese production, improving consistency and industrial efficiency.

12:00

Arnau Mayoral Macau

Automated design of ethical environments using multi-agent reinforcement learning

Reinforcement learning (RL) algorithms have seen many applications, from game playing to autonomous driving and conversational agents, demonstrating a remarkable ability to learn complex tasks. However, while these algorithms may excel in learning to optimize a given reward function, ensuring that they behave according to human moral values while doing so is a significant challenge. On my thesis, we are working on an algorithm to automatically design ethical RL environments with large state spaces in such a way that the optimal behavior for the agents to learn is value aligned. We do so by adding an extra objective, called ethical objective, with its own system of rewards that incentives ethical actions and punishes unethical behavior. Then, our algorithm computes a so called ethicalweight which will be used to linearly combine the individual task of the environment with the ethical objective. With this linear combination we transform a multi-objective problem into a single-objective problem, where agents learn with just one reward function. Our contribution is how to compute the minimal weight needed to make the ethical objective prevail over the individual objective, so when agents learn using the combined reward function, all agents learn to do their task while abiding by the moral values encoded in the ethical objective.

12:30

Salvador Pullido

Design and Evaluation of a Reinforced Learning (AR) model to improve Deliberative Quality Online

This project proposes a specific approach using neural network architectures applied to natural language processing (NLP) such as Transformers and reinforcement learning (AR) algorithms to improve deliberative quality in online discussions. The inclusion of the Deliberative Quality Index (DQI) and the design of a personalized objective function will allow obtaining a highly specific model adapted to the challenges of evaluating deliberative quality. This approach has the potential to significantly advance the understanding and application of reinforcement learning techniques in the context of digital democracy and online citizen participation. With this research work, we seek to delve deeper into the field of natural language processing (NLP), focusing on improving the deliberative quality in online discussions. A variety of NLP neural network architectures will be explored, with the aim of developing a reinforcement learning model tailored for this purpose. The aim is to define an objective and loss function that guide the model towards specific improvements in the dimensions of the Deliberative Quality Index (DQI). In addition, an experimental validation of the model will be carried out using labeled or unlabeled data sets, demonstrating its ability to adapt and improve quality progressively over time.

13:00

Lunch

14:00

Valeria Giustarini

Algebraic Constructions in Nonclassical Logics

This thesis focuses on the mathematical study of nonclassical logics, which are meant to represent and study different kinds of reasoning. For example, they allow to deal with incomplete, partial, or inconsistent information. In particular, we are interested in the framework of substructural logics, which are a broad class of nonclassical logics that includes many of the most well-known systems; e.g., many-valued logics (which allow partial truth, or degrees of truth), intuitionistic logic (interpreting mathematical constructivism), some paraconsistent logics (where the logical system can handle contradictory information), and relevance logics (which emphasize meaningful connections between premises and conclusions in deductions). All these logics can be represented by means of algebraic structures called residuated lattices, which provide a common, unifying semantical framework. Semantical methods in general, and algebraic ones in particular, have proven effective in the study of logic, as they allow one to study syntactical properties from the algebraic perspective. These methods have indeed played a central role in the development of substructural logics throughout the last century. Residuated lattices have been widely studied, and are known to have a rich and deep theory. However, large classes of these structures still lack an effective algebraic description, hindering the understanding of the corresponding logical systems. To help bridge this gap, the first part of this thesis is dedicated to the development of novel algebraic constructions for the description of those algebras that lack an effective characterization. A key method is to develop constructions that obtain new structures from known ones. To this end, we will draw on techniques belonging to the mathematical domains of universal algebra and topology, which have proven extremely fruitful in the study of residuated lattices and substructural logics. The second part of the thesis focuses on applying these constructions to the study of syntactical properties of substructural logics, which can be relevant in applications to model checking or automated proof systems (e.g., the interpolation property, or unification problems). This line of work will allow us to push the frontiers of our structural understanding of residuated lattices and deepen our semantic insight into substructural logics.

14:30

Pol Rodríguez Farrés

Artificial Intelligence Approaches for the Detection and Mitigation of Dis/Misinformation in Social Media

Disinformation and misinformation spread through social media via coordinated behavior (bots, organized networks, AI agents, etc.) and independent dissemination by both intentional disinformers and unwitting misinformed users. This thesis develops computational models and detection algorithms to identify these diverse pathways and design mitigation strategies. The research pursues complementary approaches: empirical detection methods (e.g. leveraging network patterns via co-retweet synchronization), possibly validated on labeled real-world data; theoretical modeling of social media dynamics (e.g. non-Bayesian social learning) to overcome dataset scarcity; and integration of the two previous areas (e.g. multi-agent reinforcement learning to simulate strategic coordination on a network following theoretical models). These methods combine network analysis, machine learning, and social and economic theory to  understand actor behavior, and develop detection and mitigation strategies on social media platforms.

9:00

David Gomez Guillen

Human-Like Approaches to Calibrating Simulation Models for Cost-Effectiveness Analysis via Bayesian Optimization and AI Agents

The use of mathematical simulation models of diseases in economic evaluation is an essential tool in medicine aimed at guiding healthcare decision-making. Cost-effectiveness analyses assess the balance between health benefits and the economic sustainability of different interventions. A critical aspect of these models is the accurate representation of a disease's natural history, which requires calibrating parameters such as transition probabilities and disease burden rates to fit expected outcomes. However, this process is computationally expensive, and traditional optimization methods are often time-consuming, lacking the contextual awareness needed to navigate complex health economic interactions and constraints. This thesis explores a human-aligned approach to model calibration, using Bayesian optimization and AI agent frameworks as methodologies for replicating a human expert natural reasoning. These two frameworks navigate the calibration space in fundamentally different ways. Bayesian optimization tackles high-dimensional constraints by encoding domain knowledge and using sequential block decomposition and data-driven embeddings. Meanwhile, AI agents approach the problem by replicating the heuristic logic of manual calibration, using semantic information that is usually unavailable to standard optimization algorithms. This ongoing research aims to demonstrate that embedding human domain logic into algorithmic search has the potential to accelerate the calibration process while fostering the generation of more transparent, clinically plausible, and robust economic models.