Accepted Papers

In this era of Artifical Intelligence (AI), cyber threats are no longer static, as modern attackers are intelligent, adaptive, and capable of exploiting vulnerabilities faster than traditional defences can respond. My research investigates how attack and defence strategies co-evolve in complex, adversarial environments by integrating evolutionary game theory (EGT) with AI-driven multi-agent systems (MAS). This combined framework allows me to model strategic interactions between competing actors, analyse how defensive decisions influence overall security outcomes, and identify the conditions under which defences hold or break down. A particular focus of my work is the emergence of coordinated behaviours in multi-agent AI systems that can compromise system reliability in ways that existing safety approaches do not adequately address. Through theoretical modelling and controlled multi-agent experiments, my research proposes mechanisms that reduce attack success and strengthen defensive resilience. This work contributes to the understanding of adaptive cyber conflicts and the development of more robust and trust-worthy AI-enabled digital systems.

Knowledge graphs (KGs) promise machine-readable semantics together with human-readable labels and relations, yet domain experts struggle to use them due to technical barriers, especially in automatically constructed graphs with unlabeled nodes and complex connections. While LLMs enable KG interfaces such as retrieval-augmented generation, their potential to leverage ontology definitions to improve graph readability remains unexplored. This research investigates LLMs as ontology-aware verbalization interfaces that generate human-readable and usable textual information encoded in the KG structure.

This research studies interpretability methods for deep neural network--based radio receivers in future wireless communication systems. The work focuses on representation-level analysis and training data attribution techniques that reveal how models encode physical-layer channel conditions and how training data influence receiver behavior. These insights enable validation, adaptation, and more trustworthy deployment of data-driven communication systems under realistic channel conditions, while providing deeper understanding of model behavior.

LLMs increasingly serve as the control unit of AI agents. However, current systems struggle to support personalized interactions, as they lack effective mechanisms to remember and reuse user-specific information across sessions.
To address this, our research is centered around how memory augmentation and agentic decision-making can enable personalization in long-term interactions with LLM-based agents. As a first step, we proposed CAIM, a memory framework inspired by cognitive AI principles that aims to simulate human thought processes to create more adaptable systems. Building on this, we extended CAIM with an agentic workflow and user-specific memory modules to support personalized behavior beyond static retrieval-based approaches. We conducted a user study to evaluate perceived personalization, comparing our memory-based approach to a RAG baseline.
Our ongoing research focuses on how personalization can be measured without user studies, which are costly and hard to scale.
The long-term goal of our research is to establish foundations for designing and evaluating personalized LLM-based systems.

Despite advances in the computational analysis of multiple sclerosis, the ability to accurately predict disease progression remains limited due to unimodal approaches in the context of deep learning. KG-MultiMS proposes as a multimodal deep-learning framework that integrates data from three heterogeneous sources: (1) MRI volumes using the Swin-UNETR architecture to learn both semantic embeddings and segmentation masks of brain lesions; (2) tabular clinical variables using the FT-Tokenizer architecture to learn clinical embeddings; (3) medical knowledge graph based on the Harvard-Oxford atlas using a Heterogeneous Graph Transformer to learn contextual embeddings from patient-specific subgraphs. A cross-attention mechanism will allow for leveraging multimodal contributions by dynamically weighting them based on the anatomical-functional concordance represented in the contextual embeddings.

My research initiates the use of large language models (LLMs) to automate the design of effective algorithms for challenging problems in computational social choice, namely participatory budgeting (PB) and facility location mechanism design. To this end, I propose two LLM-based frameworks that respectively generate interpretable PB rules and mechanisms fulfilling competing objectives, i.e., optimality vs. fairness or strategyproofness. Future directions include extending these frameworks to achieve provable guarantees and addressing other relevant areas such as fair division.

Foundation models’ versatile generalization has driven their broad adoption, yet their deployment is constrained by critical gaps in trustworthiness. First, we conduct a survey to present a comprehensive analysis of what constitute trustworthy foundation model and examined approaches for each requirement from a lifecycle perspective. We identify out-of-distribution (OOD) robustness as the primary technical bottleneck, where performance collapse under distribution shifts invalidates other pillars like fairness and accountability. The nature of this vulnerability is two-fold: first, a reliance on spurious background cues during inference, and second, dimensional collapse in the representation space during fine-tuning. This research aims to mitigate these failures by developing mechanistic insights and algorithmic solutions during foundation model adaptation and deployment phases of the lifecycle.

Neuro-symbolic systems are a promising paradigm for Hybrid Intelligence (HI), combining the complementary strengths of symbolic reasoning and machine learning in human-AI interactions. However, designing architectures that satisfy desired properties currently relies on informal reasoning, rather than a rigorous framework. Our research addresses this gap by developing a formal intermediate layer that links architecture design to system properties. Building on the visual pattern language of the Boxology for neuro-symbolic systems, we develop a mathematical formalization that is grounded in Category Theory, in order to enable proof-based guarantees of properties. The soundness of this formalism will be demonstrated through empirical case studies of existing Hybrid Intelligence systems. This work advances the theoretical foundations of neuro-symbolic AI by supporting principled design of HI systems.

Multimodal deep learning for clinical decision support frequently fails in real-world deployments due to severe data missingness and sensor noise. Standard fusion and Mixture-of-Experts (MoE) architectures assume all modalities are uniformly informative, causing routing collapse and miscalibrated confidence when faced with corrupted inputs. My doctoral research addresses this fragility by shifting the paradigm from generative imputation to uncertainty-guided abstention. As a foundational step, I introduce QA-MoE (Quality-Aware MoE), which explicitly decouples reliability estimation from the routing process. By quantifying epistemic uncertainty, the model dynamically filters noisy modalities before fusion, ensuring sparse and stable inference. Building on this static robustness, my ongoing and future research extends to longitudinal and structured clinical data. I am developing PathFlow-EM, which integrates Expectation-Maximization to handle hidden variable distributions in unevenly sampled time-series, and Uncertainty-Guided Dynamic Topological Fusion (U-DTF), which uses topological data analysis to gate features based on evolving patient graphs. Ultimately, this research trajectory aims to deliver robust, calibration-aware architectures that form the computational basis for reliable Patient Digital Twins.

Equivariant graph neural networks (EGNNs) have revolutionized molecular property prediction by encoding 3D structures while preserving physical symmetry. However, existing EGNNs struggle to capture fine-grained geometric features needed for tasks like understanding protein interactions, and designing molecules that modulate them are central challenges in computational drug discovery.
This thesis develops a progressive framework based on equivariant graph neural networks (EqGNNs) that respects the geometric symmetries of molecular structures.
Our contributions progress along two axes: from discriminative to generative and from basic representations to physics-informed, language-conditioned generation. In this study, we propose novel methods for both discriminative (E(Q)AGNN-PPIS, GDEGAN) and generative tasks (PhysFlow). Together, these advance symmetry-aware geometric deep learning for drug discovery.

Deep learning models for clinical time-series forecasting face significant challenges in generalizability when deployed across different clinical settings. Various choices in the deep learning pipeline can affect a model's out-of-distribution performance. This abstract introduces a comprehensive benchmarking framework to assess domain generalization and adaptation under varying distribution shifts in multivariate time-series forecasting using real-world data. Preliminary empirical results on the influence of backbone model architecture and a proxy metric, the robustness slope, are presented alongside remaining challenges.

As AI systems are increasingly deployed in high-stakes settings, two concerns become particularly important: cautiousness (awareness of model limitations for risk-aware decisions) and explainability (providing human-understandable justifications). My PhD research develops methods that enable both, from theoretical and application perspectives. On the theoretical side, I formalize cautious and interpretable models for multiple uncertainty frameworks, with algorithms for explanation and preference elicitation under uncertainty. On the application side, I implement these methods in post-stroke rehabilitation, designing AI systems that analyze multimodal clinical data to support individualized rehabilitation planning.

Renewable energy is a key alternative to fossil fuels and their environmental harms. However, sub-Saharan African countries remain behind in their implementation, due to poor investment in the appropriate resource, among other causes. This information is dispersed as unstructured knowledge across scientific papers, making the task manually demanding for decision and policy makers. In this work, we combine the strength of Large Language Models (LLMs) and structured knowledge (Knowledge Graphs) in a neuro-symbolic framework to tackle this issue. We implement this approach in the hydropower field, building a question-answering system that uses Retrieval Augmented Generation (RAG) on a knowledge graph. Using LLMs such as LLaMA 3.1 8B, LLaMA 3.3 70B, LLaMA 4 Scout 17B, and Kimi K2, we observe an improvement in exact match of 48 − 60% across all models. These results support the ability of this framework to participate in evidence-based energy planning in developing countries.

RAG systems ground LLM responses in external evidence, yet their trustworthiness remains underspecified. Retrieval, re-ranking, and generation are optimized independently with no unified account of whether the system is interpretable, grounded, and traceable. My thesis operationalizes trustworthy RAG through three current sub-goals, each at the stage where it is most vulnerable. For traceability, REASONS benchmarks LLMs on 20K papers, showing attribution failures are systematic even under retrieval augmentation. For interpretability, IMRNNs decomposes dense embeddings via learned adapters, exposing human-readable query-document matching at retrieval. For grounding, METEORA replaces opaque re-ranking with rationale-driven selection, ensuring generation operates over explicitly justified context. I unify these under Interpretable Knowledge Flow (IKnowFlow), where trustworthiness is preserved across stage boundaries rather than checked at output. IKnowFlow will next address two longstanding blind spots: negation-awareness in retrieval, where surface similarity masks semantic inversion, and provenance-aware verification through structured source graphs.

Analogical reasoning, the ability to map different domains using their structural similarity, is a core mechanism of human cognition that supports tasks such as problem-solving, argumentation, and creativity. Despite recent advances in AI systems, including Large Language Models (LLMs), they still struggle to perform analogical reasoning reliably. In this project, we get inspiration by cognitive theories of analogy and try to develop models that can better capture and perform analogical reasoning. As a key component, we investigate abstraction levels and study how to best extract and represent them. In addition, we systematically analyze the limitations of LLMs by developing a dataset annotated with analogical structure and explore fine-tuning approaches to improve their analogical reasoning abilities.

Neural networks enable high-performance decision-making, but their deployment in safety-critical settings requires formal guarantees that may become invalid after changes in the system dynamics. This research studies how such guarantees can be recovered using information from the initial verification, enabling post-deployment analysis to assess preservation, quantify degradation, and restore compliant behavior without recomputing verification from scratch.

A vision-based robot can be faced with changing friction, gravity, or motor behavior, whose effects are observable only from pixels mid-episode. Visual model-based reinforcement learning struggles in such settings because most world models assume the latent factors driving these changes stay fixed during interaction, or fold them into a single context embedding. My thesis is developing a context-adaptive Mamba-based world model that infers K context vectors from short pixel-trajectory windows. Each vector steers a disjoint slice of the Mamba backbone through block-diagonal feature-wise linear modulation (FiLM), so distinct dynamics signals travel along parallel pathways rather than collapse into a single channel. The policy trains on imagined rollouts conditioned on the inferred context, and I show that its suboptimality is bounded by √K times the context-inference error along a Lipschitz chain through FiLM, the SSM, and the decoder. Evaluation spans within-episode shifts on DeepMind Control (Walker, Quadruped) and mode/difficulty switching on Atari, against an inferred-context baseline and a privileged-context oracle.