M.Sc. Computer Science Student at University of Milan. Researching mechanistic interpretability, representation geometry, and verifiable AI safety.
Research & Projects
Prospective Collaborations & Inquiries
I am actively seeking PhD positions, pre-doctoral fellowships, and lab collaborations starting in 2026/2027. If your research group works on mechanistic interpretability, representation geometry, or verifiable AI safety, I welcome the opportunity to discuss shared research directions.
I am interested in AI safety and mechanistic interpretability. To me, interpretability is not an academic exercise for its own sake, but a pragmatic tool to make model safety measurable and verifiable on the path toward safe AGI. Right now I am particularly focused on the geometry of representations: how abstract concepts — like refusal and truth — are encoded within a model's activation space. Two questions follow from this:
Model biology: whether such concepts conform strictly to single linear directions, or rely instead on higher-dimensional, possibly non-linear manifolds.
Applied interpretability: whether representation-level edits — steering, abliteration, concept erasure — are genuinely surgical when evaluated against practical baselines, with off-target side-effects measured as rigorously as the primary behavioral change.
The two are connected, since an intervention built on the wrong picture of the geometry is precisely one whose side-effects go unmeasured.
Ongoing Research & Thesis
The Geometry of Absence, Opposite, and the Activation Manifold in Representation-Level Interventions
Ilyass Ouardi, Elisabetta Rocchetti, Alfio Ferrara • M.Sc. Thesis / ISLab, University of Milan •
Investigating how language models encode the absence of a concept versus its semantic opposite on the internal activation manifold. We evaluate representation-level interventions—including linear concept erasure (INLP, LEACE) and directional activation steering—measuring off-target collateral damage and reasoning degradation alongside targeted behavioral modifications.
The Multidimensional Geometry of Truth in LLMs: A Concept-Cone Extension of the Linear Representation Hypothesis
Ilyass Ouardi • Research Report / ISLab •
Investigates whether propositional truth is mediated by a single linear direction or by a $k$-dimensional concept cone. Adapting Targeted Directional Optimization (TDO) and Targeted Cone Optimization (TCO) from Wollschlager et al. (2025) under causal necessity, sufficiency, and KL-retention losses, we demonstrate that in larger models (e.g. Llama-3, Gemma), a single linear probe fails causal interchange tests, whereas a low-dimensional concept cone ($k=2$) successfully mediates truth judgments while remaining near-orthogonal to naive Difference-in-Means directions.
What Does a Heart-Sound Classifier Actually Learn? A Diagnostic Audit on PhysioNet CinC 2016
Ilyass Ouardi • Research Report / University of Milan •
A diagnostic refutation of shortcut learning in cardiac murmur classification. Demonstrates that CNN phonocardiogram models trained on PhysioNet CinC 2016 exploit recording-site background acoustic artifacts rather than true cardiac pathological signals, causing macro-accuracy to collapse from 0.866 to 0.457 under leave-site-out cross-validation. Proposes an explicit spectral diagnostic filter to catch spurious feature reliance prior to clinical deployment.
Shortcut LearningDiagnostic AuditAudio Pattern RecognitionHealthcare AI SafetyCNNs
The Legend Challenge: Embedding Ethical Compliance into LLMs
Ilyass Ouardi • Security of Service-Oriented Architectures •
Evaluates regulatory prompting versus narrative exemplars ("champion stories") for debiasing large language models and enforcing normative ethical constraints. Benchmarked systematically on GenderEqGLUE across varying prompt configurations and model scales.
AI AlignmentGenderEqGLUEEthical CompliancePrompt Engineering
Ilyass Ouardi, Carlo Maria Bellettini (Advisor) • B.Sc. Thesis (110/110 cum laude) & Production Deployment •
Complete digital donation and logistics management platform built for RECUP APS (Milan food rescue NGO). Deployed production backend with Supabase and integrated an experimental LangChain/LangGraph natural-language-to-SQL conversational agent allowing NGO staff to query logistics and inventory data using natural Italian queries.
