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: M.Sc. Thesis
What Does Forgetting Look Like? Where activations land after concept erasure
Concept erasure methods are usually judged by what they remove and how much else they damage, but not by where the edited activations end up. My thesis studies these post-erasure activations directly and compares them with the activations the model produces on natural inputs. A guiding question is whether making a concept unrecoverable (representational absence) also makes the model treat it like a concept it has never known (semantic absence).
Ilyass Ouardi • Preprint (draft), University of Milan •
Is propositional truth one linear direction in a model's residual stream, or a small concept cone? I optimize a single direction (TDO) and a $k$-dimensional cone (TCO) under two causal axioms and a KL-retention term, on six instruction-tuned models from three families. One direction is enough for the smaller models but fails for the larger ones; a 2-D cone closes the gap for every model, with basis vectors near-orthogonal to the Difference-in-Means probe yet each causally effective.
Ilyass Ouardi • Preprint (draft), University of Milan •
Three feature representations and three classifiers on PhysioNet/CinC 2016, with every performance claim paired with a diagnostic that could refute it. The best model reaches a modified accuracy of 0.866, but the recording site is recoverable from the same features and diagnostic performance drops to near chance under leave-one-site-out evaluation: the aggregate metric overstates what the models learned.
Fatmir Bylyshi*, Hicham El-Khazri*, Ilyass Ouardi* • Preprint (draft), SESAR Lab, University of Milan •
Does a model learn a regulation better from stories of it being followed than from the regulation itself? We fine-tune the same backbone on either corpus, holding everything else fixed, and introduce GenderEqGLUE, a five-task benchmark, plus an occlusion-based interpretability analysis. The two regimes teach complementary competences: the regulation text wins on aggregate, the legends win on selecting the compliant action. *Equal contribution.
A self-study curriculum ordered the way I would teach it: the conceptual vocabulary first, then reverse-engineering small models by hand, causal methods, and current open problems. Starred papers make a one-pass version.
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.
