Short bio
I am an M.Sc. student in Computer Science at the University of Milan, where I am writing my thesis on concept erasure in large language models with Elisabetta Rocchetti and Prof. Alfio Ferrara at ISLab. I received my B.Sc. in Computer Science from the same university in 2024, graduating 110/110 cum laude, and I am currently a teaching assistant for the first-year Python course of the Bachelor’s in Artificial Intelligence.
My research is in mechanistic interpretability and AI safety. I study the geometry of representations: how concepts such as truth and refusal are encoded in a model’s activation space, and whether interventions built on that picture, like steering and concept erasure, change only what they are meant to change.
News
- I will be at the Swiss AI Safety Days in Zurich (7–8 November). Happy to meet if you are there.
- Started as teaching assistant for Computer Programming (Python), the first-year course of the inter-university Bachelor's in Artificial Intelligence (Pavia, Milan, Milano-Bicocca), with Prof. Stefano Ferrari.
- Joined a team working on the Steerability Challenge.
- Started my M.Sc. thesis on concept erasure in LLMs with Elisabetta Rocchetti and Prof. Alfio Ferrara at ISLab.
- Joined AI Safety for Italy (AIS4I), the Italian AI safety community, as a member and organizer.
- Finished drafts of my manuscripts on the geometry of truth in LLMs, on what heart-sound classifiers learn and on the Legend Challenge.
- Started as teaching assistant for Accertamento delle Competenze Informatiche at the University of Milan.
- Joined Lead the Future, a STEM mentorship community, as a mentee.
Looking for a PhD
I am looking for PhD positions, pre-doctoral fellowships and lab collaborations starting in 2026/2027. If your group works on mechanistic interpretability, representation geometry or AI safety, I would be glad to talk.
Research interests
I see interpretability as a way to make model safety measurable, so most of my questions are about whether our picture of a model’s internals is the right one.
- Model biology. Do concepts live on single linear directions, or on higher-dimensional, possibly non-linear structures? My work on truth suggests that, in larger models, a small multi-dimensional cone describes the concept better than one direction.
- Applied interpretability. Are representation-level edits (steering, abliteration, concept erasure) really surgical? I try to measure their side effects as carefully as the behaviour they are meant to change.
The two are connected: an intervention built on the wrong picture of the geometry is exactly the one whose side effects go unmeasured.
Manuscripts
- The Multidimensional Geometry of Truth in Large Language Models: A Concept-Cone Extension of the Linear Representation Hypothesis
Ilyass Ouardi. Preprint (draft), 2026. - What Does a Heart Sound Classifier Actually Learn? A Controlled Comparison of Cepstral, Spectral and Wavelet-Packet Representations
Ilyass Ouardi. Preprint (draft), 2026. - The Legend Challenge: Embedding Ethical Compliance into LLMs through Champion Narratives
Fatmir Bylyshi*, Hicham El-Khazri*, Ilyass Ouardi*. Preprint (draft), 2026. *Equal contribution.
Ongoing: What Does Forgetting Look Like? Where activations land after concept erasure (M.Sc. thesis, with Elisabetta Rocchetti and Alfio Ferrara). More on the Research page.
Outside research
I have played football since I was a kid, from the streets up to semi-pro, and basketball is how I switch off after a long day of studying.