I’m a second-year PhD student in Explainable AI at the University of Antwerp (Belgium), supervised by Prof. David Martens.
My current focus is on:
- Mechanistic Interpretability — understanding AI model behaviour by investigating internal components (individual neurons, attention heads, layers, activations) and reverse-engineering learned computational circuits to explain how and why models produce specific outputs.
- AI Alignment — aligning model objectives and behaviours with human values and safety constraints.
- Concept Discovery in Large AI Models — understanding the concepts learned by large AI downstream models, including concepts that may be beyond current human understanding (see e.g. Schut et al., 2025).
I care about building transparent, reliable ML systems and turning model internals into understandable, human-legible explanations.
- 📄 Publications: Google Scholar
- 🧑🏫 Supervisor: Prof. David Martens
- 🧑🔬 Former advisor: Prof. Przemyslaw Biecek
- 📍 Antwerp, Belgium
Always happy to connect on collaborations around interpretability, alignment, and concept discovery.
Doc: ➡️ Key Takeaways & Write-up
What this investigates:
Trade-off prompts like “Do you accept X euros to wait an additional Y minutes?” on Llama-3.2-3B-Instruct, analyzing token-level behavior for Yes/No. It is a mechanistic interpretability follow-up on Cash or Comfort? How LLMs Value Your Inconvenience (Cedro et al., 2025)
Core findings:
- A single MLP output layer (21) steers “No”: mean-ablation flips model decisions from No → Yes.
- Within that layer, neuron 8190 strongly pushes toward “No”; zero-ablating it flips the decision (acts as a counterfactual neuron).
- MLP layers dominate logit attributions for “Yes/No” over attention heads.
- Interference emerges a few layers later: probability shifts appear ~layers 24+ after the ablation.
- Behavior shows systematic anomalies: rejections at powers of 10 (€10, €100, €1,000), occasional greediness (accepting €0.10 for 600 minutes), and no freebie dilemma.
Why it matters:
Points to compact, targetable circuits for preference-like judgments; suggests concept-level control via sparse, layer-localised edits.
- Segment Anything Model (SAM) by Meta AI on Medical Images
- Computer Vision Classification Model and XAI Explanations with PyTorch
- Computer Vision Classification Model from Scratch with PyTorch
- Binary Neural Network Classification Model from Scratch with PyTorch
- Linear Regression Model from Scratch with PyTorch
