Research

My research interests are broadly on how machine-learning systems learn, adapt, and make decisions. My specific interests include continual learning, reinforcement learning, probabilistic methods/uncertainty quantification, and mechanistic interpretability.

Continual Learning and Optimization

My MSc thesis at the University of Groningen focuses on optimization-based continual learning and is supervised by Gido van de Ven and Matthia Sabatelli. I am designing a new curvature-aware optimization framework for continual learning in a variety of scenarios (classification, reinforcement learning, etc.). I am currently in the process of finishing this by the end of the year and hope to share more soon :)

Mechanistic Interpretability and Alignment

As a Research Fellow at Apart Research (July 2025–March 2026), I co-authored The Anatomy of Alignment, which uses interpretable sparse features to study preference optimization. We found that the learned policy favored style and formatting features over honesty-related ones, while generation coherence deteriorated. The paper received a Spotlight at the NeurIPS 2025 Mechanistic Interpretability Workshop.

Selected Papers

  1. NeurIPS MI
    The Anatomy of Alignment: Decomposing Preference Optimization by Steering Sparse Features
    Jeremias Ferrao, Matthijs van der Lende, Ilija Lichkovski, and Clement Neo
    In NeurIPS Mechanistic Interpretability Workshop, 2025
    Spotlight arXiv Bib HTML PDF Code

See all publications →

Reinforcement Learning and Probabilistic Methods

My bachelor's thesis was on interpretable function approximation with Gaussian processes in value-based model-free reinforcement learning. The work shows a practical way to use Gaussian processes in reinforcement learning. It was published at NLDL 2025.

Selected Papers

  1. NLDL
    Interpretable Function Approximation with Gaussian Processes in Value-Based Model-Free Reinforcement Learning
    Matthijs van der Lende, Matthia Sabatelli, and Juan Cardenas-Cartagena
    In Proceedings of the 6th Northern Lights Deep Learning Conference (NLDL), 2025
    Bib HTML PDF Code

See all publications →