Sonia Laguna
Sonia Laguna

Research @Apple MLR
PhD in Machine Learning @ETH Zurich

Previously @Cambridge @Google @Harvard

About Me

Hi there!👋 I am currently a research intern at Apple MLR working with Marco Cuturi on LLM research, and a PhD student in Machine Learning at ETH Zurich, supervised by Prof. Julia Vogt (ETH) and Prof. Bernhard Schölkopf (Max Planck Institute). Throughout my PhD, my research has focused on the intersection of machine unlearning, large language models, model adaptation, and interpretability. I am particularly interested in how models acquire, use, and forget information, with a current focus on long-context LLMs and model adaptation. Beyond research, I am actively involved in the ML community through talks, organizing, mentoring, and academic service.

During my PhD, I have been a visiting student at Cambridge University with Prof. Mihaela Van der Shaar, working on alignment and interpretability of LLMs. Additionally, I have been a Research Intern and a Student Researcher at Google, developing 3D diffusion-based generative models in the AR&VR team, and the team co-leader of CSNOW, Computer Science Network of Women at ETH.

Prior to my doctoral studies, I obtained a MSc in the Department of Information Technology and Electrical Engineering at ETH Zurich, and spent a semester at Harvard University working on 3D generative models for super-resolution of MR images. I was lucky to be supported by two Spanish Excellence Fellowships, La Caixa and Rafel del Pino. Before that, I completed my BSc in Biomedical Engineering at Universidad Carlos III de Madrid, spent one year at Georgia Institute of Technology, and carried out an internship at ETH Zurich as an Amgen Scholar.

I am always happy to collaborate and discuss new topics, feel free to reach out! 😃💡

Interests
  • Large Language Models
  • Model Adaptation
  • Machine Unlearning
🗞️ Recent News

See all News »

(2026). DynaMiCS: Fine-tuning LLMs with Performance Constraints using Dynamic Mixtures. Preprint - Arxiv.
(2026). Rethinking Machine Unlearning: Models Designed to Forget via Key Deletion. In ICLR 2026 Workshop TTU (Oral) and extended Preprint - ArXiv.
(2026). Reference-Guided Machine Unlearning. In ICLR 2026 Workshop AIWILD and extended Preprint - ArXiv.
(2024). Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable. In NeurIPS 2024.
(2024). Stochastic Concept Bottleneck Models. In NeurIPS 2024.