Xavier Theimer-Lienhard

I'm a research associate with the LiGHT Lab (Meditron) at EPFL, embedded in the Meditron program end-to-end across training, evaluation, and deployment. I started in 2023 in the sprints that produced early Meditron finetunes. As I took on more of the stack, I was asked to lead sub-projects in Meditron-Reasoning (RLHF + reasoning) and Meditron-Synthetic (safe clinical data generation).

My work focuses on making medical LLMs efficient, interpretable, and auditable: I co-designed evaluation pipelines, led RL and reasoning post-training on HPC, and built Meditree, a multi-agent inference method that mirrors clinical reasoning. Recently, I co-led Apertus-Meditron-FO, a fully open adaptation of the Swiss Apertus models for medicine, including an open medical corpus and fine-tuning of 8B/70B variants.

I hold a Master's degree in Data Science from EPFL (GPA 5.4/6) and a Bachelor's degree in Communication Systems. When I'm not iterating on training runs, I like to indulge in a little bit of running, climbing, random wikipedia articles reading sessions, café hopping (just this year i went from tea afficionado to chai lover to black coffee extremist). I have also been dubbed king of brunch by my friends.