Niclas Dern
PhD student in Statistics at UC Berkeley

I am a first-year PhD student in Statistics at UC Berkeley. My research motivation is to make large-scale machine learning systems more robust, interpretable, and trustworthy, ideally in a principled manner.
I completed my undergraduate studies in mathematics at the Technical University of Munich. Most recently, I worked as a machine learning research intern at IMC Trading and as a software engineering intern at Tacto. Previously, I interned at the Vector Institute with Geoff Pleiss on theoretical properties of ensembles, and at the Helmholtz AI Institute with Niki Kilbertus and Elisabeth Ailer on estimating nonlinear causal effects in confounded environments. I also co-founded Mathis, an AI tutor for German high school students.
During my undergraduate degree, I was supported by the German Academic Scholarship Foundation and am an Atlas Fellow (2022) and winner of the National Student Championships in Computer Science (Germany, 2020). In my free time, I enjoy writing, reading, swimming, traveling, and meeting new people.
news
Sep 05, 2025 | New preprint online: Energy-Weighted Flow Matching: Unlocking Continuous Normalizing Flows for Efficient and Scalable Boltzmann Sampling. We introduce a novel training objective enabling continuous normalizing flows to model Boltzmann distributions using only energy function evaluations. |
---|---|
Aug 20, 2025 | Starting my PhD in Statistics at UC Berkeley! Very much looking forward to this next chapter and to working with the amazing people here. |
Jul 17, 2025 | 🎉 Our paper Theoretical Limitations of Ensembles in the Age of Overparameterization was accepted as an oral presentation at ICML 2025 (top 1% of submissions)! |
Dec 17, 2024 | Our paper Effects of Adaptive Feedback Generated by a Large Language Model: A Case Study in Teacher Education was published in Computers and Education: Artificial Intelligence. |
Oct 21, 2024 | New preprint online: Theoretical Limitations of Ensembles in the Age of Overparameterization. We show that ensembles of overparameterized models behave similarly to single, larger models. |
selected publications
- Energy-Weighted Flow Matching: Unlocking Continuous Normalizing Flows for Efficient and Scalable Boltzmann SamplingarXiv preprint arXiv:2509.03726, 2025
- Theoretical Limitations of Ensembles in the Age of OverparameterizationICML 2025, 2025Oral Presentation (top 1%)
- Sumformer: Universal approximation for efficient transformersIn Topological, Algebraic and Geometric Learning Workshops 2023, 2023