David Berghaus

I am a research scientist at Fraunhofer IAIS.

CV / Email / Google Scholar / DBLP / LinkedIn / Github

My research is focused on leveraging deep learning techniques to infer hidden dynamics in complex systems. Additionally, I am interested in the intersection of machine learning and maths. I also built LLM-based solutions for various industrial applications.

This paper introduces a foundation model to infer the vector fields of ordinary differential equations from noisy trajectory data in a zero-shot setting. These can be used to model the dynamics of various scientific systems in physics, biology, or engineering.

This paper introduces a foundation model to predict the intensity function of temporal point processes in a zero-shot setting. These can be used to predict the dynamics of various phenomena, such as earthquakes, social interactions, or chemical reactions.

This paper studies whether reasoning in continuous latent space improves multilingual capabilities. Using the CODI framework, we find that continuous chain-of-thought significantly outperforms explicit reasoning on low-resource languages in zero-shot settings, while compressing reasoning traces by roughly 29× to 50×.

This paper introduces a foundation model to predict the dynamics of stochastic differential equations in a zero-shot setting. These can be used to model various phenomena, such as financial markets, physical systems, or biological processes.

This paper presents a novel permutation invariant architecture that combines ideas from GNNs with self-attention algorithms and is tailored for Ramsey graphs. We use RL to try to find new Ramsey graphs (with no success! :D).

This paper introduces a foundation model to infer the hidden dynamics of Markov Jump Processes in a zero-shot setting. It is useful for practioners in science because they do not have to train models for each new dataset.

This paper presents a fast numerical algorithm to compute modular forms on noncongruence subgroups.

In this paper we numerically compute the Laplace eigenvalues of various shapes with dihedral symmetry, to investigate their series expansions.

In this paper we prove that the Laplace eigenvalues of regular polygons admit an expansion that involves multiple zeta values.

This website is based on the template of Jon Barron's website. Used with permission.