Florian Kark

Hey, feel free to connect if you’re interested in collaborative research! 👋

I’m a Masters student at the Heinrich-Heine-Universität Düsseldorf and the MPI of sustainable Materials.

My research interest lies in meta-learning: learning how to learn rather than optimizing individual prediction tasks. I am particularly interested in indirect forms of learning—unsupervised, self-supervised, weakly supervised, and agentic approaches—that infer underlying structure instead of relying solely on labeled data.

Coming from computer science, my fascination with indirect learning naturally led me to representation learning and generative models, particularly vision transformers and large language models, where the abundance of data often masks fundamental limitations. After moving into materials science, however, I was confronted with a very different reality: datasets are typically small, heterogeneous, noisy, and expensive to obtain. This motivated me to investigate the limits of conventional representation learning.

I explored ways to improve sample efficiency by incorporating stronger inductive biases through symmetry, invariance, equivariance, geometric priors and latent space regularizations. While these approaches certainly help, they ultimately felt like incremental improvements rather than a fundamental solution. This led me to question not only our model architectures, but also the learning paradigm itself. Foundation models excel at learning compressed statistical representations of data, yet they rarely challenge the assumptions underlying the data-generating process or construct explanatory models of the world. My interest has therefore shifted beyond representation learning toward systems that actively formulate hypotheses, validate assumptions, and iteratively refine their own understanding of an environment. In such systems, I view individual neural models and agents not as perfect learners, but as imperfect yet sufficient building blocks of a larger learning process.

My long-term vision is to develop AI systems that combine foundation models, symbolic reasoning, causal inference, and autonomous experimentation into a unified system capable of discovering compact first-principles explanations across multiple scales. Rather than learning correlations alone, such systems should continuously generate hypotheses, perform experiments, validate assumptions, and construct increasingly compressed, interpretable, and predictive models of reality.

I believe intelligence emerges from discovering efficient computational descriptions of the world. If artificial systems eventually become capable of recursively improving both their architecture and the computational substrates on which they run down to the bits and atoms, they may uncover principles that fundamentally reshape our understanding of computation, materials, and the physical universe.

Ultimately, I hope to contribute to a future in which scientific discovery becomes increasingly automated, accelerating research and enabling more sustainable technologies.

I am also a lecturer at the Heinrich-Heine-University holding a seminar on the foundations of LLMs, Agents, and their applications in research and industry twice each semester.

In my spare time, I enjoy doing sports (mostly weights and running), traveling and exploring new places and activities, and the best food spots in Düsseldorf.