Markus Herre

PhD Researcher · Machine Learning

Technische Universität Clausthal

About

Machine learning for structured data — with a focus on tabular domains.

I'm a PhD researcher at Technische Universität Clausthal, working on machine learning for structured data. My research spans representation learning, model architectures, and learning dynamics for non-sequential data.

Right now I'm particularly interested in tabular foundation models, hypernetworks, and differentiable decision tree ensembles — and in the methodological question of when these new families actually beat well-tuned gradient-boosted trees.

Publications Explaining Tabular Foundation Model Differences Through Meta-Features Accepted · ICML Workshop 2026 Currently PhD Candidate, ML Group · TU Clausthal Goslar, Germany · From March 2026

Interests

Research

  • Tabular Foundation Models
  • Meta-features & Routing
  • Hypernetworks
  • Differentiable Trees
  • Representation Learning
  • Model Architectures
  • Benchmarking
  • Reproducibility

Selected work

Latest paper

2026 ICML Workshop First-author

Explaining Tabular Foundation Model Differences Through Meta-Features

Markus Herre, Andrej Tschalzev, Sascha Marton, Christian Bartelt

With tabular foundation models now competing with well-tuned tree ensembles, it's often unclear which family will win on a given dataset. We test whether model-agnostic dataset meta-features can explain these performance gaps across the 51 TabArena datasets — and find that, under strict statistical control, global meta-feature approaches are not robust enough to offer reliable explanations.

Figure 1: Two panels comparing TabICLv2 and TabPFN-2.6. Left, ROC AUC vs. total feature count d: both models track together until d grows large, where TabICLv2 drops sharply while TabPFN-2.6 stays flat. Right, the TabICLv2-minus-TabPFN AUC gap as a function of redundant nuisance fraction: many independent columns favor TabPFN, but converting them into redundant copies closes the gap toward zero.
Details PDF arXiv OpenReview

Updates

News

  • Explaining Tabular Foundation Model Differences Through Meta-Features was accepted at the 2nd ICML Workshop on Foundation Models for Structured Data.

  • Started as ML researcher and PhD candidate at Technische Universität Clausthal.

  • Completed the M.Sc. in Business Informatics at the University of Mannheim.

  • Began exchange semester at the University of Padua as part of the M.Sc. program.