Jan Disselhoff

About

I am a Deep Learning Researcher and PhD candidate at Johannes Gutenberg University Mainz, focusing on deep learning research, LLM fine-tuning, symbolic reasoning, and test-time training methods.

ARC Prize 2024 - First Place

As part of "The ARChitects" team with Daniel Franzen and David Hartmann, we achieved a remarkable 53.5% accuracy by creatively utilising large language models (LLMs) in new ways to win the ARC Prize 2024 competition. The work focused on large language model finetuning on conceptually complex tasks and represents strong progress in artificial general intelligence benchmarks.

Experience

• Deep Learning Researcher & PhD Candidate
  Johannes Gutenberg University Mainz, 2022–Present
• ARC-Prize 2024 Winner (The ARChitects)
  Abstraction and Reasoning Corpus Competition – Solved 53.5% of tasks in kaggle competition
• Tutor
  University of Mainz – Mathematics & Computer Science, 2015–2019
• Developer, Psychology Experiments
  oTree Framework, 2020–2021

Education

• PhD Candidate, Computer Science
  Johannes Gutenberg University Mainz, 2020–Present
• Master of Science, Computer Science
  Johannes Gutenberg University Mainz, 2020
• Bachelor of Science, Computer Science & Mathematics
  Johannes Gutenberg University Mainz

Selected Publications

• "Product of Experts with LLMs: Boosting Performance on ARC Is a Matter of Perspective"
  Daniel Franzen, Jan Disselhoff, David Hartmann. ICML 2025
  📄 arXiv 💻 GitHub Page
• "The LLM ARChitect: Solving ARC-AGI Is A Matter of Perspective"
  Daniel Franzen, Jan Disselhoff, David Hartmann. Technical Report for ARC Prize 2024
  📄 PDF
• "Nonlinear Advantage: Trained Networks Might Not Be As Complex as You Think"
  Christian H. X. Ali Mehmeti-Göpel, Jan Disselhoff – CVPR 2023
  📄 Paper
• "Initialization Noise in Image Gradients and Saliency Maps"
  Ann-Christin Woerl, Jan Disselhoff, Michael Wand – CVPR 2023
  📄 Paper