available for research & ML roles
Daniel Franzen
>Deep Learning Researcher · PhD Candidate
Johannes Gutenberg University Mainz · SFB TRR 146
I build neural networks that respect symmetry, and teach language models to reason abstractly. Winner of the ARC Prize 2024, runner-up in 2025.
Achievements
ARC Prize 2024 Winner
1ST PLACE
Winning entry on the flagship abstraction & reasoning challenge (alongside Jan Disselhoff) — 53.5% accuracy with a fine-tuned 12B-parameter LLM using test-time training and a custom inference and scoring algorithm.
ARC Prize 2025 Runner-Up
2ND PLACE
Second place on the harder next-generation benchmark — using a 2D-aware masked-diffusion LLM with recursive self-refinement. Kaggle gold medal, as part of the ARChitects team (with Jan Disselhoff and David Hartmann).
Publications
ICML
2025
International Conference on Machine Learning · 2025
NeurIPS
2021
Neural Information Processing Systems · 2021
Frontiers
2021
Studying the Evolution of Neural Activation Patterns During Training of Feed-Forward ReLU Networks
↗
Frontiers in Artificial Intelligence · 2021
Phys.
Rev. App.
Rev. App.
2024
Physical Review Applied · 2024
Education
2019 — present
PhD in Computer Science
JGU Mainz · SFB TRR 146 · Equivariant Neural Networks
ongoing
2015 — 2018
M.Sc. Computer Science
JGU Mainz · Specialization: Algorithms, Physics
Thesis: 3D Neural Networks for Vertebra Detection
1.0 · top grade
2012 — 2015
B.Sc. Computer Science
JGU Mainz · Focus areas: HPC/CUDA, Graphics
Thesis: Optical Tracking at High Spatial and Temporal Resolution
Department award for outstanding thesis
1.2 · first-class
German grading scale · 1.0 is the best possible grade (range 1.0 – 5.0)