Jonathan Pirnay
Jonathan Pirnay

PhD @ TUM

Biography

I’m a machine learning researcher with a PhD from the Technical University of Munich, where I have been researching self-play and self-improvement methods for solving complex combinatorial optimization and reasoning problems (at the GrimmLab). I am particularly interested in efficient decoding methods for the underlying sequence models. My work has been published in top ML venues. Before starting my PhD, I worked as an AI consultant at Fujitsu. To pay my bills at uni, I wrote server-side code in Go, Swift, and Node.js for various clients.

I wrote my master’s thesis in Arithmetic Geometry on Tropical Cycle Classes for the Berkovich Analytification of Algebraic Varieties and my bachelor’s thesis on Tropical Hypersurfaces of Laurent Polynomials, both investigating the combinatorial shadow of algebraic varieties, the central object of study in Algebraic Geometry.

My brother Niklas is a quantum computing researcher.

Interests
  • Reinforcement Learning
  • Natural Language Generation
  • Reasoning
  • Combinatorial Optimization
Education
  • Dr. rer. nat. (summa cum laude)

    Technical University of Munich

  • M.Sc. Mathematics

    University of Regensburg

  • B.Sc. Mathematics

    University of Regensburg

  • MFA Visual Communication

    University of Art and Design Kassel

Publications
(2025). GraphXForm: Graph transformer for computer-aided molecular design. In Digital Discovery.
(2025). Deep reinforcement learning enables conceptual design of processes for separating azeotropic mixtures without prior knowledge. In Computers & Chemical Engineering.
(2024). Self-Improvement for Neural Combinatorial Optimization: Sample Without Replacement, but Improvement. In Transactions on Machine Learning Research (TMLR), 06/2024, Featured Certification.
(2023). Superior protein thermophilicity prediction with protein language model embeddings. In NAR Genomics and Bioinformatics.
(2023). Convex Envelope Method for determining liquid multi-phase equilibria in systems with arbitrary number of components. In Computers & Chemical Engineering.
(2023). Policy-Based Self-Competition for Planning Problems. ICLR 2023.
(2022). Inpainting Transformer for Anomaly Detection. ICIAP 2022.
(2021). Quantum algorithms for process parallel flexible job shop scheduling. In CIRP Journal of Manufacturing Science and Technology.
Recent & Upcoming Talks
More stuff

Some links to math stuff I wrote: