Diffusion LLMs are Natural Adversaries for any LLM
Conference on Neural Information Processing Systems, 2026
*Equal contribution
Hi, I'm Paul
I'm a PhD student in AI4Science at the Technical University of Munich, supervised by Prof. Nils Thuerey in the Physics-based Simulation group.
I work on foundation models for PDEs. The idea is to pretrain one neural surrogate on simulations from many different physical systems, so it can be adapted to a new problem from a handful of examples instead of being trained from scratch. I'm also interested in uncertainty quantification and reliable machine learning more broadly.
Before the PhD I spent nine months as an intern at Apple in the Wireless Analytics and Intelligence team. There I built agentic workflows and predictive ML models for cellular and connectivity data. Before that I worked on adversarial attacks on LLMs, which became a NeurIPS 2026 paper, and was a research assistant at LMU Munich for two years.
News
- Started my PhD with Nils Thuerey at TUM, working on foundation models for PDEs.
- Diffusion LLMs are Natural Adversaries for any LLM was accepted at NeurIPS 2026. It started as my master's thesis.
- Finished my M.Sc. in Computer Science at TUM with high distinction.
- Joined Apple's Wireless Analytics and Intelligence team in Munich as an intern.
- Our paper on automating verbal-report data collection with ML is out in Behavior Research Methods.
- Presented Robust 3D Gaussian Splatting at GCPR 2024 in Munich.
- Presenting “Using LLMs to Automate the Analysis of Verbal Reports” at MathPsych 2024 in Tilburg, Netherlands.
- Gave a talk at TeaP 2024 in Regensburg on using LLMs to automate the analysis of verbal reports.
Publications
Using machine learning to automate the collection, transcription, and analysis of verbal-report data
Behavior Research Methods, 57(10), 2025
Robust 3D Gaussian Splatting for Novel View Synthesis in Presence of Distractors
German Conference on Pattern Recognition, 2024
Experience
-
PhD student, AI4Science Technical University of Munich, Physics-based Simulation group of Prof. Nils Thuerey
-
Wireless Data Science Intern Apple, Munich. Agentic workflows and predictive ML models for cellular performance analytics in the Wireless Analytics and Intelligence team
-
Research Assistant LMU Munich, Computational Psychology chair of Prof. Chris Donkin. ML pipeline for recording and analyzing verbal reports
-
Web Developer, working student Wirth & Horn Informationssysteme, Munich
Education
-
PhD, AI4Science Technical University of Munich, supervised by Prof. Nils Thuerey. Foundation models for PDEs
-
M.Sc. Computer Science Technical University of Munich. Grade 1.1, top 5% of class. Thesis: Guiding Diffusion Models to Attack Large Language Models
-
B.Sc. Computer Science Technical University of Munich. Minor in mathematics, tandem exchange with Tsinghua Shenzhen International Graduate School. Thesis: Gaussian Processes for Light Microscopy Image Segmentation