Physics Colloquium – Moritz Münchmeyer

AI Reasoning in Theoretical Physics with the TPBench Project

Assistant Professor Moritz Münchmeyer
UW-Madison Department of Physics

Large-language models are becoming powerful enough to assist physicists with mathematical reasoning at the research level. In this talk, I will first present our dataset TPBench (tpbench.org), which was constructed to benchmark and improve AI models specifically for theoretical physics. I will then discuss how test-time scaling and symbolic verification can be used to improve their performance and reliability.

I will also show preliminary results from two new projects. In the first, we apply reinforcement learning to fine-tune models on QFT problems. In the second, we apply LLM-based code evolution (similar to Google’s AlphaEvolve) to algorithmic problems in cosmology.