AI Reasoning in Theoretical Physics with TPBench and MadEvolve
Moritz Münchmeyer (University of Wisconsin-Madison)
Abstract: Large-language models are now powerful enough to assist physicists with mathematical reasoning at 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. In the second part of my talk I will present MadEvolve, our new LLM-based system to iteratively improve scientific algorithms. I will show that MadEvolve can set state-of-the-art results for algorithms in computational cosmology, such as the reconstruction of initial conditions of the universe. Finally, I will briefly present preliminary work on reinforcement learning fine-tuning of reasoning for theoretical physics.
HEP - phenomenologyHEP - theorymathematical physics
Audience: researchers in the topic
Series comments: Description: Weekly research seminar of the NHETC at Rutgers University
Livestream link is available on the webpage.
| Organizers: | Christina Pettola*, Vivek Saxena, Nicolas Fernandez Gonzalez, Erica DiPaola* |
| *contact for this listing |