[HEX-HET] AI Reasoning in Theoretical Physics - Insights from the TPBench Project
Date and Time
Location
Broida 3302
Moritz Münchmeyer [U. Wisconsin, Madison] will present on
"AI Reasoning in Theoretical Physics - Insights from the TPBench Project".
Abstract: In this talk, I will first present our dataset TPBench (arxiv:2502.15815, tpbench.org), which was constructed to benchmark and improve AI models specifically for theoretical physics. I will then show our work applying test-time scaling techniques to TPBench, including agentic symbolic verification to boost performance. Finally, I will show some early results on fine-tuning reasoning models using RL on a narrow theoretical physics domain. More generally, we will discuss how academic researchers can contribute to these developments without access to industrial-scale computers.
