Events

Date: 5 December 2025   Time: 14:30 - 15:30

Title: AI Reasoning in Theoretical Physics - Insights from the TPBench Project

Abstract: 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. We will then discuss how test-time scaling techniques can be used to improve performance, including agentic symbolic verification to boost performance (arxiv:2506.20729). I will then show preliminary results of two new projects. In the first of those, we use GRPO reinforcement learning to fine-tune models on QFT problems. In the second, we apply LLM code evolution (similar to AlphaEvolve) to several algorithmic problems in cosmology.