Joint Theory Seminar: François Charton

Title: Language models in maths and theoretical physics

Abstract: Problem solving can be seen as a translation task, where a problem, a sequence of symbols, is translated into its solution, another sequence of symbols. Transformers, neural network architectures originally designed for machine translation, can be used to solve various problems of mathematics and physics. In particular, I will present recent results on predicting the properties of scattering amplitudes in planar N=4 supersymmetric Yang-Mills. Whereas large language models are often criticized for their tendency to hallucinate, and the lack of interpretability of their prediction, math transformer predictions can often be explained, and their errors attributed to mathematical reasons. 

Student session: 13:10