Learning Abstractions for Mathematics

Research Project, 2026 – 2030

The goal of this project is to develop new methods for automating search of new mathematical discovery and proof. In particular, we propose to use hierarchical reinforcement learning and intrinsic motivation to extend current generation of AI systems to higher level reasoning  by introducing useful abstractions, similar to the way human mathematicians do.While there have been dramatic adcances in AI systems for mathematical discoevery and proof (for example the recent demonstration by Google Deepmind´s systems to achieve near gold medal level perfoamce in the international mathematical olympiads), such systems are still very limited  - they rely more on compute and data to scale their performance. Our proposal is to investigate system that reason in the way human mathematicians do, namely by using higher level abstractions and relying on intrinsic motivation to learn a subject rather than being driven only by a fnal goal.This project will use methods from hierarchical reinforcement learning and intrinsic motivations coupling them with modern proof assiistants and theorem provers.