Seminar, Hosted by Harvard Center of Mathematical Sciences and Applications — New Technologies in Mathematics Seminar
Who Teaches the Teacher? Self-Improving Reasoning Systems
NYU
Abstract
AI reasoning has progressed remarkably fast, from competition-level problems to original mathematics, including contributions to long-standing open problems. Mathematics has a special advantage because its results can be verified, which makes self-improvement increasingly feasible: models can explore ideas, learn from failures and design curricula for other instances of themselves, building intermediate steps towards problems they cannot yet solve. Mathematics may nonetheless be the easy case, since most real-world settings lack automatic verifiers, tests or clear reward signals, leaving open what reasoning and self-improvement should mean without an external source of truth. Julia Kempe presents recent work on exploration, teacher-student training and efficient reasoning in mathematics, then turns to the broader goal of systems that learn world models, plan with them and improve through interaction with the world.