Working with LLMs to do high quality math

[mklabgr]

In this talk, Daniel Litt (University of Toronto) discusses how large language models can contribute to high-quality mathematical work when used collaboratively rather than autonomously. While frontier AI systems have already shown that they can solve or make progress on difficult open mathematical questions, Litt argues that the more important issue for working mathematicians is how these models can improve the actual process of doing mathematics—helping explore ideas, test conjectures, search for arguments, identify gaps, and refine proofs while a human mathematician remains in control of judgment and verification.
Drawing on experiments he conducted over the previous year, the talk presents LLMs less as automatic “paper-producing machines” and more as potentially powerful mathematical collaborators, while also considering how this human–AI style of research might develop in the future. 

LECTURE