Mathematics in the Age of AI

---

Open LLMs and the Data Journey of K2 Horizon

Speaker: Mikhail Yurochkin, IFM MBZUAI Silicon Valley Lab

Location: Warren Weaver Hall 109

Date: Wednesday, October 21, 2026, 3:30 p.m.

Synopsis:

Open language models are becoming increasingly capable, especially in coding, yet substantial gaps remain in areas such as formal mathematics. Why are some capabilities easier to reproduce in the open than others?

In this talk, I will use the development of K2 Horizon, a family of fully open LLMs released by our lab in September 2026, to explore one possible answer: data. Much of our work is devoted to transforming "raw" sources of human knowledge, from web pages and GitHub repositories to scientific literature, into diverse, information-rich training examples. I will discuss a variety of techniques we developed for data synthesis and enrichment, and why code repositories provide particularly fertile ground for teaching models to solve programming tasks.

In mathematics, arithmetic, computational problems with numerical answers, and olympiad-level problems all offer ample opportunities for data synthesis, while generating effective training data for formal and research-level mathematics remains challenging without domain expertise.

I will conclude with some thoughts on Open Data as a path forward to transform the landscape of AI.