Full talk · edited transcript

Analyzing Public Pedagogical Benchmarks for AI Tutors and Building Synthetic Students for Testing

Airat Azbukhanov23 May 2026

Airat Azbukhanov, head of the ML group, Yandex Education.

The speaker's team works on educational language models: math tutors for grades 5–11 and the "tutor" mode in Alice. The talk has three parts: why separate pedagogical benchmarks are needed and how they differ from ordinary ones; how the team defines the educational value of a tutor's answer; how the environment with "hardware-like students" (iron students, a Russian idiom for a scripted, controllable simulated student) is built for evaluating and training models.

Continue reading — it's free

Google or a one-time email code. No password.