Teaching · Spring 2026

AI for Mathematics and Optimization

Bachelor and master seminar · Spring Semester 2026 · ETH Zürich

Course code252-5256-00L
TermSpring 2026
LevelB.Sc. & M.Sc. seminar
Projects9

Overview

This seminar course explores how AI can support research in mathematics and optimization. Students work in small groups on an open research problem, using modern AI tools as tutor, literature assistant, coding assistant, proof critic, formalization assistant and exploration partner.

Course information

Lecturers
Prof. Dr. Niao He, Dr. Zebang Shen
Format
Group projects on open problems in optimization, each supervised by a member of the group.
Participants
35 students, most of them bachelor (18) or master (17) students without formal optimization training. The seminar tested whether AI support lets such students make research-level progress on open problems within a single semester — something that would ordinarily be unusual.
Language
English

Target of the course

  • Engage meaningfully with an open research problem in optimization, with AI tools lowering the barrier to the required background.
  • Read and reconstruct the relevant literature, and judge which results bear on the problem at hand.
  • Use AI assistance critically: as a proof critic and formalization assistant rather than an oracle.
  • Report the outcome of the project in a form that another reader can follow and check.

Project list and outcomes (FS2026)

Cycling Behavior of Polyak's Heavy-Ball Method

Project outcome

Derandomization of the Arcsine Stepsize Schedule
Making the Gradient Small with Projected Stochastic Gradient Descent
Towards the Optimal Sample Complexity of SGD under Heavy-Tailed Noise
Nonsmooth DP-SCO: Optimal Rates in Linear Time

Project outcome

Nesterov Acceleration under a Local Polyak-Łojasiewicz Condition

Project outcome

A Principled Impossibility Result for Optimistic Gradient, and a Neuro-Symbolic Search for the Natural Proof
On Oracle Complexity of Convex Optimization with Linear Memory
Closed-Form and Convergence of Fitzpatrick Functions for Entropy Subdifferentials

Project outcome