About

Research interests

A learning problem can have many solutions that fit its training objective equally well. A downstream goal gives us a reason to prefer one over another. I study how to make that choice explicit, and how the geometry of the solution set shapes the algorithms we can build.

This is central to my work on bilevel optimization, where an outer decision depends on the solution of an inner problem. In Select-then-differentiate, we study inner minimizers that form a manifold: under a local Polyak–Łojasiewicz condition, uniqueness of the best inner choice is enough for differentiability of the outer objective.

I’m interested in applications to large language models (LLMs) and reinforcement learning (RL). A practical variant of our selection method is evaluated for LLM training-data reweighting on reasoning and instruction-following tasks.

My broader interests include stochastic analysis, high-dimensional probability, robust programming, dynamical systems, and control theory. My information-theoretic work studies generalization and distribution mismatch. It examines how learning from one set of observations translates to performance on new data.

Academic background & collaborations

I began my PhD at EPFL in 2021 and am completing it under the supervision of Negar Kiyavash and Patrick Thiran. Most of my doctoral work has been in close collaboration with Niao He.

Since 2024, I have worked with Zebang Shen and Niao He on bilevel optimization. We have written two papers, Superquantile-Gibbs Relaxation and Select-then-differentiate, with two more in preparation.

From July to December 2024, I visited Georgia Tech, hosted by George Lan, and we collaborated on a paper during the visit.

Contact

For research collaborations, academic opportunities, or industry research roles, please reach me at mohammadsaeed.masiha@epfl.ch.