Mazdak Teymourian
Research

Understanding learned systems

I am interested in how intelligent behavior emerges in learned systems, how it can remain robust and efficient, and how we can understand the mechanisms behind what models learn.

Two straight perturbation paths across a smooth U-shaped decision boundaryA long straight perturbation crosses a smooth U-shaped decision boundary twice and returns to the original side, while a shorter straight perturbation crosses once and ends on the other side.

01 · Robustness / Adversarial Training

SORA

Free Second Order Attacks in Fast Adversarial Training

ICML 2026 · Accepted

Can we obtain useful second-order information for adversarial training without sacrificing the computational advantages of fast training?

Open project
Partial tour over a set of pointsA set of points with a partially completed tour connecting several points in sequence and a dashed candidate edge indicating the next decision.

02 · Interpretability / Neural Combinatorial Optimization

Understanding Learned Algorithms

Decision-Making Mechanisms in Neural Routing Solvers

AAAI 2027 · Submitted

What heuristics and algorithmic structure have neural combinatorial optimization models learned?

Open project
Transformation-based interpretability illustrationAn input structure is transformed in two ways and the resulting model responses illustrate invariance to one transformation and sensitivity to another.rotateinvariantscalesensitive

03 · INTERPRETABILITY / TRANSFORMATIONS & NTK

Transformation-Based Interpretability

Understanding Model Behavior Through Transformations and Neural Tangent Kernels

Ongoing research

Can the transformations a model is invariant to—or sensitive to—reveal the structure of what it has learned?

Research