Soheil Abbasloo

概要

Soheil is a Senior Research Scientist at Microsoft Research, where he currently works on reasoning in large language models — how to evaluate it honestly, how to train the parts that standard benchmarks quietly miss, and how to serve it without paying for capacity a model never uses. He is drawn to the places where a widely accepted metric turns out to be measuring the wrong thing, and he tends to go after those before going after the model.

He comes to this from AI-oriented design of networks and systems. Over the past decade he has built learning-based systems that had to survive contact with real infrastructure rather than a simulator — work that earned him a PhD from New York University in 2020, and continued at the University of Toronto’s Department of Computer Science, where he was a Postdoctoral Fellow and co-led the NU Lab with Prof. Ganjali until 2023. That line of work left him with a conviction he carries into language models: the hard part of a learned system is rarely the learning, it is proving the thing actually works.

Underneath all of it is a belief he keeps finding evidence for: the most complicated problems usually turn out to be the most straightforward ones. What is missing is rarely a more elaborate solution — it is the angle from which the problem was being looked at. Most of his work is really the search for that angle, and the discipline to keep looking until the simplicity that was there all along becomes visible.