Amrith Setlur

I’m a final year PhD Student in the Machine Learning Department at Carnegie Mellon University, where I am fortunate to be advised by Virginia Smith. I am also a long term visiting researcher at UC Berkeley, advised by Sergey Levine, and collaborate very closely with Aviral Kumar. My PhD is generously supported by the JP Morgan AI PhD Fellowship award.

Research Overview

I work on fundamental principles and practical recipes for building AI systems that continually adapt at test-time and solve hard problems (e.g., open conjectures in CS/math research). These systems spend additional computation on difficult instances through reasoning, search, and interaction, rather than behaving as static predictors (e.g., neural nets that fail under distribution shift). We can think of a system that adapts at test-time as one that is running an algorithm. This means I am essentially interested in learning algorithms (e.g., by training LLMs that can represent one) efficient at spending the additional computations on an input to perform intelligent test-time search needed to solve the hard problem.

More recently, I have focused on understanding the bottlenecks in training reasoning LLMs, improving exploration on hard problems where verification signals are weak, and, more broadly, developing RL training methods for LLMs that can continually interact, adapt, and update themselves in underspecified test environments when deployed.