Ryan Bahlous-Boldi (bah-LOOSE BOWL-dee)

I'm a PhD student at MIT studying how to build agents that continually learn and adapt in open-ended environments. Unlike biological learners that adapt fluidly, current AI systems are brittle. They forget past knowledge when adapting to new tasks and require massive data to achieve fragile gains. My goal is to develop architectures, algorithms, and environment generation techniques that enable smaller, more data-efficient models to be taught incrementally about the world. My research draws on reinforcement learning, evolutionary computation, and artificial life.

I am advised by Pulkit Agrawal in the Improbable AI Lab. Previously, I was an undergrad at UMass Amherst advised by Lee Spector and Scott Niekum. I've also had the pleasure of collaborating with Stefanos Nikolaidis at USC and Katia Sycara at CMU.

My work is supported by the NSF Graduate Research Fellowship.

CV  /  Scholar  /  GitHub  /  Publications  /  Blog

Email: [first name] bb [at] mit [dot] edu

Representative Papers

published under Ryan Bahlous-Boldi and Ryan Boldi

For a full list, see publications or google scholar.

* = equal contribution

TL;DR: We propose a new class of quality-diversity algorithms that are simply genetic algorithms with fitness augmentations.

TL;DR: We frame reward function inference from diverse groups of people as a multi-objective optimization problem.

TL;DR: We present an approach that uses lexicase selection to solve deceptive problems by optimizing a series of defined objectives, implicitly maintaining population diversity.

TL;DR: We develop methods to identify the most productive training cases for lexicase selection, improving computational efficiency while maintaining solution quality.

News

© 2025 Ryan Bahlous-Boldi

Last Updated: Dec 2025

Design adapted from Jon Barron