Ryan Bahlous-Boldi

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

I'm a PhD student at MIT working on open-ended and continual learning. Most machine learning commits early, optimizing one fixed objective and converging onto a single solution. I'm interested in the opposite: systems that stay non-committal, keep exploring, and let what matters emerge rather than fixing it ahead of time. My work 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 / Twitter / Publications / Blog

Email: ryanbb [at] mit [dot] edu

Representative Papers

For a full list, see publications or google scholar.

* = equal contribution

	Dynamic Compression in Recurrent Networks Jyothish Pari, Ryan Bahlous-Boldi, Pulkit Agrawal arXiv preprint, 2026 PDF TL;DR: Recurrent models normally compress history into a fixed-size state in one causal pass, so each input must be compressed before the model knows how it will be used. We let the model selectively revisit past tokens and revise its state through extra recurrent updates, which substantially reduces the state size needed for accurate reuse and scales better as the number of stored functions grows.
	Vector Policy Optimization: Training for Diversity Improves Test-Time Search Ryan Bahlous-Boldi, Isha Puri, Idan Shenfeld, Akarsh Kumar, Mehul Damani, Sebastian Risi, Omar Khattab, Zhang-Wei Hong, Pulkit Agrawal arXiv preprint, 2026 PDF / Code / Website TL;DR: We propose Vector Policy Optimization (VPO), a drop-in replacement for the GRPO advantage estimator that trains LLMs to produce diverse sets of solutions specialized to different trade-offs in a vector-valued reward space, improving test-time search (pass@k, best@k) and unlocking problems evolutionary search cannot otherwise solve.
	Digital Red Queen: Adversarial Program Evolution in Core War with LLMs Akarsh Kumar, Ryan Bahlous-Boldi, Prafull Sharma, Phillip Isola, Sebastian Risi, Yujin Tang, David Ha GECCO 2026 PDF TL;DR: We study Digital Red Queen (DRQ), a self-play algorithm in which an LLM repeatedly evolves a new Core War program to defeat every program it produced before. Across rounds the programs get more general against held-out human warriors and less behaviorally diverse across independent runs, converging toward a single general-purpose strategy.
	Dominated Novelty Search: Rethinking Local Competition in Quality-Diversity Ryan Bahlous-Boldi, Maxence Faldor, Luca Grillotti, Hannah Janmohamed, Lisa Coiffard, Lee Spector, Antoine Cully GECCO 2025 PDF / DOI TL;DR: We propose a new class of quality-diversity algorithms that are simply genetic algorithms with fitness augmentations.
	Pareto Optimal Learning from Preferences with Hidden Context Ryan Bahlous-Boldi, Li Ding, Lee Spector, and Scott Niekum Reinforcement Learning Journal, vol. 6 (RLC 2025) & Pluralistic Alignment Workshop @ NeurIPS 2024 PDF / RLJ TL;DR: We frame reward function inference from diverse groups of people as a multi-objective optimization problem.

News

Aug 18, 2026	New preprint with Jyo Pari: Dynamic Compression in Recurrent Networks is on arXiv. We let a recurrent model revisit past tokens and revise its fixed-size state, trading extra computation for a much smaller state.
May 21, 2026	New preprint: Vector Policy Optimization (VPO) is on arXiv. We train LLMs to produce diverse sets of solutions specialized to different trade-offs in a vector-valued reward space, improving test-time search and unlocking problems evolutionary search cannot otherwise solve.
Jun 13, 2025	I'm honored to have been awarded the NSF Graduate Research Fellowship! This fellowship will support my PhD research on how intelligence emerges in adaptive artificial systems.
May 9, 2025	Pareto Optimal Preference Learning (POPL) was accepted to the 2025 Reinforcement Learning Conference (RLC)!
Apr 15, 2025	I'm thrilled to announce that I have committed to the PhD in EECS at MIT!
Mar 19, 2025	Dominated Novelty Search (DNS) was accepted to the 2025 Genetic and Evolutionary Computation Conference (GECCO)!

Last Updated: Jun 2026

Design adapted from Jon Barron