Montaser Mohammedalamen, PhD

Bio

I am an Applied Research Scientist in AI Trust & Safety at the Alberta Machine Intelligence Institute (Amii), where I lead research at the intersection of reinforcement learning and AI safety. My work spans two complementary tracks: using RL as a diagnostic and offensive tool to systematically audit and defend frontier large language models against adversarial jailbreaking, and addressing foundational failure modes within RL agents themselves — including epistemic uncertainty, partial reward observability, non-stationarity, and continual learning constraints.

I completed my PhD in Computing Science at the University of Alberta, advised by Prof. Michael Bowling, where my dissertation introduced the Monitored MDP framework — a principled treatment of reinforcement learning under partially observable reward signals — along with algorithms for cautious, robust decision-making in novel and out-of-distribution environments.

Prior to academia, I spent three years as an AI Engineer at SonyAI in Tokyo, contributing to the ACE table-tennis robot project — a system that achieved professional-level play and resulted in a Nature publication. My work there spanned multi-agent simulation, sim-to-real transfer, and real-time robot control integration.

My research has been published in TMLR, AAMAS, and RLC, and I collaborate closely with the Canadian AI Safety Institute (CAISI), the Canadian Institute for Advanced Research (CIFAR), and the National Research Council of Canada (NRC).

Research Vision

Looking ahead, my research agenda centers on building agents that continuously learn and adapt from real-world experience, moving beyond the idealized assumptions of traditional MDP formulations that rarely hold outside labs and simulations. I am particularly interested in settings where partial reward observability and non-stationarity are the rule rather than the exception, and in developing the principled foundations that let agents navigate these conditions reliably. Across all of this, I hold as a core constraint that deployed agents must remain safe — not just in the moment, but throughout their operational lifetime — to themselves, to the humans who depend on them, and to the broader environment they inhabit.

Updates

• Jul 2026 Two workshop papers accepted at the Agents in the Wild: Safety, Security, and Beyond Workshop, ICML 2026: "A Systematic Investigation of The RL-Jailbreaker in LLMs" and "AI Agent Safety is a Reinforcement Learning Problem."
• Jul 2026 Paper Generalization in Monitored Markov Decision Processes (Mon-MDPs) accepted at the Reinforcement Learning Conference (RLC) 2026, Montreal, Canada.
• Apr 2026 Contributed to the ACE table-tennis robot project at SonyAI, which resulted in a Nature publication on a system capable of defeating professional players.
• Dec 2025 Completed PhD in Computing Science at the University of Alberta. Thesis: Reinforcement Learning with Partially Observable Rewards.
• Oct 2025 Paper Learning to Be Cautious published in Transactions on Machine Learning Research (TMLR).
• Jul 2025 Joined the Alberta Machine Intelligence Institute (Amii) as an Applied Research Scientist in AI Trust & Safety.
• May 2025 Received the Andrew Stewart Memorial Graduate Prize from the University of Alberta (5,000 CAD).
• May 2024 Presented Monitored Markov Decision Processes at AAMAS 2024, Auckland, New Zealand.
• Jan 2024 Received the Graduate Student Engagement Scholarship (GSES) from the University of Alberta (10,000 CAD).
• Jan 2023 Received the Graduate Student Engagement Scholarship (GSES) from the University of Alberta (10,000 CAD).
• Jan 2022 Started PhD at the Computing Science Department, University of Alberta, advised by Prof. Michael Bowling.
• Jun 2020 Joined SonyAI as an AI Engineer in Tokyo, Japan.