Mansi Maheshwari

I'm a PhD student at Mila studying Lifelong Machine Learning under Prof. Sarath Chandar.

I study how learning agents can adapt continually, even after deployment, without sacrificing safety. My past work tackled the plasticity–stability dilemma in single-agent reinforcement learning: how an agent can keep learning new things without forgetting, destabilizing, or becoming unsafe in the process.

Current obsessions:

• Continual multi-agent RL. How do we bring continual learning to settings where everyone else is changing too?
• Continual alignment. If learning never stops, safety can't be a one-off check. How do we keep every update aligned? Broadly interested in alignment.
• Plasticity and agency. Is an agent that can't change still choosing anything?

I moved from India to Seattle to pursue a Bachelor's in Electrical Engineering at the University of Washington. After graduating, I spent two years building scalable systems as a Software Engineer at Nordstrom. While I loved the practical side of tech, I felt a pull towards research — to tackle fundamental, open-ended questions.

That realization brought me to pursue a Master's at UMass Amherst, where I fell in love with Reinforcement Learning. I had the privilege of conducting research at the Autonomous Learning Lab (founded by Turing Awardee Andy Barto) under Prof. Bruno Castro da Silva. Our work was published at CoLLAs 2025 and AAMAS 2026 (oral), and I distilled the core ideas for a broad audience in my Three Minute Thesis. I also got to share that passion with high school students by teaching Intro to AI at UW, and put it to work in the real world as an AI Research Intern building autonomous driverless tractors at CNH Industrial.

Now at Mila, my research focuses on developing lifelong systems that can continuously adapt and learn safely, often drawing inspiration from natural learning systems.