Too Big too Fool: Resisting Deception in Language Models
Preprint
I'm Mo
I’m a student researcher at Google Research, while pursuing my M.Sc. at Mila - Québec AI Institute. I also co-founded Zette AI, where we’re building the next wave of high-quality, fact-checked data pipelines for AI. Before this, I was a visiting researcher at ServiceNow, interned at EPFL, and worked on AI R&D at multiple startups. I earned my B.Sc. in Computer Science from Sharif University of Technology.
I like to build agents that can plan effectively in complex situations. I’ve been exploring how LLMs form a mental model of their world, often called “world models”, and how improving world models boost agentic systems. In our “Recall to Imagine” paper, enhancing the world model for reinforcement learning led us to superhuman performance. Next up at Google, I’m integrating advanced computer vision to make these world models more practical, focusing on real-world applications like robotics.
My CV