AI agents that act competently in novel environments typically fail in one of three characteristic ways: reinforcement learning requires millions of interactions, large language models plan unreliably despite broad knowledge, and classical planners demand labor-intensive expert specification. My project aims to learn compact, human-readable symbolic world models from few environment interactions by optimizing for safety-focused planning success. The goal is an agent that plans well from only 50-200 rollouts, with reasoning that is interpretable and cautious under uncertainty.

I am participating in SuperUROP because I’m interested in developing trustworthy and capable AI systems. I hope to deepen my background in machine learning and symbolic reasoning under Dr. Kaelbling’s guidance, leaving SuperUROP with a clearer sense of what I want to continue working on.