William Jurayj

I am a PhD student at Johns Hopkins University, advised by Benjamin Van Durme. My research is centered on methods to help imperfect AI systems earn human trust. Recently, I’ve focused on making language model agents more effective at reasoning about and conveying their uncertainty, and at following complex rules and constraints faithfully.

Some questions that are currently on my mind:

• How does an agent’s uncertainty develop with increased inference budgets (Test-Time Scaling Confidence)? How can this behavior be learned (RLCM) or leveraged for early stopping (Conformal Thinking)?
• Can language models reason economically over large corpora of rules (Legal Logic Programs)? When are opinionated workflows more effective than general agents (Deontic Agentic Reasoning)?
• What other bottlenecks prevent widespread diffusion of language model agents throughout the economy?

I was a research intern at Amazon AGI working with Sapana Chaudhary and Kaj Bostrom in Summer 2026. Before I came to Hopkins, I worked at Abnormal Security as a machine learning engineer training behavioral models to detect compromised accounts. I completed my Bachelor’s and Master’s degrees at Brown, where I was fortunate to be advised by Carsten Eickhoff, Ellie Pavlick, and George Konidaris.