I am a research scientist in Meta's Applied AI organization, where I work on auto-research, active data generation, and expert-supervised improvement loops for models and agents. Previously, I was on Meta's Adaptive Experimentation team in Central Applied Science, where I led work on principled methods for modernizing Bayesian optimization, preference learning, and adaptive experimentation with AI and human feedback. This line of work asks how optimization systems can move beyond fixed numeric/vector inputs to reason over language-rich system components, natural-language goals, preferences, multiple objectives, and scarce online experiments, with applications across large-scale AI and recommender systems, infrastructure optimization, and user-facing products. I am also a core contributor to the open-source projects BoTorch and Ax.

My research develops Bayesian and probabilistic methods for modeling and learning from human behavior, preferences, and feedback, with the goal of making AI systems more data-efficient, interpretable, and responsive to human context. I study how uncertainty, expert judgment, natural-language feedback, and behavioral signals can inform active data collection, preference learning, model evaluation, and adaptive optimization in AI systems and small-sample decision settings. More broadly, I’m interested in how computational models can help us better understand human behavior and, in turn, design AI systems that learn from and interact with people in more reliable and meaningful ways. My work has been covered in the New York Times, LA Times, San Francisco Chronicle, the Verge, Washington Post, MIT Technology Review, Scientific American, and other places.

I received my Ph.D. in computer science from Stanford University advised by Sharad Goel and B.S. in computer science with a minor in probability and statistics from Georgia Institute of Technology. You can reach me at zylin@cs.stanford.edu.