CAIA Speaker: Gon Buzaglo

​Who: Gon Buzaglo, Princeton (VIRTUAL)
When: February 18, 2–3 pm PT
Where: Gates Annex B122
Zoom: https://rit.zoom.us/j/95346944675

​Title: The Hidden Game Problem

​Abstract:
Many challenges in AI alignment can be framed as language games, with AI safety via debate providing a central example. In these settings, the space of possible actions consists of sentences or arguments, which is vast and makes efficient learning extremely difficult. However, such games often exhibit hidden structure: a small set of strategies or arguments consistently perform better than the rest. In this talk, I will introduce the hidden game problem, which formalizes this idea and asks how learning dynamics can efficiently discover and focus on these strong strategies. I will present recent game-theoretic algorithms that exploit hidden structure to reach equilibrium much faster than was previously possible, while still maintaining rationality in general. This work is a first step toward developing a new algorithmic framework for alignment. Based on joint work with Noah Golowich and Elad Hazan: https://arxiv.org/abs/2510.03845

​Bio:
Gon Buzaglo is a second-year Ph.D. student in Computer Science at Princeton University, advised by Prof. Elad Hazan. His research focuses on theoretical machine learning, with current interests in AI alignment, optimization, online control, and learning in games. Before Princeton, he completed his M.Sc. at the Technion under Prof. Daniel Soudry, working on generalization in deep learning in collaboration with Prof. Nathan Srebro, and his B.Sc. in Computer Science and Physics, during which he interned with Prof. Michal Irani at the Weizmann Institute studying memorization in neural networks.

​Everyone is welcome: No specific technical background is required. Come learn and ask questions.
And yes, we will have pizza and boba.