Kush Bhatia

Kush Bhatia is a Research Scientist at Google DeepMind in San Francisco, having previously completed a postdoctoral fellowship at Stanford University under Christopher Ré. He earned his PhD in Electrical Engineering and Computer Sciences from UC Berkeley in 2022, where he was co-advised by Peter Bartlett and Anca Dragan, and his dissertation was titled "Learning when Objectives are Hard to Specify." Before Berkeley, he completed his undergraduate degree in Computer Science at IIT Delhi and spent two years as a research fellow at Microsoft Research India working with Prateek Jain and Manik Varma. His research spans statistical machine learning, high-dimensional statistics, optimization, and AI alignment, with a particular focus on problems at the intersection of human feedback and learning system objectives, including reward misspecification, reward hacking, and developing value-aligned systems. Notable works include "The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models" (ICLR 2022), "On the Sensitivity of Reward Inference to Misspecified Human Models" (ICLR 2023), and contributions to large language model prompting and training methodology. His postdoctoral work on safety in AI and value-aligned systems was supported by the Long-Term Future Fund.

No endorsements yet.

Personal Website

-

LinkedIn

LinkedIn

Twitter / X

-

LessWrong

-

EA Forum

-