Aviral Kumar is an Assistant Professor of CS and ML and CMU. His research focuses on reinforcement learning and decision-making, in particular on offline reinforcement learning, scaling up reinforcement learning, and more recently on the intersection of reinforcement learning and foundation models. He finished his PhD from UC Berkeley in September 2023, where he was awarded the CV Ramamoorthy Distinguished Research Award, given to 1 student at UC Berkeley for outstanding contributions to a new research area in CS, Apple PhD fellowship, and a Facebook PhD fellowship. Aviral was also named as a semi-finalist (top 100) for MIT Technology’s Review 35 innovators under 35 in 2023. His research was recognized by a Samsung AI Researcher of the Year Award in 2024. He previously completed his undergraduate at IIT Bombay in India, during which he worked with Geoffrey Hinton on reinforcement learning.

AI2050 Project

This project will reimagine the paradigm used to train generalist foundation models. Rather than supervising them with “what” the correct response should be, Aviral will build approaches to supervise them with “how” to arrive at the correct response. This paradigm will be called “algorithm learning”. Akin to how human scientists push the frontiers of human knowledge and make discoveries, this paradigm of “algorithm learning” should enable generalist models to think more, propose hypotheses, seek information, and default to verbalizing their uncertainty. Aviral will build novel approaches for pre-training on Internet data and subsequent fine-tuning on downstream tasks to enable this.