Nov 30, 2023

Yitao Liang

Yitao Liang (yitaol@pku.edu.cn) is an assistant professor at Peking University. He obtained his Ph.D. degree in Computer Science from UCLA, advised by Prof. Guy Van den Broeck. His research interests span knowledge reasoning and machine learning.His work has received recognition from top AI conferences; for example, the best-paper honorable mention from AAMAS 2016, the best paper from RL for Real Life workshop in ICML 2019, a best paper runner-up from the LLD workshop in NeurIPS 2017, a best paper from the TEACH workshop in ICML2023. He regularly serves as area chairs in top venues. Recently, his group is taking a neural-symbolic approach to building a generalist agents in open-world environments.

Abstract

With the advent of large language models, the debate about whether generalist agents are coming resurges. It maybe an over ambitious goal. Yet, to make any progress, we need an appropriate testing bed accompanied with principled evaluation protocols. In our past findings, we noticed that the prior testing beds for agents are mostly designed to have one specific task and goal (sometimes specified by one reward function). This greatly limits our ability to benchmark whether we are making significant progress in building a generalist agent. In this tutorial, we will introduce the comprehensive efforts from my group and a few other related prominent research labs of using open-world environments (e.g., Minecraft) to target generalist agents. We will dig into why now it is a good time to do the switch; what are the characteristics of those environments; what are the unique challenges to them and how addressing those challenges are indispensable from generalist agents; and lastly, how the latest research in this area is reshaping our community.

Slides