Talk Title: Training Language-Based Agents

Talk Abstract: Following instructions by executing actions in embodied environments is a core competency of language-using agents, but is not adequately learned from pre-training on large amounts of web text data. In this talk I will describe several approaches to training language-based embodied agents without relying on large amounts of in-domain pre-training data, including exploration-based approaches using interaction with human users and environment simulators, covering tasks like language-conditioned navigation, device control, and software engineering. Finally, I will describe how these approaches can also be used to train more accurate and efficient reasoning models.

Bio: Alane Suhr is an Assistant Professor at UC Berkeley EECS, affiliated with the Berkeley AI Research lab. Alane's work focuses on building language-using systems that communicate with and learn from human users in collaborative, situated interactions. Prior to joining Berkeley, Alane completed a PhD in Computer Science at Cornell University / Cornell Tech and spent a year afterwards as a Young Investigator at the Allen Institute for AI.