Invited Talk 2 - Diyi Yang
Abstract
Title: Human-AI Collaboration Across Turns
Abstract: Recent advances in large language models (LLMs) have transformed human-AI interaction, however, building effective multi-turn collaboration requires both rigorous evaluation frameworks and systems that truly understand us. In this talk, we first present Collaborative Gym (Co-Gym), a framework for studying human-agent collaboration across diverse tasks. Our findings reveal that collaborative agents consistently outperform fully autonomous counterparts. We then introduce General User Models (GUMs), which learn about users by observing any computer interaction and constructing propositions about user knowledge, preferences, and context. Overall, this talk demonstrates how to build AI systems that collaborate effectively within tasks and maintain understanding across turns to enable personalized human-AI collaboration.