Building Proactive Agents That Augment Human Work

Recent advances in large language models (LLMs) have transformed human-AI interaction yet building agents that augment human work requires systems that understand the people they work with. In this talk, Diyi Yang will first audit the U.S. workforce to assess the impact of automation and augmentation on the future of work, grounding agent design in workers’ perspectives. She will complement this audit with an analysis of nearly 250,000 real-world conversations, examining what work people bring to AI, how much agency they retain, and where collaboration breaks down. Building on these findings, Yang will introduce General User Models (GUMs), which learn about users by observing any computer interaction and constructing propositions about their knowledge, preferences, and context. She will further present NAP (Next Action Prediction), a framework for anticipating user actions by reasoning over rich multimodal sequences of human-computer interactions, where modeling long interaction histories produces significantly more accurate predictions. Overall, these results show how to build proactive agents that understand users and augment their work.

This talk is part of the Kempner Seminar Series, a research-level seminar series, open to the public, that covers topics related to the basis of intelligence in natural and artificial systems.