- Main
Integrating Language Models with Evolving External Knowledge
- Wu, Di
- Advisor(s): Chang, Kai-Wei
Abstract
Large language models acquire broad capabilities from public training data, but their deployment increasingly takes place in private worlds shaped by individual users, organizations, and specialized environments. These worlds contain personally and economically valuable knowledge that is local, dynamic, and accumulated through continued interaction and experience. This dissertation studies how language models can effectively integrate such evolving external knowledge. We first characterize how the knowledge requirements of deployed systems expand as applications progress from chat assistants to long-horizon agents. To address these challenges, we develop a two-sided approach to knowledge augmentation. On the model side, we study how language models can recognize the boundary of their own knowledge and selectively draw from multiple knowledge sources. On the knowledge side, we design increasingly expressive representations that make it possible to discover, organize, and reuse broader forms of domain, reasoning, and organizational knowledge. Together, these directions provide a unified framework for language models to continually acquire and use external knowledge as they adapt to evolving worlds.