Collaborative Research: III Small: Towards Large Open-World Foundation Models: Construction, Adaptation, and Deployment

About This Grant

Large language models (LLMs), such as GPT-4, also known as foundation models, represent a groundbreaking advancement in artificial intelligence, powering diverse applications such as chatbots, information retrieval engines, and scientific discovery tools. The success of LLMs is primarily attributed to their vast internal knowledge learned from massive unstructured texts. However, LLMs' internal knowledge is inherently constrained by static snapshots of closed-world texts used for training. This limitation presents two key challenges when deploying such a closed-world LLM in an open-world environment where knowledge (including unstructured texts and structured data) is rapidly evolving. First, closed-world LLMs have a limited and static view of open-world knowledge, often leading to unfaithful yet overconfident hallucinations. Second, since they primarily process unstructured text, they struggle with tasks that require reasoning over structured data, such as databases or scientific records. This project addresses key challenges in artificial intelligence by developing new algorithms, theorems, and systems to ensure the reliability of advanced foundation models. At its core, the research focuses on an open-world foundation model (OWFM), a powerful AI model designed to interact with ever-changing real-world information. This model is built on an open-world knowledge network (OKN), a flexible and expandable data structure that organizes semi-structured information from diverse sources. Moving beyond unstructured knowledge and closed-world LLMs, this project establishes a new paradigm of knowledge organization and foundation models in an open-world environment through three key thrusts. The first thrust creates a highly expandable OKN across domains and builds an OWFM with plug-and-play modular components. The second thrust develops adaptation methods that enable OWFMs to accurately answer questions on rapidly evolving topics. The third thrust deploys the OWFM in two knowledge-intensive, high-stakes applications: medical diagnostic reasoning and financial portfolio generation. Beyond advancing scientific research, this project also integrates its findings into educational activities, ensuring broad dissemination and real-world impact. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.