Yi Fang with the Computer Science and Engineering Mechanical Engineering Department has received a $50,000 award from Docomo Innovations, Inc.
As Large Language Models (LLMs) grow in scale and adoption, there is an increasing need for methods that enable efficient adaptation and precise control without the prohibitive cost of full model retraining. This project explores a unified framework combining parameter-efficient fine-tuning and targeted knowledge editing to enable lightweight, domain-specific customization of LLMs at scale. We aim to investigate how minimal updates can effectively insert, revise, or remove factual and behavioral knowledge in LLMs, while preserving overall performance and generalization. The project will develop evaluation protocols to assess factual consistency, locality of edits, task alignment, and catastrophic forgetting of LLMs. Our approach supports rapid deployment of customized LLMs across diverse applications, from personalized assistants and enterprise agents to dynamic simulation environments.