CAREER: Understanding, Correcting, and Adapting Large Language Models from a Knowledge-Oriented Perspective in NLP Applications and Beyond

University Of California-Santa Barbara, Santa Barbara CA

Investigators

Abstract

Large language models (LLMs) have been widely adopted for various tasks, including question-answering, code generation, and addressing complex challenges based on user instructions. Their extensive applications highlight their ability to store, process, and deliver knowledge, revolutionizing the fields of natural language processing and artificial intelligence. Despite the superior performance of LLMs in harnessing knowledge, three research challenges remain that prevent LLM techniques from being applied to a wider range of real-world applications and use cases: 1) understanding how knowledge interacts with LLMs’ behavior; 2) surgically and precisely correcting outdated and incorrect knowledge in LLMs without affecting other knowledge; and 3) efficiently imparting new knowledge to adapt LLMs to different tasks and domains. Limited access to LLM parameters, pre-training data, training recipes, and computational resources hinders conventional methods from addressing these challenges. Hence, there is a strong need to develop innovative algorithms that can improve the understanding, correction, and adaptation of LLMs’ knowledge despite these constraints, which is the main focus of this project. Additionally, this research will be integrated into education through new teaching modules in developing graduate LLM courses, promoting education for undergraduate research, delivering workshops and tutorials at major conferences, and outreach to students from underrepresented communities to promote their awareness and skills in utilizing and comprehending LLMs. This project is structured from a knowledge-oriented perspective and is organized into four thrust quadrants based on two dimensions: the location of knowledge (within model parameters or introduced through model inputs) and the end goal (enhancing understanding for better interpretability or updating the knowledge base for corrections or task adaptation). Specifically, thrust 1 focuses on improving the understanding of how LLMs leverage external knowledge by investigating the impact of 'task' and 'knowledge' information in in-context learning. Building upon the understanding of how knowledge is delivered in in-context examples, thrust 2 focuses on advancing in-context selection mechanisms that enable LLMs to learn from their own mistakes and to be deployable in scenarios where the data for in-context selection may change over time. Thrust 3 aims to understand where knowledge is embedded in LLM model parameters, specifically with different levels of knowledge granularity. Finally, thrust 4 aims to deliver methods to update knowledge surgically through model parameters. 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.