Understanding, Leveraging, and Improving Large Language Models

The emergence of Large Language Models (LLMs) has marked a substantial advancement in Natural Language Processing (NLP), contributing significantly to enhanced task performance both within and outside specific domains. However, amidst these achievements, three key questions remain unanswered: 1) The mechanism through which LLMs accomplish their tasks and their limitations, 2) Effectively harnessing the power of LLMs across diverse domains, and 3) Strategies for enhancing the performance of LLMs. This talk aims to delve into our research group's endeavors to address these pivotal questions. Firstly, I will outline our approach, which involves utilizing ontology-guided prompt perturbations to unravel the primary limitations of LLMs in solving mathematical problems. Moving on to the second question, we will explore the utilization of synthetic data generated by LLMs to bolster challenging downstream tasks, particularly focusing on structured prediction where LLMs face persistent challenges. I will elaborate on our initiatives aimed at improving LLMs by incorporating highly effective retrieval strategies, specifically addressing the prevalent challenge of hallucinations that often plagues contemporary LLMs. Finally, I will present a technique on LLM realignment to restore safety lost during fine-tuning.