Instructing Language Models to Be Intelligent AI Assistants

Large Language Models (LLMs) have become foundational to natural language processing (NLP). They demonstrate impressive fluency and coherence on human language after pre-training on trillions of words from the Internet. However, true language intelligence requires more than generating text—it demands the ability to interact effectively with humans. AI assistants are expected to flexibly handle a wide range of tasks specified by users, yet such interactive behavior is underrepresented in standard pre-training corpora. To address this gap, recent research has focused on instruction alignment, which aims to train LLMs to better follow human instructions and exhibit goal-directed, helpful behavior. This thesis advances the development of instruction-following LLMs to make them more reliable and helpful. My work spans both general instruction-following capabilities and specific application domains. For general capabilities, I propose methods for automatic instruction optimization and improve adherence to hierarchical instruction structures. For applications, I introduce techniques to enhance LLMs' reasoning skills, multilingual performance, and factual accuracy. Extensive experiments demonstrate that these approaches significantly improve LLMs' ability to handle complex real-world tasks and interact naturally with users, contributing to the creation of intelligent AI assistants.<p></p>