Harnessing Large Language Models for Scientific Innovation and Broader Impact

Abstract

The rapid evolution of generative models, particularly large language models (LLMs), has catalyzed significant breakthroughs across scientific research, healthcare, and other specialized fields. Despite their transformative potential, these models often struggle with reliability and controllability when applied to complex, domain-specific tasks. In high-stakes environments such as drug discovery, synthetic chemistry, and emotionally sensitive dialogue systems, even minor deviations from desired outputs can lead to profound consequences. This dissertation identifies the inherent limitations of current LLMs in meeting these stringent demands and argues for the critical role of integrating domain-specific constraints to enhance output precision, robustness, and overall performance.

To address these challenges, this research introduces novel frameworks incorporating textual guidance, multimodal approaches, and domain-informed constraints into generative model design and evaluation. Specifically, systems like InstructMol and PRESTO demonstrate improved precision and generalization in molecular science tasks. Concurrently, the MoleculeQA benchmark is presented for rigorously assessing factual accuracy in molecular LLMs, while the GuideSafe framework offers a robust defense against adversarial disruptions, collectively bolstering model trustworthiness. The principles of domain-specific control are further extended to human-centered applications, demonstrating how frameworks for emotionally sensitive dialogue can achieve greater coherence, relevance, and empathetic alignment by leveraging structured emotional context and constrained supervision. Collectively, the contributions of this dissertation advance the reliability and controllability of LLMs. By embedding domain knowledge and developing targeted evaluation and defense mechanisms, this work paves the way for their safer and more effective deployment oaf critical scientific, healthcare, and human-centered applications, ultimately driving meaningful impact across diverse fields.

Date of Award	2025
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Yuan YAO (Supervisor) & Yangqiu SONG (Supervisor)