UC Santa Barbara

This dissertation investigates critical aspects of extending and ensuring reliability in large foundation models across multiple domains. Through five comprehensive chapters, we first address fundamental challenges in knowledge extraction and domain adaptation, then we tackle content detection and model safety when the text generation ability of large foundation models is increasing for both general and specific domains. We first present PcMSP, a novel annotated dataset for materials synthesis procedures, featuring manually validated synthesis action graphs from 305 scientific articles. This contribution includes a robust annotation framework and benchmark results across four NLP tasks, establishing a foundation for materials science information extraction. Building on this, we develop an innovative knowledge extraction system for polycrystallinematerials research, processing millions of publications to create a structured knowledge base and search engine. We further demonstrate the successful domain adaptation through continued pre-training on materials science literature, creating a specialized scientific chatbot. In addressing content detection, we introduce DNA-GPT, a training-free approach that leverages text truncation and regeneration to distinguish between human and machine-generated content. Our method achieves state-of-the-art performance across multiple languages and models while providing explainable results and demonstrating resilience to revision attacks. Finally, we present the "weak-to-strong" jailbreaking attack, revealing a critical vulnerability in aligned language models. By manipulating decoding distributions using smaller models, we achieve a 99% misalignment rate across multiple LLMs. This discovery highlights urgent safety concerns and includes an initial defense strategy while emphasizing the need for more robust protection mechanisms. This research advances our understanding of both extending and securing large foundation models, providing crucial insights for their responsible deployment in specialized domains while maintaining reliability and utility.