Ruohan Zong's Preliminary Exam

Recent advances in large language models (LLMs) have significantly improved performance across a wide range of tasks, enabling AI systems to generate, reason, and interact with information at an unprecedented level. However, strong model capability does not directly translate to reliable and usable systems in real-world settings. In practice, AI systems must handle diverse and evolving tasks, align their outputs with human objectives, and present results in a form that users can understand and trust. These requirements introduce challenges that extend beyond traditional model design. This thesis studies human-centered alignment for reliable and explainable AI systems through three interconnected components: adaptive modeling, reliable alignment, and explainable interaction. First, we develop adaptive modeling approaches that enable AI systems to adjust their behavior based on task characteristics, improving performance across diverse and complex scenarios. Second, we propose alignment methods that incorporate multiple objectives, such as accuracy, fairness, and factuality, and explicitly account for potential conflicts among them to improve reliability. Third, we design explainable interaction mechanisms that present model outputs with clear reasoning and supporting evidence, enabling users to effectively interpret and use AI systems. Together, these contributions provide a unified perspective on designing AI systems that are both reliable in behavior and explainable in use, bridging the gap between strong model capability and trustworthy operation in real-world settings.

Questions? Contact Ruohan Zong.