More about HKUST
Semantic Grounding in Large Language Models: Methods, Challenges, and Applications
PhD Qualifying Examination Title: "Semantic Grounding in Large Language Models: Methods, Challenges, and Applications" by Mr. Jiayang CHENG Abstract: Semantic grounding--the process by which language models connect their outputs to verifiable knowledge sources and factual information--has emerged as a critical challenge with the advent of Large Language Models (LLMs). These models, while demonstrating remarkable generation capabilities across diverse domains, face fundamental issues with factual accuracy, hallucination, and lack of transparency in their reasoning processes. We present an in-depth analysis of core grounding tasks including knowledge base question answering, retrieval-augmented generation, and fact verification, examining the evolution from traditional symbolic systems to modern agentic approaches that incorporate sophisticated reasoning capabilities. We organize current methods across three key dimensions: parametric knowledge grounding, context-based grounding, and external knowledge grounding, analyzing how each addresses challenges such as hallucination mitigation and source attribution. Furthermore, we identify critical limitations in existing approaches and outline promising directions for future research to advance reliable and transparent language generation. Date: Thursday, 19 June 2025 Time: 9:00am - 11:00am Venue: Room 3494 Lifts 25/26 Committee Members: Dr. Yangqiu Song (Supervisor) Prof. Raymond Wong (Chairperson) Dr. Dan Xu