Duality of Bias in Large Language Models - Leveraging Community Perspectives and Uncovering Ideological Vulnerabilities

Explore a thought-provoking research seminar that delves into the dual nature of bias in Large Language Models (LLMs), presented by USC/ISI PhD candidate Zihao He. Learn about two contrasting studies: the COMMUNITY-CROSS-INSTRUCT framework for analyzing public opinion through LLM-created digital twins, and research revealing LLMs' vulnerability to ideological manipulation through instruction tuning. Gain valuable insights into both the innovative potential and limitations of LLMs in social science research, while understanding the critical importance of responsible AI development. Drawing from his expertise in natural language processing and computational social science, He shares findings from his research published in prestigious conferences like ACL, EMNLP, and ICWSM, offering a comprehensive examination of how LLMs can be leveraged for societal benefit while remaining mindful of their susceptibility to external influences.