TIME
2024年12月17日(周二)14:00 – 15:00
VENUE
信管学院308会议室
SPEAKER
Jun Zhao(赵君) is a third-year PhD student in the Natural Language Processing Lab at Fudan University (FDU), advised by Prof. Xuanjing Huang and Prof. Qi Zhang. His research interests focus on Trustworthy Large Language Models, particularly in the areas of Open-domain Information Extraction, LLM Interpretability, and Alignment. He has published 8 first-author papers in top-tier Natural Language Processing conferences including ACL, EMNLP, and COLING, and received a Best Paper Award nomination at CCL 2022. He serves as an Area Chair for ACL 2024 and NAACL 2024, and regularly reviews for prestigious conferences such as ACL, AAAI, ICML, and ICLR.
TITLE
Towards Trustworthy and Responsible Large Language Models
ABSTRACT
Large language models (LLMs) have demonstrated remarkable performance on knowledge reasoning tasks, owing to their implicit knowledge derived from extensive pretraining data. However, LLMs still face several critical challenges: they may generate false or misleading information (hallucination), their internal decision-making mechanisms remain difficult to interpret, and their alignment with human intent is insufficient. These inherent limitations severely restrict the large-scale application of LLMs in critical domains such as medical diagnosis, legal consultation, and financial decision-making, where reliability and transparency requirements are exceptionally high. In this talk, I will discuss my perspectives on building trustworthy and responsible large language models, focusing on three key aspects: (1) how to effectively mine constantly evolving knowledge from open-domain sources; (2) the fundamental basis of LLMs' linguistic capabilities and knowledge storage; and (3) how to efficiently and unbiasedly align with human intent while expanding LLMs' capability boundaries.
编审:唐志皓 江波
欢迎 关注!