On August 26, 2025, our school hosted a special lecture entitled Frontiers and Applications of Knowledge Editing in Large Models at Room X124, Nanyong Building West, Suzhou Campus of Nanjing University. The lecture was delivered by Professor Xiangnan He and hosted by Professor Wenbin Li.
During the lecture, Xiangnan He highlighted the research progress in Knowledge Editing for Large Models. He pointed out that knowledge editing serves as a critical approach for updating knowledge in large models, offering advantages such as precision and efficiency while effectively complementing post-training methods. The presentation showcased the team's latest achievements in this field, primarily addressing two key challenges: Firstly, how to prevent existing knowledge from being tampered with. To this end, the research achieved a balance between old and new knowledge through null-space constraints, successfully applied in AlphaEdit and ACE. Secondly, how to handle complex, unstructured long-form knowledge. The team proposed an auto-regressive editing method, AnyEdit, enabling iterative updates of knowledge with arbitrary length and format.
He emphasized that knowledge editing can not only be used for correcting and updating model knowledge but also holds significant expansive potential. He further outlined its broad prospects in areas such as model capability control, safe content generation, lifelong learning after model deployment, and reasoning chain editing.
During the Q&A session, attending students and faculty actively raised questions regarding multimodal signals, real-time updates, fine-tuning of large models, and theoretical foundations. Xiangnan He addressed each inquiry in detail, fostering a lively atmosphere and positive feedback throughout the venue.
The lecture concluded successfully. Professor Wang Xiang from the University of Science and Technology of China and Fan Qi from the School of Artificial Intelligence Science and Technology were also in attendance.
