Laura Wynter (IBM Research) – How to collaboratively add new knowledge to an LLM

How to collaboratively add new knowledge to an LLM

November 26, 2024 2:00 PM Singapore (Registration starts at 1:50 PM)

Abstract

LLM tuning aims at giving an LLM new knowledge or skills. The LLM should not, however, lose its ability to perform well on the tasks it was previously trained for, including the general knowledge from pre-training and skills like instruction following. This tug-of-war is further amplified when the LLM is successively subjected to more tuning, over time. We show what happens when an LLM is successively tuned and then show how to avoid the forgetting of old knowledge whilst benefiting from recent knowledge acquisition.

About the Speaker

Dr. Laura Wynter is the Director of a world-class Industrial Research Lab with 20 years of experience in conceiving, designing, and delivering cutting-edge AI and ML projects. She is a proven leader with a decade of director-level expertise in guiding high-performing teams to develop transformative research innovations that drive business success, leveraging advanced technologies to create impactful solutions and foster strategic growth.