A team of researchers from Zhejiang University, the National University of Singapore, the University of California, Ant Group, and Alibaba Group have conducted a study focusing on knowledge editing for Large Language Models (LLMs). LLMs have recently showcased their impressive capability to process and memorize extensive amounts of information, surpassing human capacity.
To ensure the fairness and safety of Artificial Intelligence (AI) systems, it is crucial to understand how LLMs display and process information. This study aims to survey the history and current state of knowledge editing techniques for LLMs. The researchers provide an overview of LLMs’ design, how knowledge is stored, and related approaches such as parameter-efficient fine-tuning, knowledge augmentation, continuing learning, and machine unlearning.
The researchers classify knowledge editing strategies for LLMs into three categories: editing internal knowledge methods, merging knowledge into the model, and resorting to external knowledge. These strategies draw inspiration from human cognitive processes, such as recognition, association, and mastery phases of learning.
The study includes experiments conducted on twelve natural language processing datasets, carefully considering performance, usability, underlying mechanisms, and other factors. The researchers create a benchmark called KnowEdit to evaluate information insertion, modification, and erasure using state-of-the-art LLM knowledge editing techniques.
The findings demonstrate how knowledge editing affects general tasks and multi-task knowledge editing, showcasing that it successfully updates facts without significantly impacting the model’s cognitive abilities and adaptability in different knowledge domains. Additionally, the researchers explore the limitations and potential repercussions of knowledge editing for LLMs.
Moreover, the study discusses the wide range of applications for knowledge editing, including trustworthy AI, efficient machine learning, AI-generated content, and individualized agents in human-computer interaction. The researchers hope that this research will inspire further exploration into LLMs, focusing on both efficiency and creativity.
The researchers have made all their resources, including codes, data splits, and trained model checkpoints, publicly available to encourage more study in this area.