Mass Editing Memory in a Transformer

In this paper, the authors present a method called MEMIT, which allows for the direct updating of a language model with multiple memories. This is a significant advancement as previous work in this area has primarily focused on updating single associations. By applying MEMIT, the language models GPT-J (6B) and GPT-NeoX (20B) were able to handle thousands of associations, far surpassing previous methods by a considerable margin. The authors provide code and data for replication purposes, which can be found at the given URL. The submission history shows that this paper was first submitted on October 13, 2022, with a revised version submitted on August 1, 2023.

To top