Large language models demonstrate remarkable performance across various downstream tasks, but errors can occur and some information may become outdated over time.
For example, the president of a country 10 years ago may be different from the current one — yet updating a large language model entirely just to correct such details is neither time-efficient nor computationally practical.
How is a model edited?
As the world evolves, updating Large Language Models (LLMs) while avoiding the computational burden of retraining entirely new models becomes crucial. Model Editing emerges as a promising approach to meet this demand, allowing for efficient modification of the model’s behavior.
Specifically, Model Editing enables changes to the model’s outputs within specific areas of interest without adversely affecting other inputs. Various methods have been proposed to date, with one prominent approach involving the integration of auxiliary networks with the original, unaltered model, or manipulating the model’s outputs for specific cases by altering the parameters associated with unwanted outputs.
paper: link
Future of Model Editing
Indeed, current methods primarily focus on editing factual information that can be clearly distinguished as true or false, followed by subsequent evaluation to ensure that the edits are correct. However, in the future, the scope of model editing may broaden to include editing emotional aspects of specific opinions where it’s difficult to clearly distinguish between truth and falsehood.
One important consideration is the need for objective metrics to evaluate whether model editing has been performed correctly after the edits are made.
more about model editing: link
- related post: link