In LLMs, the ability to update factual knowledge of the model is an important and attractive ability, due to the need to deal with everchanging nature of information in the world.
However, predicting whether an edit applied to a LLM will be successful or not is difficult. In this work, we suggest two metrics that can predict the editing success:
(1) where the knowledge is stored in the parameters as reflected by the logit-lens technique;
(2) the probability that the model assigns to the correct output.
We find a correlation between the location of the knowledge and the optimal layer for editing, as well as between the output probability and the success. Moreover, we found a differential relation between the output probability and each component of the success.
