Twice is Nice! Improving MEMIT for Knowledge Editing
Abstract: In this ever changing world, Large Language Models (LLMs) pose a major challenge on frequently retraining as they consume enormous resources. Direct knowledge editing emerged as an efficient alternative, where it locates a stale factual knowledge in the LLM's layers and edits those layers' weights in order for the LLM to generate new factual knowledge. We observed that MEMIT, state-of-the art knowledge editing algorithm is not used at its full potential.
In this paper, we empirically demonstrated the limitations of executing only one single MEMIT update. We then proposed an intuitive and straightforward solution, Running MEMIT twice, and showed its effectiveness over two knowledge editing datasets compared to strong baselines.
We conducted extensive analysis to understand the effectiveness of our solution. In particular we analyzed multiple runs of MEMIT and found out the performance to plateaus at second run of MEMIT. To discern the reason we analyzed the gradients between each run and found negligible change in gradients between second and third run of MEMIT.
Paper Type: Long
Research Area: Machine Learning for NLP
Research Area Keywords: Kowledge Editing, Optimization
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 2441
Loading