Lifelong Sequential Knowledge Editing without Model Degradation
Akshat Gupta, Phudish Prateepamornkul, Maochuan Lu, Ahmed Alaa, Thomas Hartvigsen, Gopala Anumanchipalli
2025-02-04
Summary
This paper talks about ENCORE, a new method that allows AI models to be updated with new information many times without losing their ability to perform well on other tasks. It focuses on solving problems that happen when models are edited repeatedly.
What's the problem?
When AI models are updated with new facts or corrections, they often start to overfit, meaning they focus too much on the new information and lose their general abilities. This problem gets worse when models are updated many times, making it hard to keep them accurate and useful in the long run.
What's the solution?
The researchers created ENCORE, which stands for Early stopping and Norm-Constrained Robust knowledge Editing. This method prevents overfitting by controlling how much the model changes during each update and limiting the growth of certain internal values that can cause problems. Using ENCORE, they were able to update an AI model up to 10,000 times without it losing its overall performance. ENCORE is also faster than other methods like MEMIT and AlphaEdit.
Why it matters?
This research is important because it makes AI systems more reliable and adaptable. By allowing models to learn new information repeatedly without degrading their performance, ENCORE helps create smarter AI that can stay up-to-date and useful in real-world applications like customer support, education, and healthcare.
Abstract
Prior work in parameter-modifying knowledge editing has shown that large-scale sequential editing leads to significant model degradation. In this paper, we study the reasons behind this and scale sequential knowledge editing to 10,000 sequential edits, while maintaining the downstream performance of the original model. We first show that locate-then-edit knowledge editing methods lead to overfitting on the edited facts. We also show that continuous knowledge editing using these methods leads to disproportionate growth in the norm of the edited matrix. We then provide a crucial insight into the inner workings of locate-then-edit methods. We show that norm-growth is a hidden trick employed by these methods that gives larger importance to the output activations produced from the edited layers. With this "importance hacking", the edited layers provide a much larger contributions to the model's output. To mitigate these issues, we present ENCORE - Early stopping and Norm-Constrained Robust knowledge Editing. ENCORE controls for overfitting and the disproportionate norm-growth to enable long-term sequential editing, where we are able to perform up to 10,000 sequential edits without loss of downstream performance. ENCORE is also 61% faster than MEMIT and 64% faster than AlphaEdit on Llama3-8B.