Reasons and Solutions for the Decline in Model Performance after Editing

TL;DR

A Dump for Sequence (D4S) method is proposed, which successfully overcomes the previous editing bottleneck by reducing the L1-norm of the editing layer, allowing users to perform multiple effective edits and minimizing model damage.

Abstract

Knowledge editing technology has received widespread attention for low-cost updates of incorrect or outdated knowledge in large-scale language models. However, recent research has found that edited models often exhibit varying degrees of performance degradation. The reasons behind this phenomenon and potential solutions have not yet been provided. In order to investigate the reasons for the performance decline of the edited model and optimize the editing method, this work explores the underlying reasons from both data and model perspectives. Specifically, 1) from a data perspective, to clarify the impact of data on the performance of editing models, this paper first constructs a Multi-Question Dataset (MQD) to evaluate the impact of different types of editing data on model performance. The performance of the editing model is mainly affected by the diversity of editing targets and sequence length, as determined through experiments. 2) From a model perspective, this article explores the factors that affect the performance of editing models. The results indicate a strong correlation between the L1-norm of the editing model layer and the editing accuracy, and clarify that this is an important factor leading to the bottleneck of editing performance. Finally, in order to improve the performance of the editing model, this paper further proposes a Dump for Sequence (D4S) method, which successfully overcomes the previous editing bottleneck by reducing the L1-norm of the editing layer, allowing users to perform multiple effective edits and minimizing model damage. Our code is available at https://github.com/nlpkeg/D4S.

Figures (8)

• Figure 1: This framework outlines the comprehensive approach to understanding the performance decline of edited models. On the left, traditional knowledge editing tasks are categorized into different types, each with distinct editing objectives: yes/no, a/b/c/d, and entity/event. On the right, our experiments are structured from both data and model perspectives. From the data perspective, we conduct three experiments: (a) a comprehensive performance evaluation of the model, (b) the construction of a Multi-Question Dataset (MQD), and (c) an assessment of the impact of editing different target outputs on model performance. From the model perspective, we design four experiments: (d) an evaluation of the edited model's forgetting ability, (e) an identification of the current knowledge editing method's bottleneck and an exploration of the correlation between editing probability values and parameter layer norms, and (f) a proposal of a sequence editing method, which effectively enhances the performance of the edited model.
• Figure 2: Evaluation results of different editing methods on various types of datasets. The horizontal axis in the image represents the number of edited samples, and the vertical axis represents the performance of the edited model.
• Figure 3: The performance of the model after editing data for different question types.
• Figure 4: Assessment of forgetting ability of models.
• Figure 5: The bottleneck and L1-norm correspondence in sequence editing are illustrated in the figure. In subfigure (a), the horizontal axis represents the number of edited samples, and the vertical axis represents the probability value of the edited object. The blue line represents the ROME method, while the green line represents the MEMIT method. In subfigures (b) and (c), the horizontal axis represents the number of edits, and the vertical axis represents the L1-norm value of the editing layer.