Should We Really Edit Language Models? On the Evaluation of Edited Language Models

TL;DR

This paper evaluates various methods for editing language models, revealing that current techniques often impair overall performance and safety, especially with larger models or numerous edits, highlighting the need for more reliable editing approaches.

Contribution

It provides a comprehensive evaluation of existing editing methods across different models, exposing their limitations and the impact on model performance and safety.

Findings

Editing methods cause performance decline on general benchmarks.

Larger models are more resistant to editing.

Model safety is significantly weakened after editing.

Abstract

Model editing has become an increasingly popular alternative for efficiently updating knowledge within language models. Current methods mainly focus on reliability, generalization, and locality, with many methods excelling across these criteria. Some recent works disclose the pitfalls of these editing methods such as knowledge distortion or conflict. However, the general abilities of post-edited language models remain unexplored. In this paper, we perform a comprehensive evaluation on various editing methods and different language models, and have following findings. (1) Existing editing methods lead to inevitable performance deterioration on general benchmarks, indicating that existing editing methods maintain the general abilities of the model within only a few dozen edits. When the number of edits is slightly large, the intrinsic knowledge structure of the model is disrupted or even completely damaged. (2) Instruction-tuned models are more robust to editing, showing less performance drop on general knowledge after editing. (3) Language model with large scale is more resistant to editing compared to small model. (4) The safety of the edited model, is significantly weakened, even for those safety-aligned models. Our findings indicate that current editing methods are only suitable for small-scale knowledge updates within language models, which motivates further research on more practical and reliable editing methods. The details of code and reproduction can be found in https://github.com/lqinfdim/EditingEvaluation.