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

TL;DR

A comprehensive evaluation on various editing methods and different language models indicates 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.

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.

Figures (12)

• Figure 1: Illustration about the model editing and its pitfalls in retaining edited knowledge. Left panel: model editing methods can efficiently update knowledge within language models; Right panel: when scaling editing to thousands, the model can't retain edited knowledge, see rome for details.
• Figure 2: Performance trends of evaluating edited Llama2-7B base model across different benchmarks using six editing methods. Results reveal that PMET and MEND can effectively preserve the model's abilities across all tasks. While KN drastically drops even less than ten edits.
• Figure 3: Performance trends of assessing edited Llama2-chat-7B across different benchmarks using 6 editing methods. Results reveal that PMET and MEND can effectively preserve the model's abilities across all tasks. While KN drastically drops even less than ten edits.
• Figure 4: Quantitative results of exploring the impact of model scale on edited language models. We perform editing with different-size models in Pythia model families with ROME (the first row) and MEMIT (the second row), and then these models are evaluated across diverse benchmarks.
• Figure 5: Evaluation of different kinds of general capabilities of edited language models. Results reveal that PMET and MEND can effectively preserve the model's abilities across all tasks.