Constraining Sequential Model Editing with Editing Anchor Compression
Summary: arXiv:2503.00035v2 Announce Type: replace-cross
Abstract: Large language models (LLMs) struggle with hallucinations due to false or outdated knowledge. Given the high resource demands of retraining these models, there is an increasing focus on developing model editing. However, the general abilities of LLMs across downstream tasks are prone to significant degradation during sequential editing. This paper statistically observes that the parameter matrix after editing exhibits a significant deviation compared to its previous state as the number of edits increases. This serious deviation affects the original knowledge associations within LLMs and leads to the degradation of their general abilities. To this end, a framework termed Editing Anchor Compression (EAC) is proposed to constrain the deviation of the parameter matrix during sequential editing. It compresses the editing information by selecting editing anchors that are important in encoding new relations without deviating too much from the original matrix, thereby preserving the general abilities.
Introduction
The emergence of large language models (LLMs) has revolutionized natural language processing tasks. However, these models are not without their challenges, particularly when it comes to their tendency to produce hallucinations—outputs that are factually incorrect or based on outdated knowledge. Traditional methods of addressing these issues often involve retraining the models, a process that demands extensive computational resources and time.
Challenges in Model Editing
As researchers explore the potential of model editing as an alternative to retraining, they face a significant hurdle: the degradation of the model’s general abilities across a variety of downstream tasks. Sequential editing, while allowing for targeted updates, has been shown to result in substantial deviations in the parameter matrices of the models. This deviation compromises the knowledge associations that the model has previously established, leading to a decline in its overall performance.
Editing Anchor Compression Framework
In response to these challenges, the authors propose a novel framework known as Editing Anchor Compression (EAC). This approach aims to mitigate the adverse effects of sequential editing by strategically compressing the editing information. The key features of EAC include:
- Selection of Editing Anchors: The framework identifies and prioritizes editing anchors that are crucial for encoding new relationships.
- Minimization of Deviation: By focusing on essential anchors, EAC reduces the extent to which the parameter matrix deviates from its original state.
- Preservation of General Abilities: The overarching goal of EAC is to maintain the model’s general capabilities while effectively integrating new knowledge.
Experimental Validation
The efficacy of the Editing Anchor Compression framework was tested through experiments applying EAC to two widely used editing methods, utilizing three different LLMs across four distinct tasks. The results demonstrated that EAC significantly reduces unreasonable deviations caused by model editing, achieving the remarkable feat of preserving over 70% of the model’s general abilities. Furthermore, EAC proved to be more effective in retaining the newly edited knowledge compared to traditional methods.
Conclusion
The Editing Anchor Compression framework marks a significant advancement in the field of model editing for large language models. By addressing the critical issue of knowledge degradation during sequential edits, EAC provides a promising solution that balances the integration of new information while safeguarding the model’s overall performance. As the demand for reliable and efficient language models continues to grow, innovations such as EAC will play a pivotal role in advancing the capabilities of artificial intelligence.