Knowledge Editing for Large Language Model with Knowledge Neuronal Ensemble

TL;DR

Knowledge editing in large language models faces localization coupling, imprecise localization, and weak cross-layer interaction. The authors introduce Knowledge Neuronal Ensemble (KNE), which uses token-level gradient attribution to assemble a targeted set of neurons across layers and updates only this ensemble, enabling dynamic, coordinated edits with a reduced parameter footprint. Across three datasets and two models, KNE achieves higher edit accuracy while maintaining portability and locality, and shows that editing key FFN layers can yield strong results with distributed knowledge storage. The approach significantly reduces computational cost and supports batch editing, offering a practical path for scalable, precise knowledge updates in evolving LLMs.

Abstract

As real-world knowledge is constantly evolving, ensuring the timeliness and accuracy of a model's knowledge is crucial. This has made knowledge editing in large language models increasingly important. However, existing knowledge editing methods face several challenges, including parameter localization coupling, imprecise localization, and a lack of dynamic interaction across layers. In this paper, we propose a novel knowledge editing method called Knowledge Neuronal Ensemble (KNE). A knowledge neuronal ensemble represents a group of neurons encoding specific knowledge, thus mitigating the issue of frequent parameter modification caused by coupling in parameter localization. The KNE method enhances the precision and accuracy of parameter localization by computing gradient attribution scores for each parameter at each layer. During the editing process, only the gradients and losses associated with the knowledge neuronal ensemble are computed, with error backpropagation performed accordingly, ensuring dynamic interaction and collaborative updates among parameters. Experimental results on three widely used knowledge editing datasets show that the KNE method significantly improves the accuracy of knowledge editing and achieves, or even exceeds, the performance of the best baseline methods in portability and locality metrics.

Figures (5)

• Figure 1: The framework of KNE method.
• Figure 2: Visualization of Knowledge Storage Deviations in Large Language Models
• Figure B.3: Effects of Localized Knowledge versus Full Dataset Localization
• Figure B.4: Performance Metrics Across Varying Parameter Settings for Knowledge Editing
• Figure B.5: Performance Metrics Across Different Batch Sizes