The paper introduces a novel method called GLAME (Knowledge Graph Augmented Model Editing) for enhancing large language models (LLMs) through knowledge graph integration. GLAME addresses the limitations of existing model editing methods by effectively capturing and incorporating changes in associated knowledge due to edits. The method consists of two main components: Knowledge Graph Augmentation (KGA) and Graph-based Knowledge Edit (GKE). 1. **Knowledge Graph Augmentation (KGA)**: This component uses external knowledge graphs to construct a subgraph that captures the altered knowledge associated with the edited knowledge. The LLM is then used to extract hidden representations of entities and relations within this subgraph, which serve as initial representations for the subgraph. 2. **Graph-based Knowledge Edit (GKE)**: This component integrates the new knowledge associations from the subgraph into the LLM's parameter editing process using a relational graph neural network (RGNN). The RGNN performs message propagation and aggregation operations on the subgraph to enhance the subject's representation with the newly constructed associated knowledge. The updated parameters are then used to edit the LLM's knowledge. The paper evaluates GLAME on two datasets, COUNTERFACT and COUNTERFACTPLUS, and compares it with state-of-the-art editing methods such as ROME, MEMIT, and MEND. Experimental results show that GLAME significantly improves the generalization capabilities of post-edit LLMs in utilizing edited knowledge, outperforming other methods in various evaluation metrics. The paper also discusses the limitations of GLAME, such as the reliance on the availability and quality of relevant knowledge graphs, and suggests future directions for improving subgraph quality and sampling strategies. Ethical considerations are addressed, emphasizing the importance of responsible development and use of generative LLMs.The paper introduces a novel method called GLAME (Knowledge Graph Augmented Model Editing) for enhancing large language models (LLMs) through knowledge graph integration. GLAME addresses the limitations of existing model editing methods by effectively capturing and incorporating changes in associated knowledge due to edits. The method consists of two main components: Knowledge Graph Augmentation (KGA) and Graph-based Knowledge Edit (GKE). 1. **Knowledge Graph Augmentation (KGA)**: This component uses external knowledge graphs to construct a subgraph that captures the altered knowledge associated with the edited knowledge. The LLM is then used to extract hidden representations of entities and relations within this subgraph, which serve as initial representations for the subgraph. 2. **Graph-based Knowledge Edit (GKE)**: This component integrates the new knowledge associations from the subgraph into the LLM's parameter editing process using a relational graph neural network (RGNN). The RGNN performs message propagation and aggregation operations on the subgraph to enhance the subject's representation with the newly constructed associated knowledge. The updated parameters are then used to edit the LLM's knowledge. The paper evaluates GLAME on two datasets, COUNTERFACT and COUNTERFACTPLUS, and compares it with state-of-the-art editing methods such as ROME, MEMIT, and MEND. Experimental results show that GLAME significantly improves the generalization capabilities of post-edit LLMs in utilizing edited knowledge, outperforming other methods in various evaluation metrics. The paper also discusses the limitations of GLAME, such as the reliance on the availability and quality of relevant knowledge graphs, and suggests future directions for improving subgraph quality and sampling strategies. Ethical considerations are addressed, emphasizing the importance of responsible development and use of generative LLMs.