GraphEdit is a novel approach that leverages large language models (LLMs) to learn complex node relationships in graph-structured data. The method addresses the limitations of existing graph structure learning (GSL) methods that rely on explicit graph structures as supervision signals, which can be affected by data noise and sparsity. By enhancing the reasoning capabilities of LLMs through instruction-tuning over graph structures, GraphEdit aims to improve the reliability of graph structure learning. The model effectively denoises noisy connections and identifies node-wise dependencies from a global perspective, providing a comprehensive understanding of the graph structure. GraphEdit uses a lightweight edge predictor to identify candidate edges and then incorporates the reasoning capabilities of LLMs to refine the graph structure. The model is evaluated on multiple benchmark datasets, including Cora, PubMed, and Citeseer, demonstrating its effectiveness and robustness across various settings. The results show that GraphEdit outperforms existing GSL methods in terms of accuracy and robustness, particularly in noisy environments. The model's ability to capture implicit global dependencies and denoise connections leads to improved graph representations, which enhance the performance of downstream tasks such as node classification. The study also highlights the importance of incorporating both edge deletion and addition strategies, along with the reasoning capabilities of LLMs, to optimize the original graph structures. GraphEdit's performance is further validated through ablation studies and robustness analysis, showing that it can effectively handle noisy and incomplete data. The model's ability to restructure graphs in a way that facilitates node classification is demonstrated through visual analysis and case studies. Overall, GraphEdit provides a promising solution for graph structure learning by leveraging the reasoning capabilities of LLMs to enhance the quality and reliability of graph representations.GraphEdit is a novel approach that leverages large language models (LLMs) to learn complex node relationships in graph-structured data. The method addresses the limitations of existing graph structure learning (GSL) methods that rely on explicit graph structures as supervision signals, which can be affected by data noise and sparsity. By enhancing the reasoning capabilities of LLMs through instruction-tuning over graph structures, GraphEdit aims to improve the reliability of graph structure learning. The model effectively denoises noisy connections and identifies node-wise dependencies from a global perspective, providing a comprehensive understanding of the graph structure. GraphEdit uses a lightweight edge predictor to identify candidate edges and then incorporates the reasoning capabilities of LLMs to refine the graph structure. The model is evaluated on multiple benchmark datasets, including Cora, PubMed, and Citeseer, demonstrating its effectiveness and robustness across various settings. The results show that GraphEdit outperforms existing GSL methods in terms of accuracy and robustness, particularly in noisy environments. The model's ability to capture implicit global dependencies and denoise connections leads to improved graph representations, which enhance the performance of downstream tasks such as node classification. The study also highlights the importance of incorporating both edge deletion and addition strategies, along with the reasoning capabilities of LLMs, to optimize the original graph structures. GraphEdit's performance is further validated through ablation studies and robustness analysis, showing that it can effectively handle noisy and incomplete data. The model's ability to restructure graphs in a way that facilitates node classification is demonstrated through visual analysis and case studies. Overall, GraphEdit provides a promising solution for graph structure learning by leveraging the reasoning capabilities of LLMs to enhance the quality and reliability of graph representations.