Large language models (LLMs) have demonstrated impressive capabilities in various natural language processing tasks, yet they often struggle with deep and reliable reasoning, particularly when external knowledge is required. To address this limitation, recent approaches have integrated knowledge graphs (KGs) with LLMs to enhance reasoning accuracy and mitigate hallucinations. In this paper, we propose Global Relation Pruning (GRP), a novel method for knowledge graph reasoning that eliminates the traditional entity pruning step and instead treats entities connected by each relation as unified entity sets. This approach prevents premature elimination of potentially useful entities, thereby expanding the search breadth while reducing redundant computations. GRP further aggregates relations across entity sets and performs global relation pruning, significantly improving reasoning efficiency without compromising depth. Experimental results on the WebQSP and CWQ datasets demonstrate the superiority of GRP, achieving Hit@1 scores of 86.0 and 70.6, respectively, outperforming existing state-of-the-art methods. Our findings highlight the effectiveness of GRP in enhancing multi-hop reasoning and its potential for scalable integration of KGs with LLMs.The code is publicly available at https://github.com/baiyi111/GRP .