Large language models (LLMs) have shown impressive abilities in various tasks, but how well can they truly reason, especially when it comes to structured knowledge? A new research paper explores this question by examining how LLMs perform on knowledge graph question answering (KGQA). Knowledge graphs organize information as interconnected entities and relationships, offering a rich source for complex reasoning tasks. Traditional KGQA methods often struggled with retrieving relevant subgraphs and making accurate inferences. This new research introduces a novel approach called LLM-based Discriminative Reasoning (LDR), which breaks down the KGQA process into three key steps: subgraph searching, subgraph pruning, and answer inference. Instead of generating entire reasoning paths like some previous methods, LDR employs a discriminative strategy, having the LLM choose the best next step at each stage based on a set of options. This approach helps to avoid hallucinations or illogical jumps in reasoning. In experiments on popular benchmarks like WebQSP and CWQ, LDR outperformed existing state-of-the-art methods. LDR was more accurate in finding the right subgraphs and less sensitive to the size of the retrieved information, demonstrating a more focused and efficient reasoning process. This research sheds light on the potential for LLMs to perform complex reasoning on structured knowledge, highlighting the importance of designing appropriate interaction mechanisms to guide their decision-making. While LDR shows promising results, the researchers identify future directions such as scaling the approach to even larger LLMs and developing more fine-grained explanations for the reasoning steps. These advancements will be crucial for building more robust and transparent AI systems capable of true knowledge-based reasoning.