This work evaluates the capabilities of Large Language Models (LLMs) in semantic reasoning tasks. We construct a knowledge graph that represents a real-world Internet of Things (IoT) environment and define various reasoning rules for device identification use cases. We test the performance of three LLMs: Llama-3.1, Qwen-2.5, GPT-4o, based on different levels of fine-grained rule descriptions in the input context and the increasing rule complexity. We use a rule-based reasoner, Apache Jena, to generate the ground truth data for each reasoning rule. The results show that while LLMs are effective in retrieving direct links in the knowledge graph, they struggle with highly customized, complex reasoning rules involving multiple triple patterns. This work provides an application and a baseline for using LLMs to enable semantic reasoning for device identification in IoT environments.