Pest detection in agriculture is critical for protecting farmland, and You Only Look Once (YOLO) has been widely adopted due to its efficiency and accuracy. However, despite the proliferation of YOLO-related research, the field remains highly fragmented. This fragmentation poses significant challenges for researchers in pest detection when selecting models. Notably, to the best of our knowledge, there has been no study that thoroughly investigates the performance of mainstream YOLO algorithms on pest datasets. To bridge this gap, this study conducts a systematic performance evaluation of multiple YOLO variants across diverse datasets and assesses the effectiveness of mainstream improvement in algorithms. Experimental results demonstrate that different YOLO series exhibit distinct characteristics on pest datasets, with some prioritizing accuracy while others emphasize speed. Among these models, YOLOv12 demonstrates superior accuracy while maintaining competitive real-time performance.