The paper introduces a novel framework called HIMap (HybrId Representation Learning for End-to-end Vectorized HD Map Construction) to improve the accuracy and completeness of vectorized High-Definition (HD) map construction. Traditional methods often rely on point-level representation learning, which can lead to limitations in capturing element-level information and handling element-level failures. HIMap addresses these issues by introducing a hybrid representation called HIQuery, which integrates both point-level and element-level information. The framework includes a point-element interactor to extract and encode hybrid information, and a point-element consistency constraint to enhance the consistency between point-level and element-level information. Extensive experiments on the nuScenes and Argoverse2 datasets demonstrate that HIMap outperforms previous state-of-the-art methods, achieving 77.8 mAP on the nuScenes dataset, an improvement of at least 8.3 mAP over the best previous method. The paper also discusses the limitations and future directions, including model acceleration, 3D HD map construction, and temporal consistency in HD maps.The paper introduces a novel framework called HIMap (HybrId Representation Learning for End-to-end Vectorized HD Map Construction) to improve the accuracy and completeness of vectorized High-Definition (HD) map construction. Traditional methods often rely on point-level representation learning, which can lead to limitations in capturing element-level information and handling element-level failures. HIMap addresses these issues by introducing a hybrid representation called HIQuery, which integrates both point-level and element-level information. The framework includes a point-element interactor to extract and encode hybrid information, and a point-element consistency constraint to enhance the consistency between point-level and element-level information. Extensive experiments on the nuScenes and Argoverse2 datasets demonstrate that HIMap outperforms previous state-of-the-art methods, achieving 77.8 mAP on the nuScenes dataset, an improvement of at least 8.3 mAP over the best previous method. The paper also discusses the limitations and future directions, including model acceleration, 3D HD map construction, and temporal consistency in HD maps.