HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention **Abstract:** Predicting the trajectories of road agents is crucial for autonomous driving systems. Current methods often use a static paradigm, predicting future trajectories based on a fixed duration of historical frames, which can lead to instability and temporal inconsistency. This paper introduces HPNet, a novel dynamic trajectory forecasting method that leverages both historical frames and historical predictions. Specifically, HPNet includes a Historical Prediction Attention module to encode the dynamic relationship between successive predictions, extending the attention range beyond the visible window. The method is evaluated on the Argoverse and INTERACTION datasets, demonstrating state-of-the-art performance in generating accurate and stable future trajectories. **Introduction:** Accurate trajectory prediction is essential for autonomous driving systems. Recent methods have achieved notable results by integrating heterogeneous information, but they often treat trajectory prediction as a static task, using a fixed number of historical frames. This approach can lead to instability and temporal inconsistency in successive predictions. HPNet addresses this by modeling the dynamic relationship between successive predictions, improving both stability and accuracy. **Method:** HPNet consists of three main components: Spatio-Temporal Context Encoding, Triple Factorized Attention, and Multimodal Output. It first aggregates agent and lane features with mode queries to generate prediction embeddings. Then, Triple Factorized Attention, comprising Agent Attention, Historical Prediction Attention, and Mode Attention, models interactions between agents, predictions, and modes. Finally, the prediction embeddings are decoded into multimodal future trajectories. **Experiments:** HPNet is evaluated on the Argoverse and INTERACTION datasets. Results show that HPNet achieves state-of-the-art performance, outperforming existing methods in terms of accuracy and stability. Ablation studies confirm the effectiveness of each component, particularly the Historical Prediction Attention, which enhances both accuracy and stability without increasing computational overhead. **Conclusion:** HPNet introduces a novel dynamic trajectory forecasting method that leverages historical predictions to improve the accuracy and stability of trajectory forecasts. Experiments demonstrate its superior performance on real-world datasets, validating the effectiveness of the proposed approach.HPNet: Dynamic Trajectory Forecasting with Historical Prediction Attention **Abstract:** Predicting the trajectories of road agents is crucial for autonomous driving systems. Current methods often use a static paradigm, predicting future trajectories based on a fixed duration of historical frames, which can lead to instability and temporal inconsistency. This paper introduces HPNet, a novel dynamic trajectory forecasting method that leverages both historical frames and historical predictions. Specifically, HPNet includes a Historical Prediction Attention module to encode the dynamic relationship between successive predictions, extending the attention range beyond the visible window. The method is evaluated on the Argoverse and INTERACTION datasets, demonstrating state-of-the-art performance in generating accurate and stable future trajectories. **Introduction:** Accurate trajectory prediction is essential for autonomous driving systems. Recent methods have achieved notable results by integrating heterogeneous information, but they often treat trajectory prediction as a static task, using a fixed number of historical frames. This approach can lead to instability and temporal inconsistency in successive predictions. HPNet addresses this by modeling the dynamic relationship between successive predictions, improving both stability and accuracy. **Method:** HPNet consists of three main components: Spatio-Temporal Context Encoding, Triple Factorized Attention, and Multimodal Output. It first aggregates agent and lane features with mode queries to generate prediction embeddings. Then, Triple Factorized Attention, comprising Agent Attention, Historical Prediction Attention, and Mode Attention, models interactions between agents, predictions, and modes. Finally, the prediction embeddings are decoded into multimodal future trajectories. **Experiments:** HPNet is evaluated on the Argoverse and INTERACTION datasets. Results show that HPNet achieves state-of-the-art performance, outperforming existing methods in terms of accuracy and stability. Ablation studies confirm the effectiveness of each component, particularly the Historical Prediction Attention, which enhances both accuracy and stability without increasing computational overhead. **Conclusion:** HPNet introduces a novel dynamic trajectory forecasting method that leverages historical predictions to improve the accuracy and stability of trajectory forecasts. Experiments demonstrate its superior performance on real-world datasets, validating the effectiveness of the proposed approach.