Abstract: To address the problems of mode collapse and error accumulation in long-horizon multimodal vehicle trajectory prediction, a trajectory prediction framework integrating temporal decoupling and dynamic refinement was proposed. First, the trajectory proposal network was constructed, adopting a temporal decoupling strategy to divide the prediction horizon into consecutive segments and then utilizing a segment-wise interactive attention mechanism to model both local and global spatiotemporal dependencies, thereby generating initial trajectory proposals. Subsequently, the dynamic refinement network recursively fused scene context with these proposals, progressively correcting their geometric deviations to achieve fine-grained trajectory generation. Experimental validation on large-scale real-world driving dataset Argoverse 2 demonstrated that the proposed framework achieved a minimum average displacement error (minADE) of 0.65 m, a minimum final displacement error (minFDE) of 1.30 m, a miss rate (MR) of 0.17, and an average inference time of 93 ms in multimodal prediction tasks. Compared with various mainstream methods, this framework demonstrates a comprehensive advantage in both prediction accuracy and computational efficiency, providing an effective solution for long-horizon trajectory prediction in complex and dynamic scenarios.