Abstract: Ship trajectory prediction and behavior recognition can help effectively assess navigational risks and provide an important basis for decision-making in collision avoidance and traffic management. To improve the accuracy of ship trajectory prediction and behavior recognition, this paper studies a multi-task Informer model for simultaneous trajectory prediction and behavior recognition. Based on the Informer framework, the model incorporates a multi-task learning approach. It addresses the issue that inaccurate ship behavior records in AIS data cannot be directly used as model inputs by designing a multi-task loss function that jointly trains behavior recognition and trajectory prediction in parallel. During training, an adaptive updating strategy for the loss function-based on homoscedastic uncertainty-is designed to automatically allocate weights to the losses of the two tasks. Evaluated using real AIS data from the Taicang sector waters, the multi-task Informer model reduces trajectory prediction loss by 40. 2% and 14. 7% compared to LSTM and Informer models, respectively. In behavior recognition, the multi-task model improves accuracy by 11. 7% and 5. 95% compared to LSTM and Informer models, respectively. The results demonstrate that the multi-task model effectively enhances the performance of ship trajectory prediction while achieving accurate recognition of ship behavior.