With the popularity of GPS-enable smart devices and the development of wireless network, Spatial Crowdsourcing (SC), as a framework for assigning location-sensitive tasks to moving workers, has received wide attention in recent years. In real-world scenarios, some complex tasks exist that may not be completed by a single worker. In this case, the tasks are often assigned to multiple workers, which is called group task assignment. However, the assignment of tasks that satisfy all group members in an even way remains a challenge. To this end, we propose a novel preference-aware group task assignment framework that includes two components: Mutual Information-based Preference Modeling (MIPM) and Preference-aware Group Task Assignment (PGTA). Specifically, MIPM learns the preferences of worker groups by maximizing the mutual information among workers based on the worker-task interaction data and the group-task interaction data, where an attention mechanism is used. PGTA adopts an optimal task assignment algorithm based on tree decomposition to assign tasks to appropriate worker groups, which aims to maximize the overall number of assigned tasks while giving priority to the groups of workers that are more interested in the tasks. Finally, extensive experiments are conducted, verifying the effectiveness and practicality of the proposed solutions.