This paper presents an extension to a decentralized multi-UAV task allocation algorithm, consensus-based bundle algorithm (CBBA), to improve the communication efficiency in the plan consensus process. The presented algorithm termed grouped consensus-based bundle algorithm (G-CBBA) provides a systematic way of grouping the UAVs based on their task preference represented by the initial guess created by the UAVs. G-CBBA features a nested iteration between two layers of plan consensus: local and global. The local stage aims at plan consensus over the list of tasks shared within a group, while the global stage ensures agreement across the entire network over the whole task set. The proposed two-layer scheme can improve the communication efficiency by reducing the need for propagating irrelevant bids while preserving robust convergence property of CBBA. Numerical experiments demonstrate reduction in the number of messages to be communicated without incurring substantial performance degradation compared to the baseline CBBA.