Unmanned Aerial Vehicles (UAVs) are widely applied for processing collaborative tasks due to their terrain-crossing capabilities, low infrastructure needs, and growing intelligence. To address the challenges posed by the delay sensitivity of randomly arriving collaborative tasks in unpredictable environments, dynamic task allocation among multi-UAVs has become a pivotal research focus. This paper addresses the decentralized dynamic task allocation problem, aiming to enhance collaboration among UAVs, improve system robustness, and eliminate reliance on a central UAV. Traditional decentralized methods often yield suboptimal results due to limited local observation, while centralized approaches impose high communication burdens and lack scalability. To overcome these challenges, we propose a Multi-Agent Deep Reinforcement Learning (MADRL) algorithm based on neighborhood cognition, which replaces global knowledge with localized information exchange among neighbors to develop robust and long-term decentralized task allocation strategies. By integrating a reward prediction module, our method redistributes delayed task rewards to stabilize policy iteration. Moreover, we incorporate Lyapunov optimization to ensure system queue stability while minimizing energy cost and task failure rates. Experiments demonstrate that our decentralized approach outperforms baseline methods in terms of energy cost and task failure rates, while showcasing excellent scalability across systems ranging from 10 to 100 UAVs.