Abstract: Large Language Models (LLMs) have demonstrated strong capabilities in semantic understanding and complex rea-soning, showing great potential for multi-UAV cooperative task planning. However, existing LLM-driven planning ap-proaches still face significant challenges, including constraint violations, semantic drift during long-horizon execution, and limited adaptability to dynamic environments due to insufficient structured planning mechanisms and constrained context capacity. To address these issues, this paper proposes an LLM-driven collaborative task planning framework for multi-UAV systems based on a multi-agent architecture. The framework integrates task decision, situational awareness analysis, and UAV-level execution into a unified cooperative process. High-level semantic instructions are transformed into graph-structured atomic task representations, enabling a structured mapping from abstract intentions to executable actions. A context-aware closed-loop mechanism is introduced to continuously monitor environmental changes and support incremental task adjustment under unexpected events, thereby maintaining logical consistency and improving dynamic responsiveness during long-horizon missions. Furthermore, a two-stage alignment strategy combining supervised fine-tuning and reinforcement learning is adopted to enhance the reasoning capability and deci-sion reliability of the LLM in multi-UAV task scenarios. Simulation results demonstrate that, compared with the classi-cal LLM-based planning framework ReAct, the proposed method improves task success rate by approximately 30% and achieves a 95.2% tool invocation success rate, effectively realizing a complete closed-loop process from semantic task input to autonomous multi-UAV cooperative execution.

Key words: large language models, multi-agent system, task planning, UAV swarm, model fine-tuning