Large or Small? Cloud or Edge? Investigating the Use of Language Models for Smart Drone Services

Due to the requirements of various drone applications and their complex operational environments, smart drone services often require powerful reasoning capabilities for autonomous decision making. While large language models (LLMs) have proven their strong reasoning capabilities, an open question remains is that whether we should use LLMs in the cloud, or adapt and deploy small language models (SLMs) at the edge, to support the special needs of smart drone services.

To answer this question, we create a set of structured drone reasoning tasks and adapt a small student model using teacher-generated supervision with QLoRA-based parameter-efficient fine-tuning. Our experiments are designed to evaluate model performance from three perspectives: accuracy, key constraint coverage, and generalization ability. The results show that task-oriented adaptation can improve the performance of edge SLMs for specialized drone reasoning tasks, but their performance on general reasoning tasks is less effective compared with cloud LLMs.