Microgrids play a vital role in integrating distributed renewable energy and balancing local loads, but optimal load allocation remains challenging due to the stochastic nature of renewable generation and the complexity of system operation. The rise of large language model (LLM) offers new possibilities for addressing these challenges. This paper reviews the current research on applying LLM to microgrid load optimization. It first introduces the Transformer architecture and major training paradigms of LLM, followed by a review of conventional microgrid optimization techniques. Building on the standard structure of energy management systems, load allocation is discussed through four key components: generation scheduling, load forecasting, energy storage management, and demand response. Recent studies show that LLM can achieve accurate short-term load forecasting through prompt design and sequence generation, provide dispatch suggestions comparable to human experts, and demonstrate potential advantages in battery life prediction and user behavior modeling. The paper further analyzes major technical challenges—interpretability, multimodal fusion, real-time performance, and data privacy—and summarizes corresponding improvement strategies. Finally, future directions are outlined, including domain-specific energy LLM, physics-guided reasoning, benchmark development, broader energy applications, and the societal impact of AI adoption. Overall, integrating LLM into microgrid optimization introduces a new research paradigm that could significantly enhance the intelligence, efficiency, and reliability of future power systems.