Personalized exercise recommendation is crucial for effective adaptive learning systems. However, existing data-driven approaches are often limited by their reliance on extensive student-exercise interaction data, which is frequently unavailable or expensive to acquire. To address this challenge, we propose a zero-shot exercise recommendation framework powered by Large Language Models (LLMs). Inspired by human educators, our three-stage framework utilizes LLMs to predict knowledge states, identify weak concepts, and recommend relevant exercises without requiring domain-specific training. Experiments on a real-world educational dataset show that our zero-shot approach achieves comparable performance to data-driven methods, highlighting its potential for practical and scalable adaptive learning, particularly in data-limited scenarios.