Addressing the issues of high randomness in seed generation and low utilization of feedback in traditional fuzzing methods for IoT protocols, this paper proposes an enhanced fuzzing framework integrating Large Language Models (LLMs) and Reinforcement Learning (RL). Firstly, the LLM is employed to parse protocol specifications and generate structured initial seeds. Subsequently, a reward function is designed through RL mechanisms, leveraging execution state feedback of seeds to achieve adaptive optimization of mutation strategies. Experimental results indicate that the proposed fuzzing system demonstrates superior performance in IoT protocol testing, significantly improving vulnerability detection efficiency and coverage rate.