Abstract: The success of large language models (LLMs) depends not only on their vast scale but, more crucially, on alignment techniques that conform their behavior to human expectations. Reinforcement learning from human feedback (RLHF) is the core paradigm for achieving this alignment. This article reviews the developmental trajectory of reinforcement learning techniques for LLMs: First, it analyzes the challenges faced by traditional RLHF methods, represented by the proximal policy optimization (PPO) algorithm, such as high complexity and significant computational overhead; subsequently, it analyzes the challenges such as high complexity and large computational overhead faced by traditional RLHF methods represented by the PPO algorithm, and discusses innovations in reinforcement learning algorithms tailored to the characteristics of large language models, which successfully enhance training efficiency significantly while preserving the advantages of online learning, thereby ushering in a new wave of LLM-specific reinforcement learning; finally, it looks forward to how directions such as reinforcement learning based on artificial intelligence feedback can drive models to achieve self-improvement, and describes the future trends of symbiotic evolution and mutual promotion between reinforcement learning and large models.