The integration of Large Language Models (LLMs) into recommender systems (RS) is rapidly evolving, with a notable shift towards enhancing system capabilities without increasing inference costs. Recent advancements focus on leveraging LLMs for knowledge enhancement, interaction enhancement, and model enhancement, addressing challenges such as scalability, cold-start scenarios, and personalized user behavior prediction. Notably, LLMs are being used to bridge user-side knowledge gaps in knowledge-aware recommendations and to provide intuitive exploration strategies for knowledge graphs, improving recommendation performance in cold-start situations. Additionally, the use of LLM-enhanced logs and personalized prompts is showing promise in predicting user behavior in smart spaces, enhancing the intelligence of systems like smart homes and vehicles. The field is also exploring novel frameworks that harmonize traditional recommendation models with LLMs, achieving semantic convergence through two-stage alignment and behavioral semantic tokenization. These developments collectively push the boundaries of recommender systems, offering more nuanced, scalable, and personalized recommendations.