Integrating Large Language Models with Reinforcement Learning: A Survey of LLM-RL Synergistic Recommendation

The interactive and dynamic nature of recommendation systems makes reinforcement learning (RL) a dominant modeling paradigm. Unlike static models, RL offers superior modeling fidelity by capturing sequential user interactions and directly optimizing long-term objectives and discrete metrics. However, RLbased recommendation systems face challenges regarding limited state representation, difficulty in action modeling, non-trivial reward design, sparse and delayed feedback, and unreliable user simulation. The emergence of large language models (LLMs) offers a solution to these bottlenecks. With their extensive world knowledge, robust reasoning capabilities, and superior semantic understanding, LLMs not only equip RL agents with enhanced state understanding and decision-making intelligence but also serve as high-fidelity environment simulators for interactive training. This integration fosters a new paradigm of LLM-RL synergistic recommendation, leading to systems that are more adaptive, robust, trustworthy, and interpretable. To this end, this paper presents a comprehensive survey of this emerging field. We propose a taxonomy based on the functional roles of LLMs within the RL pipeline, systematically categorizing existing methodologies into five distinct paradigms: LLM as policy, reasoner, representer, explainer, and simulator. Furthermore, we summarize relevant evaluation protocols, delve into key opportunities, and analyze critical challenges to provide guidelines for future research directions.