The field of large language models is witnessing significant advancements in the integration of retrieval-augmented generation (RAG) and reinforcement learning (RL). Researchers are actively exploring ways to unify these two paradigms to improve knowledge grounding and complex reasoning abilities. A key challenge in this area is the development of frameworks that can dynamically coordinate between retrieval and reasoning, enabling adaptability across diverse tasks. Recent innovations have introduced novel approaches, such as difficulty-aware curriculum training and hybrid knowledge access strategies, to address this challenge. These advancements have led to significant performance improvements in various benchmarks, including open-domain question answering and multimodal question answering. Noteworthy papers in this area include: UR$^2$, which proposes a general framework that unifies retrieval and reasoning through reinforcement learning, achieving state-of-the-art results on several benchmarks. REX-RAG, which introduces a novel framework that explores alternative reasoning paths while maintaining rigorous policy learning, demonstrating competitive results across multiple datasets. Part I: Tricks or Traps, which provides a systematic review of RL techniques for LLM reasoning, offering clear guidelines for selecting techniques and a reliable roadmap for practitioners. A Curriculum Learning Approach to Reinforcement Learning, which describes a comprehensive retrieval-augmented generation system for multimodal question answering, leveraging curriculum learning strategies to guide reinforcement learning and achieving top results in a challenge.