Towards transparent 6G AI-RAN: A survey on explainable deep reinforcement learning for intelligent network slicing

The advent of the sixth generation (6G) wireless networks envisions an artificial intelligence (AI)-native radio access network (AI-RAN), where deep reinforcement learning (DRL) emerges as a key enabler for intelligent and autonomous network slicing. Despite the demonstrated performance gains of DRL-based solutions in dynamic resource allocation and slice orchestration, their opaque decision-making nature raises critical concerns regarding trust, accountability, and operational deployment. To bridge this gap, explainable deep reinforcement learning (XDRL) has recently attracted significant attention as a means to enhance transparency, interpretability, and controllability of AI-RAN slicing policies. This survey provides a comprehensive overview of the state of the art in explainable DRL for intelligent network slicing. We first review the fundamental principles of DRL in the context of RAN slicing and identify the unique explainability challenges posed by high-dimensional, multi-slice environments. We then categorize existing XDRL approaches into post-hoc explanation, symbolic abstraction, and human-in-the-loop steering, analyzing their methodologies, strengths, and limitations. Furthermore, we highlight benchmark environments and experimental testbeds that have been employed to evaluate XDRL in realistic network scenarios. Finally, we outline key open challenges, including scalability, generalization across traffic patterns, integration with large language models (LLMs), and alignment with intent-based networking, and discuss promising research directions toward achieving transparent, trustworthy, and human-centric AI-RAN in 6G. • This paper provides a comprehensive survey of explainable deep reinforcement learning (XDRL) techniques for intelligent network slicing in 6G AI-driven Radio Access Networks (AI-RAN). • It identifies key challenges and outlines promising research directions to achieve transparent, trustworthy, and human-centric AI-RAN systems for future 6G networks.