Emerging sixth-generation (6G) solutions demand learning agents that are not only autonomous and efficient, but also robust in dynamic environments and transparent in decision making. However, conventional deep reinforcement learning (DRL) approaches for mobile networks lack sufficient explainability and suffer catastrophic forgetting under dynamic conditions. In this paper, we propose a Bayesian reasoning-based explainable deep active inference (BRAIN) model, the embodied AI-inspired approach applied to mobile networks. BRAIN employs a deep generative model and minimizes variational free energy to unify perception (Bayesian state estimation) and action (resource allocation) in a single framework. Unlike DRL baselines, our agent inherently eliminates catastrophic forgetting through continuous belief updates without retraining and provides built-in explainability by exposing posterior beliefs and free energy components at runtime. Deployed as an eXtended application (xApp) on an Open, GPU-accelerated, AI-RAN testbed; BRAIN demonstrates (i) causal reasoning for slice-specific quality of service (QoS) adherence (throughput/latency/reliability), (ii) 28.3% higher robustness to traffic shifts versus DRL baselines, and (iii) real-time interpretability of resource-allocation decisions via operator-accessible belief states.

Traditional rule-based or static management approaches struggle to cope with the dynamic, multi-layered nature of 5G/6G networks, creating a strong motivation for AI-native solutions – management systems built from the ground up with artificial intelligence – to enable autonomous, real-time network control. In this work, we introduce GNN-ATIVE, an AI-native orchestration framework that leverages Graph Neural Networks (GNNs) and knowledge graphs (KGs) in a unified graph-based paradigm for network management. GNN-ATIVE uses a semantic knowledge graph to represent the network’s state and context, employing standard ontologies to ensure consistency and interoperability. Building on this foundation, we design Knowledge Graph enabled Generative Pretrained Transformer (KG-GPT), a novel graph-to-graph Transformer model that performs knowledge-driven reasoning on the KG. KG ingests the structured network state (nodes, links, and attributes) and infers optimal configurations or management actions, serving as a high-level decision engine for the orchestrator. We implement and evaluate GNN-ATIVE on an Optical Transport Network (OTN) testbed using real network components. The results demonstrate that GNN-ATIVE can effectively manage OTN resources and adapt to network changes while achieving low-latency inference for decision making.

Network management systems for beyond 5G (B5G) and 6G networks today require efficient approaches for handling increased heterogeneity, network-function dis-aggregation, performance requirements, and optimizing networks to support highly diverse use cases. While the Open-Radio Access Network (O-RAN) Service Management and Orchestration (SMO) frameworks efficiently manage RAN and cloud infrastructure, the fragmentation of management platforms across RAN, Core Network (CN), and Transport Network (TN) introduces operational inefficiencies, particularly in Non-Public Networks (NPNs) deployments. This paper proposes a novel extension to the O-RAN SMO architecture, integrating CN and TN management functionalities into a unified control framework. By exploiting AI/ML-driven$\mathrm{x} / \text{rApps}$and a converged data analytics pipeline, the proposed architecture enhances SMO's fault management, resource optimization, and service continuity capabilities. Our implementation validates the feasibility of the proposed SMO extension, demonstrating subscriber-specific QoS assurance through O-RAN-based mobility management mechanisms. The proposed approach successfully shows how RAN-/Core-converged SMOs enable significant enhancements to O-RAN's x/rApps, allowing for subscriber-specific as well as application-specific differentiated QoS assurance.