<p>With the evolution of 5G and the emerging 6G networks, Virtual Network Embedding (VNE) plays a critical role in enabling network slicing, ultra-low latency, and dynamic, multi-tenant resource sharing over a common physical infrastructure. However, the VNE problem mapping virtual nodes and links to substrate network resources is NP-hard, and conventional approaches rely on a sequential strategy that first maps nodes and then connects links. This disjoint process often results in inefficient embeddings, high rejection rates, and suboptimal Quality of Service (QoS), especially under dynamic and high-demand 5G/6G scenarios. To overcome these challenges, we propose a novel Graph attention network-Proximal Policy Optimization (PPO)-based adaptive VNE framework with joint node-link mapping. Unlike traditional methods, our approach integrates GAT to model complex spatial and topological correlations, and PPO for intelligent, policy-driven decision-making in dynamic environments. The key innovation lies in early-stage joint embedding, where link mapping is initiated immediately after placing a minimal subset of nodes (e.g., two or three), ensuring that feasibility and resource constraints are evaluated holistically from the outset. This joint node-link mapping mechanism is particularly well-suited for 5G/6G network slicing environments, where heterogeneous service requirements (e.g., eMBB, URLLC, and mMTC) demand flexible, efficient, and adaptive embedding strategies. Our framework dynamically adapts to network load variations and topological changes, thereby improving resource utilization, service isolation, latency control, and the acceptance ratio of VNRs. Experimental evaluations on realistic topologies demonstrate significant performance gains over state-of-the-art baselines in terms of embedding cost, long-term revenue, and scalability highlighting the framework potential to support the high-performance demands of next-generation mobile networks.</p>