Knowledge-Defined and Twin-Assisted Network Management for 6G

The increasing complexity, dynamism, and heterogeneity of 6G networks demand management systems that can reason proactively and generalize beyond pre-defined cases. In this paper, we propose a modular, knowledge-defined architecture that integrates Digital Twin models with semantic reasoning and zero-shot learning to enable autonomous decision-making for previously unseen network scenarios. Real-time data streams are used to maintain synchronized virtual replicas of the physical network, which also forecast short-term state transitions. These predictions feed into a knowledge plane that constructs and updates a graph-based abstraction of the network, enabling context-aware intent generation via graph neural reasoning. To ensure adaptability without retraining, the management plane performs zero-shot policy matching by semantically embedding candidate intents and selecting suitable pre-learned actions. The selected decisions are translated and enforced through the control plane, while a closed-loop feedback mechanism continuously refines predictions, knowledge, and policies over time. Simulation results confirm that the proposed framework observes notable improvements in policy response time, SLA compliance rate, and intent matching accuracy.