SysUM: NDT-Assisted System-Level Scenario-Adaptive Universal Model for AI-6G Networks

Abstract

Existing works on artificial intelligence (AI)-empowered wireless communication primarily focus on monolithic usage scenarios and isolated tasks with limited labeled data or capacity-driven modular optimization, which become inadequate and unscalable with the proliferation of task classes and heterogeneous requirements envisioned for future 6G systems. Although several universal models have been developed, they fail to address system-optimal performance in multi-cell deployments. Therefore, in this letter, we propose a system-level scenario-adaptive universal model termed SysUM for flexible and real-time multi-task decision-making across diverse objectives, trained via end-to-end reinforcement learning. The emergence of network digital twin (NDT) facilitates efficient channel acquisition and realistic system evaluation. Specifically, multi-cell user equipment (UE) channels are constructed to capture mutual coupling relationships. A masking mechanism is introduced to emulate UE or service changes. Three learnable embeddings are incorporated for knowledge injection. Based on the universal representations extracted by the backbone network, an actor-critic network is employed for policy sampling and value estimation. These decisions are then configured on the NDT system-level platform to obtain final rewards. Experimental results demonstrate that the proposed SysUM can adapt to unseen UE counts, extract scenario-related features, and outperform AI/non-AI benchmarks in exemplary subcarrier scheduling, power control, and link adaptation tasks.