Large Models Empowered End-to-End AI for Split Learning in Native AI Future Networks

The emergence of sixth-generation mobile communication networks (6G) marks a paradigm shift toward native artificial intelligence (AI) architectures, where AI is deeply embedded in the network infrastructure. At the core of this transformation is End-to-End (E2E) AI, which enables autonomous monitoring, orchestration, and reasoning across distributed AI workloads. A key enabler of scalable E2E AI is split learning, where deep neural networks are partitioned across edge and cloud nodes to meet the performance and efficiency demands of 6G applications. However, the reliability of split learning is particularly vulnerable to performance degradation arising from communication delays, resource dynamics, and model drift. These issues pose significant challenges that traditional monitoring and maintenance techniques fail to address effectively. To tackle these challenges, this paper proposes a Large Models Empowered Proactive Maintenance Framework tailored for split learning in native AI future networks. The framework integrates Large Forecasting Models (LFMs) for continuous segment-level performance monitoring and Large Language Models (LLMs) for root cause reasoning and adaptive recovery planning. By aligning with the principles of E2E AI, the proposed system enables anticipatory anomaly detection, interpretable diagnosis, and automated corrective actions, thus enhancing the resilience, scalability, and autonomy of AI-native 6G networks. Experimental results demonstrate the framework's ability to maintain stable split model performance under dynamic system conditions, offering a practical path toward intelligent and self-sustaining AI service delivery in next-generation networks.