Generative Foundation Models for Autonomous 6G Edge Intelligence

Abstract The global transition toward decentralized industrial automation in 6G networks demands a complete redesign of wireless architecture to ensure ultra-reliable and low-latency communications. However, traditional optimization approaches are not effective in addressing the stochastic characteristics of 6G networks, where sensing, computing, and communication must be integrated. In this study, a novel paradigm for generative foundation models in autonomous 6G edge intelligence is proposed. This paradigm shifts beyond traditional task-oriented reinforcement learning, establishing a universal control plane capable of managing heterogeneous network devices. The proposed architecture assimilates physics-informed models and generative pre-training to ensure that digital twins are synchronized with hardware states in reconfigurable intelligent surface (RIS)-assisted networks. Moreover, multi-objective optimization techniques are incorporated to ensure a fair trade-off between the rate-energy Pareto frontier in space-air-ground integrated networks. The results show that the proposed generative foundation model improves zero-shot adaptability by 45% over unseen network topologies and reduces system overhead by 30% using semantic data compression. The proposed paradigm also offers better energy-neutral performance compared to traditional multi-agent approaches. The proposed paradigm offers a promising pathway toward achieving autonomous intelligence at the network edge, closing the gap between 6G theory and industrial practice.