As AI increasingly migrates to edge environments characterized by decentralization, limited resources, and real-time demands, the need for autonomous governance mechanisms has become paramount. This study introduces Generative Policy Models (GPMs), a novel class of transformer-based generative reinforcement learning frameworks designed for self-evolving policy generation in edge AI settings. By synthesizing policies without reliance on labeled data or central supervision, GPMs enable autonomous swarms, adaptive IoT networks, and mission-critical edge systems to operate efficiently and intelligently. Furthermore, three simulated environments, UAV swarms, smart traffic control, and IoT resource allocation, were used to evaluate GPM performance. Results demonstrate that GPMs surpass traditional RL baselines in decision latency, adaptability, and policy novelty, confirming their suitability for real-world decentralized systems. This work fills a critical gap in the literature by merging generative AI with edge autonomy and paves the way for resilient, explainable, and self-governing AI infrastructures.