AI-Native O-RAN Architectures for 6G: Towards Real-Time Adaptation, Conflict Resolution, and Efficient Resource Management

The emergence of Open Radio Access Network (O-RAN) introduces a modular and intelligent approach to orchestrating wireless networks, enabling AI-driven control through open interfaces and disaggregated architectures. However, as sixth-generation (6G) systems demand more responsive, autonomous, and intent-aligned orchestration, critical challenges emerge in ensuring performance, consistency, and adaptability across distributed control agents. This paper proposes a novel AI-native orchestration framework that integrates four original advancements in next-generation RAN control: (i) dynamic management of real-time decision conflicts among distributed xApps and rApps; (ii) low-latency, context-coherent inference through a data locality-aware design; (iii) semantic translation of high-level operator intent into actionable policies using Large Language Models (LLMs); and (iv) closed-loop integration of Reconfigurable Intelligent Surfaces (RISs) for adaptive radio environment optimization. The framework is structured around a Deep Reinforcement Learning-powered Conflict Management Engine (CME) within the Near-RT RIC, a Context Sharing Layer (CSL) for synchronized multi-source telemetry, an LLMbased orchestration module in the Non-RT RIC, and an RIS control xApp leveraging real-time key performance indicators. Collectively, these innovations form an integrated, intelligent control plane that advances O-RAN toward the 6G vision of realtime adaptability, conflict-resilient coordination, latency-sensitive optimization, and intent-aware automation.