o-ran architecture

To address the lack of intelligent, adaptive, and automated orchestration for end-to-end edge AI services in 6G Open Radio Access Network (O-RAN), this paper proposes the first Large Language Model (LLM)-driven, O-RAN-native edge AI service orchestration framework. The framework deploys a lightweight LLM agent on the RAN Intelligent Controller (RIC) platform to directly parse user natural-language requests into deployable AI services and corresponding network configurations. It introduces a novel rApp-integrated architecture that unifies AI model selection, service deployment, adaptive network resource scheduling, and real-time xApp-based monitoring in a closed-loop manner. A prototype—built upon OpenAirInterface, FlexRIC, and Hugging Face models—demonstrates significant improvements in service provisioning efficiency and human-AI interaction usability. The framework establishes a scalable, interpretable paradigm for intelligent network orchestration in 6G.

To address the explosive demand for edge AI in 6G networks, this work tackles the limitation of conventional RAN architectures—designed only for AI-assisted optimization—by enabling native support for distributed AI workloads. Method: (1) We extend the O-RAN SMO framework with a lightweight AI-RAN orchestrator for cross-domain orchestration of communication and AI resources; (2) we design distributed AI-RAN sites featuring multi-tier latency awareness and geographically precise scheduling; (3) leveraging modular, cloud-native Open RAN, we enable co-deployment of real-time and batch AI tasks alongside multi-vendor interoperability. Contribution/Results: This is the first architecture to provide native AI compute support atop RAN infrastructure—without requiring new hardware—thus repurposing existing investments. It transforms the RAN from a connectivity pipeline into an edge intelligence-enabling platform, significantly enhancing telecom operators’ AI monetization capabilities.

This work addresses the inefficiencies in developing and deploying AI applications (xApps/rApps) in O-RAN and the inability of large language models (LLMs) to meet the stringent real-time and deterministic inference requirements of radio access networks. To bridge this gap, the authors propose the Dual-Brain architecture, which uniquely integrates an LLM’s natural language understanding and code generation capabilities with NeuralSmith—a lightweight AutoML engine capable of on-demand model training—to enable an end-to-end pipeline from natural language intent to automated AI service deployment. Serving as an orchestrator, the LLM coordinates with a containerized O-RAN 5G standalone testbed via an API-driven framework, demonstrating full automation of data collection, model training, and deployment in a real-world environment. This approach significantly enhances AI service provisioning efficiency while ensuring both security and practicality.

This work addresses the challenge of deploying AI inference under stringent 10-ms latency constraints in O-RAN near-real-time RIC by proposing an embedded lightweight AI xApp. The approach exports logistic regression and shallow MLP models as deterministic C code, which is directly compiled into the xApp binary—eliminating dependencies on external machine learning runtimes. A synthetic dataset is constructed using cross-layer features including MAC, RLC, PDCP, GTP, and UE count. Experimental evaluation on OpenAirInterface and FlexRIC demonstrates inference latencies of only 1–25 microseconds, end-to-end service latency below 4 ms, model accuracy between 0.88 and 0.90, and over 95% of execution cycles meeting the 10-ms deadline. This study presents the first validation of deterministic embedded AI within a near-real-time RIC closed loop and releases the RIC Workbench to enable reproducible research.

To address the high entry barrier, fragmented and outdated documentation, and lack of practical validation environments for xApp development in the O-RAN ecosystem, this paper proposes the first comprehensive, end-to-end xApp development paradigm. The paradigm systematically covers architectural analysis, design and configuration, lifecycle management, E2 interface invocation, RAN control, and debugging. Grounded in O-RAN Alliance specifications, it integrates the Near-Real-Time RIC architecture, xApp SDK, containerized deployment, and closed-loop real-time policy control. The resulting workflow is fully reproducible and validated. This approach significantly reduces learning and trial-and-error costs for third-party developers, accelerates algorithm prototyping and deployment, and bridges the critical gap between academia and industry by providing the first practical, implementation-oriented xApp development guide. It thereby advances the adoption of intelligent, programmable O-RAN network applications.

This work addresses the high energy consumption of mobile networks by proposing BeGREEN, an AI-driven intelligent plane within the O-RAN architecture to enable autonomous, energy-efficient radio access network management. By integrating an AI engine with rApps/xApps coordination mechanisms into the O-RAN intelligent plane for the first time, the approach establishes an end-to-end energy efficiency optimization loop that dynamically controls the operational states of simulated cells. Leveraging AI/ML algorithms, the O-RAN intelligent plane framework, and cell state management techniques, the proposed method significantly reduces base station energy consumption in simulation environments, thereby demonstrating the feasibility and effectiveness of AI-driven energy optimization in O-RAN networks.

This work addresses the inefficiencies in autonomous network management caused by unintended interactions among multiple control loops and independent applications due to the programmability of O-RAN. To this end, the paper proposes a multi-scale agent-based AI framework that deploys coordinated intelligent agents across O-RAN’s non-real-time, near-real-time, and real-time layers. It presents the first integration of hierarchical large language models (LLMs), small language models (SLMs), and wireless physical foundation models (WPFMs) to establish an intent-driven, cross-timescale autonomous architecture. Cross-layer coordination is achieved through standard O-RAN interfaces and telemetry data, and a prototype system is implemented using open-source models. The framework demonstrates end-to-end autonomous network control capabilities in two scenarios: robust operation under non-stationary environments and intent-driven network slicing resource orchestration.

This work addresses the operational complexity challenges in secure and auditable multi-tenant, multi-objective Open RAN deployments. The authors propose an agent-based primitive framework for O-RAN, integrating planning–execution–observation–reflection cycles, skill-as-tool abstractions, memory and evidence modules, and self-managed gating mechanisms to realize an intent-driven, explainable, and self-evolving RAN intelligent control system. The framework focuses on three core tasks: network slice lifecycle management, closed-loop wireless resource management, and security compliance enforcement. Evaluated in a multi-cell simulation environment across three representative slice types, the system reduces average resource overhead by 8.83% compared to conventional baselines and ablation variants, demonstrating its efficacy and compatibility with existing standards.

This work addresses the severe performance degradation of AI models in O-RAN caused by abrupt shifts in data distribution during dynamic radio access network (RAN) reconfiguration, which conventional passive retraining strategies fail to mitigate without prolonged service disruption. To overcome this limitation, the authors propose RANPilot, a novel framework that introduces the first proactive AI adaptation mechanism tailored for O-RAN. RANPilot employs a lightweight, trajectory-driven virtual O-RAN simulator to generate high-fidelity synthetic data reflecting the target configuration prior to physical reconfiguration, thereby enabling preemptive model adaptation. This paradigm shift—from reactive retraining to anticipatory preparation—significantly enhances service continuity. Experimental validation on a real-world 5G testbed demonstrates an 85%–94% reduction in AI service interruption time, achieving near-seamless functional transitions.

This work addresses the surge in energy consumption arising from the deep integration of artificial intelligence (AI) and radio access networks (RAN) in the 6G era, a challenge exacerbated by the lack of cross-application adaptive energy-saving coordination mechanisms in current O-RAN architectures. To bridge this gap, the paper proposes an AI-native RAN architecture that, for the first time, introduces agent-based paradigms and semantic intent abstraction into RAN control. By harmonizing O-RAN’s structured framework with the unified vision of AI-RAN, the proposed approach leverages a large language model (LLM)-driven coordination mechanism to enable adaptive orchestration of heterogeneous workloads, multi-objective optimization, and resolution of cross-application conflicts. Experimental results demonstrate that this method significantly enhances resource utilization efficiency and effectively reduces RAN energy consumption, offering a key enabler for sustainable 6G networks.