Talk 8 - Beyond LLMs: Why World Models are the Key to AI-Native RAN

The evolution of AI within telecommunications has progressed through three distinct eras: from exploratory experiments and opportunistic applications to the current pursuit of native scalability.
In the early stages, traditional machine learning was confined to isolated use cases like traffic prediction and basic load balancing. The subsequent rise of Big Data and Deep Learning enabled more sophisticated RAN functions, including channel estimation, anomaly detection, and Self-Optimizing Networks (SON). However, these "black-box" models often suffered from a lack of explainability, leading to functional conflicts that hindered large-scale deployment.
While the recent emergence of Large Language Models (LLMs) and Agentic AI has expanded the horizons of autonomous networks, they face critical bottlenecks in a RAN environment:
•    Lack of physical grounding: LLMs often fail to capture the fundamental physical laws of radio propagation, leading to suboptimal or erroneous decisions.
•    Infrastructure & latency: The massive computational overhead and high inference latency of LLMs are incompatible with the stringent near-real-time requirements of RAN control.
•    Static digital twins: Traditional Digital Twins (such as AODT) provide valuable pre-deployment training but often struggle to capture real-world dynamics or scale efficiently for real-time decision-making.
In this session, we introduce a paradigm shift: AI-Native RAN based on World Models. Unlike static models, a World Model inherently learns the underlying dynamics of the system, enabling it to predict outcomes and make reasoned decisions with high precision. By utilizing architectures ranging from tiny to medium-sized, this approach delivers the agility needed for both near-real-time and non-real-time control without the need for prohibitive hardware investments. We will demonstrate the superior performance and operational efficiency of World Models integrated within our RIC platform.