The shift toward open, programmable O-RAN and AI-RAN 6G networks presents new opportunities for Intent-Based Networking (IBN) to dynamically optimize RAN operations. However, applying IBN to the RAN scheduler—a key component for resource allocation—remains challenging. Existing methods rely on coarse network slicing, lacking the granularity to adapt to individual user conditions and traffic. While many scheduling algorithms exist, their use is limited by implementation complexity, lack of systematic evaluation, and deployment challenges. To overcome this, we introduce ALLSTaR, a novel framework that uses LLMs to automate intent-driven scheduler design, implementation, and evaluation. ALLSTaR translates natural language intents, extracts scheduler logic from academic literature using OCR and LLMs, and deploys them as O-RAN dApps on a production-grade, multi-vendor 5G testbed. This enables the largest OTA comparison of 18 synthesized schedulers and forms the basis for an Intent-Based Scheduling system that dynamically selects and deploys the optimal scheduler. Our approach supports use cases beyond current slicing techniques, enabling fine-grained control based on buffer status, PHY conditions, and diverse traffic types.
Preprint: https://arxiv.org/pdf/2505.18389