Disturbance-Resilient Agentic AI Control for Age-Fair SPS in 5G V2I Networks

Abstract

In fifth-generation (5G) New Radio (NR) vehicle-to-infrastructure (V2I) networks, semi-persistent scheduling (SPS) reduces signaling but may retain an unsuitable resource reservation interval (RRI) after the local load changes. We study age-fair keep/reselect control and distinguish a reward-aligned age-fair utility (RAAFU) surrogate from direct age-of-information (AoI), Jain freshness, and SPS radio-resource collision endpoints. The Full agent combines deterministic Rule/model-predictive-control (MPC) evaluation and a safety gate with an event-triggered large-language-model (LLM) proposal path. Across 128 independent scenario families, Full achieved the highest displayed Jain freshness, 0.9386, and a 1.009% collision rate; it also slightly reduced mean/q95 AoI relative to the same-stack controls. In an independent 256-case shock study, the LLM path reduced q95 AoI and collision relative to a budget-matched non-LLM selector, showing a measurable component contribution, although the stronger Rule remained competitive. The main trade-off is a higher absolute AoI than some external comparators, and measured provider latency is on the order of seconds rather than slots. These findings support only a simulation-based slow-path role for the LLM; they do not establish real-time deployment.