Researchers want to put language models in charge of 5G network scheduling — just not in the way most AI hype implies.
The paper introduces Agentic-V2X, a two-layer architecture designed for vehicle-to-everything (V2X) networks. A small, locally deployed language model acts as a periodic policy creator, generating structured rules around service priorities, weight limits, and safety constraints. A separate lightweight controller then takes those validated rules and enforces them in the scheduler — the LLM never touches live packet decisions. The system was tested across 126 simulation runs in ns-3/ns3-ai, covering services like teleoperated driving, HD map sharing, and sensor data.
The separation of duties matters. Full-scale large language models are too slow and too unreliable for near-real-time network control — they hallucinate, they lag, and they offer no control guarantees. By pushing the LLM into a slower, non-real-time policy role and keeping fast execution in a deterministic controller, the architecture sidesteps those failure modes. The paper also includes a validator that checks and repairs LLM output before anything gets enforced.
The results are honest: Agentic-V2X generates valid, executable policies and edges out proportional fair scheduling on critical reliability at high vehicle density, but it does not beat the strongest static expert policies overall. The authors call it a safe architecture, not a dominant one — which is a more defensible claim than most AI networking papers bother to make.