Advanced LLM-Enhanced Intent-Based 5G Network Management using Dynamic Semantic Routes

As 5G and next-generation networks become increasingly distributed, complex, and dynamic, traditional human-centric management approaches become infeasible. The ultimate goal for network automation is Zero-touch network and Service Management, but an intermediate transition stage is Human-on-the-Loop intent-based networking, where operators control the network using natural language. This paper addresses the challenge of translating unstructured natural language operator prompts into structured, machine-executable actions for 5G core networks without suffering from LLM hallucination or reliance on proprietary models.

The authors propose an end-to-end intent extraction model built around three primary subsystems: the user input, route selection via a semantic router, and dynamic route execution. The semantic router ingests the operator's prompt, uses an encoder to generate similarity scores against predefined thresholds, and routes the message to the appropriate intent category. Specifically, the system evaluates Regular Notification Requests and Intent Report Requests, along with a default None route to act as a guardrail against irrelevant prompts.

Once a static route is triggered, a dynamic route executes a predefined function utilizing an open-source Large Language Model (Mistral 7B) to parse key parameters from the prompt. This extraction process translates conversational expressions—such as update frequencies or positional references to previous intents—into standardized key-value pairs and JSON schemas suitable for direct operational use.

The authors evaluate five different text encoders for intent extraction and route selection accuracy, testing their performance both before and after training on a diverse set of augmented human-written prompts. The results indicate that both static and dynamic routing mechanisms are highly effective at correctly identifying operator intents, extracting fine-grained parameters, and formatting them into required schema structures. Furthermore, the incorporation of the None route successfully prevents out-of-scope or irrelevant messages from erroneously triggering network actions.