LLM-Enabled NWDAF: A Step Toward AI-Native 6G Network Intelligence

## The Core Claim

Network management in 5G and beyond requires specialized interfaces that only trained engineers can operate. This paper shows that adding a large language model (LLM) layer to the Network Data Analytics Function (NWDAF)—the component responsible for real-time network monitoring—lets non-expert operators manage complex network tasks through plain English conversation instead of technical commands.

## How It Works

The authors built an open-source NWDAF system integrated with Free5GC, an open-source 5G core network platform. The innovation lies in three layers sitting between a human operator and the underlying network:

**Natural language input.** An operator describes what they want in conversational English—for example, "Show me all active sessions in the network" or "Alert me when packet loss exceeds 5%."

**Intent classification.** The system converts the user's text into semantic embeddings (numerical representations of meaning) using an embedding model, then maps that meaning to one of seven predefined intent categories. This bridges the gap between freeform language and structured network commands.

**Network action.** Once classified, the intent triggers either an analytics query or an event subscription command targeting specific network functions—specifically the Access and Management Function (AMF) and Session Management Function (SMF), which handle user connections and session data. Results flow back through Prometheus (a monitoring platform) and are presented conversationally to the user.

The system collects real-time data by subscribing to events from various network functions, making it possible to monitor and manage the network without manual polling or complex API calls.

## Why It Matters

NWDAF is foundational for "zero-touch" network management—the goal of keeping 5G networks running with minimal human intervention through automation and analytics. However, current NWDAF tools require operators to understand technical interfaces and command structures, creating a barrier to broader adoption.

By wrapping NWDAF in an LLM-powered conversational interface, the authors lower that barrier significantly. Non-specialist staff can now interact with network analytics and subscription systems as easily as chatting with a colleague. This matters for 6G because the network intelligence stack will only become more complex; making it accessible through natural language is a practical step toward "AI-native" networks where intelligence and usability are built in from the start.

The authors also released their code and datasets publicly, addressing a broader problem: open-source NWDAF implementations have been limited in scope. This contribution provides the community with a concrete, extensible reference implementation.

The seven predefined intents represent a simplifying assumption—the system works well for common tasks but doesn't yet handle arbitrary operator requests. Still, the design is modular enough that adding new intent categories should be straightforward, making this a practical foundation rather than a finished product.