Seven intent categories. That’s the entire abstraction layer an operator needs to go from “show me session failures in the last 5 minutes” to a live Prometheus query against AMF and SMF event subscriptions. The authors built this as a drop-in open-source NWDAF for Free5GC, and they published the whole thing on GitHub.
How an LLM Bridges Intent and Network Telemetry
The Network Data Analytics Function (NWDAF) is 5G’s brain for zero-touch closed-loop automation, but most open-source implementations are half-baked or locked behind proprietary APIs. This one takes a different route: it accepts natural language, encodes the user’s intent via a semantic embedding model, and maps that embedding to one of seven predefined categories—think “subscriber analytics,” “network load,” “event subscription.” No YAML editing, no curl incantations. Once the intent is classified, the system triggers the appropriate analytics query or subscription command against the live network functions.
Seven categories isn’t a lot, but it’s exactly the right number to cover the common analytics workflows in a 5G core: monitor PDU session establishment, track registration failures, subscribe to SMF events, and so on. The interface sits on top of Prometheus for real-time metric retrieval, so operators can ask “what’s the AMF load right now?” and get a time-series response in plain text.
Why This Matters for 6G’s AI-Native Ambitions
Free5GC is already the go-to open-source 5G core for research testbeds. Wiring an LLM interface directly into its NWDAF means any lab can now experiment with intent-driven network management without building the plumbing from scratch. The paper’s architecture—semantic encoder → intent classifier → analytics executor—is simple enough to extend far beyond the seven categories. Add new intents by training on a few more examples; the embedding model does the heavy lifting.
More importantly, this is a tangible artifact for the “AI-native 6G” vision that’s been floating through 3GPP and ITU-T. Instead of another white paper, we get a working repo (github.com/HenokDanielbfg/testbed) that shows how LLMs can replace scripted interfaces for network operations. The next step is obvious: take those seven intents and let the LLM negotiate directly with network slices, not just read analytics.
Source: LLM-Enabled NWDAF: A Step Toward AI-Native 6G Network Intelligence
Domain: arxiv.org
Comments load interactively on the live page.