LLM-Based Emulation of the Radio Resource Control Layer: Towards AI-Native RAN Protocols

TL;DR

The paper investigates embedding AI-native capabilities into 6G RAN by emulating the Radio Resource Control layer with a decoder-only Large AI Model, fine-tuned via Low-Rank Adaptation on multi-operator traces. It introduces segmentation-safe, ASN.1-preserving QA datasets and schema-bounded prompting to maintain standards conformance, and evaluates across 120 configurations using ASN.1 conformance, field coverage, and UL→DL state-machine checks, plus semantic similarity and latency metrics. Results show that an 8B LLaMA-based model with LoRA attains a median semantic similarity around 0.97 and strong conformance, with constrained decoding and INT4 quantization significantly reducing latency, though sub-100 ms targets remain out of reach under current setups. The work demonstrates the feasibility of AI-native RRC orchestration and outlines key engineering steps—including faster inference techniques, longer context modeling, and protocol-aware evaluation—to realize real-time, standards-compliant AI-driven control planes for future RANs.

Abstract

Integrating Large AI Models (LAMs) into 6G mobile networks is a key enabler of the AI-Native Air Interface (AI-AI), where protocol intelligence must scale beyond handcrafted logic. This paper presents, to our knowledge, the first standards-compliant emulation of the Radio Resource Control (RRC) layer using a decoder-only LAM (LLAMA-class) fine-tuned with Low-Rank Adaptation (LoRA) on a multi-vendor corpus of real-world traces spanning both 5G and 4G systems. We treat RRC as a domain-specific language and construct a segmentation-safe, question--answer (Question-and-Answer (QA)) dataset that preserves Abstract Syntax Notation (ASN.1) structure through linearization prior to Byte Pair Encoding (BPE) tokenization. The proposed approach combines parameter-efficient adaptation with schema-bounded prompting to ensure syntactic and procedural fidelity. Evaluation introduces a standards-aware triad -- ASN.1 conformance, field-level coverage analysis, and uplink-to-downlink state-machine checks -- alongside semantic similarity and latency profiling across 120 configurations. On 30k 5G request--response pairs plus an additional 4.8k QA turns from 4G sessions, our 8B model achieves a median cosine similarity of 0.97, a 61% relative gain over a zero-shot baseline, while sustaining high conformance rates. These results demonstrate that LAMs, when augmented with protocol-aware reasoning, can directly orchestrate control-plane procedures, laying the foundation for the future Artificial Intelligence (AI)-native Radio Access Network (RAN).

Figures (9)

• Figure 1: High-level concept of an AI-native gNB-side RRC layer powered by a , illustrating its core inputs, outputs, learning source, and interaction with the protocol stack.
• Figure 2: Low‑level architecture of an NR gNB disaggregating an LLM‑based RRC layer.
• Figure 3: Data flow for fine-tuning and inference in the LLM-based RRC system.
• Figure 4: Example of a pre-processed trace segment. Messages are structured into Question (Q) / Answer (A) pairs for LLM training, where 'Q' denotes the input RRC message(s) (X) and 'A' the target response (Y). Note the consolidation of multiple physical messages (e.g., rrcReconfigurationComplete and measurementReport) into a single logical request, labeled here as Q4(concat) and Q6(concat).
• Figure 5: Training loss trajectory on the NR dataset (proprietary corpus). The pale blue trace is the raw loss, and the dark blue curve is its exponential moving average (EMA). The vertical red dashed line marks the epoch selected for downstream evaluation.