LLM-RAN Operator: Semantic Control in 6G RAN

• LLM-RAN Operator is a control construct where large language models convert high-level intents and network states into structured RAN actions using a validated command DSL.
• It is architecturally layered, spanning direct RRC emulation in the protocol stack to supervisory, non-real-time orchestration in the management plane.
• Key design patterns include structured intermediate representations, template-constrained outputs, and hybrid LLM-RL integration that improve performance and reliability.

An LLM-RAN Operator is a large-language-model-centered control and management construct for the Radio Access Network in which an LLM, or a broader agentic stack organized around an LLM, maps high-level intents and current network state into structured RAN actions, policies, protocol messages, or deployable control logic. The term is made explicit in “The LLM as a Network Operator: A Vision for Generative AI in the 6G Radio Access Network” (Giwa et al., 27 Aug 2025), but adjacent work already instantiates narrower forms of the same idea: protocol-native Radio Resource Control generation inside the gNB (Liu et al., 22 May 2025), intent-based O-RAN management over O1/NETCONF (Bimo et al., 17 Jul 2025), hierarchical non-RT/near-RT guidance (Bao et al., 25 Apr 2025), and automated xApp or AI-service provisioning from natural-language requirements (Natanzi et al., 22 May 2026, Li et al., 19 Mar 2026). In current literature, the term therefore denotes not a single product or reference implementation, but a family of architectures in which LLMs move from advisory analytics toward operational participation in the RAN.

1. Formal concept and scope

The most explicit formalization treats the operator as an intent-conditioned control map. In that framing, the LLM-RAN Operator consumes an intent $i \in \mathcal{I}$ and a network state $s_t \in \mathcal{S}$, produces an action $a_t = \mathcal{O}_{\rm LLM}(i,s_t)$, and induces a closed-loop transition $s_{t+1} = f_{\rm env}(s_t, a_t)$ (Giwa et al., 27 Aug 2025). The same work further restricts actions to a finite RAN Command DSL $\mathcal{L}$ and requires all emissions to pass a validator $V:\mathcal{L}\to\{0,1\}$, which enforces units, bounds, and safety before compilation to O-RAN interfaces such as A1, E2, or O1 (Giwa et al., 27 Aug 2025). This formulation makes the operator a cognitive policy generator rather than a generic conversational assistant.

Across the literature, the operational meaning of “operator” varies. In the narrowest interpretation, it is a protocol-native actuation primitive that translates one control-plane event into another; the RRC-layer emulation work places a large model directly in the gNB CU-CP and treats uplink RRC messages as tokenized inputs $X$ and downlink RRC responses as autoregressively generated outputs $Y$ under $P(Y\mid X;\theta)$ (Liu et al., 22 May 2025). In a broader management-plane interpretation, the operator is an intent-driven orchestration layer that translates JSON-formalized objectives into configuration changes over O-RAN O1 using NETCONF \<get-config\> and \<edit-config\> (Bimo et al., 17 Jul 2025). Still other systems define the operator as a supervisory semantic layer that provisions new xApps, deploys AI services, or reasons over telemetry while leaving deterministic execution to smaller models or conventional controllers (Natanzi et al., 22 May 2026, Ngo et al., 14 Mar 2026).

This diversity implies an important boundary condition. An LLM-RAN Operator is not necessarily a fully autonomous, end-to-end replacement for existing RAN control and OSS/BSS functions. In the surveyed works, it may instead appear as a control-plane micro-agent, a semantic planner at the SMO, a non-real-time policy guider, a provisioning engine, or a human-in-the-loop reasoning copilot. The common denominator is that the LLM participates in operational decision translation, not merely in post hoc explanation.

2. Architectural placement in O-RAN and AI-RAN

The literature suggests several recurrent placements for LLM functionality within the RAN stack. One placement is inside the protocol stack itself. The RRC-emulation system embeds a Large RRC Model in the gNB CU-CP, where the RRC layer normally resides; uplink RRC arrives from the UE through the DU and F1-C into the CU-CP, the model generates the downlink ASN.1-encoded RRC response, and decisions affecting data radio bearers can also be signaled over E1 toward the CU-UP (Liu et al., 22 May 2025). This is a control-plane actuation role inside L3 signaling rather than a management-plane analytics role.

A second placement is above the RAN in the O-RAN management hierarchy. The intent-based RAN management system runs over O-RAN O1, continuously retrieves PM data, uses an agentic LLM layer built with LangGraph and NVIDIA NIM, and issues CM operations through NETCONF (Bimo et al., 17 Jul 2025). The E-ARC architecture similarly places semantic coordination at the SMO and Non-RT RIC, where a Semantic Coordination agent, a Digital Twin agent, and a Configuration, Deployment, and Monitoring agent together select and validate rApps before deployment (Aroua et al., 20 Jun 2026). In these systems, the LLM is not a fast radio scheduler; it is a supervisory non-real-time coordinator.

A third placement is hierarchical coordination across timescales. LLM-hRIC assigns the LLM to the non-RT RIC as an rApp that provides strategic guidance using global information, while a DDPG-based xApp in the near-RT RIC consumes that guidance together with local observations to make 100 ms decisions (Bao et al., 25 Apr 2025). AgentRAN generalizes this pattern further into a self-organizing hierarchy that decomposes intents across timescales from sub-millisecond to minutes, spatial domains from cell to network-wide, and protocol layers from PHY/MAC to RRC, with rApps, xApps, and dApps cooperating through A2A and MCP over JSON-RPC (Elkael et al., 25 Aug 2025).

A fourth placement is service lifecycle automation rather than direct control. The Dual-Brain architecture keeps the LLM in the Non-RT RIC as the ZTO-Agent, where it translates operator intents into telemetry policies, labeling rules, model requests, and template-constrained deployment artifacts, while a lightweight ONNX classifier executes later in the Near-RT RIC (Natanzi et al., 22 May 2026). AutoRAN and AutORAN extend this orchestration role to zero-touch deployment pipelines and natural-language programming for xApp development, respectively (Maxenti et al., 15 Apr 2025, Li et al., 19 Mar 2026).

Taken together, these placements indicate that current research does not converge on a single canonical layer for LLM insertion. A plausible implication is that “LLM-RAN Operator” is best understood as a layered pattern: semantic reasoning and orchestration usually sit at the SMO or Non-RT RIC, while direct protocol-native generation and real-time execution, when present at all, remain tightly constrained and task-specific.

3. Core design patterns and implementation strategies

Several design patterns recur across otherwise different implementations. The first is structured intermediate representation. The RRC-emulation work treats RRC signaling as a domain-specific language by linearizing ASN.1 message structure into text while preserving hierarchy before BPE tokenization and LoRA fine-tuning of LLaMA3-8B (Liu et al., 22 May 2025). The formal vision paper likewise constrains LLM outputs to a finite command language $\mathcal{L}$ validated before compilation (Giwa et al., 27 Aug 2025). Intent-based RAN management represents objectives in JSON aligned with 3GPP TS 28.312 and uses a five-section prompt structure consisting of Instruction, Intent, Current Observation, Configuration Constraints, and Output Format (Bimo et al., 17 Jul 2025).

The second is hard-bounded tool use or template use instead of unrestricted generation. The Net Analyzer rApp exposes only tool-gated log-query and configuration-query operations, separates reasoning from actuation, and requires explicit operator approval before configuration change (Ngo et al., 14 Mar 2026). The Dual-Brain provisioning system never lets the LLM write arbitrary xApp logic; it fills pre-verified Jinja2 templates and delegates near-real-time execution to lightweight classifiers (Natanzi et al., 22 May 2026). AutoRAN similarly constrains intent translation through compatibility graphs, tool-calling, iterative validation loops, and JSON outputs that feed Tekton and ArgoCD rather than direct cluster mutation (Maxenti et al., 15 Apr 2025).

The third is domain grounding. ORANSight-2.0 creates RANSTRUCT, a RAG-based instruction-tuning dataset over 116 O-RAN documents and 4,980 srsRAN C++ files, then fine-tunes 18 open-source models via QLoRA to create O-RAN-specialized foundation models (Gajjar et al., 7 Mar 2025). Contextual RAG improves O-RAN question answering by retrieving against query variants augmented with answer choices, rather than the question alone, while avoiding model fine-tuning in a rapidly evolving standards environment (Conger et al., 18 Dec 2025). BLINC grounds causal learning by asking an LLM to identify mandatory and prohibited edges in a Bayesian Network over configuration parameters, measurements, and KPIs, thereby constraining the hypothesis space for structure learning (Prasad et al., 29 Apr 2026).

The fourth is hybridization rather than LLM monolithism. LLM-hRIC combines LLM guidance at the non-RT layer with RL execution at the near-RT layer (Bao et al., 25 Apr 2025). PA-MRL uses ORANSight as a state-contextualization engine whose prompt embeddings are optimized jointly with MARL for slicing control (Lotfi et al., 31 May 2025). ReLLM separates SLA-risk estimation from resource reallocation so that the expensive allocator agent is invoked only when necessary (Salan et al., 28 Nov 2025). E-ARC separates semantic reasoning, DT validation, and deployment monitoring into SC, DT, and CDM agents (Aroua et al., 20 Jun 2026). This suggests that the dominant implementation style is an LLM-centered composite controller rather than an unconstrained single-model policy.

4. Empirical demonstrations

The strongest evidence for protocol-native control comes from the RRC emulation study. On 2,524 RRC sessions yielding 30,247 dialog pairs from a real 5G network, the fine-tuned 8B model achieved a median cosine similarity of 0.97 against ground-truth downlink RRC messages, versus 0.60 for zero-shot LLaMA 3-8B, described as an absolute gain of 0.37 and a relative gain of approximately 61% (Liu et al., 22 May 2025). The same paper reports median inference time per RRC message of 6.9 seconds and average 10.4 seconds, explicitly stating that this is not real-time and falls far short of mobile-network timer requirements (Liu et al., 22 May 2025).

Intent-based management results are narrower but operationally concrete. In the O1/NETCONF closed loop, TxPower was iteratively reduced from 30 dBm to 11 dBm, while PEE.EnergyEfficiency increased from 765493 bits/Joule to 803257 bits/Joule; the target intent example was RANEnergyEfficiency IS_GREATER_THAN 810000 bit/joule, but the reported iterations did not reach that threshold (Bimo et al., 17 Jul 2025). ReLLM, which controls slice resource allocation on an OAI/FlexRIC testbed, reports that it maintains the latency of the high-priority slice under 10 ms and keeps the low-priority slice drop ratio around 0.5%, whereas a fixed 70-30 allocation can produce packet drops for the low-priority slice of up to 50% (Salan et al., 28 Nov 2025).

For configuration optimization, BLINC reports a throughput improvement of 63.5% with 19.7% reduction on block error rate over data-only baselines through joint optimization of uplink power control and link adaptation parameters (Prasad et al., 29 Apr 2026). In structure learning, “Hill climb + BIC + LLM” with Claude 4.5 Sonnet achieves 10 correct edges, 2 misses, 2 reversals, direction score 0.83, and recall 0.71, compared with 5 correct, 4 misses, 5 reversals, direction 0.50, and recall 0.36 for data-only hill climbing with BIC (Prasad et al., 29 Apr 2026). In LLM-augmented MARL for slicing, ORANSight PA-MRL reports QoS improvements of 55.4% for eMBB, 47.05% for mMTC, 23.66% for URLLC, and convergence improvement of 8.97%, outperforming GPT PA-MRL and ORANSight MARL in the reported ablation table (Lotfi et al., 31 May 2025).

Provisioning and development pipelines show a different kind of evidence: reduced time from intent to deployable control artifact. The Dual-Brain O-RAN system reports 384 ms warm steady-state orchestration latency for Llama-3.1-8B, with a 49 kB ONNX LightGBM model achieving 97.7% accuracy, F1 = 0.975, and <1 ms CPU inference in the congestion-management use case (Natanzi et al., 22 May 2026). AutORAN reports that generated xApps can achieve similar or better performance than the best known hand-crafted baselines and measures full control-loop execution times of 20–110 ms, 55–90 ms, and 250–420 ms for representative generated xApps on a real O-RAN stack (Li et al., 19 Mar 2026).

Knowledge-centric operator functions have also been benchmarked. ORANSight-2.0 reports improvements over ChatGPT-4o and Gemini of 5.421% on ORANBench and 18.465% on srsRANBench, while its best RAG-augmented cumulative score is 0.854 for ORANSight-Qwen-2.5 32B with RAG (Gajjar et al., 7 Mar 2025). Contextual RAG further shows that retrieval design matters: Qwen 3.0 4B with Contextual RAG reaches 0.8824 / 0.8718 / 0.8261 on Easy/Medium/Hard ORAN-Bench-13K in Direct Q&A mode, while the same paper states that Contextual RAG consistently improves accuracy over standard RAG and base prompting (Conger et al., 18 Dec 2025).

Finally, deployment studies in AI-RAN quantify where inference can actually run. On a 5G SA AI-RAN testbed, on-device execution remained multi-second and missed both the 0.5 s and 1.0 s SLA budgets, while at the RAN edge 7B-AWQ achieved 99.0% [email protected] and cloud execution on the measured WAN path achieved at best 32.9% [email protected] but 100% [email protected] for all evaluated variants (Yet et al., 27 Feb 2026). Under saturated downlink traffic and up to $s_t \in \mathcal{S}$0 concurrent inference clients, MIG isolation preserved SlotInd and U-plane timing-health proxies, supporting safe co-location under fixed partitioning (Yet et al., 27 Feb 2026).

5. Safety, limitations, and recurrent misconceptions

A recurrent misconception is that current systems already realize a fully autonomous, carrier-grade LLM operator. The evidence does not support that interpretation. The RRC-emulation study explicitly states that it is a stepping stone toward AI-native RAN control, but not a real-time live gNB control-plane operator, because inference latency is measured in seconds and evaluation relies on cosine similarity rather than exact parse-validity, conformance, state-machine consistency, or interoperability metrics (Liu et al., 22 May 2025). The formal vision paper offers propositions on expressiveness and a theorem on convergence to stable fixed points, but these guarantees are conditional on a restricted DSL, validator enforcement, and contraction properties that are acknowledged to be difficult to establish in realistic wireless systems (Giwa et al., 27 Aug 2025).

Hallucination and grounding remain major concerns. In the intent-based O1 management system, the authors explicitly report hallucinations and inconsistent interpretation when comprehensive PM data are fed directly to the LLM, and warn that inaccurate, incomplete, or delayed PM data can induce a domino effect in the closed loop (Bimo et al., 17 Jul 2025). In the RRC setting, a medium-similarity case showed syntactically legal but semantically unnecessary insertion of securityConfig { ... }, illustrating that “valid but unnecessary” protocol content can still be harmful (Liu et al., 22 May 2025). In knowledge and explanation systems, the prevailing mitigation is architectural: tool-gated access, explicit approval gates, and strict separation between reasoning and actuation (Ngo et al., 14 Mar 2026).

Another recurrent limit is timescale mismatch. LLM-hRIC places the LLM in the non-RT RIC precisely because near-RT deadlines are too tight for modern LLM inference (Bao et al., 25 Apr 2025). The AI-RAN deployment study reaches a similar conclusion empirically: Premium-class 0.5 s inference is edge-feasible only for selected quantized variants, while cloud is essentially a 1.0 s-class medium-latency tier on the measured path (Yet et al., 27 Feb 2026). This supports a now-common architectural rule: semantic reasoning and orchestration may be LLM-native, but hard real-time execution is usually delegated to RL agents, deterministic controllers, dApps, or compact classifiers.

The literature also shows that generic foundation models are often insufficient without specialization or constraint. Public LLMs in the RRC benchmark added explanatory wraparound text or drifted out of the “base station” role (Liu et al., 22 May 2025). ORANSight-2.0 and Contextual RAG are motivated precisely by the inadequacy of generic pretrained knowledge for evolving O-RAN standards and codebases (Gajjar et al., 7 Mar 2025, Conger et al., 18 Dec 2025). A plausible implication is that domain adaptation and constrained I/O contracts are not optional refinements but core requirements for any operational LLM-RAN Operator.

6. Development trajectory and open research directions

The next research stage is increasingly clear in the surveyed work. For protocol-native control, the missing pieces are exact ASN.1 parse-validity, schema-validity, state-machine consistency, interoperability testing with real UEs and gNB stacks, latency reduction to sub-second and eventually millisecond operation, and confidence or guardrail mechanisms that prevent unauthorized or state-inconsistent message emission (Liu et al., 22 May 2025). For management-plane orchestration, the open issues include stronger domain grounding through RDF or ontology, robustness to noisy PM data, and resilience mechanisms beyond prompt constraints, with reinforcement learning suggested as one future direction for handling inaccurate or delayed telemetry (Bimo et al., 17 Jul 2025).

For configuration intelligence, BLINC points toward a division of labor in which the causal graph structure is relatively stable while CPDs are continually updated through $s_t \in \mathcal{S}$1, with occasional full relearning after major network changes (Prasad et al., 29 Apr 2026). Agentic architectures extend this toward self-improving control fabrics: AgentRAN’s AI-RAN Factory proposes code generation, model distillation, fine-tuning, and hybrid agent creation driven by a data lake of KPIs, decisions, and reasoning traces (Elkael et al., 25 Aug 2025). E-ARC similarly suggests that future operators will need semantic coordination, Digital Twin validation, and application-context enrichment to orchestrate both AI-for-RAN and AI-on-RAN under shared energy and QoE objectives (Aroua et al., 20 Jun 2026).

The broader trajectory therefore points toward a layered operator architecture rather than a monolithic one. The surveyed literature suggests a likely stack composed of domain-specialized foundation models, structured intent schemas, validator-backed command languages, retrieval or tool-grounded reasoning, deterministic lower-layer executors, and Digital Twin or sandbox validation before deployment. In that sense, the LLM-RAN Operator is emerging less as a single model that “runs the RAN” and more as a semantic control layer that interprets intent, coordinates specialized modules, and progressively narrows the gap between human operational goals and machine-executable radio behavior.