PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic AI Empowered Intent-Based Networking for 6G

Genze Jiang, Kezhi Wang, Xiaomin Chen, Yizhou Huang

arXiv:2601.06640v1cs.AIcs.NI

TL;DR

Existing 6G intent-based networking approaches have difficulty handling linguistic variation, interpretability, and operational constraints. This paper introduces a hierarchical multi-agent framework that outperforms rule-based systems, monolithic agents, and direct LLM prompting across benchmark scenarios.

  • Problem

    Existing intent-based networking approaches struggle with linguistic variation, generalisation, interpretability, and autonomous multi-step orchestration of natural-language intents.

  • Method

    The paper uses an orchestrator coordinating RAN and Core Network specialist agents through iterative ReAct reasoning, evaluated with semantic accuracy and engineering utility metrics.

  • Results

    The multi-agent architecture outperforms monolithic agents, rule-based systems, and direct LLM prompting, achieving 8% higher semantic understanding than monolithic agents.

  • Takeaways & Limitations

    The findings support tiered deployment that assigns complex or ambiguous intents to multi-agent processing and simpler intents to lightweight agents.

  • Takeaways & Limitations

    The multi-agent system requires 7× more tokens and 5× longer execution time than monolithic agents.

Abstract

from arXiv · show

The transition towards sixth-generation (6G) wireless networks necessitates autonomous orchestration mechanisms capable of translating high-level operational intents into executable network configurations. Existing approaches to Intent-Based Networking (IBN) rely upon either rule-based systems that struggle with linguistic variation or end-to-end neural models that lack interpretability and fail to enforce operational constraints. This paper presents a hierarchical multi-agent framework where Large Language Model (LLM) based agents autonomously decompose natural language intents, consult domain-specific specialists, and synthesise technically feasible network slice configurations through iterative reasoning-action (ReAct) cycles. The proposed architecture employs an orchestrator agent coordinating two specialist agents, i.e., Radio Access Network (RAN) and Core Network agents, via ReAct-style reasoning, grounded in structured network state representations. Experimental evaluation across diverse benchmark scenarios shows that the proposed system outperforms rule-based systems and direct LLM prompting, with architectural principles applicable to Open RAN (O-RAN) deployments. The results also demonstrate that whilst contemporary LLMs possess general telecommunications knowledge, network automation requires careful prompt engineering to encode context-dependent decision thresholds, advancing autonomous orchestration capabilities for next-generation wireless systems.

I. INTRODUCTION … C. Large Language Models in Communications

The paper motivates hierarchical agentic AI for translating natural-language intents into feasible 6G network-slice configurations, addressing the semantic and operational limits of existing IBN, machine-learning, and direct-LLM approaches. It positions domain-specialised agents, iterative reasoning, and prompt refinement as key mechanisms for autonomous orchestration, while situating the work within prior IBN, network-optimisation, and communications-LLM research.

  • I. INTRODUCTION: 6G’s zero-touch vision requires autonomous orchestration that translates high-level natural-language objectives into network configurations.Intent-Based Networking provides this policy-driven management paradigm, but existing approaches face limitations in natural-language handling.
  • I. INTRODUCTION: LLMs support telecommunications tasks including configuration generation and fault diagnosis, but their direct use in autonomous orchestration is constrained by fundamental limitations.The paper therefore treats LLMs as components within broader agentic architectures rather than as standalone orchestration systems.
  • I. INTRODUCTION: The proposed framework coordinates one orchestrator and two domain-specialist agents to translate natural-language intents into technically feasible network-slice configurations through iterative reasoning.The architecture decomposes complex orchestration across agents with domain-specific expertise encoded through engineered system prompts.
  • I. INTRODUCTION: Experimental validation shows the multi-agent architecture outperforms monolithic agents, rule-based systems, and direct LLM prompting approaches.Evaluation also identified systematic prompt-induced biases, which targeted prompt refinement resolved, establishing prompt engineering as architecturally important.
  • II. RELATED WORKS: Prior research spans IBN, machine learning, and LLM-enabled communications, motivating the research gap addressed by the hierarchical multi-agent architecture.The related-work discussion connects these domains to autonomous network orchestration and identifies the need for semantic interpretation of operational intents.
  • A. Intent-Based Networking and Network Automation: Existing IBN and O-RAN orchestration methods automate policy handling and numerical resource optimisation but generally lack semantic reasoning for unstructured natural-language requirements.The proposed work positions LLMs as a semantic interpretation layer while retaining automated orchestration principles.
  • B. Machine Learning for Network Optimisation: Machine-learning approaches such as deep reinforcement learning and graph-based allocation rely on numerical state representations and extensive training interaction, limiting direct natural-language processing and safety-critical deployment.These limitations distinguish the proposed language-driven orchestration approach from conventional network-optimisation methods.
  • C. Large Language Models in Communications: Communications-focused LLM research has established telecom surveys, networking adaptation, configuration-generation evaluation, and autonomous-wireless-network visions as foundations for LLM-based orchestration.NetLLM reported 10 to 40 percent improvements over stateof-the-art methods on diverse tasks, while other work developed telecom-specific training pipelines and intent-lifecycle architectures.

D. Agentic AI and Multi-Agent Systems … B. Network System Model

The paper positions agentic, multi-agent LLM orchestration as a response to IBN’s limitations, then formalises 6G service requirements and a network-state-based provisioning problem. The framework must translate natural-language intents into feasible configurations while respecting QoS and resource constraints.

  • D. Agentic AI and Multi-Agent Systems: ReAct interleaves reasoning traces with tool actions, enabling iterative task refinement and interpretable decision traces for agentic AI systems.The architecture literature also identifies autonomous agents through profiling, memory, planning, and action modules.
  • D. Agentic AI and Multi-Agent Systems: LLM-based autonomous-agent architectures combine profiling, memory, planning, and action modules, while role-based prompting supports cooperation but can cause looping and role drift.These mechanisms motivate orchestrator and specialist agent designs.
  • E. Research Gap and Positioning: IBN remains constrained by structured policy languages, while numerical machine-learning methods lack natural-language intent processing and interpretability.Recent telecommunications LLM applications have mainly used single-shot prompting or configuration assistance rather than autonomous multi-step orchestration.
  • E. Research Gap and Positioning: Applying agentic AI to telecommunications IBN requires addressing safety-critical operation, strict latency, continuous adaptation to dynamic network states, and regulatory auditability.Existing agentic frameworks have demonstrated effectiveness in other domains but have not been systematically applied to IBN orchestration.
  • A. 6G Use Case Categories and QoS Requirements: The system model uses three foundational 6G service categories—URLLC, eMBB, and mMTC—to represent distinct QoS requirements for evaluation.The paper adopts these categories although the ITU has expanded 5G’s three scenarios into six 6G categories.
  • A. 6G Use Case Categories and QoS Requirements: URLLC requires below 10 milliseconds end-to-end latency and above 99.999 percent reliability, whereas eMBB targets multigigabit-per-second throughput with 20 to 50 milliseconds latency.URLLC uses edge-tier compute and guaranteed resources, while eMBB can use regional-tier data centres.
  • A. 6G Use Case Categories and QoS Requirements: mMTC supports extremely high device densities, low per-device data rates, and seconds-to-minutes latency tolerance, favouring wide-area, energy-efficient, often low-band transmission.These service distinctions enable context-dependent metric weighting and mapping from implicit natural-language QoS requirements to technical configurations.
  • B. Network System Model: The network state comprises RAN sectors with load, active users, and spectrum availability; core nodes with compute load and tier; and a latency matrix linking each sector-node pair.The objective is to synthesize a provisioning tuple selecting a RAN sector, spectrum band, and UPF node that satisfies intent-implied QoS under state-encoded resource constraints.

C. Evaluation Framework · 1) Semantic Accuracy, i.e., Golden Standard: · 2) Engineering Utility Function:

The evaluation framework combines qualitative Semantic Accuracy against expert-defined Golden Standards with quantitative Engineering Utility for technical optimality. Utility scores candidate configurations using weighted latency, resource-efficiency, and congestion criteria whose weights depend on intent context.

  • C. Evaluation Framework: The framework evaluates IBN agents with two complementary metrics: Semantic Accuracy for expert alignment and Engineering Utility for technical optimality.This hybrid design addresses the absence of a single objectively correct configuration for each intent.
  • 1) Semantic Accuracy, i.e., Golden Standard:: Benchmark scenarios pair natural-language intents with expert-defined ideal configurations called Golden Standards.Domain experts specify these configurations using knowledge of RAN resource allocation and core-network topology.
  • 1) Semantic Accuracy, i.e., Golden Standard:: Semantic Accuracy compares each agent-generated configuration A* with its corresponding Golden Standard.The score uses three semantic levels: Perfect Match, Partial Match, and No Match.
  • 1) Semantic Accuracy, i.e., Golden Standard:: Perfect Match scores 1.0 for agreement on sector, spectrum, and UPF node; Partial Match scores 0.5 for two dimensions; No Match scores 0.0.These levels distinguish exact agreement, partial alignment, and fundamental misalignment with expert expectations.
  • 2) Engineering Utility Function:: Engineering Utility U scores candidate configurations against network performance indicators using a weighted sum of normalized latency, resource-efficiency, and congestion subscores.The formulation is U(A, I, S) = w_l · S_latency + w_r · S_resource + w_c · S_congestion.
  • 2) Engineering Utility Function:: The utility function evaluates configuration A in the context of intent I and network state S, with non-negative weights satisfying w_l + w_r + w_c = 1.The intent determines context-dependent weighting, while the network state provides resource-availability information.
  • 2) Engineering Utility Function:: Latency scoring evaluates end-to-end delay between the selected RAN sector and core node, assigning 1.0 below 10 ms, 0.5 from 10 ms to under 30 ms, and 0.0 at 30 ms or above.This subscore measures whether UPF selection satisfies the intent’s latency requirements.
  • 2) Engineering Utility Function:: Intent category dynamically sets utility priorities: URLLC uses w_l = 0.8, eMBB uses w_c = 0.6, and mMTC uses w_r = 0.6.These settings prioritize latency compliance for URLLC, congestion avoidance and resource efficiency for eMBB, and spectrum selection for mMTC.

3) Hybrid Evaluation Protocol: · D. Problem Formulation

The evaluation protocol defines success through semantic accuracy and engineering utility thresholds, while the problem formulation casts IBN orchestration as utility maximisation under operational constraints. Its central difficulty is translating unstructured natural-language intents into explicit, network-state-dependent technical requirements.

  • 3) Hybrid Evaluation Protocol:: A deployment succeeds when Semantic Accuracy A ≥0.5 and Engineering Utility U ≥0.7.The dual-metric protocol tests both alignment with expert expectations and feasible implementation under prevailing network conditions.
  • D. Problem Formulation: The IBN orchestration problem seeks an optimal configuration that maximises utility while satisfying operational constraints.The formal statement builds on the network system model and evaluation framework introduced earlier.
  • D. Problem Formulation: The formulation uses U(A, I, S) as its utility function and constrains load, available spectrum, computation capacity, and intent-derived latency.Lmax denotes the sector load threshold, Br∗ available spectrum bands, κmax UPF computation capacity, and τreq(I) the extracted latency threshold.
  • D. Problem Formulation: Unlike classical optimisation, IBN begins with unstructured natural-language intents that must be mapped to technical constraints and resolved against network state S.Examples include translating “ultra-low latency” and “high density” into concrete requirements.
  • D. Problem Formulation: The primary challenge is extracting relevant constraints and implicit numerical thresholds from language rather than solving optimisation with known parameters.Traditional optimisation assumes explicit decision variables, objectives, and constraints, whereas IBN must determine which requirements matter, such as latency versus capacity.
  • D. Problem Formulation: This semantic gap motivates an agentic AI architecture in which LLMs bridge natural-language intents and formal technical specifications.The LLMs serve as the reasoning layer for the architecture presented in Section IV.

IV. AGENTIC AI ARCHITECTURE AND IMPLEMENTATION … C. Agent Prompt Engineering

The paper implements a hierarchical, ReAct-inspired multi-agent architecture that translates high-level intents into feasible network configurations through orchestrated specialist reasoning. Sequential coordination, agent-specific prompts, structured state and outputs, and explicit operational constraints support interpretable, modular, and auditable automation.

  • IV. AGENTIC AI ARCHITECTURE AND IMPLEMENTATION: The implementation combines agent organisation, coordination, prompt engineering, network-state representation, and iterative reasoning to enable autonomous intent translation.
  • A. Hierarchical Multi-Agent Architecture: Three LLM-based agents form a two-tier hierarchy: one central orchestrator and two domain-specific specialist agents.
  • 1) Orchestrator Agent:: The orchestrator parses intents into domain-specific sub-problems, delegates them, synthesises recommendations, executes configurations, and iteratively refines plans through Thought-Action-Observation cycles.
  • 2) Specialist Agents:: Specialists reduce monolithic-model context and reasoning bottlenecks: the RAN agent recommends spectrum policies from sector state, while the Core agent evaluates UPF placement against latency and capacity constraints.Specialist knowledge can be refined independently without retraining the orchestrator.
  • B. Sequential Coordination Protocol: The sequential protocol parses requirements, consults RAN for physical-layer feasibility, conditions Core consultation on the RAN decision, and performs a final consistency check over the recommendations.Core selection uses latency profiles from matrix T and compute availability κc to select the optimal UPF node c∗.
  • B. Sequential Coordination Protocol: Sequential coordination preserves reasoning traces for auditability, permits independent specialist upgrades, and propagates constraints from RAN decisions into Core decisions.
  • C. Agent Prompt Engineering: Agent-specific system prompts assign expert roles and encode operational constraints, while the orchestrator prompt enforces specialist delegation, tool use, ReAct formatting, and pre-provisioning synthesis.The orchestrator is defined as a “6G Intent Coordinator” and uses call_agent and provision_slice tools.
  • C. Agent Prompt Engineering: Specialist prompts encode quantitative RAN and Core policies, require structured recommendations with reasons, and remain external Markdown files for rapid iteration and version control.Rules include prioritising millimetre-wave spectrum above 10k active users, selecting MEC for URLLC below 10 milliseconds, and verifying UPF compute load below 85 percent.

D. Network State Representation · E. ReAct Loop Implementation

The system represents network conditions as structured JSON and injects this verified state into specialist-agent contexts. An orchestrator then applies a bounded, deterministic ReAct loop to consult specialists, provision slices, or terminate with a failure report.

  • D. Network State Representation: The experimental network state is implemented as a structured JSON object that captures essential system-model characteristics for controlled agentic reasoning evaluation.This mock representation avoids the complexity of interfacing with production network-management platforms.
  • D. Network State Representation: The JSON state contains sector and node mappings, with sector entries specifying active users, load percentage, and per-band spectrum availability.Spectrum availability is represented for mmWave, mid_band, and low_band frequencies.
  • D. Network State Representation: When specialists are invoked, the system loads their Markdown system prompt, serializes the network state, and appends it to the orchestrator’s request.The specialist receives the task description, current JSON state, and a request for an expert recommendation in one context.
  • D. Network State Representation: Injecting the complete environmental state grounds specialist recommendations in verified conditions and helps prevent hallucinated resource availability without iterative tool calls.The stated purpose is to ensure recommendations respect physical constraints.
  • E. ReAct Loop Implementation: The orchestrator initializes conversation history with system and intent messages, then repeatedly queries the LLM in a loop limited to 10 iterations.The iteration limit prevents infinite loops.
  • E. ReAct Loop Implementation: The loop parses each response for CALL_AGENT, PROVISION_SLICE, or FINISH actions, consulting specialists and appending observations, executing extracted configurations, or terminating respectively.Specialist consultation receives the network state, while provisioning extracts configuration parameters before execution.
  • E. ReAct Loop Implementation: LLM calls use temperature 0.0 with Cerebras Cloud’s Llama-3.3-70B model, while token usage is tracked and aggregated for computational-efficiency analysis.The deterministic setting is intended to support reproducibility across experimental runs.

V. EXPERIMENTS · A. Experimental Setup · B. Performance on Benchmark Scenarios

The experiments evaluate the hierarchical agentic AI framework on 12 intent-based 6G network-slice scenarios using repeated runs and Semantic Accuracy and Engineering Utility metrics. The system achieves feasible provisioning with efficient reasoning, while prompt revisions modestly improve semantic accuracy and expose deployment trade-offs in prompt specificity.

  • A. Experimental Setup: The evaluation covers 12 intent-based network-slice provisioning scenarios spanning three foundational 6G use-case categories.The benchmark includes four URLLC, four eMBB, and four mMTC scenarios.
  • A. Experimental Setup: The mock deployment models a metropolitan-scale 6G network with heterogeneous RAN sectors and a three-tier core compute architecture.The topology includes five sectors and edge-tier nodes with differing latency and compute-load characteristics.
  • A. Experimental Setup: Each scenario is executed three times, with results averaged using Semantic Accuracy and Engineering Utility metrics.
  • B. Performance on Benchmark Scenarios: The proposed architecture reaches mean Semantic Accuracy of 0.667 and mean Engineering Utility of 0.747 across the benchmark.Four scenarios receive Perfect Match scores of 1.0, seven receive Partial Match scores of 0.5, and Engineering Utility exceeds the 0.70 feasibility threshold.
  • B. Performance on Benchmark Scenarios: 13,281 tokens per decision and 3.8 seconds mean latency characterize computational efficiency, with convergence within 3.0 ReAct iterations.
  • B. Performance on Benchmark Scenarios: Lower accuracy for regional placement of delay-tolerant traffic results from insufficiently precise Core-specialist policy encoding.The specialist consistently selected the lowest-latency option because the prompt did not precisely distinguish traffic-placement requirements.
  • B. Performance on Benchmark Scenarios: Revising the prompt to encode traffic-aware placement policies increases Semantic Accuracy from 0.667 to 0.708.The revision reserves edge nodes for safety-critical URLLC, prefers regional centres for eMBB streaming at 30–50 ms tolerance, and mandates regional placement for high-tolerance mMTC.
  • B. Performance on Benchmark Scenarios: Prompt design must balance ambiguity against computational cost, contradictory instructions, brittleness, and limited generalisation across deployment contexts.The optimal level of prompt detail remains an open research question varying with domains, model capabilities, and deployment contexts.

C. Qualitative Analysis: System Reasoning Behaviour … D. Comparative Analysis Against Baseline Approaches

The qualitative trace shows bounded, specialist-coordinated reasoning that infers implicit URLLC requirements and produces auditable network decisions. Comparative results show higher semantic accuracy than monolithic, rule-based, and direct-prompting baselines, while those baselines retain efficiency or constraint-adherence advantages.

  • 2) Analysis and Architectural Validation:: The trace demonstrates intent parsing beyond keyword matching by inferring URLLC requirements from safety-critical operations and generating traffic-type-aware specialist queries.This supports the claimed architectural properties of bounded reasoning, sequential consultation, and constraint propagation.
  • C. Qualitative Analysis: System Reasoning Behaviour: The industrial automation trace tests implicit URLLC recognition and edge-node selection for a safety-critical robotic assembly line requiring latency below 5ms and high reliability.The standard specifies mid-band spectrum and edge-node UPF placement for the sub-5ms requirement.
  • 1) Execution Summary:: The system converges in three iterations through sequential orchestrator coordination, with RAN consulted before Core to propagate constraints.The trace is presented as a condensed execution summary, with complete output available in Appendix A.
  • 2) Analysis and Architectural Validation:: Semantic Accuracy is 1.0 (Perfect Match), while Engineering Utility is 0.94, comprising Slatency = 1.0 and Sresource = 1.0.Scongestion = 0.40.
  • 2) Analysis and Architectural Validation:: 13,573 total tokens and 4 seconds decision latency indicate reasoning depth associated with constraint complexity, while explicit justifications support operator auditing.Examples include trade-offs for mmWave interference and rejection of a 15ms-latency edge node.
  • D. Comparative Analysis Against Baseline Approaches: The multi-agent system achieves 0.67 Semantic Accuracy versus 0.62 for the monolithic baseline, an 8% improvement, while the monolithic approach reduces tokens by 86%.The monolithic agent shows higher decision variance despite identical network-state information and tool access.
  • D. Comparative Analysis Against Baseline Approaches: The rule-based system achieves 0.75 Engineering Utility but only 0.46 Semantic Accuracy, a 31% degradation relative to the multi-agent system, because it struggles with language variation.Its competitive utility reflects strict adherence to predefined engineering constraints.
  • D. Comparative Analysis Against Baseline Approaches: The direct LLM prompting baseline records 0.58 Semantic Accuracy and 0.66 Utility, but offers 1,157 tokens and 0.6 seconds through less structured reasoning.Its lack of iterative refinement and specialist consultation leads to premature decisions that frequently violate feasibility constraints.

E. Ablation Study Results · VI. DISCUSSIONS · VII. CONCLUSION

Ablations identify ReAct iteration as the largest contributor to performance, followed by engineered prompts and specialist decomposition, while discussions highlight accuracy–efficiency tradeoffs, prompt-induced biases, and deployment limitations. The conclusion presents hierarchical multi-agent orchestration as a promising, interpretable approach for translating 6G intents into executable network configurations.

  • E. Ablation Study Results: Removing engineered system prompts reduces Semantic Accuracy by −0.09 and causes inconsistent threshold application and missed cost-optimisation opportunities.The prompts encode quantitative rules such as >10k users requiring mmWave and load >80% triggering warnings.
  • E. Ablation Study Results: Removing specialist agents reduces Semantic Accuracy by 0.05, showing measurable but smaller benefits from decomposing RAN and Core expertise.The limited effect is attributed to the tested scenarios having relatively independent RAN and Core decisions.
  • E. Ablation Study Results: Removing ReAct iteration causes the largest degradation, reducing Semantic Accuracy by −0.13 and Utility by −0.20.Utility falls from 0.75 to 0.55 because single-pass decisions cannot incorporate specialist feedback or refine constraint-violating configurations.
  • VI. DISCUSSIONS: The corrected Core specialist “latency greedy” behaviour shows that prompt-induced biases can create systematic, compounding failure patterns rather than random errors.These tendencies differ from traditional software bugs that cause immediate crashes.
  • VI. DISCUSSIONS: The multi-agent system achieves +8% semantic understanding over monolithic agents but requires 7× more tokens and 5× longer execution time.The discussion proposes routing complex intents to multi-agent systems, straightforward intents to lightweight monolithic agents, and repetitive patterns to cached rule-based configurations.
  • VI. DISCUSSIONS: The ablations rank ReAct refinement first, engineered prompts second, and specialist decomposition third, although the ordering may not generalise beyond moderate-complexity intents with largely decoupled RAN/Core decisions.The findings therefore prioritise reasoning protocols and precise prompt engineering over organisational structure.
  • VI. DISCUSSIONS: The proposed system uses 13,281 tokens versus 1,894 for the monolithic system, costing approximately 0.013 USD per provisioning decision at current Cerebras pricing.Suggested cost-reduction directions include prompt compression, knowledge distillation, and caching strategies.
  • VII. CONCLUSION: The paper concludes that hierarchical multi-agent agents can translate natural-language 6G intents into executable network-wide configurations while preserving transparency through interpretable reasoning traces.The architecture coordinates an orchestrator with two domain-specific specialists through iterative ReAct reasoning, with O-RAN identified as a promising deployment target despite evaluation using mock network data.

APPENDIX A. Complete Reasoning Trace for URLLC Industrial Automation Scenario

The trace converts an implicit URLLC requirement for robotic assembly into a network slice using mid-band spectrum and an edge UPF placement. It enforces <5ms latency and high reliability using industrial context and network-state evidence.

  • Requirements: The robotic assembly slice at industrial_park_a requires ultra-low latency (<5ms) and high reliability for safety-critical real-time control.These requirements are inferred from the industrial automation context.
  • Golden Standard Configuration: The golden configuration selects mid_band spectrum and mec_industrial_1 as the UPF node.Mid-band balances reliability and capacity, while the co-located edge node achieves 3ms latency.
  • Network State Context: mec_industrial_1 provides 3ms latency, satisfying the <5ms requirement for the industrial slice.The latency matrix lists mec_stadium_1 at 15ms, metro_agg_hub at 10ms, and regional_dc_north at 35ms.
  • Network State Context: The industrial_park_a state includes 500 active users, 60% load, and available spectrum of 200 MHz mmWave, 100 MHz mid-band, and 10 MHz low-band.The sector is a high-interference factory zone with critical reliability requirements.
  • Complete Execution Trace: The first reasoning iteration consults RAN and Core specialists to determine spectrum allocation and UPF placement for the latency and reliability requirements.The RAN consultation explicitly compares mmWave, mid-band, and low-band options for industrial_park_a.

**DECISION PROCESS**

The decision process recommends mid-band for industrial_park_a because it balances capacity, penetration, and reliability under moderate density and 200 MHz mmWave availability. It places the UPF at mec_industrial_1 for 3ms latency, provisions the slice, and reports a perfect semantic match with 0.94 engineering utility.

  • RAN decision: Mid-band is recommended for industrial_park_a because its capacity and penetration balance better supports reliability than mmWave’s limited 200 MHz availability.The recommendation targets an automated robotic assembly line with ultra-low latency and high reliability requirements.
  • RAN decision: Mid-band’s 100 MHz availability should support 500 active users, but the sector’s 60% load requires monitoring to avoid congestion and possible spectrum reassignment or upgrades.Efficient bandwidth management is assumed, while mmWave range and penetration limitations remain concerns in the industrial environment.
  • Core decision: 3ms latency from mec_industrial_1 is the only listed option meeting the automated assembly line’s <5ms requirement, so the UPF is deployed there.Other listed latencies are 15ms, 10ms, and 35ms, and the recommendation accounts for reliability and current network state.
  • Slice provisioning: The configured slice industrial_autonomy_001 uses mid-band at industrial_park_a and UPF@mec_industrial_1, with monitoring recommended at 60% sector load.The configuration combines RAN reliability and capacity with Core latency performance for the automated robotic assembly line.
  • Evaluation: Semantic Accuracy: 1.0 and U = 0.8 × 1.0 + 0.1 × 1.0 + 0.1 × 0.40 = 0.94, with decision latency of 4 seconds and 2 specialist consultations.The golden-standard comparison identifies a perfect match across all three dimensions; consultations comprised one RAN and one Core specialist.
Loading 2601.06640v1…