Source-linked AI summary
Advanced LLM-Enhanced Intent-Based 5G Network Management using Dynamic Semantic Routes
Thomas Benton Townsend, Dimitrios Michael Manias
TL;DR
The paper addresses the challenge of translating natural-language operator intents into machine-readable actions for increasingly complex 5G+ networks. It introduces a semantic router with dynamic routes for intent identification and detail extraction, evaluating multiple encoders and extraction performance. Static route selection exceeds 99% collective accuracy, while RNR and IRR detail extraction reach 97.4% and 95% accuracy, respectively.
Problem
Increasingly distributed and dynamic 5G+ networks make human-centric management difficult, motivating natural-language intent translation for Human-on-the-Loop networking.
Method
The paper uses a semantic router with dynamic routes to identify 5GC intents and extract relevant details into machine-readable policy schemas.
Results
Static route selection achieved collective accuracy exceeding 99%, with individual accuracies reaching 100%.
Takeaways & Limitations
The system effectively generates schema outputs and accurately extracts relevant information for LLM-enhanced intent-based networking in the 5G+ Core.
Abstract
from arXiv · showhide
As the use of Artificial Intelligence (AI) and Large Language Models (LLMs) is becoming common in everyday applications, their ability to interpret natural language has increased significantly. An emerging application of AI is integration with network management and orchestration practices. An instance of this integration is LLM-enhanced intent-based networking, where network operators will control a network using natural language. This work presents the use of dynamic routes with a semantic router to identify an intent from a network operator's prompt and extract necessary details for intent fulfillment in intent-based 5G+ core networks. Furthermore, the performance of static route selection is assessed by evaluating multiple encoders and dynamic route detail extraction accuracy against a series of realistic operator prompts. The presented results show that static and dynamic routes are successful in detail extraction and schema formatting.
I. INTRODUCTION
Modern 5G+ networks are increasingly distributed, dynamic, and complex, making human-centered management difficult to sustain. The paper advances LLM-enhanced intent-based networking by using dynamic routing to translate operator intents into machine-readable policies.
- Motivation: 5G+ networks are expanding in users, devices, performance requirements, and distribution, increasing the difficulty of human-centric management.The paper identifies network scale and complexity as central management challenges.
- Motivation: Zero-touch network and Service Management aims for self-configuration, self-optimization, and self-healing, but current systems must progress through Human-on-the-Loop management.In this intermediate model, operators convey goals and requirements as intents while the network acts on them.
- Motivation: Human-on-the-Loop networking requires translating natural-language intents into machine-readable actions.LLMs provide an opportunity to support this transition from Human-in-the-Loop management toward zero-touch automation.
- Contribution: The paper advances prior 5G+ Core networking work by improving intent identification and replacing static semantic routes with dynamic routes.Dynamic routes extract and format key intent parameters into policy schemas for direct network use.
- Paper organization: The paper is organized into state-of-the-art review, system model, methodology, results, and conclusion sections.The section sequence is specified by the authors.
II. RELATED WORK
Prior work positions LLMs across network design, optimization, security, and automation, with intent-based networking emerging as a major application. This paper extends static semantic routing with dynamic routes that produce machine-readable schemas for 5G+ Core management.
- LLM applications: LLM applications in networking span design, optimization, security, and automation, including configuration generation, algorithm design, performance enhancement, and intrusion detection.The cited synthesis organizes network-related LLM integration into four broad categories.
- Intent-based networking: Intent-based networking research covers intent profiling, translation, resolution, activation, and assurance within a closed-loop automation cycle.Approaches include templates, graphical interfaces, natural-language processing, intent-based languages, and machine learning.
- Prior work: Earlier 5GC work identified six intent types and highlighted stochastic LLM outputs and dependence on proprietary models as limitations.Subsequent work addressed these issues with a semantic router, static routes, and open-source LLMs.
- This work: The paper converts semantic-router static routes into dynamic routes that translate natural-language intents into machine-executable policies.The stated target is progress toward zero-touch management and real-time provisioning and enforcement.
- Contributions: The proposed system combines an end-to-end semantic-router model with dynamic routes for 5GC intent extraction.Its contributions include route functions that generate machine-readable schema output.
- Contributions: A comparative analysis evaluates five encoder options for intent extraction and route selection.The evaluated options include four standalone encoders and a hybrid encoder.
III. SYSTEM MODEL
The system has user, route-selection, and route-execution subsystems: a semantic router identifies an intent, then a dynamic route extracts details and formats them into a schema. Evaluation focuses on RNR and IRR intents under explicit input assumptions.
- System architecture: The system model contains user, route-selection, and route-execution subsystems initiated by a network operator’s message.The operator message is sent to the semantic router before route selection and execution.
- Target intents: The system focuses on Regular Notification Request and Intent Report Request intents as foundations for more complex intent and schema structures.These intents demonstrate the proposed approach’s functionality.
- User Input: User inputs are assumed to contain the details required to create a schema, including a subject and frequency for RNR requests.IRR inputs must provide an intent identifier or a continuous positional statement such as “latest intent” or “last 5 intents”.
- Route Selection: The semantic router selects a route from the operator’s message using an encoder-generated similarity score and route-layer thresholds.The system evaluates HF, AB, MS, OA, and HY encoders before and after training across varied prompts.
- System architecture: Figure 1 presents the high-level system overview with subsystem labeling.The labeled architecture corresponds to the user, route-selection, and route-execution stages described in the system model.
- Route Execution: Dynamic routes execute functions that retrieve key information from user messages after route selection.RNR extraction includes the target network element or network partition and notification frequency, while IRR extraction identifies an intent ID or positional statement.
- Route Execution: Figure 2 presents examples of user inputs and resulting outputs for the dynamic-route extraction process.The examples include intent identifiers and associated user prompts.
IV. METHODOLOGY
The methodology combines augmented operator prompts, semantic-router route selection, and dynamic functions that extract intent details into machine-readable outputs for RNR, IRR, and None routes.
- Prompt Generation: Prompts were human-written and augmented to vary length, phrasing, punctuation, and spelling across RNR, IRR, and None route examples.The resulting prompts were used for route definitions, encoder training, and performance testing.
- Semantic Router Implementation: Each route uses utterances and encoder-specific similarity thresholds to determine whether an input triggers RNR, IRR, or the default None route.Multiple route layers were created with HF, Alibaba Qwen, Mistral, and OpenAI encoders.
- Dynamic Route Implementation: Dynamic routes execute predefined functions that extract key information from the user message and return key-value schema results.The selected static route automatically invokes its associated dynamic route.
- Dynamic Route Implementation: The dynamic-route functions use the Mistral 7B model from Ollama to follow function schemas and generate the desired extraction output.The LLM receives a function name and schema describing the required result.
- Performance Evaluation: RNR extraction is evaluated by frequency and notification subject, whereas IRR extraction is evaluated by intent identifiers or positional identifiers for previous intents.RNR frequency values may require time conversion; IRR positional identifiers are represented as negative integers.
V. RESULTS AND ANALYSIS
The results section compares encoder accuracy and thresholds before and after training across RNR, IRR, and None route tests.
- Results Overview: Accuracy and thresholds were compared before and after training for each encoder.The analyses were divided into RNR, IRR, and None route performance tests.
- Results Overview: The evaluation covers three route categories: RNR, IRR, and None.Each category represents a separate performance test.
- Results Overview: Post-training results are organized around both encoder accuracy and route-selection thresholds.The comparison examines changes associated with training across the three tests.
A. Static Route Performance
Static-route performance was evaluated across RNR, IRR, and None tests using pre- and post-training encoder accuracy and thresholds. OpenAI showed the strongest combination of accuracy and stability across the tests.
- RNR Performance: HF improved significantly after training in the RNR test, while MS and AB declined after their thresholds became more restrictive.HY performed slightly better than OA, but OA was more consistent because of its lower standard deviation.
- IRR Performance: In IRR, HF improved after training but became less stable, while OA achieved the highest post-training performance and remained consistent.MS and AB also declined, and HY lost accuracy after its threshold increased.
- None Performance: HF, AB, and OA maintained high None-route accuracy before and after training, whereas MS and HY remained comparatively inaccurate and unstable.None-route accuracy matters because misclassified irrelevant prompts could pass inaccurate values to 5GC fulfillment functions.
- Overall Static Route Performance: OpenAI maintained high accuracy and high stability across every static-route performance test.Its lower error and consistency made it the strongest overall option among the evaluated encoders.
B. Dynamic Route Performance
Dynamic-route extraction was tested with 40 RNR and 40 IRR prompts using the OA encoder and Mistral 7B. Extraction accuracy was high, with errors concentrated in time conversion and positional intent intervals.
- RNR Performance: 97.4% of RNR samples were extracted correctly in full, with 38 of 39 valid route samples succeeding.One misclassified sample was excluded because the experiment targeted parameter extraction rather than route selection.
- RNR Performance: The sole RNR extraction error converted 60 minutes to minutes:60 instead of hours:1 and minutes:0.The paper states that post-processing checks could identify and correct such conversion errors.
- IRR Performance: IRR extraction achieved 95% accuracy, with two unexpected outputs among 40 prompts.Both errors involved positional statements, while intent-ID extraction was 100% successful across relevant samples.
- IRR Performance: The IRR positional errors involved incomplete or shifted negative-integer intervals for previous intents.Future post-processing is intended to identify non-sequential positional parameters and correct them.
C. Comparison to SOTA
Compared with prior static-route evaluation, this work broadens validation and adds dynamic routes for parameter extraction. Static route selection exceeds 99% collective accuracy, with individual accuracies reaching 100%.
- 94% accuracy in prior work was measured across a limited validation study using the best-performing encoder.
- The expanded evaluation measures route-selection performance individually rather than only collectively and increases the semantic router’s base performance.
- Collective static route-selection accuracy exceeds 99%, while individual accuracies reach 100%.
- Dynamic routes address a prior limitation by enabling reliable parameter extraction alongside improved static route selection.
VI. CONCLUSION
The paper introduces dynamic semantic routing into an LLM-enhanced 5G+ Core intent-extraction architecture to extract and format prompt details. The system accurately extracts relevant information and generates schema outputs, while intent fulfillment and live-network integration remain future work.
- Dynamic semantic routing extends the prior intent-extraction architecture for 5G+ Core management by extracting and formatting relevant prompt details.
- The system was tested with multiple encoders, including a hybrid encoder, for intent-extraction performance.
- The system effectively generates schema outputs and accurately extracts relevant information.
- The study focuses on schema output generation; implementing intent fulfillment, dedicated virtualized infrastructure, and live 5G+ integration remain future requirements.