PaperlayerView exact revision on arXiv

Source-linked AI summary

Intent2Tc: Automated Intent-to-Traffic Control Translation with Language Models

Andrea Masini, Sudipta Acharya, Paolo Bellavista, Luca Foschini, Burak Kantarci

arXiv:2609.31397v1cs.NIcs.AIcs.CL

TL;DR

Intent2Tc addresses the difficult, error-prone translation of high-level service intents into executable network configurations. It uses a closed-loop language-model framework to generate Linux tc rules, and evaluation on 100 RFC 9315-compliant intents reports high translation quality alongside efficiency gains from RAG.

  • Problem

    Translating business-level intents into executable network configurations remains complex, error-prone, and difficult to automate.

    Source: §I · p. 1

  • Method

    Intent2Tc combines an AQM-based digital-twin semantic model, metadata extraction, critique-driven refinement, and RAG-based knowledge reuse to translate intents into deployable Linux tc rules.

    Source: §I · p. 1, §III · p. 3

  • Results

    Claude Sonnet-4.6 achieves 0.98 semantic similarity, 1.0 semantic unit coverage, and 0.045 normalized edit distance on the 100-intent evaluation.

    Source: §I · p. 1

  • Takeaways & Limitations

    RAG reduces token consumption and inference latency, while compact models such as Phi-4-mini achieve performance comparable to larger LLMs in sub-intent and tc generation.

    Source: §V · p. 6

Abstract

from arXiv · show

Automated and highly usable Quality-of-Service (QoS) enforcement requires translating high-level service intents into deployable traffic-management policies. Although intent-based networking (IBN) has simplified policy specification, bridging the gap between business-level intents and executable network configurations remains complex, error-prone, and difficult to automate. This paper presents Intent2Tc, a closed-loop language-model-driven framework that translates business-level traffic-shaping intents into declarative sub-intents and subsequently into validated, executable Linux traffic control (tc) configurations. The framework integrates an Active Queue Management (AQM)-based digital twin (DT) semantic model, automated metadata extraction, critique-driven refinement, and Retrieval-Augmented Generation (RAG)-based knowledge reuse to improve semantic consistency and configuration reliability. We evaluate multiple open-source large language models (LLMs) and small language models (SLMs), together with Claude Sonnet-4.6, on 100 Request for Comments (RFC) 9315-compliant traffic-shaping intents. Across both translation stages, Intent2Tc achieves high semantic fidelity, configuration accuracy, and deployment readiness, with Claude Sonnet-4.6 reaching 0.98 semantic similarity, 1.0 semantic unit coverage, and 0.045 normalized edit distance. Furthermore, RAG reduces token consumption and inference latency while enabling compact models such as Phi-4-mini to approach the performance of substantially larger models. Linux tc serves as the target configuration platform, demonstrating the practical applicability of the proposed framework.

I. INTRODUCTION

Intent2Tc targets the largely manual terminal translation from high-level intents to Linux tc rules. It combines semantic modeling, critique-driven correction, and retrieval of validated prior translations.

I. INTRODUCTION

  • Intent2Tc addresses the largely manual terminal translation phase of intent-based networking by translating natural-language traffic-shaping intents into Linux tc rules.Earlier LM-based intent translation still relied on static keyword matching and a single-pass critique.Source: §I · p. 1
  • The workflow uses an AQM-based digital twin as a semantic model, then extracts metadata, generates sub-intents and tc rules, and critiques each translation stage.Corrected intents, sub-intents, and tc rules are added to a RAG database indexed by traffic profile and time bounds.Source: §I · p. 1
  • The framework contributes a scalable AQM priority-queue digital twin, a database of validated translations, and a closed-loop translation pipeline.The pipeline combines dynamic traffic-profile matching, confidence-based metadata extraction, critique-driven correction, and RAG-assisted reuse.Source: §I · p. 1
  • On 100 RFC 9315-compliant intents, Claude Sonnet-4.6 achieves 0.98 semantic similarity, 1.0 semantic unit coverage, and 0.045 normalized edit distance.The evaluation includes open-source LLMs and SLMs, with and without RAG; the reported results also show reduced token use and inference latency with RAG.Source: §I · p. 1

II. RELATED WORK

Prior intent-to-tc work leaves limited room for iterative refinement and knowledge reuse. Intent2Tc uses a digital twin, staged language-model translation, critique, and a RAG database to produce deployment-ready rules.

II. RELATED WORK

  • Existing approaches improve intent interpretation and configuration generation, but largely use single-pass translation; also identifies iterative learning as a missing component.Linux tc’s strict syntax and kernel-level complexity make manual configuration difficult, while technically plausible but incorrect LM configurations remain a concern.Source: §II · p. 2

III. METHODOLOGY

  • The digital twin models a finite-capacity, non-preemptive priority queue calibrated with queueing theory, using CoDel to favor dropping low-priority packets under overload.The simulation uses Poisson arrivals and service rates tuned to target utilization; its performance metrics link traffic objectives to feasible system behavior.Source: §III · p. 2
  • The language-model pipeline uses the intent, confidence-based lookup table, semantic model, and traffic-profile taxonomy to extract metadata and generate declarative sub-intents.Sub-intents encode performance, priority, AQM, and temporal requirements; RAG retrieves two profile-matched examples when enabled, otherwise the system uses generic few-shot examples.Source: §III · p. 2
  • A deterministic critique module removes invalid or inconsistent sub-intents, fills missing requirements, and merges redundant entries before tc-rule generation.The language model then generates rules using traffic-profile attributes, the semantic model, and retrieved or few-shot examples; a second critique enforces tc-specific constraints and syntax.Source: §III · p. 3
  • When enabled, RAG adds a corrected translation only if its Token-F1 exceeds at least one existing record, progressively populating the database with validated mappings.The final output is a validated, deployment-ready set of Linux tc rules.Source: §III · p. 3

A. Dataset and Evaluation Metrics

The benchmark contains 100 transformed, RFC 9315-compliant intents with validated ground-truth sub-intents and Linux tc configurations. Evaluation measures translation quality and deployment readiness.

A. Dataset and Evaluation Metrics

  • The 100-intent benchmark spans traffic profiles, QoS objectives, and temporal requirements, with ground-truth sub-intents and tc configurations validated by network administrators and traffic engineers.The intents were transformed from a business-intent dataset while preserving operational objectives and abstracting implementation-specific parameters.Source: A. Dataset and Evaluation Metrics · p. 3
  • Dataset statistics cover latency, bandwidth, priority, time- or load-aware control, and packet-drop objectives across diverse application domains and single- or multi-objective policies.The objectives emphasize latency and priority; policies include both always-active and time-sensitive requirements.Source: A. Dataset and Evaluation Metrics · p. 3
  • Evaluation uses semantic, lexical, and deployment-readiness metrics for sub-intent and tc-rule generation, with automated FEACI counterparts supplementing measures requiring human assessment.Sub-intent metrics include SBERT, ROUGE-L, Token-F1, and METEOR; tc-rule metrics include semantic unit coverage, Token-F1, and normalized edit distance.Source: A. Dataset and Evaluation Metrics · pp. 3–4

B. Case Study: Stabilize Vehicular Traffic Data Transmission

The vehicle-to-everything case study shows critique-driven corrections to both sub-intents and tc rules before deployment.

B. Case Study: Stabilize Vehicular Traffic Data Transmission

  • The critique module splits a malformed sub-intent into separate profile/time-bound detection and priority-assignment sub-intents.This correction occurs in Phase 2-B.Source: B. Case Study: Stabilize Vehicular… · p. 4
  • For tc rules, the critique module removes a hallucinated filter for low-priority HTB class 1:11 and inserts the missing TIME_WINDOW application annotation.These corrections are made before deployment in Phase 2-C.Source: B. Case Study: Stabilize Vehicular… · p. 4

C. Analysis of Results

The evaluation uses static few-shot and dynamic RAG-based prompts for sub-intent generation and translation into Linux tc rules.

C. Analysis of Results

  • The evaluations compare static few-shot and dynamic RAG-based prompting for generating sub-intents and translating them into Linux tc rules.Source: C. Analysis of Results · p. 4

1) Dynamic Metadata Extraction Analysis:

Dynamic metadata extraction is compared with the static keyword-matching logic of using Claude Sonnet-4.6.

1) Dynamic Metadata Extraction Analysis:

  • 95% traffic-profile identification and 94% time-sensitive constraint detection accuracy, versus 70% and 59% with static keyword matching, respectively.The evaluation uses Claude Sonnet-4.6 against ground-truth metadata.Source: 1) Dynamic Metadata Extraction Analysis · p. 4

2) Sub-Intent Generation Results:

Intent-to-sub-intent generation performs strongly across models, and tc-rule generation is also strong, with smaller models competitive and RAG generally beneficial.

2) Sub-Intent Generation Results:

4) Token Usage and Inference Time Analysis:

The framework improves both translation stages over the single-pass pipeline of while reducing resource use; RAG also lowers token consumption and latency against the Base setting.

4) Token Usage and Inference Time Analysis:

  • For Claude Sonnet-4.6, RAG lowers total simulation cost from approximately $1.87 to $1.72, averaging about 375 fewer tokens and 1.25 seconds less per intent.Hermes-2-Pro and Phi-4-mini show similar efficiency gains while maintaining output quality.Source: 4) Token Usage and Inference Time… · p. 6
  • 0.07, 0.06, and 0.06 gains in Phase 2-B ROUGE-L, Token-F1, and SBERT, plus 0.065 Token-F1 gain and 0.218 NED reduction in Phase 2-C, versus.The comparison uses Claude Sonnet-4.6.Source: 4) Token Usage and Inference Time… · p. 6
  • Approximately 23,500 fewer tokens per intent, 2.2 seconds lower latency, and $6 lower full-dataset cost versus.Source: 4) Token Usage and Inference Time… · p. 6

V. CONCLUSION

Intent2Tc translates traffic-shaping intents into deployable Linux tc configurations through digital-twin semantic modeling, critique-driven correction, and RAG-based knowledge reuse. The conclusion reports high translation quality, greater efficiency with RAG, and competitive results from compact models.

V. CONCLUSION

  • Across five independent runs, Intent2Tc demonstrated strong semantic fidelity, configuration accuracy, and deployment readiness.The reported results are average values across the runs.Source: §V · p. 6
  • RAG augmentation reduced token consumption and inference latency while maintaining or improving translation quality.Source: §V · p. 6
  • 33% lower inference time and approximately 50% smaller model size let Phi-4-mini perform comparably to larger models in sub-intent and tc generation.The conclusion also reports lower RAM and VRAM usage for Phi-4-mini.Source: §V · p. 6
  • Future work includes telemetry-driven semantic modeling, persistent multirun RAG knowledge bases, larger real-world evaluations, and critique-module extensions for edge-case rule conflicts.Source: §V · p. 6

Related papers

Loading 2609.31397v1…