Text2Net: Transforming Plain-text To A Dynamic Interactive Network Simulation Environment

Abbaas Alif Mohamed Nishar; Alireza Marefat; Ashwin Ashok

REVIEW 4 major objections 5 minor 13 references

Text2Net: Transforming Plain-text To A Dynamic Interactive Network Simulation Environment

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Text2Net is a system that converts plain-English network descriptions into working emulated topologies, and its three-test comparison reports a 45 to 58 percent reduction in configuration time over manual setup.

desk verdict Useful LLM-to-EVE-NG proof-of-concept with a clean architecture, but the time-savings claim rests on single-run timings and no accuracy data; worth refereeing with major revision. read the letter →

arxiv 2502.15754 v1 pith:HI6CWUOM submitted 2025-02-10 cs.NI cs.AIcs.LG

Alireza Marefat , Abbaas Alif Mohamed Nishar , Ashwin Ashok This is my paper · ORCID

classification cs.NIcs.AIcs.LG

keywords text-to-networksimulationlargelanguagemodelsnaturalprocessingnetworkconfigurationautomationstaticroutingemulationeducationaltechnology

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces Text2Net, a pipeline that takes a plain-English description of a network topology, uses an instructed large language model to turn it into structured command strings, parses those into JSON, and feeds the JSON to a network emulator to provision the topology automatically. The central claim is that this makes deploying network simulations much faster than manually configuring the emulator: in three static-routing scenarios, deployment took 110, 250, and 310 seconds versus 200, 510, and 730 seconds manually, a 45 to 58 percent saving. The authors also report a qualitative survey rating ease of use and educational value at 4.66 out of 5. The point of the paper is that students and professionals can focus on network design rather than vendor-specific syntax.

What carries the argument

The central mechanism is the Structured Command String (SCS), a short command segment containing one or a few key-value pairs that an instructed large language model produces from a plain-text scenario description. A regex-based extractor then converts these SCS segments into a JSON blueprint for the emulator, and this two-stage design is what lets different wordings of the same topology converge to the same configuration.

What would settle it

Give Text2Net 50 plain-English topology descriptions that vary in wording, detail level, and phrasing, including ambiguous or partially invalid IP addresses, and count how many produce a complete deployable topology without human correction; if the success rate is far below perfect, the timing comparison overstates real-world savings.

Watch

Extended reading notes

Core claim

Text2Net demonstrates that a natural-language description of devices, interfaces, IP addresses, and static routes can be carried through an LLM-and-regex pipeline into a working emulated topology. The load-bearing result is the timing comparison: across all three test scenarios, Text2Net needed fewer steps and less time than the manual best case, with the gap widening as complexity grows from one to three routers. The system also detects missing information, such as absent static-route details, and asks the user for it before generating command strings.

Load-bearing premise

The whole speed advantage assumes that the language model will always produce command strings the regex parser can turn into valid configuration JSON, and this conversion is never measured for accuracy anywhere in the paper.

Editorial extensions

If this is right

For static-routing configurations, plain-text input alone is sufficient to produce a deployable emulated topology with no manual CLI work.
Manual configuration time grows roughly linearly as nodes are added, while Text2Net's time grows more slowly, so the advantage widens with topology complexity.
Users can describe the same network in different narrative styles and still get the same structured output, which supports use in classrooms with varied student phrasing.
Incomplete descriptions do not fail blindly: the system detects missing static-route details and asks the user for them before proceeding.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the manual timings deliberately exclude thinking and troubleshooting time, real-world manual deployments are likely to take even longer, so the actual time savings in practice may exceed 58 percent.
The unmeasured end-to-end success rate of the LLM-to-SCS-to-JSON conversion is the main robustness risk; a small pilot with diverse user phrasings could turn the timing comparison into a reliability claim.
The pipeline's clean split between extraction and emulator provisioning means a future version could add a JSON validation step to catch malformed command strings before calling the emulator.
If the paper's planned retrieval-augmented generation replaces the regex parser, the same architecture should extend beyond static routing to Layer 2 protocols and NAT/VPN scenarios.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a coherent little proof-of-concept for using an LLM to provision EVE-NG simulations from plain English, but the time-savings claim is built on three single-run measurements and no accuracy data. Treat the 45–58% numbers as illustrative, not established.

What's new: the specific pipeline — prompted LLM → Structured Command Strings → regex → JSON → EVE-NG API — is a real integration that I haven't seen in the cited work, which mostly targets configuration generation for real devices. The SCS intermediate format is a sensible design choice because it constrains the LLM's output into something the deterministic parser can handle. I also give credit for the robustness check where two different phrasings of the same topology yield the same SCS, and for the conservative manual baseline that excludes thinking and troubleshooting time. The authors are open about the static-routing scope.

Soft spots: the evaluation is the weak link. Timings are single runs per scenario — no repetition, no variance, no statistical test. So the word "significantly" does a lot of work. More important, there's no end-to-end accuracy measurement. We never learn how often the LLM produces well-formed SCS or how often the regex parser fails. If a user's phrasing produces non-conforming SCS, the pipeline either regenerates (adding latency) or fails silently. The three scenarios are hand-picked, and one deliberately incomplete, which isn't a substitute for a robustness study. The "first initiative" claim in Section II is overstated relative to CONFPILOT, ABC, and NetCR; the real novelty is the simulation provisioning angle, not LLM-to-config translation. There are no public artifacts, which makes reproduction harder. These are fixable issues, but they sit exactly on the paper's central claim.

Bottom line: the paper is a reasonable prototype description for a networking-education venue. The architecture makes sense and the limitations are acknowledged. A serious referee should ask for repeated trials, an accuracy study with varied phrasings, and a clearer success metric before the time-savings claim can be trusted.

Recommendation: send it to peer review, but expect heavy revision.

Referee Report

4 major / 5 minor

Summary. The paper introduces Text2Net, a system that takes plain-English descriptions of network topologies and, through an instructed LLM (ChatGPT-4T) and a regex/SpaCy-based extractor, produces JSON configuration files that are used to provision live network topologies in the EVE-NG emulator. The system is demonstrated on three static-routing scenarios of increasing complexity. The authors evaluate Text2Net by comparing task completion time and number of steps with manual EVE-NG configuration, and by collecting qualitative feedback from 15 participants. They report that Text2Net reduces deployment time from 200/510/730 seconds to 110/250/310 seconds across the three scenarios and claim that the system significantly reduces the time and effort required to deploy network scenarios.

Significance. If the quantitative and reliability claims were fully supported, Text2Net would be a useful contribution to network-education tooling by lowering the barrier to using realistic emulators such as EVE-NG. The paper's strengths are that it describes a concrete, end-to-end prototype with a clear system flow, it explicitly states its scope limitation to static routing, and its manual baseline is conservative in the sense that thinking and troubleshooting time are excluded. However, the central efficiency claim rests on single measurements per condition with no variance or statistical testing, and the pipeline's end-to-end correctness is never measured. The significance is therefore conditional on additional evaluation evidence.

major comments (4)

[Section IV, Figs. 6-8] The central quantitative claim is based on one timing measurement per scenario per method: Text2Net at 110/250/310 seconds versus manual EVE-NG at 200/510/730 seconds. No repeated trials, standard deviations, confidence intervals, or statistical tests are reported, so the phrase "significantly reduces" in the abstract and Section IV is not supported by the data. Please report multiple runs per scenario, describe how start and end times are defined for both conditions (including whether Text2Net times include LLM API latency, parsing, and EVE-NG provisioning), and provide a paired statistical comparison or per-run distributions.
[Sections III.B-III.D and Section IV] The evaluation contains no end-to-end accuracy or reliability metric. The pipeline's correctness depends on the instructed LLM emitting the expected SCS format and on Algorithm 1 parsing it correctly, but the only supporting evidence is three hand-written scenarios, one of which is deliberately incomplete. If a user utterance produces non-conforming SCS, the system may either regenerate (adding latency) or silently generate a wrong JSON, and such outcomes are not reflected in the reported timings. Please add an accuracy evaluation on a held-out set of paraphrased topology descriptions, reporting at least the first-attempt success rate, exact-match JSON validity, and failure-recovery behavior.
[Algorithm 1, Section III.C] Algorithm 1 is not fully specified and appears to contain errors: line 16 reads "if ace<- ExtractInterfaceDetails(l)" without a defined variable or comparison, lines 17-18 then refer to "if ace", and line 6's condition "if k has no comma" is unexplained. Because this algorithm is the core extraction routine that converts SCS into the JSON used for provisioning, please provide corrected, unambiguous pseudocode or the actual parser logic so that the construction of key-value pairs can be reproduced.
[Section IV and Abstract] The abstract and Section IV claim that the results "showcase its scalability across various network complexities," but the evaluation covers only three static-routing scenarios with up to three routers. This is too narrow to support a general scalability claim. Please either temper the statement to reflect the demonstrated scope or add experiments with larger topologies and more diverse configuration tasks.

minor comments (5)

[Section III.A] The text says the model "was trained" to interpret and generate SCS, but no fine-tuning or training procedure is described; it appears to be prompt engineering or in-context instruction. Please clarify whether any model weights were updated or whether the behavior is obtained solely through prompting.
[Section III.B] The sentence "as well as a static router on 'R-3'" should read "a static route on 'R-3'."
[Section IV] The definition of "steps" is not uniform: manual steps include launching the simulator, logging in, and creating a lab, while Text2Net's two steps are not enumerated. Please clarify what counts as a step for each method and whether Text2Net's step count includes typing the prompt and waiting for deployment.
[Section IV, Fig. 9] The qualitative survey is summarized only by an average score of 4.66 out of 5; no questionnaire items, response scales, or per-item results are provided. Please include the survey instrument and a breakdown of responses.
[General] The paper contains inconsistent spelling of the regular-expression component ("RegEX" and "regex") and refers to the model as both "ChatGPT-4T" and "GPT-4T"; please standardize the terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the time/step comparison is an empirical measurement, not a prediction reduced from fitted inputs.

full rationale

The paper's central claim is that Text2Net lowers deployment time and steps versus manual EVE-NG configuration, supported by measured timings (110/250/310 seconds vs. 200/510/730 seconds) for three scenarios. These numbers are direct observations of system runs, not quantities derived from assumptions or fitted parameters. The SCS extraction and JSON conversion pipeline is a design whose correctness is assumed rather than independently benchmarked; that is an evidence gap about reliability, not circularity. No equation in the paper defines the reported times in terms of the LLM prompt, and no fitted parameter is renamed as a prediction. The references to prior work are contextual and none is load-bearing for the evaluation; no uniqueness theorem or ansatz is imported from the authors' own prior work. The only quasi-internal step is that GPT-4T's output format is defined by prompt engineering and then parsed by the authors' regex rules, so a failed parse could invalidate a timing run in principle; however, the paper does not claim to derive the timing advantage from that definition, and the reported advantages are empirical. Therefore, under the hard rules, there is no exhibited reduction of a claimed result to its own inputs, and the correct finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claims rest on unverified assumptions about LLM output stability, parser correctness, user input format, and EVE-NG API reliability. These are typical engineering assumptions for an LLM-based pipeline, but the paper does not quantify failure rates or test robustness.

assumptions (4)

domain assumption The instructed LLM (OpenAI ChatGPT-4T) will consistently produce Structured Command Strings in the expected format for natural-language topology descriptions.
Invoked in Section III.A and shown in Figure 4 for three hand-written scenarios; no held-out accuracy test is reported. The regex parser in Section III.B depends on this format stability.
domain assumption The regex and pattern-matching extraction (Algorithm 1) correctly parses every SCS the LLM returns.
Section III.B and Algorithm 1 assume SCSs decompose cleanly into key-value pairs; malformed or unexpected strings would break provisioning. The paper does not report failure cases.
domain assumption Users will describe topologies in the standard input format established by a qualitative survey (Section III.A), so the LLM receives inputs similar to the tested styles.
Section III.A notes standardization as a challenge and says a survey established a standard format; the system's robustness to out-of-format descriptions is untested.
domain assumption The EVE-NG API operations (node creation, linking, execution) succeed when called by the provisioning script.
Section III.D assumes the EVE-NG API behaves as documented; no error handling beyond logging is described.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Text2Net: Transforming Plain-text To A Dynamic Interactive Network Simulation Environment." pith.science (2026). https://pith.science/paper/HI6CWUOM

@misc{pith2026250215754,
  author       = {Pith},
  title        = {Pith review of: Text2Net: Transforming Plain-text To A Dynamic Interactive Network Simulation Environment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HI6CWUOM}},
  note         = {Machine review of arXiv:2502.15754}
}

read the original abstract

This paper introduces Text2Net, an innovative text-based network simulation engine that leverages natural language processing (NLP) and large language models (LLMs) to transform plain-text descriptions of network topologies into dynamic, interactive simulations. Text2Net simplifies the process of configuring network simulations, eliminating the need for users to master vendor-specific syntaxes or navigate complex graphical interfaces. Through qualitative and quantitative evaluations, we demonstrate Text2Net's ability to significantly reduce the time and effort required to deploy network scenarios compared to traditional simulators like EVE-NG. By automating repetitive tasks and enabling intuitive interaction, Text2Net enhances accessibility for students, educators, and professionals. The system facilitates hands-on learning experiences for students that bridge the gap between theoretical knowledge and practical application. The results showcase its scalability across various network complexities, marking a significant step toward revolutionizing network education and professional use cases, such as proof-of-concept testing.

Figures

Figures reproduced from arXiv: 2502.15754 by the authors.

**Figure 1.** Text2Net system model and pipeline computing resources more deeply [9]. Similarly, NetLM [10] introduces an AI-driven architecture to enhance autonomous capabilities in network management, notably in complex 6G environments. The system leverages multi-modal representation learning to integrate diverse network data, aiming to refine network intents and autonomously manage network operations. ABC (Automatic Bottom-up … view at source ↗

**Figure 3.** Static-route Scenario Network Topology value pairs essential for network topology provisioning. The model recognizes detailed textual descriptions of network setups, outputting accurate command strings without superfluous content. For valid, complete inputs, the model confirms with returning the phrase ”Understood”, moving to the next phase. For inputs that are empty, incomplete, or incorrect, it prompts the user t… view at source ↗

**Figure 4.** Structured Command Strings (SCSs) [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: key-value pairs output Algorithm 1 facilitates the structured extraction and processing of network topology data. D = {d1, d2, . . . , dn} represents the set of all devices, where in {d1, d2, . . . , dn} each element is a tuple containing key-value pairs (k, v) that d…

**Figure 6.** Figure 6: Steps and time comparison for Text2Net and EVE-NG Network [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Steps and time comparison for Text2Net and EVE-NG Network [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Steps and time comparison for Text2Net and EVE-NG Network [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 9.** Figure 9: Consolidated Benefits and Ratings for Text2Net [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

Visual learning tools for teach- ing/learning computer networks: Cisco networking academy and packet tracer,

J. Janitor, F. Jakab, and K. Kniewald, “Visual learning tools for teach- ing/learning computer networks: Cisco networking academy and packet tracer,” in 2010 Sixth international conference on networking and services. IEEE, 2010, pp. 351–355

work page 2010
[2]

Simulation-based learning via cisco packet tracer to enhance the teaching of computer networks,

J. Allison, “Simulation-based learning via cisco packet tracer to enhance the teaching of computer networks,” in Proceedings of the 27th ACM Con- ference on on Innovation and Technology in Computer Science Education Vol. 1, 2022, pp. 68–74

work page 2022
[3]

J. C. Neumann, The book of GNS3: build virtual network labs using Cisco, Juniper, and more. No Starch Press, 2015

work page 2015
[4]

Emulated Virtual Environment - Next Generation,

Uldis Dzerkals, “Emulated Virtual Environment - Next Generation,” 2024, [Online; accessed 1-June-2024]. [Online]. Available: https: //www.eve-ng.net/

work page 2024
[5]

Comparison of packet tracer and eve-ng tools for efficient network design,

A. Sharma, A. Kumar, and M. Kumar, “Comparison of packet tracer and eve-ng tools for efficient network design,” in 2024 11th International Con- ference on Computing for Sustainable Global Development (INDIACom) . IEEE, 2024, pp. 917–919

work page 2024
[6]

Teaching of computer networks with using simulation and emulation environments,

A. Siersze ´n, L. Sturgulewski, S. Przyłucki, and D. Czerwi ´nski, “Teaching of computer networks with using simulation and emulation environments,” in ICERI2017 Proceedings. IATED, 2017, pp. 5173–5179

work page 2017
[7]

Simulation for network education: Transferring networking skills between simulated to physical environ- ments,

J. Marquardson and D. Gomillion, “Simulation for network education: Transferring networking skills between simulated to physical environ- ments,” Information Systems Education Journal , vol. 17, no. 1, p. 28, 2019

work page 2019
[8]

Netgpt: An ai-native network architecture for provisioning beyond per- sonalized generative services,

Y . Chen, R. Li, Z. Zhao, C. Peng, J. Wu, E. Hossain, and H. Zhang, “Netgpt: An ai-native network architecture for provisioning beyond per- sonalized generative services,” IEEE Network, 2024

work page 2024

Show all 13 references

[9]

Ten issues of netgpt,

W. Tong, C. Peng, T. Yang, F. Wang, J. Deng, R. Li, L. Yang, H. Zhang, D. Wang, M. Ai et al., “Ten issues of netgpt,” arXiv preprint arXiv:2311.13106, 2023

2023 arXiv
[10]

Network meets chatgpt: Intent autonomous management, control and operation,

J. Wang, L. Zhang, Y . Yang, Z. Zhuang, Q. Qi, H. Sun, L. Lu, J. Feng, and J. Liao, “Network meets chatgpt: Intent autonomous management, control and operation,” Journal of Communications and Information Networks , vol. 8, no. 3, pp. 239–255, 2023

2023
[11]

Abc: Automatic bottom-up construction of configuration knowledge base for multi-vendor networks,

W. Ding, L. Liu, L. Chen, and H. Xu, “Abc: Automatic bottom-up construction of configuration knowledge base for multi-vendor networks,” in 2023 IEEE 5th International Conference on Cognitive Machine Intel- ligence (CogMI). IEEE Computer Society, 2023, pp. 135–140

2023
[12]

Confpilot: A pilot for faster configuration by learning from device manuals,

J. Zhao, H. Sun, J. Wang, Q. Qi, Z. Zhuang, S. Tao, and J. Liao, “Confpilot: A pilot for faster configuration by learning from device manuals,” in 2023 IEEE 43rd International Conference on Distributed Computing Systems (ICDCS) . IEEE, 2023, pp. 108–119

2023
[13]

Netcr: Knowledge graph based recommendation framework for manual network configuration,

Z. Guo, F. Li, J. Shen, and X. Wang, “Netcr: Knowledge graph based recommendation framework for manual network configuration,” IEEE Internet of Things Journal , 2023. Fig. 9. Consolidated Benefits and Ratings for Text2Net

2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Visual learning tools for teach- ing/learning computer networks: Cisco networking academy and packet tracer,

J. Janitor, F. Jakab, and K. Kniewald, “Visual learning tools for teach- ing/learning computer networks: Cisco networking academy and packet tracer,” in 2010 Sixth international conference on networking and services. IEEE, 2010, pp. 351–355

work page 2010

[2] [2]

Simulation-based learning via cisco packet tracer to enhance the teaching of computer networks,

J. Allison, “Simulation-based learning via cisco packet tracer to enhance the teaching of computer networks,” in Proceedings of the 27th ACM Con- ference on on Innovation and Technology in Computer Science Education Vol. 1, 2022, pp. 68–74

work page 2022

[3] [3]

J. C. Neumann, The book of GNS3: build virtual network labs using Cisco, Juniper, and more. No Starch Press, 2015

work page 2015

[4] [4]

Emulated Virtual Environment - Next Generation,

Uldis Dzerkals, “Emulated Virtual Environment - Next Generation,” 2024, [Online; accessed 1-June-2024]. [Online]. Available: https: //www.eve-ng.net/

work page 2024

[5] [5]

Comparison of packet tracer and eve-ng tools for efficient network design,

A. Sharma, A. Kumar, and M. Kumar, “Comparison of packet tracer and eve-ng tools for efficient network design,” in 2024 11th International Con- ference on Computing for Sustainable Global Development (INDIACom) . IEEE, 2024, pp. 917–919

work page 2024

[6] [6]

Teaching of computer networks with using simulation and emulation environments,

A. Siersze ´n, L. Sturgulewski, S. Przyłucki, and D. Czerwi ´nski, “Teaching of computer networks with using simulation and emulation environments,” in ICERI2017 Proceedings. IATED, 2017, pp. 5173–5179

work page 2017

[7] [7]

Simulation for network education: Transferring networking skills between simulated to physical environ- ments,

J. Marquardson and D. Gomillion, “Simulation for network education: Transferring networking skills between simulated to physical environ- ments,” Information Systems Education Journal , vol. 17, no. 1, p. 28, 2019

work page 2019

[8] [8]

Netgpt: An ai-native network architecture for provisioning beyond per- sonalized generative services,

Y . Chen, R. Li, Z. Zhao, C. Peng, J. Wu, E. Hossain, and H. Zhang, “Netgpt: An ai-native network architecture for provisioning beyond per- sonalized generative services,” IEEE Network, 2024

work page 2024