Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Large Language Model Agent for Structural Drawing Generation Using ReAct Prompt Engineering and Retrieval Augmented Generation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A six-LLM chain converts natural-language beam descriptions into AutoCAD drawing code.

desk verdict A credible modular LLM-agent pipeline for structural drawing code, undercut by an evaluation that never executes the code and a workload claim that is never measured. read the letter →

arxiv 2507.19771 v1 pith:SCDA43W3 submitted 2025-07-26 cs.LG cs.AI

Xin Zhang , Lissette Iturburu , Juan Nicolas Villamizar , Xiaoyu Liu , Manuel Salmeron , Shirley J.Dyke , Julio Ramirez This is my paper
classification cs.LGcs.AI
keywords structuraldrawingslargelanguagemodelsLLMagentReActpromptingretrieval-augmentedgenerationAutoCADautomationpyautocadcodegenerativeAIincivilengineering
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

This paper tries to establish that the labor-intensive step of turning an engineer's verbal description of a structural drawing into an AutoCAD drawing can be largely automated by a chain of six large language models, each responsible for one subtask. The first model identifies the drawing type, the second refines the user's description, the third fills in and calculates mandatory geometric details, the fourth sets workspace defaults, the fifth formats everything into a JSON schema, and the sixth generates pyautocad code that recreates the section in AutoCAD. The claimed payoff is that an engineer could type something like "24x14in RC cross-section with No 4 closed stirrups at 5 in" and get a first-pass drawing script, cutting repetitive drafting effort and making iterative design changes cheaper. The paper supports the claim with three case studies covering reinforced concrete, steel, and precast beams, reporting per-step success rates of 77 to 100 percent based on 100 runs per case.

What carries the argument

The load-bearing mechanism is a six-step LLM agent pipeline built on ReAct prompting and retrieval-augmented generation. ReAct is a prompting pattern that makes each model output alternating Thought, Action, and Observation steps before a final answer, so the model reasons about what it knows before acting; RAG supplies the model with human-curated external facts, organized as five Info categories per drawing type, instead of relying on parametric memory alone. The pipeline's division of labor carries the argument: the early, simple steps can use a lightweight LLM while the calculation, formatting, and code-generation steps use a stronger LLM, and the reinforced concrete case splits step 3 into three substeps to respect token limits. The formatting step also matters because it converts heterogeneous information into a fixed JSON schema that makes code generation more stable and the reasoning transparent to a human checker.

What would settle it

Take the 100 recorded outputs for the reinforced concrete beam, execute each generated step-6 script in AutoCAD or an equivalent CAD engine, and compare the rendered geometry, including vertex positions, rebar circles, stirrup lines, arcs, and hook lines, against the specified 24x14 inch section; if a significant share of scripts rated successful at the text level produce misplaced, missing, or overlapping elements, the central conversion claim fails. A complementary check is to measure how long the same drawings take an experienced drafter from scratch.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that structural drawing generation becomes tractable for an LLM when it is decomposed into six narrow tasks, each with its own prompt, its own retrieval-augmented background knowledge, and its own reasoning-and-acting trace. The pipeline converts a free-text user description into a structured JSON specification and then into Python code that uses the pyautocad interface to draw the cross-section. Each drawing type is supported by a human-curated database entry specifying useful information, mandatory information, calculation procedures, a required output format, and step-by-step coding instructions; this external knowledge is what keeps the LLM on a reliable path instead of hallucinating details. The paper reports per-step success rates of 77 to 100 percent over 100 runs for the three beam cross-section types, with the hardest geometry-calculating substeps on the reinforced concrete beam at 85, 77, and 81 percent and the final code-generation step at 83 percent for that case.

Load-bearing premise

The load-bearing assumption is that an output counts as successful when its generated text and structure are complete and error-free on inspection; the paper never executes the generated Python code in AutoCAD or verifies the geometry of the resulting drawing, so the claim that drawings are produced rests on text-level correctness being a valid proxy for a usable drawing.

Editorial extensions

If this is right

  • An engineer can describe a standard beam cross-section in plain language and receive draft AutoCAD code, with per-step success rates of 77 to 100 percent on the tested cases.
  • A new member type can be added by writing a new database entry with useful information, mandatory information, calculation methods, and coding steps, without retraining the models.
  • Larger-context foundation models could allow substeps to be merged, reducing the error propagation that the current six-step chain is designed to avoid.
  • Human engineers remain the final check: the reported accuracy is on step outputs, not on executed drawings, so the practical value of the workflow depends on a human review stage.

Reading between the lines

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

  • The reported accuracy is a text-level measure; running each generated script in AutoCAD and measuring geometric errors would almost certainly lower the numbers, so the workload-reduction claim should be read as conditional on executed-code validation.
  • Because the database entries carry the domain knowledge, the real engineering cost of this approach shifts from drafting to curating and maintaining that knowledge base for each structure type and code edition.
  • A testable extension would be to measure end-to-end wall-clock time from prompt to final checked drawing against manual drafting for the same three sections; the paper reports workload reduction qualitatively, not through a time comparison.
  • The scheme is most plausible for standard, parametrically defined sections like the ones demonstrated; novel structures would require the database to be extended first, so the method's generality is bounded by database coverage.
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.

Referee Report

5 major / 5 minor

Summary. The paper introduces an LLM-agent pipeline that converts natural-language descriptions of structural drawings into Python code targeting AutoCAD. The pipeline comprises six steps: drawing-type identification, useful-information extraction, mandatory-information calculation, workspace-detail extraction, JSON formatting, and code generation, with ReAct prompting and RAG using author-curated external knowledge. Three case studies are presented: a reinforced-concrete beam cross-section, a steel beam cross-section, and a precast beam cross-section. The authors report per-step completion accuracies over 100 runs, ranging from 77% to 100%, and claim that the approach enables direct and efficient conversion of descriptions into AutoCAD drawings while significantly reducing manual drafting workload.

Significance. If the evaluation supported the claims, the work would be a useful proof-of-concept for LLM-assisted drafting in civil engineering. The modular six-step pipeline, the use of ReAct and RAG, and the three diverse case studies are reasonable contributions, and the authors are transparent about several limitations, including model 'laziness', missing key information, and the need for human final checks. However, the current evidence is not sufficient for the central claims: the evaluation measures text-level correctness of intermediate steps rather than the correctness of the produced AutoCAD drawing, no generated code is executed in AutoCAD, no end-to-end success rate is reported, the workload-reduction claim is unquantified, and no ablation isolates the contribution of RAG or ReAct. The strengths are the clarity of the pipeline design and the honest discussion of failure modes, but the empirical validation needs substantial strengthening before the claimed benefits can be accepted.

major comments (5)
  1. [§4.4, Table 5] The success criterion in §4.4 is defined as fulfilling a step's task requirements 'without any omissions or errors in the generated text or structure', but the paper never executes the generated Python code in AutoCAD, never inspects a resulting .dwg file, and reports no end-to-end success rate. The abstract and introduction claim 'direct conversion of a structural drawing's natural language description into an AutoCAD drawing', yet the per-step accuracies in Table 5 support conclusions only about text and code formatting, not about the correctness of the final drawing. Please add an evaluation that executes the code in AutoCAD (or a simulator), verifies geometric entities against the specification, and reports the fraction of runs that produce a usable drawing. Also clarify whether Figures 9-11 were produced by executing the generated code; the caption says the sample drawing is 'simplified just for demonstration', which is ambiguous about whether the code was actually run.
  2. [Table 2, Step 3-1] In the RC beam example, Step 3-1 lists the top-right No. 8 rebar center as (13, 21). With a 14-inch width, 2-inch clear cover, No. 8 radius 0.5 inch, and No. 4 stirrup diameter 0.5 inch, the correct x-coordinate is 14 - (2 + 0.5 + 0.5) = 11. Step 3-2 and all subsequent steps correctly use (11, 21). Under the §4.4 success definition, Step 3-1 should be counted as a failure; either the evaluation tolerates a coordinate error that would produce an incorrect drawing, or later LLM steps silently repair an earlier error, which means the reported per-step percentages do not reflect what the abstract claims. Please state which situation holds and adjust the reported accuracies and the evaluation methodology accordingly.
  3. [§4.2, §4.3] For the steel and precast beam case studies, the LLM does not generate the section geometry; it copies from preset .dwg files (e.g., 'steelBeamDrawingSet/W1100X390.dwg' and 'Preset_Prestressed_Concrete/I-Beam_I.dwg'). The existence and content of these source files are never shipped or otherwise made available, and the generated SendCommand strings ('SELECT ALL ', 'COPYCLIP ', 'PASTECLIP') are not verified against real AutoCAD behavior. The demonstration for these cases is therefore contingent on unverified external assets, and the claimed conversion to an AutoCAD drawing is not directly evidenced. Please provide the preset files or describe their contents, and test the generated file-copying workflow in an actual AutoCAD session.
  4. [Abstract, §5] The abstract states that the method enables 'significantly reducing the workload compared to current working process associated with manual drawing production', but the paper contains no measurement of manual drafting time, no user study, no baseline comparison, and no workload metric. The workload-reduction claim is central to the paper's motivation and cannot be supported by the current evaluation. Either add a quantitative comparison (e.g., time and effort to draft the same drawings manually versus using the pipeline) or temper the claim to describe the method as a potential assistive tool whose workload benefits remain to be demonstrated.
  5. [§2, §4] The paper attributes the pipeline's performance to ReAct prompt engineering and RAG, but it reports no ablation or baseline experiments without these components. The per-step accuracies in Table 5 evaluate the full pipeline only. Because the injected 'external knowledge' is authored by the same team and prescribes the exact geometry rules, JSON schema, and code commands, the results show that the LLM can follow these prescriptions, but they do not establish that RAG or ReAct improves accuracy or reliability over a simpler prompting approach. Please include baselines (e.g., zero-shot or few-shot prompting without RAG, or a single-LLM pipeline) to support the claimed benefits of the proposed techniques.
minor comments (5)
  1. [Throughout] There are numerous typographical errors, including 'Lanugage' in §2.2, 'Corss-section' in §3.3, 'BEam' in §4.1, and 'Manual Salmeron' versus 'Manuel Salmeron' in the author contributions. Please proofread the manuscript carefully.
  2. [Table 5] The table does not include confidence intervals for the reported percentages, and the row for 'Step 3' in the RC beam case uses sub-steps 3-1, 3-2, 3-3 without explaining in the caption how these relate to the overall six-step pipeline. Please add confidence intervals, state the number of runs in the caption, and clarify the sub-step notation.
  3. [§3.2.2, Appendix A] The steel beam designation is inconsistent: the example uses 'W1100X390' while Appendix A uses 'HP360X174'. Standardize the notation and, if 'W1100X390' is a non-standard or metric designation, provide a reference or explanation.
  4. [Appendices C, H, K] The appendix prompt templates contain several formatting and spelling issues (e.g., 'ALW AYS', 'infomation', 'mathematic') and inconsistent use of uppercase and spacing in code commands. Please clean up the prompts and ensure they exactly match the text described in Sections 2 and 3.
  5. [§4.4] The list of observed error types is useful, but the paper does not report the frequency of each error type or how errors at one step propagate to later steps. A short error analysis with examples from failed runs would make the reliability discussion more concrete.

Circularity Check

2 steps flagged · score 6.0 of 10

The evaluation is self-referential: step 'success' is defined as conforming to author-written prompt requirements, and steel/precast 'generation' is file copying by construction.

  1. self definitional [Section 4.4 (Performance Evaluation); requirements defined in Section 2.3.2 and Appendices G–K]
    "A successful completion was defined as fulfilling the task requirements of a specific step without any omissions or errors in the generated text or structure."

    The metric's target is supplied by the method itself. Section 2.3.2 says 'we leverage a database curated by human experts to fetch pertinent background details', and Appendices G–K list the exact mandatory-information formulas, JSON keys, and pyautocad commands that define each step. Section 4.4 then scores success as 'fulfilling the task requirements' of those prompts. Thus a correct answer is, by construction, one that conforms to the authors' injected instructions; the reported accuracy is a self-consistency score, not a measure of whether a valid AutoCAD drawing was produced. No code is executed and no drawing is checked. The abstract's claim of 'direct conversion ... into an AutoCAD drawing' is therefore not derived from an independent target.

  2. other [Sections 3.2.2, 3.3.2, 4.2 and Appendix K (Step 6 external knowledge for steel and precast beams)]
    "The sample drawing is shown in figure 10 and since we directly copy the drawing from source files, users can regulate source files based on their need."

    For steel and precast beam case studies, Step 6 code does not generate geometry. Appendix K instructs LLM6 to open a pre-existing source file (e.g., os.path.join(os.getcwd(),'steelBeamDrawingSet','W1100X390.dwg')) and to copy/paste it into a target drawing; the paper admits 'we directly copy the drawing from source files'. The 'generated structural drawing' therefore equals the source .dwg asset, an input selected by catalog name. Presenting this as 'generative AI-based method for generating structural drawings' renames retrieval/copying as generation; the claimed conversion reduces to a lookup by construction.

full rationale

The paper's central evaluation is not an independent test of drawing generation. Section 4.4 defines success as fulfilling each step's 'task requirements', and those requirements are exactly the author-written 'external knowledge' in Appendices G–K (mandatory-information formulas, JSON schema, pyautocad command sequences). The reported 77–100% accuracies therefore measure instruction-following with respect to the method's own prompts; no generated code is executed in AutoCAD and no drawing file is independently checked. This is a self-definitional metric: the target is the input. Additionally, for two of the three case studies (steel and precast beams), the 'generated' drawing is produced by opening a pre-existing .dwg file and copying it, so the claimed natural-language-to-drawing conversion reduces to a lookup. The RC example in Table 2 further shows the stated success criterion was not actually enforced: Step 3-1 places the top-right No. 8 bar at x=13 in a 14-inch section with 2-in clear cover and No. 4 stirrups, while Step 3-2/5/6 use x=11; under §4.4, Step 3-1 should have been counted as an error. No significant self-citation circularity was found; the circularity is in the evaluation definition and in the copy-as-generation label.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The pipeline introduces no fitted numeric parameters and no new physical entities. It relies on hand-authored prompts, ACI/AISC-derived geometry rules, and pre-existing DWG templates, which are domain assumptions rather than free parameters because the paper treats them as given inputs.

assumptions (3)
  • domain assumption Reinforced concrete geometry rules embedded in Appendix H are correct and sufficient.
    All RC drawing outputs depend on formulas for rebar positions, stirrup lines, and hooks injected into the prompt; the paper does not independently verify them against a drawing standard.
  • domain assumption The source DWG files for steel and precast beams contain correct representations of the named sections.
    Steel and precast case studies copy from pre-existing files in steelBeamDrawingSet and Preset_Prestressed_Concrete; these files are not shipped or audited.
  • domain assumption LLM outputs can be reliably parsed and executed from the constrained '<result>' format.
    Every pipeline step assumes the model emits parseable output; Table 5 shows this fails in a non-negligible fraction of runs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Model Agent for Structural Drawing Generation Using ReAct Prompt Engineering and Retrieval Augmented Generation." pith.science (2026). https://pith.science/paper/SCDA43W3

@misc{pith2026250719771,
  author       = {Pith},
  title        = {Pith review of: Large Language Model Agent for Structural Drawing Generation Using ReAct Prompt Engineering and Retrieval Augmented Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SCDA43W3}},
  note         = {Machine review of arXiv:2507.19771}
}
read the original abstract

Structural drawings are widely used in many fields, e.g., mechanical engineering, civil engineering, etc. In civil engineering, structural drawings serve as the main communication tool between architects, engineers, and builders to avoid conflicts, act as legal documentation, and provide a reference for future maintenance or evaluation needs. They are often organized using key elements such as title/subtitle blocks, scales, plan views, elevation view, sections, and detailed sections, which are annotated with standardized symbols and line types for interpretation by engineers and contractors. Despite advances in software capabilities, the task of generating a structural drawing remains labor-intensive and time-consuming for structural engineers. Here we introduce a novel generative AI-based method for generating structural drawings employing a large language model (LLM) agent. The method incorporates a retrieval-augmented generation (RAG) technique using externally-sourced facts to enhance the accuracy and reliability of the language model. This method is capable of understanding varied natural language descriptions, processing these to extract necessary information, and generating code to produce the desired structural drawing in AutoCAD. The approach developed, demonstrated and evaluated herein enables the efficient and direct conversion of a structural drawing's natural language description into an AutoCAD drawing, significantly reducing the workload compared to current working process associated with manual drawing production, facilitating the typical iterative process of engineers for expressing design ideas in a simplified way.

Figures

Figures reproduced from arXiv: 2507.19771 by the authors.

Figure 1
Figure 1. A General Process of RAG 2.3 Design of an LLM-based Structural Drawing Generation Workflow 2.3.1 Tasks to be Solved and Challenges To establish an LLM-based workflow for generating structural drawings, it’s important to delineate the specific tasks that transition natural language descriptions into structured drawings. These tasks encompass the following specific steps: • Identifying the Type of Structural Drawing: … view at source ↗
Figure 2
Figure 2. Set up of Step 1 As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Set up of Step 2 There will be considerable variability in how individuals describe the same object to be drawn using natural language. Such differences can lead to inconsistencies in the outputs from an LLM. Therefore, step 2 is designed to refine the user’s input and the requirement for refining the input is embedded in the specifically designed prompt). This refinement process involves the LLM determining which p… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Set up of Step 3 have been introduced. This enables LLM3 to utilize these tools for more precise computations. By the end of step 3, all mandatory information must be acquired. The setup for this step is depicted in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Set up of Step 4 To generate structural drawing, additional information that isn’t directly related to the objects being drawn may also play a role. For instance, details such as the unit of measurement to be included in the AutoCAD workspace or whether the user intend…
Figure 6
Figure 6. Figure 6: Set up of Step 5 After securing all mandatory and relevant information necessary for the requested structural drawing, it is beneficial to arrange this information in a specified format. Providing this organizational structure for the information aids in delineating an…
Figure 7
Figure 7. Figure 7: Set up of Step 6 The final step involves generating Python code that can interact with the AutoCAD workspace. In this step, an LLM designated as LLM6 utilizes the organized information (output5) as input to create code based on these details. Additionally, this step in…
Figure 8
Figure 8. Figure 8: Overall Workflow for LLM Agent 3 Illustrative Case Studies 3.1 Drawing of Rectangular Reinforced Concrete Beam Cross-section 3.1.1 Logic of the Case Study In general, the structural design of reinforced concrete members is focused on determining the required steel rein…
Figure 9
Figure 9. Figure 9: Concrete beam cross section sample (annotations are added by human not LLMs) [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Steel beam cross section sample 4.3 Generation Results for Precast Beam Cross-section The example generation results for the precast beam cross-section are displayed in [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Precast beam cross section sample 4.4 Performance Evaluation Evaluating performance presents a challenge for LLM text generation tasks due to the diverse requirements of different tasks. In response to this challenge, our approach involved using the same prompt for ea…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Automating structural reliability analysis with a multi-agent large language model framework

    stat.AP 2026-07 conditional novelty 6.0 of 10

    A multi-agent LLM framework automates structural reliability analysis from natural-language input to interpreted β and Pf, with a QLoRA-tuned method-planning agent.

Reference graph

Works this paper leans on

67 extracted references · 44 canonical work pages · cited by 1 Pith paper

  1. [1]

    Drawing in the engineering design process: Learning from the first 150 years of modern engineering

    E Brito, R Póvoas, and P Providência. Drawing in the engineering design process: Learning from the first 150 years of modern engineering. InStructures and Architecture, pages 871–878. CRC Press, 2016

    work page 2016

  2. [2]

    Cad software industry trends and directions.The Engineering Design Graphics Journal, 63(1), 1999

    Patrick E Connolly. Cad software industry trends and directions.The Engineering Design Graphics Journal, 63(1), 1999

    work page 1999

  3. [3]

    Raymond J. A. Buhr, Gerald M. Karam, Carol J. Hayes, and C. Murray Woodside. Software cad: A revolutionary approach.IEEE Transactions on Software Engineering, 15(3):235–249, 1989

    work page 1989

  4. [4]

    Effectiveness of autocad 3d software as a learning support tool

    Fatariah Zakaria, Ahmad Othman, et al. Effectiveness of autocad 3d software as a learning support tool. International Journal of Emerging Technologies in Learning (iJET), 7(2):57–60, 2012

    work page 2012

  5. [5]

    Building information modeling (bim) for existing build- ings—literature review and future needs.Automation in construction, 38:109–127, 2014

    Rebekka V olk, Julian Stengel, and Frank Schultmann. Building information modeling (bim) for existing build- ings—literature review and future needs.Automation in construction, 38:109–127, 2014

    work page 2014

  6. [6]

    Construction management with autocad.Journal of Management in Engineering, 7(3):267–278, 1991

    Amir Tavakoli and Kenneth L Klika. Construction management with autocad.Journal of Management in Engineering, 7(3):267–278, 1991

    work page 1991

  7. [7]

    Crack and noncrack classification from concrete surface images using machine learning.Structural Health Monitoring, 18(3):725–738, 2019

    Hyunjun Kim, Eunjong Ahn, Myoungsu Shin, and Sung-Han Sim. Crack and noncrack classification from concrete surface images using machine learning.Structural Health Monitoring, 18(3):725–738, 2019

    work page 2019

  8. [8]

    Rapid, automated post-event image classification and documentation

    Chul Min Yeum, Shirley J Dyke, Bedrich Benes, Thomas Hacker, Julio Ramirez, Alana Lund, and Santiago Pujol. Rapid, automated post-event image classification and documentation. In7th Int’l Conference on Advances in Experimental Structural Engineering, 2017. 23 arXivA PREPRINT

    work page 2017

Show all 67 references
  1. [9]

    Mixed training of deep convolutional neural network for bridge deterioration detection with uav and inspection report sourced images

    J Dang and P Chun. Mixed training of deep convolutional neural network for bridge deterioration detection with uav and inspection report sourced images. InBridge Maintenance, Safety, Management, Life-Cycle Sustainability and Innovations, pages 308–312. CRC Press, 2021

    2021

  2. [10]

    Vision-based structural inspection using multiscale deep convolutional neural networks.arXiv preprint arXiv:1805.01055, 2018

    Vedhus Hoskere, Yasutaka Narazaki, Tu Hoang, and BillieF Spencer Jr. Vision-based structural inspection using multiscale deep convolutional neural networks.arXiv preprint arXiv:1805.01055, 2018

    2018 arXiv

  3. [11]

    Deep learning-based crack damage detection using convolutional neural networks.Computer-Aided Civil and Infrastructure Engineering, 32(5):361–378, 2017

    Young-Jin Cha, Wooram Choi, and Oral Büyüköztürk. Deep learning-based crack damage detection using convolutional neural networks.Computer-Aided Civil and Infrastructure Engineering, 32(5):361–378, 2017

    2017

  4. [12]

    Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information processing systems, 28, 2015

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks.Advances in neural information processing systems, 28, 2015

    2015

  5. [13]

    Deep learning–based fully automated pavement crack detection on 3d asphalt surfaces with an improved cracknet

    Allen Zhang, Kelvin CP Wang, Yue Fei, Yang Liu, Siyu Tao, Cheng Chen, Joshua Q Li, and Baoxian Li. Deep learning–based fully automated pavement crack detection on 3d asphalt surfaces with an improved cracknet. Journal of Computing in Civil Engineering, 32(5):04018041, 2018

    2018

  6. [14]

    Towards rapid and automated vulnerability classification of concrete buildings.Earthquake Engineering and Engineering Vibration, 22(2):309–332, 2023

    Lissette Iturburu, Jean Kwannandar, Shirley J Dyke, Xiaoyu Liu, Xin Zhang, and Julio Ramirez. Towards rapid and automated vulnerability classification of concrete buildings.Earthquake Engineering and Engineering Vibration, 22(2):309–332, 2023

    2023

  7. [15]

    Optimal policy for structure maintenance: A deep reinforcement learning framework.Structural Safety, 83:101906, 2020

    Shiyin Wei, Yuequan Bao, and Hui Li. Optimal policy for structure maintenance: A deep reinforcement learning framework.Structural Safety, 83:101906, 2020

    2020

  8. [16]

    Reinforcement learning-based bridge inspection management.STRUCTURAL HEALTH MONITORING 2023, 2023

    XIN ZHANG, MANUEL SALMERON, BENJAMIN E WOGEN, XIAOYU LIU, LISSETTE ITURBURU, and SHIRLEY DYKE. Reinforcement learning-based bridge inspection management.STRUCTURAL HEALTH MONITORING 2023, 2023

    2023

  9. [17]

    Generative adversarial networks for labeled acceleration data augmentation for structural damage detection.Journal of Civil Structural Health Monitoring, 13(1):181–198, 2023

    Furkan Luleci, F Necati Catbas, and Onur Avci. Generative adversarial networks for labeled acceleration data augmentation for structural damage detection.Journal of Civil Structural Health Monitoring, 13(1):181–198, 2023

    2023

  10. [18]

    Generative adversarial networks review in earthquake-related engineering fields.Bulletin of Earthquake Engineering, 22(7):3511–3562, 2024

    Giuseppe Carlo Marano, Marco Martino Rosso, Angelo Aloisio, and Giansalvo Cirrincione. Generative adversarial networks review in earthquake-related engineering fields.Bulletin of Earthquake Engineering, 22(7):3511–3562, 2024

    2024

  11. [19]

    Generative adversarial network for damage identification in civil structures.Shock and Vibration, 2021(1):3987835, 2021

    Zahra Rastin, Gholamreza Ghodrati Amiri, and Ehsan Darvishan. Generative adversarial network for damage identification in civil structures.Shock and Vibration, 2021(1):3987835, 2021

    2021

  12. [20]

    Generative ai: The new geotechnical assistant?Journal of Geotechnical and Geoenvironmental Engineering, 149(10):02823004, 2023

    Travis A Shoemaker, Charbel Beaino, Dylan M Centella R, Wendi Zhao, Carine Tanissa, Jack Lawrence, and Youssef MA Hashash. Generative ai: The new geotechnical assistant?Journal of Geotechnical and Geoenvironmental Engineering, 149(10):02823004, 2023

    2023

  13. [21]

    Intelligent generative design for shear wall cross-sectional size using rule-embedded generative adversarial network.Journal of Structural Engineering, 149(11):04023161, 2023

    Yitian Feng, Yifan Fei, Yuanqing Lin, Wenjie Liao, and Xinzheng Lu. Intelligent generative design for shear wall cross-sectional size using rule-embedded generative adversarial network.Journal of Structural Engineering, 149(11):04023161, 2023

    2023

  14. [22]

    Opportunities and challenges of generative ai in construc- tion industry: Focusing on adoption of text-based models.Buildings, 14(1):220, 2024

    Prashnna Ghimire, Kyungki Kim, and Manoj Acharya. Opportunities and challenges of generative ai in construc- tion industry: Focusing on adoption of text-based models.Buildings, 14(1):220, 2024

    2024

  15. [23]

    Attention is all you need.Advances in neural information processing systems, 30:I, 2017

    Vaswani Ashish. Attention is all you need.Advances in neural information processing systems, 30:I, 2017

    2017

  16. [24]

    Detection and classification of surface defects on hot-rolled steel using vision transformers.Heliyon, 10(19), 2024

    Vinod Vasan, Naveen Venkatesh Sridharan, Sugumaran Vaithiyanathan, and Mohammadreza Aghaei. Detection and classification of surface defects on hot-rolled steel using vision transformers.Heliyon, 10(19), 2024

    2024

  17. [25]

    A survey of large language models.arXiv preprint arXiv:2303.18223, 1(2), 2023

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models.arXiv preprint arXiv:2303.18223, 1(2), 2023

    2023 arXiv

  18. [26]

    Chatgpt and open-ai models: A preliminary review.Future Internet, 15(6):192, 2023

    Konstantinos I Roumeliotis and Nikolaos D Tselikas. Chatgpt and open-ai models: A preliminary review.Future Internet, 15(6):192, 2023

    2023

  19. [27]

    Application of chatgpt in civil engineering.East African Journal of Engineering, 6(1):104–112, 2023

    Martin Aluga. Application of chatgpt in civil engineering.East African Journal of Engineering, 6(1):104–112, 2023

    2023

  20. [28]

    Nitin Rane, Saurabh Choudhary, and Jayesh Rane. Integrating building information modelling (bim) with chatgpt, bard, and similar generative artificial intelligence in the architecture, engineering, and construction industry: applications, a novel framework, challenges, and fut...

    2023

  21. [29]

    Prompt engineering with chatgpt: a guide for academic writers.Annals of biomedical engineering, 51(12):2629–2633, 2023

    Louie Giray. Prompt engineering with chatgpt: a guide for academic writers.Annals of biomedical engineering, 51(12):2629–2633, 2023. 24 arXivA PREPRINT

    2023

  22. [30]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  23. [31]

    Llm lies: Hallucinations are not bugs, but features as adversarial examples.arXiv preprint arXiv:2310.01469, 2023

    Jia-Yu Yao, Kun-Peng Ning, Zhen-Hui Liu, Mu-Nan Ning, Yu-Yang Liu, and Li Yuan. Llm lies: Hallucinations are not bugs, but features as adversarial examples.arXiv preprint arXiv:2310.01469, 2023

    2023 arXiv

  24. [32]

    Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems...

    2020

  25. [33]

    Retrieval augmented generation using engineering design knowledge.Knowledge- Based Systems, 303:112410, 2024

    L Siddharth and Jianxi Luo. Retrieval augmented generation using engineering design knowledge.Knowledge- Based Systems, 303:112410, 2024

    2024

  26. [34]

    Aios: Llm agent operating system.arXiv preprint arXiv:2403.16971, 2024

    Kai Mei, Xi Zhu, Wujiang Xu, Wenyue Hua, Mingyu Jin, Zelong Li, Shuyuan Xu, Ruosong Ye, Yingqiang Ge, and Yongfeng Zhang. Aios: Llm agent operating system.arXiv preprint arXiv:2403.16971, 2024

    2024 arXiv

  27. [35]

    Data interpreter: An llm agent for data science.arXiv preprint arXiv:2402.18679, 2024

    Sirui Hong, Yizhang Lin, Bang Liu, Bangbang Liu, Binhao Wu, Ceyao Zhang, Chenxing Wei, Danyang Li, Jiaqi Chen, Jiayi Zhang, et al. Data interpreter: An llm agent for data science.arXiv preprint arXiv:2402.18679, 2024

    2024 arXiv

  28. [36]

    AutoCAD Software

    Autodesk Inc. AutoCAD Software. https://www.autodesk.com/products/autocad/overview, 2024. Accessed: 2025-07-18

    2024

  29. [37]

    Revit Software

    Autodesk Inc. Revit Software. https://www.autodesk.com/products/revit/overview, 2024. Accessed: 2025-07-18

    2024

  30. [38]

    SketchUp Software.https://www.sketchup.com/en, 2024

    Trimble Inc. SketchUp Software.https://www.sketchup.com/en, 2024. Accessed: 2025-07-18

    2024

  31. [39]

    Hidden markov models in speech and language processing

    K Knill and S Young. Hidden markov models in speech and language processing. InCorpus-based methods in language and speech processing, pages 27–68. Springer, 1997

    1997

  32. [40]

    Gaussian mixture models

    Douglas Reynolds. Gaussian mixture models. InEncyclopedia of biometrics, pages 827–832. Springer, 2015

    2015

  33. [41]

    A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt.arXiv preprint arXiv:2303.04226, 2023

    Yihan Cao, Siyu Li, Yixin Liu, Zhiling Yan, Yutong Dai, Philip S Yu, and Lichao Sun. A comprehensive survey of ai-generated content (aigc): A history of generative ai from gan to chatgpt.arXiv preprint arXiv:2303.04226, 2023

    2023 arXiv

  34. [42]

    Faster and smaller n-gram language models

    Adam Pauls and Dan Klein. Faster and smaller n-gram language models. InProceedings of the 49th annual meeting of the Association for Computational Linguistics: Human Language Technologies, pages 258–267, 2011

    2011

  35. [43]

    Recurrent neural networks.Design and applications, 5(64-67):2, 2001

    Larry R Medsker, Lakhmi Jain, et al. Recurrent neural networks.Design and applications, 5(64-67):2, 2001

    2001

  36. [44]

    Long short-term memory.Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012

    Alex Graves. Long short-term memory.Supervised sequence labelling with recurrent neural networks, pages 37–45, 2012

    2012

  37. [45]

    Empirical evaluation of gated recurrent neural networks on sequence modeling.arXiv preprint arXiv:1412.3555, 2014

    Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. Empirical evaluation of gated recurrent neural networks on sequence modeling.arXiv preprint arXiv:1412.3555, 2014

    2014 arXiv

  38. [46]

    Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Communications of the ACM, 63(11):139–144, 2020

    2020

  39. [47]

    An introduction to variational autoencoders.F oundations and Trends® in Machine Learning, 12(4):307–392, 2019

    Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders.F oundations and Trends® in Machine Learning, 12(4):307–392, 2019

    2019

  40. [48]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  41. [49]

    A survey on vision transformer.IEEE transactions on pattern analysis and machine intelligence, 45(1):87–110, 2022

    Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chunjing Xu, Yixing Xu, et al. A survey on vision transformer.IEEE transactions on pattern analysis and machine intelligence, 45(1):87–110, 2022

    2022

  42. [50]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  43. [51]

    ChatGPT (Feb 18 version) [Large language model]

    OpenAI. ChatGPT (Feb 18 version) [Large language model]. https://openai.com/chatgpt, 2025. Accessed: 2025-07-18

    2025

  44. [52]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human langua...

    2019

  45. [53]

    Distraction generation using google T5: A study

    Fernando Gonzalez Adauto and Kevin Golan. Distraction generation using google T5: A study. InUnpublished manuscript, 2022. Available from: https://feradauto.github.io/files/distraction_generation. pdf(Accessed: 2025-07-18)

    2022

  46. [54]

    Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

    2023 arXiv

  47. [55]

    A prompt pattern catalog to enhance prompt engineering with chatgpt

    Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C Schmidt. A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382, 2023

    2023 arXiv

  48. [56]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  49. [57]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    2023

  50. [58]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    2022 arXiv

  51. [59]

    Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2(1), 2023

    2023 arXiv

  52. [60]

    Knowing before seeing: Incor- porating post-retrieval information into pre-retrieval query intention classification

    Xueqing Ma, Xiaochi Wei, Yixing Gao, Runyang Feng, Dawei Yin, and Yi Chang. Knowing before seeing: Incor- porating post-retrieval information into pre-retrieval query intention classification. InInternational Conference on Knowledge Science, Engineering and Management, pages 3...

    2023

  53. [61]

    Llm-based custom chatbot using langchain

    Devangi Das, Rajeev Lochan Rath, Tarundeep Singh, Sushruta Mishra, Vinita Malik, Rajeev Sobti, and Biswajit Brahma. Llm-based custom chatbot using langchain. InInternational Conference On Innovative Computing And Communication, pages 257–267. Springer, 2024

    2024

  54. [62]

    American Institute of Steel Construction, Chicago, IL, 15th edition, 2017

    American Institute of Steel Construction.Steel Construction Manual. American Institute of Steel Construction, Chicago, IL, 15th edition, 2017. 26 arXivA PREPRINT Appendix A: Prompt Template for Step 1 You are an excellent assistant with finding what type of structural drawing ...

    2017

  55. [63]

    Precast beam cross-section is standard but the position of the strands are defined by a user

    Usually, steel beam cross-section is standard while the rectangular concrete beam cross-section is specifically designed. Precast beam cross-section is standard but the position of the strands are defined by a user

  56. [64]

    - Type of Structure: name of type

    Standard steel beam cross-section and precast beam cross-section usually has a specific name indicating the size of the steel beam, e.g., HP360X174 You should follow this format: <format> Description: the natural language description of the user about a specific type of struct...

  57. [66]

    Unit: unit name

    If user specified the unit for this structural drawing, record in the final output in the format of "Unit: unit name". Else, record in the final output in the format of "Unit: Millimeter". To complete this task, you need to follow the following format: <format> Description: th...

  58. [67]

    Save: Path

    Identify if user wanted to save the structural drawing to specific path. If so, record in the final output in the format of "Save: Path". If not, record in the final output in the format of "Save: False"

  59. [68]

    Unit: unit name

    If user specified the unit for this structural drawing, record in the final output in the format of "Unit: unit name". Else, record in the final output in the format of "Unit: Millimeter". 36 arXivA PREPRINT Appendix J: External Knowledge for Step 5 (JSON Format) "rectangular ...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.