Pith. sign in

REVIEW 2 major objections 3 minor 44 references

The paper claims that a multi-agent LLM pipeline can take a natural-language structural reliability problem and return reliability index and failure probability estimates within 0.1 of reference solutions, without expert intervention.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 20:31 UTC pith:BXMHRGQL

load-bearing objection A sensible LLM-agent pipeline for structural reliability with deterministic solvers, but the thin, partially unacknowledged empirical support doesn't yet back the trustworthiness claim. the 2 major comments →

arxiv 2607.16580 v1 pith:BXMHRGQL submitted 2026-07-18 stat.AP cs.CEcs.SE

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

Jaehwan Jeon , Chang Hee Lee , Taeyong Kim This is my paper
classification stat.AP cs.CEcs.SE
keywords structural reliabilitymulti-agent LLMreliability indexfailure probabilitymethod selectionQLoRA fine-tuninghuman-in-the-loopsubset simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a coordinated team of LLM agents can carry out the entire component-level structural reliability workflow—parsing a plain-language problem statement, selecting an analysis method, generating and executing solver code, and interpreting the resulting reliability index β and failure probability Pf—well enough that non-experts can obtain trustworthy estimates. The key move is to keep LLMs away from arithmetic: numbers are computed only by deterministic solvers, while agents handle structuring, method choice, code, and explanation. If true, this would lower the expertise barrier for reliability assessment in structural engineering. The authors also claim that fine-tuning just the method-selection agent with QLoRA on 150 examples beats larger untuned models at a priori method-category selection.

Core claim

The central claim is that an end-to-end, multi-agent LLM system can perform structural reliability analysis from natural language to interpreted results with accuracy comparable to expert-run analysis. On twenty held-out benchmark problems, the pipeline executed successfully on all twenty, with reliability-index estimates generally within Δβ≤0.1 of reference solutions obtained from Monte Carlo simulation or, for rare events, subset simulation. The one large discrepancy (p16) arises from method selection, not code execution, and the system includes a human-in-the-loop refinement path for such cases. The paper further claims that a compact open-weight model fine-tuned only for method-category

What carries the argument

Six components—Orchestrator, Problem Analyst, Method Planner, Code Engineer, Runner, and Result Interpreter—connected by typed data contracts that enforce structured hand-offs. The Runner is a non-LLM deterministic execution component that runs generated code in an isolated subprocess and parses β and Pf from fixed-format output, so numerical results are never hallucinated. The Method Planner, the only fine-tuned agent, is a QLoRA-adapted Gemma 4 E4B model trained on 150 synthetic problems to select a reliability-method category a priori from the structured problem specification alone, using a deterministic expert-defined labeling rule (rare failures route to Advanced; otherwise FORM if with

Load-bearing premise

The evaluation assumes the reference solutions—Monte Carlo simulation for most problems and subset simulation for rare events—are accurate ground truth, but for the rare-event problems the paper does not report the convergence error of the subset-simulation estimate, so the claimed accuracy is measured against an unverified reference.

What would settle it

Compute a high-confidence reference for a rare-event benchmark problem (e.g., p10 or p17) using an independent rare-event estimator with a known convergence guarantee, then compare the framework's predicted β; if the difference exceeds 0.1, the end-to-end accuracy claim for rare events fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Structural reliability estimates become accessible to engineers who lack probabilistic-computing expertise, provided they can describe the problem in natural language.
  • Because numerical results come from validated deterministic solvers, the outputs are reproducible and traceable back through the data contracts to the problem specification and generated code.
  • Method-category selection, the judgment-heavy step, can be embedded in a small fine-tuned model rather than requiring a large general-purpose model or closed APIs.
  • The workflow's human-in-the-loop gates let non-experts confirm method category and choose among advanced sampling methods without understanding the underlying theory.
  • If the approach scales to finite-element-coupled limit states and system reliability, it could bring automated reliability assessment to a much wider class of engineering problems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 60% category-selection accuracy still means the Planner misroutes 40% of benchmark problems; the end-to-end Δβ≤0.1 claim may not survive on a broader, less curated problem set.
  • The deterministic labeling rule (β>3.5 → Advanced; 5% FORM-vs-MCS threshold) is a single expert policy; the paper's own ablation shows the rule is configurable, so the learned behavior is only as general as that policy.
  • A natural extension is to replace the fixed reference with a calibrated rare-event estimator (e.g., importance sampling with known design points) to validate the †-marked problems; this is directly testable.
  • The architecture—LLMs structuring and interpreting, deterministic code computing—could be adapted to other engineering analysis domains where hallucinated numbers are unacceptable.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The paper presents a multi-agent LLM framework for component-level structural reliability analysis. A chain of agents (Orchestrator, Problem Analyst, Method Planner, Code Engineer, Runner, Result Interpreter) converts a natural-language problem statement into estimates of the reliability index β and failure probability Pf. The Method Planner is fine-tuned with QLoRA to select among three method categories (FORM-based, MCS, Advanced), while numerical results are computed only by deterministic solver execution. The framework is evaluated on 20 held-out benchmark problems for end-to-end execution and on the Planner's method-category accuracy. The authors claim the framework lowers the expertise barrier for structural reliability analysis while preserving computational trustworthiness.

Significance. If the empirical claims hold, this would be a useful contribution to the emerging area of LLM-based engineering analysis. The architectural separation between LLM reasoning and deterministic numerical computation is a sound and welcome design, as are the structured data contracts, traceability, open-weight local execution, and the decision to fine-tune only the method-selection step. The 20-problem benchmark is a reasonable first testbed. However, the current evidence does not yet establish the central 'computational trustworthiness' claim: several end-to-end results deviate substantially from the stated Δβ≤0.1 bound, and the rare-event reference solutions are not validated with convergence diagnostics.

major comments (2)
  1. [§3.2, Table 6] The claim that estimates are 'generally within Δβ≤0.1' is contradicted by the reported rows. p09 gives Δβ≈0.126, p10 gives Δβ≈0.461 with Pf differing by a factor of 15 (1.8e-8 vs 1.2e-9), and p16 gives Δβ≈1.03. Only p16 is discussed as a discrepancy. For p10, the reference is subset simulation because 'the MCS reference does not reach the target coefficient of variation' (§3.1), but no convergence diagnostics for the SuS reference are reported. Without a known error bound on the reference, the comparison cannot distinguish pipeline error from reference error. If the reference is correct, the framework silently returns an order-of-magnitude Pf error for a rare event; if it is wrong, the accuracy evaluation is not anchored. The trustworthiness claim in the abstract and §3.2 needs either additional validation or a substantially qualified statement.
  2. [§3.3, Fig. 10] The Planner comparison is based on n=20 with a single evaluation, and the headline improvement is 60% vs 50% for the larger untuned model — a difference of two problems. No confidence intervals, repeated trials, or statistical tests are provided, so the conclusion that fine-tuning improves a priori method selection is fragile. Moreover, the held-out labels are generated by the same deterministic rule used to create the training labels (§2.2), so the experiment measures the model's fidelity to that rule, not the rule's correctness as a method-selection criterion. The rule is admittedly a replaceable design choice; the paper should state this more explicitly and quantify the uncertainty in the 60% accuracy.
minor comments (3)
  1. [Table 6 header] The column header 'Predicted Reference Proposed' is confusing. Please group the columns clearly as 'Reference' and 'Proposed', with the selected method shown separately.
  2. [§1.4 and §2.1.4] The claim that the framework works 'without expert intervention' is somewhat in tension with the human-in-the-loop confirmation gates. Consider clarifying that no reliability expert is required, but user confirmation at the defined gates is still part of the workflow.
  3. [Data availability] The statement says code and weights 'will be made publicly available upon acceptance'. Making the repository available during review would materially strengthen reproducibility and allow independent verification of the solver results.

Circularity Check

0 steps flagged

No significant circularity: numeric β/Pf come from deterministic solvers, and the Planner’s same-rule labels are a disclosed consistency check rather than a forced identity.

full rationale

The main claim—end-to-end reliability estimates while preserving computational trustworthiness—is not circular. The paper explicitly separates language-based reasoning from numeric computation: Section 2.1.2 states that β and Pf are produced only by the Runner, a deterministic non-LLM component, and Section 3.2 compares the pipeline outputs against MCS/subset-simulation reference sweeps. That agreement is empirical rather than constructed; indeed, Table 6 contains real mismatches (p10 and p16), showing the comparison is falsifiable. The only potentially self-referential element is the Method Planner evaluation: Section 3.1 assigns benchmark labels “using the deterministic labeling criterion described in Section 2.2”, i.e., the same rule that generated the 150 training labels. This makes the reported 60% category accuracy a measure of fidelity to the authors’ own explicit, replaceable expert criterion rather than an external standard. However, this is not a circular derivation: the model does not receive the reference labels at inference, the labels are not fitted parameters that force the prediction, and 40% of the held-out cases are misclassified, including p10 being routed to FORM. The unverified subset-simulation reference for the † rare-event problems and the unacknowledged order-of-magnitude Pf discrepancy at p10 are validity/accuracy concerns, not circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The framework's numeric outputs rest on the correctness of the cited reliability solvers and on the reference solutions used for validation. The method-selection contribution rests on a self-authored labeling rule whose thresholds (β=3.5, 5%) are free parameters. No new physical or formal entities are introduced; the six agents are software components rather than scientific postulates.

free parameters (2)
  • Rarity threshold β_thresh = 3.5
    Problems with β>3.5 are labeled Advanced regardless of nonlinearity; this value is a replaceable design choice (Section 2.2).
  • FORM-MCS agreement threshold = 5%
    FORM-based label assigned when FORM's β is within 5% of MCS reference; this threshold is chosen by the authors and can be regenerated (Section 2.2).
axioms (3)
  • domain assumption The implemented deterministic solvers (FORM, SORM, MCS, subset simulation, SIS, CLS) in the in-house reliability_llm.solvers package are correct implementations of the cited methods.
    The framework delegates all numeric results to these solvers (Section 2.1.2); if a solver is buggy, the central reliability estimates are wrong.
  • domain assumption The reference solutions used for evaluation (MCS, or subset simulation for rare events) are treated as ground truth.
    Section 3.1 states the reference is the MCS result of the per-problem sweep, replaced by subset simulation when MCS does not reach target CoV; no error bars are given for these references.
  • ad hoc to paper The expert-defined labeling rule (β>3.5 → Advanced; FORM within 5% of MCS → FORM-based; else MCS) is the correct criteria for method selection.
    Section 2.2 defines this rule as a 'deterministic reference standard' and uses it both to generate training labels and to score the Planner, so the Planner's evaluation is relative to this self-authored criterion.

pith-pipeline@v1.3.0-alltime-deepseek · 20019 in / 14425 out tokens · 136119 ms · 2026-08-01T20:31:26.959634+00:00 · methodology

0 comments
read the original abstract

Structural reliability analysis supports the design and safety assessment of buildings and civil infrastructure but requires specialized expertise throughout the workflow. This study presents a multi-agent large language model framework that automates component-level reliability analysis from a natural-language problem statement to interpreted estimates of the reliability index and failure probability. Specialized agents handle problem formulation, method planning, code generation, execution, and result interpretation, with human confirmation at key decision points. The Method Planner is fine-tuned using QLoRA for a priori reliability-method category selection. Analysis results are not generated directly by an LLM; instead, validated deterministic solvers compute the reliability estimates, improving reproducibility and reducing hallucination risk. The framework uses open-weight models and supports local execution without closed APIs. Results show that it lowers the expertise barrier to structural reliability assessment while preserving computational trustworthiness.

Figures

Figures reproduced from arXiv: 2607.16580 by Chang Hee Lee, Jaehwan Jeon, Taeyong Kim.

Figure 1
Figure 1. Figure 1: Failure domain and the limit-state surface [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Multi-agent framework for structural reliability analysis. Green marks the [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A representative human-in-the-loop interaction for the running-example column [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The deterministic reference criterion used to assign training labels. Rare failures [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training versus inference for the Planner. (a) During training, each problem is [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Chat-based web interface of the framework, shown for the running-example col [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Closing reply of the analysis turn for the running example: the engineering [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Structured ProblemSpec produced by the Problem Analyst from the natural￾language input of the running example: four random variables with their distributions and parameters, and the limit-state expression. Using the ProblemSpec, the Planner recommends the FORM-based cat￾egory. This recommendation is consistent with the problem characteristics: the limit-state function is smooth and differentiable, and the … view at source ↗
Figure 9
Figure 9. Figure 9: Core of the self-contained Python script generated by the Code Engineer for the [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Confusion matrices of method-category selection on the held-out benchmark [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 3 canonical work pages

  1. [1]

    Der Kiureghian, Structural and system reliability, Cambridge Uni- versity Press, 2022

    A. Der Kiureghian, Structural and system reliability, Cambridge Uni- versity Press, 2022

    2022

  2. [2]

    B. R. Ellingwood, T. V. Galambos, J. G. MacGregor, Development of a probability based load criterion for American National Standard A58: Building code requirements for minimum design loads in buildings and other structures, volume 577, US Department of Commerce, National Bureau of Standards, 1980

    1980

  3. [3]

    A. M. Hasofer, N. C. Lind, Exact and invariant second-moment code format, Journal of the Engineering Mechanics Division 100 (1974) 111– 121

    1974

  4. [4]

    Rackwitz, B

    R. Rackwitz, B. Fiessler, Structural reliability under combined random load sequences, Computers & Structures 9 (1978) 489–494

    1978

  5. [5]

    S. K. Au, J. L. Beck, Estimation of small failure probabilities in high dimensions by subset simulation, Probabilistic Engineering Mechanics 16 (2001) 263–277

    2001

  6. [6]

    Geyer, I

    S. Geyer, I. Papaioannou, D. Straub, Cross entropy-based importance sampling using gaussian densities revisited, Structural Safety 76 (2019) 15–27. doi:10.1016/j.strusafe.2018.07.001

    doi:10.1016/j.strusafe.2018.07.001 2019

  7. [7]

    Papaioannou, C

    I. Papaioannou, C. Papadimitriou, D. Straub, Sequential importance sampling for structural reliability analysis, Structural Safety 62 (2016) 66–75. doi:10.1016/j.strusafe.2016.06.002. 37

    doi:10.1016/j.strusafe.2016.06.002 2016

  8. [8]

    Papaioannou, D

    I. Papaioannou, D. Straub, Combination line sampling for structural reliability analysis, Structural Safety 88 (2021) 102025. doi:10.1016/j. strusafe.2020.102025

    arXiv 2021

  9. [9]

    Wagner, S

    P.-R. Wagner, S. Marelli, I. Papaioannou, D. Straub, B. Sudret, Rare event estimation using stochastic spectral embedding, Structural Safety 96 (2022) 102179. doi:10.1016/j.strusafe.2021.102179

    arXiv 2022

  10. [10]

    Bourinet, et al., FERUM 4.1: Finite Element Reliability Using MATLAB, 2010

    J.-M. Bourinet, et al., FERUM 4.1: Finite Element Reliability Using MATLAB, 2010. Software, IFMA, France

    2010

  11. [11]

    E. R. A. Group, ERA: Software for reliability and risk analysis,https: //www.cee.ed.tum.de/era/software/, 2024. Technical University of Munich

    2024

  12. [12]

    M. Zhu, F. McKenna, M. H. Scott, Openseespy: Python library for the opensees finite element framework, SoftwareX 7 (2018) 6–11

    2018

  13. [13]

    Liang, M

    H. Liang, M. Talebi Kalaleh, Q. Mei, Integrating large language mod- els for automated structural analysis, arXiv preprint arXiv:2504.09754 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    J. Liu, Z. Geng, R. Cao, L. Cheng, P. Bocchini, M. Cheng, A large lan- guage model-empowered agent for reliable and robust structural analy- sis, arXiv preprint arXiv:2507.02938 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    Zhang, L

    X. Zhang, L. Iturburu, J. N. Villamizar, X. Liu, M. Salmeron, S. J. Dyke, J. Ramirez, Large language model agent for structural drawing generation using ReAct prompt engineering and retrieval augmented generation, arXiv preprint arXiv:2507.19771 (2025)

    Pith/arXiv arXiv 2025

  16. [16]

    D. Kim, T. Kim, Y. Kim, Y.-H. Byun, T. S. Yun, A ChatGPT- MATLAB framework for numerical modeling in geotechnical engineer- ing applications, Computers and Geotechnics 169 (2024) 106237. doi:10.1016/j.compgeo.2024.106237

    arXiv 2024

  17. [17]

    T. Kim, T. S. Yun, H. S. Suh, Can ChatGPT implement finite element models for geotechnical engineering applications?, International Journal for Numerical and Analytical Methods in Geomechanics 49 (2025) 1747–

    2025

  18. [18]

    Z. Geng, J. Liu, R. Cao, L. Cheng, D. M. Frangopol, M. Cheng, A novel multi-agent architecture to reduce hallucinations of large language mod- els in multi-step structural modeling, arXiv preprint arXiv:2603.07728 (2026)

    arXiv 2026

  19. [19]

    Youwai, D

    S. Youwai, D. Phim, V. G. Murcia, R. C. Onas, Large language model- based multi-agent systems for automated foundation design: router- driven task classification and expert selection framework, AI in Civil Engineering 5 (2026). doi:10.1007/s43503-026-00088-8

    work page doi:10.1007/s43503-026-00088-8 2026

  20. [20]

    J. Chen, Y. Bao, Multi-agent large language model framework for code- compliant automated design of reinforced concrete structures, Automa- tion in Construction 177 (2025) 106331. doi:10.1016/j.autcon.2025. 106331

    doi:10.1016/j.autcon.2025 2025

  21. [21]

    Y. Dong, Z. Zhan, Y. Hu, D. M. Doe, Z. Han, AI BIM coordinator for non-expert interaction in building design using LLM-driven multi-agent systems, Automation in Construction 180 (2025) 106563. doi:10.1016/ j.autcon.2025.106563

    arXiv 2025

  22. [22]

    Liang, Y

    H. Liang, Y. Zhou, M. Talebi-Kalaleh, Q. Mei, Automating structural engineering workflows with large language model agents, arXiv preprint arXiv:2510.11004 (2025). MASSE

    arXiv 2025

  23. [23]

    Z. Geng, J. Liu, R. Cao, L. Cheng, H. Wang, M. Cheng, A lightweight large language model-based multi-agent system for 2D frame structural analysis, arXiv preprint arXiv:2510.05414 (2025)

    arXiv 2025

  24. [24]

    Z. Geng, I. Franklin, S. Martinez, J. Liu, Y. Zhao, M. Cheng, Agentic large language models for automated structural analysis of 3D frame systems, arXiv preprint arXiv:2606.06525 (2026)

    Pith/arXiv arXiv 2026

  25. [25]

    H.-R. Xu, N. Zhang, Z.-Y. Yin, P. G. Atangana Njock, Multimodal framework integrating multiple large language model agents for intel- ligent geotechnical design, Automation in Construction 176 (2025) 106257. doi:10.1016/j.autcon.2025.106257

    arXiv 2025

  26. [26]

    L. Wang, J. Yan, J. Yan, Y. Liu, Towards intelligent safety assess- ment under blast loading: A large language model-powered multi-agent framework, Reliability Engineering & System Safety 272 (2026) 112560. doi:10.1016/j.ress.2026.112560. 39

    arXiv 2026

  27. [27]

    J. W. L. Shi, W. Solihin, J. K. W. Yeoh, Fine-tuning a large language model for automated code compliance of building regulations, Advanced Engineering Informatics 68 (2025) 103676. doi:10.1016/j.aei.2025. 103676

    doi:10.1016/j.aei.2025 2025

  28. [28]

    Jiang, J

    G. Jiang, J. Chen, Efficient fine-tuning of large language models for automated building energy modeling in complex cases, Automation in Construction 175 (2025) 106223. doi:10.1016/j.autcon.2025.106223

    arXiv 2025

  29. [29]

    Z. Dong, Z. Lu, Y. Yang, Fine-tuning a large language model for automating computational fluid dynamics simulations, arXiv preprint arXiv:2504.09602 (2025)

    arXiv 2025

  30. [30]

    Zhang, C

    M. Zhang, C. Xu, Y. Gan, Y. Wang, Y. Fu, Y. Chen, Automating construction contract question answering using large language model and fine-tuning, Expert Systems with Applications 297 (2026) 129493. doi:10.1016/j.eswa.2025.129493

    arXiv 2026

  31. [31]

    M. Aqib, M. Hamza, Q. Mei, Y. H. Chui, Fine-tuning large language models and evaluating retrieval methods for improved question answer- ing on building codes, arXiv preprint arXiv:2505.04666 (2025)

    Pith/arXiv arXiv 2025

  32. [32]

    Dettmers, A

    T. Dettmers, A. Pagnoni, A. Holtzman, L. Zettlemoyer, Qlora: Efficient finetuning of quantized llms, Advances in neural information processing systems 36 (2023) 10088–10115

    2023

  33. [33]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al., Lora: Low-rank adaptation of large language models., Iclr 1 (2022) 3

    2022

  34. [34]

    Gemma Team, Gemma 4 technical report, arXiv preprint arXiv:2607.02770 (2026)

    Pith/arXiv arXiv 2026

  35. [35]

    design by contract

    B. Meyer, Applying “design by contract”, Computer 25 (1992) 40–51. doi:10.1109/2.161279

    doi:10.1109/2.161279 1992

  36. [36]

    X. Wu, L. Xiao, Y. Sun, J. Zhang, T. Ma, L. He, A survey of human- in-the-loop for machine learning, Future Generation Computer Systems 135 (2022) 364–381. doi:10.1016/j.future.2022.05.014. 40

    doi:10.1016/j.future.2022.05.014 2022

  37. [37]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, in: International Conference on Learning Representations (ICLR), 2019

    2019

  38. [38]

    D. Han, M. Han, Unsloth team, Unsloth,https://github.com/ unslothai/unsloth, 2023

    2023

  39. [39]

    Streamlit Inc., Streamlit: A faster way to build and share data apps, https://github.com/streamlit/streamlit, 2019

    2019

  40. [40]

    Song, W.-H

    J. Song, W.-H. Kang, Y.-J. Lee, J. Chun, Structural system reliability: Overview of theories and applications to optimization, ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Part A: Civil Engineering 7 (2021). doi:10.1061/AJRUA6.0001122

    work page doi:10.1061/ajrua6.0001122 2021

  41. [41]

    ISBN 978-3-909386-79- 6

    Joint Committee on Structural Safety (JCSS), JCSS Probabilistic Model Code,https://www.jcss-lc.org/ jcss-probabilistic-model-code/, 2001. ISBN 978-3-909386-79- 6

    2001

  42. [42]

    Brussels, Belgium

    European Committee for Standardization (CEN), EN 1990:2002 Eu- rocode — Basis of structural design, 2002. Brussels, Belgium

    1990

  43. [43]

    M. Uhm, J. Kim, S. Ahn, H. Jeong, H. Kim, Effectiveness of retrieval augmented generation-based large language models for generating con- struction safety information, Automation in Construction 170 (2025) 105926. doi:10.1016/j.autcon.2024.105926. 41

    arXiv 2025

  44. [1766]

    doi:10.1002/nag.3956. 38

    work page doi:10.1002/nag.3956