The reviewed record of science sign in
Pith

arxiv: 2606.06525 · v2 · pith:G7F3MEDF · submitted 2026-06-02 · cs.GR · cs.AI

Agentic Large Language Models for Automated Structural Analysis of 3D Frame Systems

Ziheng Geng , Ian Franklin , Santiago Martinez , Jiachen Liu , Yunhe Zhao , Minghui Cheng This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-28 07:12 UTCgrok-4.3pith:G7F3MEDFrecord.jsonopen to challenge →

classification cs.GR cs.AI
keywords agentic LLMs3D frame analysisstructural engineeringmulti-agent systemsnatural language to modelautomated structural analysisSAP2000 scriptinggeometric representation
0
0 comments X

The pith

An agentic LLM framework converts natural language descriptions of 3D frames into SAP2000 analysis scripts via 2D projections and a multi-agent pipeline.

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

The paper proposes that large language models can act autonomously as agents to perform end-to-end structural analysis of three-dimensional frames starting from plain text. The approach first projects irregular 3D geometries onto a 2D plan using orthogonal gridlines and a matrix that records the number of stories per cell. This representation then feeds a sequence of specialized agents that parse the input, decompose floors, assemble nodes and members, assign supports and loads, and finally emit executable code. The authors test the resulting system on ten representative frames and report an average 90 percent accuracy across repeated runs. A reader would care because the method removes the need for manual modeling steps that currently dominate structural engineering workflows.

Core claim

Irregular 3D frames can be captured by orthogonal 2D gridline projections together with a story-count matrix; this representation supports a multi-agent LLM pipeline that parses natural-language inputs into structured JSON, derives per-floor layouts, assembles geometry through node/girder/slab/column agents, assigns boundary and load conditions, and produces SAP2000 scripts, yielding 90 percent average accuracy on ten test frames.

What carries the argument

The 2D projection representation of 3D frames that encodes spatial coordinates with orthogonal gridlines and vertical extrusion with a matrix of story counts, which enables the subsequent multi-agent decomposition.

If this is right

  • Natural language inputs can be turned directly into complete structural models and analysis scripts without intermediate manual modeling.
  • The projection method preserves topological consistency across irregular 3D geometries during automated assembly.
  • Task decomposition into distinct agents for parsing, geometry, loading, and code generation produces coordinated outputs.
  • The generated scripts execute in commercial software such as SAP2000 to complete the analysis step.
  • Repeated trials on representative frames establish 90 percent average accuracy as a baseline performance level.

Where Pith is reading between the lines

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

  • The same pipeline structure could be retargeted to other finite-element packages by swapping only the final code-translation agent.
  • Iterative text-based refinement loops could be added so that an engineer corrects an initial model by issuing further natural-language instructions.
  • The projection technique might generalize to other volumetric modeling tasks where full 3D coordinate entry is cumbersome.
  • Scaling tests on frames with hundreds of members would show whether token limits or long-horizon reasoning errors become dominant failure modes.

Load-bearing premise

The 2D projection representation using orthogonal gridlines and a matrix of number of stories can adequately capture irregular 3D frame geometries and maintain topological consistency for automated analysis.

What would settle it

Apply the framework to a set of 3D frames whose layouts contain non-orthogonal bays or curved boundaries and check whether the generated node and member coordinates match the input descriptions within the claimed accuracy.

Figures

Figures reproduced from arXiv: 2606.06525 by Ian Franklin, Jiachen Liu, Minghui Cheng, Santiago Martinez, Yunhe Zhao, Ziheng Geng.

Figure 1
Figure 1. Figure 1: Textual description template for automated structural analysis of 3D frame systems. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Benchmark problems comprising ten representative 3D frame systems with irregular geometric configurations. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multi-agent architecture of the proposed agentic LLMs for automated structural analysis. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Floor decomposition agent for converting the matrix of number of stories to stacked 2D floor plans. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Parallel node, girder, and slab agents for in-plane geometry generation. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Column agent for establishing inter-story connectivity between adjacent floors. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Support and load agents for assigning boundary and loading conditions to corresponding structural components. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Code translation stage for generating executable SAP2000 scripts through geometry translation agent and code [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance comparison between the proposed agentic LLMs and state-of-the-art general-purpose LLMs on the [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of the generated structural model and analysis results in SAP2000. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Illustrative examples of inconsistent structural geometry generated by GPT-5.4. [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Illustrative examples of inconsistent structural geometry generated by Gemini-3.1 Pro. [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Runtime and cost comparison between the proposed agentic LLMs and state-of-the-art general-purpose LLMs. [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
read the original abstract

Large language models (LLMs) have emerged as powerful foundation models with strong reasoning capabilities across domains. Beyond reactive text generation, agentic LLMs enable autonomous workflow execution through modular task decomposition and coordinated tool use. In structural engineering, recent efforts have developed agentic LLMs for automated analysis of plane frames. However, their extension to 3D frames remains underexplored due to challenges in irregular geometric representation, topological consistency, and long-horizon reasoning. This paper proposes an agentic LLM framework for automated structural analysis of 3D frames from natural language inputs. Irregular 3D frames are represented by projection onto a 2D plan, where orthogonal gridlines define spatial coordinates and a matrix of number of stories encodes vertical extrusion of each grid cell. Building on this representation, the framework establishes a multi-agent pipeline: a problem analysis agent parses input into structured JSON; a floor decomposition agent derives the spatial layout of each floor; the 3D geometry is assembled by node, girder, slab, and column agents; support and load agents assign boundary and loading conditions, and code translation agents generate executable SAP2000 script. Evaluated on ten representative 3D frames, the proposed framework achieves an average accuracy of 90% across repeated trials, demonstrating consistent and reliable performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes an agentic LLM framework for automated structural analysis of 3D frames from natural language inputs. Irregular 3D geometries are represented by 2D projection onto orthogonal gridlines (defining XY coordinates) plus a per-cell story matrix (encoding vertical extrusion). A multi-agent pipeline then parses the input to JSON, decomposes floor layouts, assembles 3D geometry via node/girder/slab/column agents, assigns supports and loads, and generates executable SAP2000 scripts. The central empirical claim is that the framework achieves 90% average accuracy across repeated trials on ten representative 3D frames.

Significance. If the 90% accuracy claim can be substantiated with explicit metrics, diverse test cases, and validation against ground-truth connectivity, the work would demonstrate a viable multi-agent decomposition for long-horizon structural engineering tasks and a compact representation for 3D frames. The empirical evaluation on multiple frames is a positive step toward reproducible automation; the projection-plus-matrix approach, if shown to preserve topology, offers a practical alternative to direct 3D mesh handling.

major comments (2)
  1. [Abstract] Abstract: the claim of 'average accuracy of 90%' provides no definition of the accuracy metric (e.g., script syntax validity, successful SAP2000 execution, or match to ground-truth nodal connectivity and member forces), no description of the ten test frames, and no per-component error rates or statistical tests. This information is required to evaluate the central performance assertion.
  2. [Abstract] Abstract (representation paragraph): the orthogonal-grid + story-matrix method is stated to handle 'irregular 3D frames,' yet the manuscript supplies no evidence that any of the ten test frames contain non-grid-aligned members, staggered columns, or floor-specific irregularities. Without such cases, the subsequent JSON parsing and node/girder/column agents cannot be shown to maintain topological consistency, which is load-bearing for the pipeline's correctness.
minor comments (1)
  1. The abstract would be clearer if it briefly indicated the mechanism used for inter-agent coordination and error recovery.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will make revisions to improve clarity and substantiation of the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of 'average accuracy of 90%' provides no definition of the accuracy metric (e.g., script syntax validity, successful SAP2000 execution, or match to ground-truth nodal connectivity and member forces), no description of the ten test frames, and no per-component error rates or statistical tests. This information is required to evaluate the central performance assertion.

    Authors: We agree that the abstract should more explicitly define the accuracy metric and reference the test cases. The metric, detailed in the Experiments section, measures the fraction of trials in which the generated SAP2000 script executes without errors and yields results matching ground-truth nodal connectivity and member forces within engineering tolerances. The ten frames are described there as representative 3D systems. In revision we will expand the abstract to include a concise definition of the metric, a brief characterization of the test frames, and a pointer to the per-component results and statistics already reported in the body. revision: yes

  2. Referee: [Abstract] Abstract (representation paragraph): the orthogonal-grid + story-matrix method is stated to handle 'irregular 3D frames,' yet the manuscript supplies no evidence that any of the ten test frames contain non-grid-aligned members, staggered columns, or floor-specific irregularities. Without such cases, the subsequent JSON parsing and node/girder/column agents cannot be shown to maintain topological consistency, which is load-bearing for the pipeline's correctness.

    Authors: The projection-plus-matrix representation encodes irregularities via per-cell story counts, which directly supports staggered columns and floor-specific extrusions while preserving topology. We acknowledge, however, that the manuscript does not explicitly document which of the ten frames exhibit these features. We will add a table or subsection in the Experiments section that lists the geometric characteristics of each test frame, including the presence or absence of non-grid-aligned elements and floor irregularities, thereby demonstrating that the agents maintain consistency on the evaluated cases. Should the current set prove limited, we will qualify the claim accordingly. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical evaluation on independent test frames

full rationale

The paper describes a multi-agent LLM pipeline for converting natural language descriptions of 3D frames into SAP2000 scripts via a 2D orthogonal grid projection and story matrix. The central result is an empirical accuracy metric (90% average on ten frames) obtained from repeated trials. No equations, fitted parameters, or derivations are present that reduce by construction to the inputs. The representation choice is an explicit modeling assumption rather than a self-definitional step, and the reported performance is measured against external ground truth rather than being statistically forced by the method itself. No self-citation chains or uniqueness theorems are invoked as load-bearing elements. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no specific free parameters, axioms, or invented entities can be identified from the provided information.

pith-pipeline@v0.9.1-grok · 5775 in / 1182 out tokens · 29517 ms · 2026-06-28T07:12:30.894394+00:00 · methodology

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

13 extracted references · 8 canonical work pages · 3 internal anchors

  1. [1]

    Multi-agent large language model framework for code-compliant automated design of reinforced concrete structures

    “Multi-agent large language model framework for code-compliant automated design of reinforced concrete structures.”Automation in Construction, 177: 106331. 15 APREPRINT- JUNE12, 2026 Chen, J. and Bao, Y

    2026

  2. [2]

    Empowering llms with logical reasoning: A comprehensive survey

    “Empowering llms with logical reasoning: A comprehensive survey.”arXiv preprint arXiv:2502.15652. Cheng, M., Wang, D., Yu, S., Li, Q., Ouyang, J., Luo, Y ., Zhang, Y ., Liu, Q., and Chen, E

    work page arXiv

  3. [3]

    Retrieval-Augmented Generation for Large Language Models: A Survey

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

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    A novel multi-agent architecture to reduce hallucinations of large language models in multi-step structural modeling

    Geng, Z., Liu, J., Cao, R., Cheng, L., Frangopol, D. M., and Cheng, M. 2026a. “A novel multi-agent architecture to reduce hallucinations of large language models in multi-step structural modeling.”arXiv preprint arXiv:2603.07728. Geng, Z., Liu, J., Cao, R., Cheng, L., Wang, H., and Cheng, M

    work page arXiv

  5. [5]

    A lightweight large language model-based multi-agent system for 2D frame structural analysis

    “A lightweight large language model-based multi-agent system for 2d frame structural analysis.”arXiv preprint arXiv:2510.05414. Geng, Z., Liu, J., Franklin, I., Cao, R., Frangopol, D. M., and Cheng, M. 2026b. “Automating structural analysis across multiple software platforms using large language models.”arXiv preprint arXiv:2604.09866. Google DeepMind

    work page arXiv

  6. [6]

    Livecodebench: Holistic and contamination free evaluation of large language models for code

    “Livecodebench: Holistic and contamination free evaluation of large language models for code.”International Conference on Learning Representations, V ol. 2025, 58791–58831. Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., and Narasimhan, K

    2025

  7. [7]

    Integrating large language models for automated structural analysis

    “Swe-bench: Can language models resolve real-world github issues?.”International Conference on Learning Representations, V ol. 2024, 54107–54157. Liang, H., Kalaleh, M. T., and Mei, Q. 2025a. “Integrating large language models for automated structural analysis.” arXiv preprint arXiv:2504.09754. Liang, H., Zhou, Y ., Kalaleh, M. T., and Mei, Q. 2025b. “Aut...

    work page arXiv 2024

  8. [8]

    Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models

    “Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models.”International Conference on Learning Representations, V ol. 2025, 94743–94765. OpenAI

    2025

  9. [9]

    Infobench: Evaluating instruction following ability in large language models

    “Infobench: Evaluating instruction following ability in large language models.”Findings of the Association for Computational Linguistics: ACL 2024, 13025–13048. Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., and Scialom, T

    2024

  10. [10]

    SoM- 1K: A Thousand-Problem Benchmark Dataset for Strength of Materials

    “Som-1k: A thousand-problem benchmark dataset for strength of materials.”arXiv preprint arXiv:2509.21079. Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y ., et al

    work page arXiv

  11. [11]

    Self-Consistency Improves Chain of Thought Reasoning in Language Models

    “Self-consistency improves chain of thought reasoning in language models.”arXiv preprint arXiv:2203.11171. 16 APREPRINT- JUNE12, 2026 Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V ., Zhou, D., et al

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  12. [12]

    ReAct: Synergizing Reasoning and Acting in Language Models

    “React: Synergizing reasoning and acting in language models.”arXiv preprint arXiv:2210.03629. Zeng, Z., Yu, J., Gao, T., Meng, Y ., Goyal, T., and Chen, D

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    Evaluating large language models at evaluating instruction following

    “Evaluating large language models at evaluating instruction following.”International Conference on Learning Representations, V ol. 2024, 40193–40219. 17

    2024