Pith. sign in

REVIEW 3 major objections 6 minor 76 references

Off-the-shelf multimodal LLMs can localize stress-critical regions on 3D CAD parts from physics-guided prompts and steer mesh refinement more precisely than geometry-only heuristics under the same point budget.

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 · grok-4.5

2026-07-13 06:41 UTC pith:NR3PCMKQ

load-bearing objection Solid empirical systems paper that cleanly measures zero-shot MLLM spatial-physical grounding for volumetric mesh refinement; the precision edge is real under matched budget but is scaffolded by an engineered feature taxonomy and CV pipeline rather than pure autonomous physics. the 3 major comments →

arxiv 2607.08798 v1 pith:NR3PCMKQ submitted 2026-07-08 cs.GR cs.CV

GReFEM: Multimodal LLMs as Zero-Shot Semantic Assistants for Physics-Guided 3D Mesh Refinement

Kartik Bali , Mahish K. Guru , Christian J Cyron , Roland Aydin This is my paper
classification cs.GR cs.CV
keywords multimodal large language modelsadaptive meshingfinite element methodzero-shot semantic groundingphysics-guided promptingorthographic view selectionCAD geometrystress concentration
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.

Adaptive finite-element meshing usually either solves the PDE repeatedly to find where resolution is needed or trains a data-hungry surrogate on large simulation archives. This paper asks whether a frozen multimodal LLM, given only images of a part and a natural-language description of the load, can already mark the places that matter. The authors build GReFEM: orthographic views are ranked by a lightweight scorer, the LLM proposes grid cells that should contain stress-critical features, classical contour operators refine those cells to pixels, and multi-view ray casting lifts the pixels into 3D anchor points that drive a distance-based size field. Across thirty CAD shapes and five loading cases the top LLM variants achieve higher precision than a matched-budget geometric heuristic that simply flags every concave edge and hole, showing that the models can obey explicit physical instructions rather than merely detecting geometry. The result matters because it opens a solver-free, zero-shot route from human engineering intuition to volumetric discretization.

Core claim

Under a fixed refinement-point budget, state-of-the-art multimodal LLMs guided by physics-aware prompts isolate stress-critical geometric features more precisely than a classical mesh heuristic that ignores load direction, and the resulting meshes measurably reduce energy and L2 displacement error relative to a coarse baseline.

What carries the argument

GReFEM plus orthoViews: a view-ranking network that selects the most informative orthographic projections, followed by grid-cell region proposals from an MLLM, classical 2D feature extraction, and multi-view ray-segment sampling that converts those detections into 3D surface anchors for a distance-based element-size field.

Load-bearing premise

The claim that the reported precision reflects true physical understanding rests on the untested assumption that the classical contour-and-ray pipeline faithfully recovers the three-dimensional locations the language model intended.

What would settle it

Replace the multi-view ray-segment sampler with ground-truth surface annotations of the same grid cells; if the precision and F1 gaps versus the geometric heuristic collapse, the gains were an artifact of the 2D-to-3D pipeline rather than of the model’s semantic grounding.

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

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

3 major / 6 minor

Summary. The paper asks whether off-the-shelf multimodal LLMs can serve as zero-shot semantic assistants for physics-guided volumetric FEM mesh refinement. It introduces GReFEM, which (i) ranks orthographic views with a trained orthoViews module (DINOv2 + self-attention), (ii) prompts MLLMs to mark grid cells containing stress-critical geometric features under load-specific text instructions, (iii) recovers 2D feature pixels via classical contour/depth operators, and (iv) lifts them to 3D surface anchors with multi-view ray–segment sampling (MV-RaySeg) for a distance-based sizing field. Evaluation covers 30 CAD parts × 5 loading cases, four SOTA MLLMs, three prompt templates, and a matched-budget geometric heuristic, reporting localization P/R/F1 against a 99th-percentile ZZ stress field plus normalized energy and L2 errors for view-selection ablations. The central empirical claim is that physics-guided MLLM proposals achieve higher precision (and competitive F1@top5) than a blind geometric heuristic under a matched refinement budget, while Load-Only prompts degrade precision.

Significance. If the results hold under clearer isolation of the MLLM’s contribution, the work is a useful empirical map of foundation models as instruction-following interfaces for CAE discretization—solver-free, domain-agnostic via natural language, and complementary to supervised GNN surrogates that require PDE-specific training data. Strengths include a carefully matched refinement budget, multi-model and multi-prompt ablations, an external ZZ reference that never enters the MLLM loop, and explicit acknowledgment that autonomous physical reasoning remains out of reach. The combination of localization metrics and (partial) downstream energy/L2 norms is more informative than pure visual grounding benchmarks. The contribution is primarily evaluative and systems-oriented rather than a new meshing theory; its lasting value is in defining the precision–recall frontier and the dependence on engineered feature taxonomies.

major comments (3)
  1. §4.3, Algorithm 2, and the Load+Features prompt template: the strongest claim—that MLLMs isolate stress-relevant features with higher precision than a matched-budget geometric heuristic (Table 1, F1@top5)—is not cleanly attributable to MLLM physical grounding. The prompts enumerate the same four feature classes (I.C.E., E.C./P.C., T.H., F) that the Feature Detection stage is hard-coded to extract; the MLLM only nominates grid cells, after which classical CV and MV-RaySeg produce the anchors. The Load-Only ablation’s sharp precision drop already shows heavy dependence on this engineered taxonomy. The manuscript should either (a) quantify how often cell proposals fail to contain the intended feature after CV, or (b) reframe claims as “instruction-following over a fixed geometric vocabulary” rather than zero-shot physical understanding, and report an ablation that removes or randomizes the
  2. §5 and Figure 6 vs Table 1: the engineering objective is reduced energy/L2 error under a fixed refinement budget, yet the matched-budget MLLM-vs-heuristic comparison is reported only on localization P/R/F1 against the ZZ field. Figure 6 shows δE,norm and δL2,norm only for orthoViews vs random views (Gemini, Load+Features), not for GReFEM anchors vs the geometric heuristic under the same point budget and size-field parameters (Eqs. 6–7). Without that comparison, the claim that higher precision is “more valuable for downstream FEA” remains an inference from localization density rather than a demonstrated solver-level gain. Please add energy and L2 error tables (or boxplots) for MLLM variants vs the subsampled heuristic across the 30×5 suite.
  3. Table 1: several top-line F1@top5 advantages over the heuristic (0.5362) are small (e.g., Gemini Load+Features 0.5418; Qwen Load+Features+Heuristics 0.5449) and are reported without uncertainty (std across geometries/loads, bootstrap CIs, or paired tests). Given free parameters (grid resolution, r = Dmax, ZZ percentile, top-K, density filter ρ) and the sensitivity analysis in Table 3 (F1 std ~0.03–0.05 across dimensions), it is unclear whether the precision–recall “frontier” is statistically robust. Report per-geometry paired differences and confidence intervals for Mean P / F1@top5, or justify why the ranking is stable under reasonable threshold perturbations.
minor comments (6)
  1. Abstract and §1: phrases such as “robust zero-shot capacity” and “physical grounding” sit in tension with §6’s admission of gross dependence on CV guard-rails and prompt-driven feature lists. Align abstract wording with the Load-Only results and the conclusion’s more cautious framing.
  2. §4.2 / Appendix A.2: orthoViews is trained with STEP-derived feature counts (Algorithm 1) on 5k CAD parts, while inference is surface-mesh only. State clearly in the main text that view ranking is supervised with B-Rep labels and is not itself zero-shot or CAD-agnostic at training time.
  3. Figure 5 caption and row layout: specify which MLLM and prompt variant produced each GReFEM row; currently only the global caption mentions Gemini Load+Features.
  4. Eqs. (4)–(5) and §5: normalize errors by the coarse baseline, but do not report absolute element counts or wall-clock meshing/solve times for refined meshes. A brief cost table would help readers judge the practical budget claim.
  5. Appendix prompts: several model responses (e.g., Claude, GPT) include free-form headings or explanations despite “DO NOT Add explanations.” Note whether parsing is robust to format drift and how invalid cells are filtered.
  6. Related work: briefly situate against industrial a priori sizing (e.g., curvature/proximity-based commercial meshing) beyond academic geometric heuristics, so the baseline is not understated.

Circularity Check

0 steps flagged

Empirical zero-shot evaluation against external ZZ reference; no derivation reduces to its own inputs by construction.

full rationale

The paper's load-bearing claims are empirical comparisons (Table 1 precision/recall/F1 under matched point budget; Figure 5 qualitative alignment; Figure 6 error vs. views) of off-the-shelf MLLM region proposals against a classical geometric heuristic and against an external Zienkiewicz–Zhu stress indicator computed solely on a fine reference mesh that never enters the GReFEM pipeline (Section 3, A.1). No parameters are fitted to target quantities and then re-presented as predictions; prompt ablations (Load+Features vs. Load-Only) and orthoViews ranking are reported as sensitivity analyses rather than circular justifications. orthoViews itself is supervised on independent CAD feature counts from STEP files (Algorithm 1), not on the downstream FEM error metrics. There are no self-citations that carry uniqueness theorems or force the central result, and no renaming of a known closed-form result. The engineered feature taxonomy and CV/MV-RaySeg pipeline (Section 4.3, Algorithm 2) constitute a methodological assumption about faithful 2D-to-3D recovery, not a circular reduction of the reported precision gains. Hence the derivation chain is self-contained against external benchmarks; residual score of 1 reflects only the ordinary dependence of any vision pipeline on its own post-processing operators, which does not meet the criteria for fitted-input or self-definitional circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

The work is an empirical systems paper; its load-bearing content rests on standard continuum-mechanics assumptions, a classical a-posteriori indicator treated as ground truth, a small set of hand-chosen geometric feature types, and several engineering hyperparameters that control the size field and evaluation radius. No new physical entities are postulated.

free parameters (5)
  • h_min / h_max ratio = 0.2
    Fixed at 0.2 for all refined meshes; directly controls how aggressively anchors refine the mesh and therefore the reported error reductions.
  • D_min, D_max (size-field transition) = 2 h_min / 4 h_min
    Set to 2 h_min and 4 h_min; defines the volumetric buffer used both for sizing and for Precision/Recall@r.
  • ZZ percentile threshold = 99th
    99th percentile of the reference ZZ field defines the ground-truth stress-critical points against which all localization metrics are computed.
  • orthoViews cosine alignment threshold τ = 0.95
    τ = 0.95 used to score feature–view alignment during training-label generation.
  • number of top-K views and grid resolution
    K and the uniform grid overlay size are free design choices that affect both MLLM cell counts and final point budgets.
axioms (4)
  • domain assumption Linear elasticity with displacement-driven boundary conditions is an adequate model for the stress concentrations of interest.
    Governing equations (1) and all experiments assume small-strain linear elasticity; no material nonlinearity or contact.
  • domain assumption The Zienkiewicz–Zhu recovered-stress indicator on a fine uniform mesh correctly identifies the regions that should be refined.
    Used exclusively as external validation signal (Appendix A.1); never fed back into the MLLM.
  • ad hoc to paper The four geometric feature classes (internal corner edges, extruded/protruding contours, through-holes, concave fillets) are the dominant stress raisers under the tested loads.
    Explicitly enumerated in the Load+Features prompts and in the feature-detection stage; the paper’s precision gains rest on the MLLM correctly selecting among these classes.
  • domain assumption Orthographic multi-view projections plus ray–mesh intersection suffice to lift 2D grid cells to 3D surface anchors.
    Core of the MV-RaySeg pipeline (Section 4.3); assumes watertight meshes and known camera parameters.
invented entities (3)
  • GReFEM framework no independent evidence
    purpose: End-to-end pipeline that turns MLLM grid-cell proposals into volumetric mesh-size fields.
    New system-level assembly of existing components; no independent physical existence outside the paper.
  • orthoViews no independent evidence
    purpose: Learned ranking of orthographic views by geometric saliency using DINOv2 + self-attention.
    Trained module introduced to maximize feature observability; evaluated by ablation against random views.
  • MV-RaySeg no independent evidence
    purpose: Multi-view ray-segment sampling that converts 2D detections into depth-aware 3D anchor sets.
    Technical subroutine required for internal cavities and edges; not claimed as a general 3D vision advance.

pith-pipeline@v1.1.0-grok45 · 34697 in / 3141 out tokens · 30662 ms · 2026-07-13T06:41:52.963145+00:00 · methodology

0 comments
read the original abstract

Adaptive volumetric finite element meshing is a critical step in computer-aided engineering and analysis that dictates the computational budget of a given problem. It traditionally requires iterative PDE solvers or heavily supervised, data-driven surrogates trained on large-scale simulation data. While Multimodal Large Language Models (MLLMs) excel in 2D visual tasks, their zero-shot capability to semantically ground regions based on geometric understanding and physics remains an open question. Overall, this study explores a significant question: can the high-level semantic understanding of off-the-shelf MLLMs serve as a viable, zero-shot geometric proxy for finite element mesh refinement? To investigate this, we introduce GReFEM (Geometric Reasoning Enhanced Multimodal LLMs for Finite Element Meshing), a framework that utilizes MLLMs to visually localize stress-critical regions based on physics-guided textual prompts. To bridge the gap between 2D MLLM pre-training and 3D geometries, we introduce orthoViews, a view-selection module that maximizes the observability of key geometric features. We conduct an in-depth empirical evaluation across diverse CAD geometries, loading cases, and SOTA MLLMs, comparing them against a tuned geometric heuristic under a strict, matched refinement budget. Our findings reveal that MLLMs demonstrate robust zero-shot capacity to accurately follow complex spatial-physical instructions, isolating stress-relevant features with higher precision than blind heuristics. By mapping both the successes and current limitations of MLLMs in physical grounding, this study defines the frontier of foundation models as semantic assistants in automated simulation workflows.

Figures

Figures reproduced from arXiv: 2607.08798 by Christian J Cyron, Kartik Bali, Mahish K. Guru, Roland Aydin.

Figure 1
Figure 1. Figure 1: GReFEM Pipeline: We process the top views V ∗ , via the Region Proposal and Feature Detection Pipelines. Once the points Q are obtained on the views, we use our Surface Projection (2D to 3D) pipeline to obtain anchor points P for mesh generation. 1 Introduction Recent large language models (LLMs) have demonstrated strong capabilities in reasoning and visual perception, enabling their application across sci… view at source ↗
Figure 2
Figure 2. Figure 2: Localization behavior of MLLMs Left: Direct pixel-level localization is unreliable, even when the model correctly identifies the stress-critical feature semantically. Right: Grid-based region localization in GReFEM views yields slightly more accurate results: Predicted green grid cells are concentrated around stress-critical re-entrant edges. in the prompt shifts the models from blind geometric feature det… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative evaluation of GReFEM. Predicted refinement points (Red) align with stress￾critical regions (Blue) identified by the ground truth ZZ stress error indicator field computed using the fine reference solution uref . Blue points within the geometry denote stress critical areas in the 99th percentile of the ηZZ field. ℓ(d) = hmin + d − Dmin Dmax − Dmin (hmax − hmin). (7) 4 Method We propose GReFEM, a … view at source ↗
Figure 4
Figure 4. Figure 4: orthoViews performs self-attention across the V views of each CAD geometry, allowing each view embedding to aggregate information from all other views before being scored by a per-view MLP. 4.3 Inference Pipeline [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative visual comparison across experimental loading cases from the test set (30X5) [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of view selection on mesh quality. Increasing the number of GReFEM views consistently reduces both energy and L 2 errors, outperforming random perspective views and leading to higher-quality refined meshes. We use the GReFEM with Gemini-3-flash and Load+Feature prompt variant for both view types (random and orthoViews). spatial hints), and Load-Only (providing only loading conditions and asking for … view at source ↗
Figure 7
Figure 7. Figure 7: Overall evaluation results. (a) shows that MLLMs in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Five images provided jointly with example prompt 1. Top row: non-gridded perspective [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Five images provided jointly with example prompt 2. Top row: non-gridded perspective [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Five images provided jointly with example prompt 3. Top row: non-gridded perspective [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Five images provided jointly with example prompt 4. Top row: non-gridded perspective [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Five images provided jointly with example prompt 5. Top row: non-gridded perspective [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The curated evaluation benchmark dataset consisting of diverse CAD geometries. The [PITH_FULL_IMAGE:figures/full_fig_p033_13.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

76 extracted references · 15 linked inside Pith

  1. [1]

    Engineering with computers , volume=

    Particle systems for adaptive, isotropic meshing of CAD models , author=. Engineering with computers , volume=. 2012 , publisher=

    2012

  2. [2]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Structured 3d latents for scalable and versatile 3d generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  3. [3]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Meshllm: Empowering large language models to progressively understand and generate 3d mesh , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  4. [4]

    arXiv preprint arXiv:2411.09595 , year=

    Llama-mesh: Unifying 3d mesh generation with language models , author=. arXiv preprint arXiv:2411.09595 , year=

    Pith/arXiv arXiv

  5. [5]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Meshgpt: Generating triangle meshes with decoder-only transformers , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  6. [6]

    International journal for numerical methods in engineering , volume=

    A-posteriori error estimates for the finite element method , author=. International journal for numerical methods in engineering , volume=. 1978 , publisher=

    1978

  7. [7]

    SIAM Journal on Numerical Analysis , volume=

    Error estimates for adaptive finite element computations , author=. SIAM Journal on Numerical Analysis , volume=. 1978 , publisher=

    1978

  8. [8]

    arXiv preprint arXiv:2412.14203 , year=

    BlenderLLM: Training Large Language Models for Computer-Aided Design with Self-improvement , author=. arXiv preprint arXiv:2412.14203 , year=

    Pith/arXiv arXiv

  9. [9]

    Computer-Aided Design , pages=

    A method of surface mesh generation for industrial CAD models by constructing conforming discrete representation , author=. Computer-Aided Design , pages=. 2025 , publisher=

    2025

  10. [10]

    arXiv preprint arXiv:2306.03908 , year=

    Sam3d: Segment anything in 3d scenes , author=. arXiv preprint arXiv:2306.03908 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2505.22143 , year=

    3D Question Answering via only 2D Vision-Language Models , author=. arXiv preprint arXiv:2505.22143 , year=

    arXiv

  12. [12]

    Enhancing Machining Feature Recognition in CAD Models through Transfer Learning on BRepNet , author=. 35. Forum Bauinformatik, fbi 2024 , pages=. 2024 , organization=

    2024

  13. [13]

    arXiv preprint arXiv:1709.06158 , year=

    Matterport3d: Learning from rgb-d data in indoor environments , author=. arXiv preprint arXiv:1709.06158 , year=

    Pith/arXiv arXiv

  14. [14]

    2011 , publisher=

    Finite Element Method, The: Its Fundamentals And Applications In Engineering , author=. 2011 , publisher=

    2011

  15. [15]

    and Remacle, J.-F

    Geuzaine, C. and Remacle, J.-F. , title =. International Journal for Numerical Methods in Engineering , volume =. 2009 , doi =

    2009

  16. [16]

    arXiv preprint arXiv:2511.07403 , year=

    SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards , author=. arXiv preprint arXiv:2511.07403 , year=

    Pith/arXiv arXiv

  17. [17]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Lerf: Language embedded radiance fields , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [18]

    European Conference on Computer Vision , pages=

    Pointllm: Empowering large language models to understand point clouds , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  19. [19]

    Advances in Neural Information Processing Systems , volume=

    3d-llm: Injecting the 3d world into large language models , author=. Advances in Neural Information Processing Systems , volume=

  20. [20]

    arXiv preprint arXiv:2403.11401 , year=

    Scene-llm: Extending language model for 3d visual understanding and reasoning , author=. arXiv preprint arXiv:2403.11401 , year=

    Pith/arXiv arXiv

  21. [21]

    European conference on computer vision , pages=

    Scanrefer: 3d object localization in rgb-d scans using natural language , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  22. [22]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scan2cap: Context-aware dense captioning in rgb-d scans , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  23. [23]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Spatialvlm: Endowing vision-language models with spatial reasoning capabilities , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  24. [24]

    Computer-Aided Design , pages=

    CADialogue: A multimodal LLM-powered conversational assistant for intuitive parametric CAD modeling , author=. Computer-Aided Design , pages=. 2025 , publisher=

    2025

  25. [25]

    Proceedings of the neural information processing systems conference

    Cad-llm: Large language model for cad generation , author=. Proceedings of the neural information processing systems conference. neurIPS , volume=

  26. [26]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Video-3d llm: Learning position-aware video representation for 3d scene understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  27. [27]

    2026 , howpublished =

    2026

  28. [28]

    Freecadspec2002 , year=

    FreeCAD , author=. Freecadspec2002 , year=

  29. [29]

    Advances in neural information processing systems , volume=

    Spatialpin: Enhancing spatial reasoning capabilities of vision-language models through prompting and interacting 3d priors , author=. Advances in neural information processing systems , volume=

  30. [30]

    Computer methods in applied mechanics and engineering , volume=

    The superconvergent patch recovery , author=. Computer methods in applied mechanics and engineering , volume=. 1992 , publisher=

    1992

  31. [31]

    arXiv preprint arXiv:2405.03685 , year=

    Language-image models with 3d understanding , author=. arXiv preprint arXiv:2405.03685 , year=

    Pith/arXiv arXiv

  32. [32]

    IEEE Transactions on Multimedia , year=

    LLMI3D: MLLM-based 3D Perception from a Single 2D Image , author=. IEEE Transactions on Multimedia , year=

  33. [33]

    International journal for numerical methods in Engineering , volume=

    Algorithms for refining triangular grids suitable for adaptive and multigrid techniques , author=. International journal for numerical methods in Engineering , volume=. 1984 , publisher=

    1984

  34. [34]

    Proceedings of the fourteenth annual symposium on Computational geometry , pages=

    Tetrahedral mesh generation by Delaunay refinement , author=. Proceedings of the fourteenth annual symposium on Computational geometry , pages=

  35. [35]

    Shewchuk, J. R. , title =. Computational Geometry , volume =. 2002 , doi =

    2002

  36. [36]

    Localizing and Matching CAD Model in Point Cloud Using Semantic Registration Method , author=

  37. [37]

    arXiv preprint arXiv:2508.14879 , year=

    MeshCoder: LLM-Powered Structured Mesh Code Generation from Point Clouds , author=. arXiv preprint arXiv:2508.14879 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2506.07491 , year=

    SpatialLM: Training Large Language Models for Structured Indoor Modeling , author=. arXiv preprint arXiv:2506.07491 , year=

    arXiv

  39. [39]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    CAD-Llama: leveraging large language models for computer-aided design parametric 3D model generation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  40. [40]

    arXiv preprint arXiv:2512.01223 , year=

    Boosting Spatial Reasoning Capability of MLLMs for 3D Visual Grounding , author=. arXiv preprint arXiv:2512.01223 , year=

    arXiv

  41. [41]

    2021 , url=

    Localizing and Matching CAD Model in Point Cloud Using Keypoints , author=. 2021 , url=

    2021

  42. [42]

    IEEE Transactions on Robotics , volume=

    Immesh: An immediate lidar localization and meshing framework , author=. IEEE Transactions on Robotics , volume=. 2023 , publisher=

    2023

  43. [43]

    2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    QueryCAD: Grounded question answering for CAD models , author=. 2025 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2025 , organization=

    2025

  44. [44]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Ppfnet: Global context aware local features for robust 3d point matching , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  45. [45]

    International journal for numerical methods in engineering , volume=

    A simple error estimator and adaptive procedure for practical engineerng analysis , author=. International journal for numerical methods in engineering , volume=. 1987 , publisher=

    1987

  46. [46]

    IEEE transactions on pattern analysis and machine intelligence , year=

    Surface reconstruction from point clouds: A survey and a benchmark , author=. IEEE transactions on pattern analysis and machine intelligence , year=

  47. [47]

    IEEE Transactions on pattern analysis and machine intelligence , volume=

    Using spin images for efficient object recognition in cluttered 3D scenes , author=. IEEE Transactions on pattern analysis and machine intelligence , volume=. 2002 , publisher=

    2002

  48. [48]

    2009 IEEE international conference on robotics and automation , pages=

    Fast point feature histograms (FPFH) for 3D registration , author=. 2009 IEEE international conference on robotics and automation , pages=. 2009 , organization=

    2009

  49. [49]

    arXiv preprint arXiv:2410.05340 , year=

    Generating cad code with vision-language models for 3d designs , author=. arXiv preprint arXiv:2410.05340 , year=

    Pith/arXiv arXiv

  50. [50]

    Information Fusion , year=

    GLUE3D: General Language Understanding Evaluation for 3D , author=. Information Fusion , year=

  51. [51]

    arXiv preprint arXiv:2503.13111 , year=

    MM-Spatial: Exploring 3D Spatial Understanding in Multimodal LLMs , author=. arXiv preprint arXiv:2503.13111 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2506.11148 , year=

    LLM-to-Phy3D: Physically Conform Online 3D Object Generation with LLMs , author=. arXiv preprint arXiv:2506.11148 , year=

    Pith/arXiv arXiv

  53. [53]

    Project report , year=

    Iterative 3D Modeling using LLM Agents with Vision , author=. Project report , year=

  54. [54]

    IJCAI 2025 , year=

    PointLLM: Point Cloud + Large Language Model for 3D Understanding , author=. IJCAI 2025 , year=

    2025

  55. [55]

    ACM Transactions on Graphics , year=

    Large Language Model for 3D Reconstruction with Primitives , author=. ACM Transactions on Graphics , year=

  56. [56]

    ICLR 2025 , year=

    POINT-SAM: Promptable 3D Segmentation Model , author=. ICLR 2025 , year=

    2025

  57. [57]

    IJCAI Proceedings 2025 , year=

    How to Enable LLM with 3D Capacity? A Survey of Spatial Intelligence , author=. IJCAI Proceedings 2025 , year=

    2025

  58. [58]

    Nature Communications , year=

    LLM-driven multimodal target volume contouring in radiation oncology , author=. Nature Communications , year=

  59. [59]

    EmergentMind Survey , year=

    3D-GPT: Integrating LLMs with 3D Data , author=. EmergentMind Survey , year=

  60. [60]

    CVPR 2023 , year=

    PointCLIP V2: Towards Universal 3D Representation , author=. CVPR 2023 , year=

    2023

  61. [61]

    arXiv preprint arXiv:2306.02245 , year=

    Sam3d: Zero-shot 3d object detection via segment anything model , author=. arXiv preprint arXiv:2306.02245 , year=

    Pith/arXiv arXiv

  62. [62]

    arXiv preprint , year=

    PartSLIP: Low-Shot Part Segmentation of 3D Point Clouds , author=. arXiv preprint , year=

  63. [63]

    arXiv preprint , year=

    Zero-Shot 3D Shape Segmentation via Text Descriptions , author=. arXiv preprint , year=

  64. [64]

    arXiv preprint , year=

    MiniGPT-3D: Efficient 3D Understanding with LLMs , author=. arXiv preprint , year=

  65. [65]

    arXiv preprint arXiv:2501.18594 , year=

    Foundational Models for 3D Point Clouds: A Survey , author=. arXiv preprint arXiv:2501.18594 , year=

    Pith/arXiv arXiv

  66. [66]

    2023 , url=

    Awesome-LLM-3D: Curated List of Multi-modal LLMs in 3D , author=. 2023 , url=

    2023

  67. [67]

    arXiv preprint , year=

    3DAxisPrompt: 3D Grounding and Reasoning in MLLMs , author=. arXiv preprint , year=

  68. [68]

    arXiv preprint , year=

    LLaVA-3D: 3D Multimodal Large Language Model , author=. arXiv preprint , year=

  69. [69]

    arXiv preprint , year=

    Shikra: Grounded Multimodal Large Language Model , author=. arXiv preprint , year=

  70. [70]

    arXiv preprint , year=

    Point-E: A System for Generating 3D Point Clouds from Text , author=. arXiv preprint , year=

  71. [71]

    arXiv preprint , year=

    Text2Mesh: Text-Driven Neural Stylization for Meshes , author=. arXiv preprint , year=

  72. [72]

    International Conference on Computational Science , pages=

    GraphMesh: geometrically generalized mesh refinement using GNNs , author=. International Conference on Computational Science , pages=. 2024 , organization=

    2024

  73. [73]

    arXiv preprint arXiv:2505.23663 , year=

    AMBER: Adaptive Mesh Generation by Iterative Mesh Resolution Prediction , author=. arXiv preprint arXiv:2505.23663 , year=

    arXiv

  74. [74]

    arXiv preprint arXiv:2407.04516 , year=

    G-Adaptivity: optimised graph-based mesh relocation for finite element methods , author=. arXiv preprint arXiv:2407.04516 , year=

    Pith/arXiv arXiv

  75. [75]

    arXiv preprint arXiv:2010.03409 , year=

    Learning mesh-based simulation with graph networks , author=. arXiv preprint arXiv:2010.03409 , year=

    Pith/arXiv arXiv 2010

  76. [76]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Abc: A big cad model dataset for geometric deep learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=