Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A Solver-Aided Hierarchical Language for LLM-Driven CAD Design

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A solver-aided hierarchical CAD language lets an untuned LLM beat OpenSCAD in few-shot design.

desk verdict Real DSL synthesis, but the OpenSCAD comparison is not controlled enough to support the headline claim. read the letter →

arxiv 2502.09819 v1 pith:D4TI4OBS submitted 2025-02-13 cs.CV cs.AIcs.GRcs.LGcs.PL

classification cs.CVcs.AIcs.GRcs.LGcs.PL
keywords CADgenerationdomain-specificlanguageconstraintsolvinglargemodelstext-to-CADneurosymbolicprogramminghierarchicaldesign
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 aims to show that the programming language an LLM writes geometry in can do most of the work of making the result precise and editable. The authors design AIDL, a hierarchical CAD DSL that lets the model name parts and relationships while a geometric constraint solver performs the exact placement arithmetic. With only a syntax description and six hand-written example programs, an untuned LLM produced AIDL designs whose renderings scored higher on prompt alignment than its OpenSCAD designs, even though OpenSCAD is in the model's training data. If the claim holds, improving DSL design is a practical alternative to fine-tuning or prompt engineering for neurosymbolic geometry generation.

What carries the argument

The load-bearing mechanism is the AIDL language together with its recursive, iterative-deepening constraint solver. AIDL compiles a model into a tree of structures; constraints are solved bottom-up, first keeping substructures fixed, then allowing translation, then full parameter solving, which keeps each solve minimal and protects local structure. The solver uses an iterated Newton's method with branch pruning for min and max expressions, and boolean operations are applied afterward using a standard CAD modeling kernel. This offloads spatial reasoning from the LLM, which is what lets the model focus on high-level intent.

What would settle it

Re-run the 36-prompt comparison giving OpenSCAD the same six few-shot examples and the same validate-until-correct retry loop; if the CLIP gap (28.90 versus 27.32) disappears or reverses, the central claim fails. A second check is whether AIDL's advantage survives lifting the ban on referencing boolean results, since that ban's expressivity argument rests on a single cited result.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that the failure of LLMs at CAD is largely a language problem, not only a model problem. AIDL achieves this by combining four design goals in one language: dependencies (references to previously constructed geometry and parameters), constraints (named geometric relationships such as Coincident and Equal), semantics (intuitive and synonymous operation names), and hierarchy (structures that contain substructures). Boolean operations are applied only after constraint solving, and their results cannot be referenced; the paper argues, citing a prior result, that this restriction costs no geometric expressivity. The reported consequence is that AIDL programs from an untuned general-purpose LLM score 28.90 on average CLIP versus 27.32 for OpenSCAD, while retaining modular structure that makes local edits possible.

Load-bearing premise

The central comparison assumes that OpenSCAD and AIDL were prompted under matched conditions, so the reported advantage comes from the language design rather than from AIDL's extra syntax description and six example programs.

Editorial extensions

If this is right

  • Untuned LLMs can generate precise, editable CAD from text if the DSL offloads spatial reasoning to a solver.
  • Banning references to boolean-operation results does not sacrifice geometric expressivity, because any design can be flattened to one geometry per structure, and the ban avoids query ambiguity that LLMs handle poorly.
  • A recursive solver that deepens from fixed substructures to translations to full parameters keeps hierarchical constraint systems tractable and protects local structure from global edits.
  • AIDL programs remain semantically structured and editable, whereas OpenSCAD programs often degrade into explicit polygon vertex coordinates that are difficult to reason about or modify.
  • Language design alone can move LLM-driven CAD generation, complementing model training and prompt engineering rather than requiring them.

Reading between the lines

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

  • We infer that the same pattern should transfer to other geometry-generation domains: a DSL with semantically named entities, an external solver for precision, and a hierarchy for editability should improve few-shot LLM output for 3D CAD, scene graphs, or diagram generation.
  • The success-rate numbers (AIDL 64%, OpenSCAD 79%) suggest AIDL trades syntactic reliability for semantic quality; fine-tuning on AIDL code or better error-recovery prompting is the natural next experiment, and the paper itself notes this.
  • The ablations show CLIP scores barely move when constraints or hierarchy are removed (28.89 and 28.64 versus 28.90), while editability clearly suffers, so CLIP-based evaluation understates the language's value; future work should measure edit-success directly.
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

4 major / 5 minor

Summary. The paper introduces AIDL, a solver-aided hierarchical domain-specific language for LLM-driven CAD generation. The key design choices are to expose semantically named geometry references and constraints while hiding boolean-operation results, to organize geometry into hierarchical structures, and to offload spatial reasoning to a recursive constraint solver. The authors evaluate AIDL against OpenSCAD and two ablated variants on 36 prompts with 10 LLM runs each, reporting CLIP scores, success rates, a perceptual study, and qualitative editability demonstrations. The central claim is that, in the few-shot regime, AIDL outperforms OpenSCAD despite OpenSCAD being in the LLM's training data, both in visual prompt alignment and in post-processability/editability.

Significance. If the comparative claim is established, the paper makes a valuable contribution: it demonstrates that language design alone can materially improve LLM-driven CAD generation, and it offers a reusable solver-aided pattern for neurosymbolic geometry. The system itself is thoughtfully designed: the hierarchical recursive solver, the explicit treatment of boolean-operation references, the use of synonymous operators to accommodate LLM imprecision, and the validate-until-correct pipeline are all concrete, well-motivated engineering choices. The paper also includes useful ablations and a substantial set of qualitative results. However, the headline quantitative claim is currently supported by a small CLIP gap with no significance testing, a possibly unmatched baseline, and a perceptual study that the authors themselves describe as compromised by a renderer bug. The constructive contribution of the DSL is credible; the comparative claim needs stronger evidence before it can be accepted.

major comments (4)
  1. [Section 4, Comparisons] The OpenSCAD baseline is not matched to the AIDL pipeline. AIDL receives a detailed syntax description, six hand-designed example programs, and a retry loop with error feedback, while the paper states only that the LLM is "directly prompt[ed]" to generate OpenSCAD code. The reported CLIP advantage (28.90 vs. 27.32 in Table 2) may therefore reflect differences in prompt engineering, few-shot examples, or selection effects rather than language design. Please run OpenSCAD with the same elements: syntax documentation, six matched few-shot examples, and the same N=5 retry loop, and report the resulting scores.
  2. [Table 2 and Section 4, Results Across Multiple Runs] The average CLIP scores are computed only over the valid runs per method, while success rates differ substantially: AIDL 64%, AIDL-no-constraints 94%, AIDL-no-hierarchy 77%, and OpenSCAD 79%. If AIDL fails disproportionately on harder prompts, averaging over its smaller valid subset inflates its score relative to methods that succeed on a broader and potentially harder set of prompts. Please report per-prompt paired comparisons (with clear handling of missing values), confidence intervals, and a significance test such as a paired t-test or Wilcoxon signed-rank test.
  3. [Appendix C, Perceptual Study] The perceptual study as described cannot support the visual-quality claim. It collected only 32 responses total, about 8 per method, used a between-subjects block design, and the authors state that a renderer bug "removed some lines from the images" and "compromised the results slightly." With 8 responses per method, the study has very low power, and the renderer bug directly affects the visual comparison being judged. Please either rerun the study with corrected renderings and a within-subjects or properly powered design, or restrict the paper's claims to editability and CLIP-based evidence.
  4. [Section 3.2, Key Challenges and DSL Design Decisions] The claim that disabling references to boolean-operation results "does not affect geometric expressivity" rests on the assertion, cited to Cascaval et al. (2023), that query-based CAD operators can be expressed as combinations of constructive and boolean operations, and on the observation that one geometry per structure reduces to CSG. Since the cited work has overlapping authorship and the specific expressive equivalence for the full AIDL language is not demonstrated here, this is a load-bearing point rather than an ancillary remark. Please either provide a direct proof or construction showing that the restricted reference model preserves the expressivity needed for the paper's claims, or soften the expressivity claim to what is actually established.
minor comments (5)
  1. [Figure 4 caption] The word "exihibit" should be "exhibit".
  2. [Section 4, Implementation] The phrase "it is prompted to generated the full AIDL program" should be "it is prompted to generate the full AIDL program".
  3. [Section 5, Conclusion] The phrase "compliments the shortcomings of LLMs" should likely be "complements the shortcomings of LLMs".
  4. [Section 3.3, Hierarchical designs] The text contains a duplicated conjunction: "each structure defines its own geometry and and constraints".
  5. [Section 4, Results Across Multiple Runs] Describing the difference between 64% and 79% success rates as "only slightly lower" is inaccurate; 15 percentage points is a substantial gap and should be acknowledged and discussed.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the headline claim is an external empirical comparison, and the only self-citation (Cascaval et al.) is supported by an independent worst-case argument in the same paragraph.

full rationale

The central claim of the paper is empirical rather than derivational: AIDL is compared against OpenSCAD and two ablations using CLIP scores, success rates, and a perceptual study on 36 prompts (Section 4, Table 2). No fitted parameter is renamed as a prediction, and no output quantity is defined in terms of the quantity it is said to predict. The design goals in Section 3.1 are motivated by prior observations about LLM behavior, not by the evaluation results. The one point that involves a self-citation is the expressivity justification in Section 3.2: the paper cites Cascaval et al. (2023), which shares an author with the present paper, to assert that query-based CAD DSLs are no more expressive than CSG since CAD operators can be expressed as constructive plus boolean operations. This is invoked to justify not allowing references to boolean results. However, the same paragraph immediately gives a citation-independent fallback: 'in the worst case, you can have one geometry per structure, achieving the same expressiveness as CSG.' That argument alone is sufficient for the language's geometric expressivity claim, so the self-citation is not load-bearing. The paper also reports that FeatureScript and the Cascaval DSL failed for LLM generation, which is an additional external, negative result. The comparison concerns that a careful reader should weigh are experimental fairness issues, not circularity: OpenSCAD is described only as 'directly prompt[ed]' while AIDL receives a detailed syntax description, six example programs, and a validate-until-correct retry loop; and the CLIP averages in Table 2 are computed only over runs that produced valid programs while success rates differ across methods. The Appendix C renderer bug is likewise a limitation of the perceptual study. These are threats to the strength of the empirical comparison, but they do not make the claimed result equivalent to its inputs by construction. No equation-level circular step was found.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical system claim, not a derivation; no physical constants are fitted. The only hand-chosen numeric parameter is the retry limit N=5. The key domain assumptions are external citations, LLM-solver division of labor, CSG/query expressivity equivalence, and the untested assumption that GPT-4o can reliably express design intent in AIDL. No new physical or theoretical entities are introduced.

free parameters (1)
  • Retry limit N = 5
    Hand-chosen cap on validate-until-correct attempts; affects both success rate and the observed loss of design intent when the LLM removes constraints or structures to clear errors (Section 4).
assumptions (4)
  • domain assumption LLMs benefit from external solvers and short programs
    Motivates solver-aided design; grounded in Bubeck et al. (2023) and Makatura et al. (2023) as cited in Section 3.1, but not re-validated in this paper.
  • domain assumption Query-based CAD DSLs have no more geometric expressivity than CSG (Cascaval et al., 2023)
    Used in Section 3.2 to justify banning references to boolean-operation results; the cited paper is by overlapping authors and is not machine-checked.
  • domain assumption GPT-4o can produce valid AIDL code from syntax descriptions plus six examples
    Required by all experiments; success rate is 64%, so the assumption holds only partially and the paper discards invalid runs.
  • domain assumption Constraint solver converges to intended geometry for valid programs, with iterative deepening and branch pruning
    Solver correctness is described in Appendix B but no formal guarantees are given; failures are possible when constraint systems are inconsistent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Solver-Aided Hierarchical Language for LLM-Driven CAD Design." pith.science (2026). https://pith.science/paper/D4TI4OBS

@misc{pith2026250209819,
  author       = {Pith},
  title        = {Pith review of: A Solver-Aided Hierarchical Language for LLM-Driven CAD Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4TI4OBS}},
  note         = {Machine review of arXiv:2502.09819}
}
read the original abstract

Large language models (LLMs) have been enormously successful in solving a wide variety of structured and unstructured generative tasks, but they struggle to generate procedural geometry in Computer Aided Design (CAD). These difficulties arise from an inability to do spatial reasoning and the necessity to guide a model through complex, long range planning to generate complex geometry. We enable generative CAD Design with LLMs through the introduction of a solver-aided, hierarchical domain specific language (DSL) called AIDL, which offloads the spatial reasoning requirements to a geometric constraint solver. Additionally, we show that in the few-shot regime, AIDL outperforms even a language with in-training data (OpenSCAD), both in terms of generating visual results closer to the prompt and creating objects that are easier to post-process and reason about.

Figures

Figures reproduced from arXiv: 2502.09819 by the authors.

Figure 1
Figure 1. A 2D CAD program in AIDL, generated using the prompt “old-school telephone”. The LLM generates AIDL code in a hierarchical fashion, adding constraints using naturally named operators. AIDL’s backend solver produces the final CAD shape rendered on the right. 2 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. AIDL allows LLMs to express constraints using semantically meaningful operators. This [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. A sample of LLM-guided 2D CAD generations using AIDL. An untuned general pur￾pose LLM is able to generate a diverse range of objects with accuracy after being prompted by the AIDL language syntax and a few example programs. Ablations We ablate our language design choices by comparing AIDL against two variants: AIDLno hierarchy and AIDLno constraints, which disable hierarchy and constraints respectively. In AIDLno hi… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison and Ablation. For the task of text-to-CAD, we compare our language to OpenSCAD and ablate on our language design choices. (Top Left) In particular, generated Open￾SCAD programs exihibit manually drawn polygons with explicit vertex positions which are difficu…
Figure 5
Figure 5. Figure 5: Types and operations of AIDL. τ represents the union type (struc￾ture—parameter—geometry). [θ] is the notation used to represent an array or list of θ. B SOLVER DETAILS Iterative Deepening Recursive Solve Constraint problems in AIDL are solved recursively over the stru…
Figure 6
Figure 6. Figure 6: Constraint solving order for an AIDL model. (Left) The recursive solve order of the entire model. (Right) Iterative deepening of the constraint solver’s scope for the root node (5 on left), in two stages, first translation deepening, then geometric deepening. Letters i…
Figure 7
Figure 7. Figure 7: Editability of AIDL. Programs generated with AIDL have semantically meaningful parts. By changing the geometry of a single part in the original ”lighthouse” (left), we can modify the entire appearance of the CAD shape in various ways to produce a wide variety of semant…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. BrepLLM: Enabling Large Language Models to Understand Boundary Representations

    cs.CV 2025-12 conditional novelty 6.0 of 10

    BrepLLM feeds boundary-representation CAD graphs directly into an LLM and reports state-of-the-art CAD captioning and generative classification over point-cloud 3D-LLM baselines.

Reference graph

Works this paper leans on

53 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    VageeshD, Arun Shankar Iyer, Suresh Parthasarathy, S

    Ramakrishna Bairi, Atharv Sonwane, Aditya Kanade, C. VageeshD, Arun Shankar Iyer, Suresh Parthasarathy, S. Rajamani, B. Ashok, and Shashank P. Shet. CodePlan : Repository -level Coding using LLMs and Planning . September 2023. URL https://www.semanticscholar.org/paper/f81a1b4510631d14b5b565c4701ee056f8d5c72f

  3. [3]

    Sparks of artificial general intelligence: Early experiments with gpt-4

    S \'e bastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023

  4. [4]

    Cadquery

    CadQuery. Cadquery. https://github.com/CadQuery/cadquery, 2024

  5. [5]

    A lineage-based referencing dsl for computer-aided design

    Dan Cascaval, Rastislav Bodik, and Adriana Schulz. A lineage-based referencing dsl for computer-aided design. Proceedings of the ACM on Programming Languages, 7 0 (PLDI): 0 76--99, 2023

  6. [6]

    Self-collaboration Code Generation via ChatGPT

    Yihong Dong, Xue Jiang, Zhi Jin, and Ge Li. Self-collaboration Code Generation via ChatGPT . 2023. doi:10.48550/ARXIV.2304.07590. URL https://arxiv.org/abs/2304.07590. Publisher: arXiv Version Number: 2

  7. [7]

    InverseCSG : Automatic Conversion of 3D Models to CSG Trees

    Tao Du, Jeevana Priya Inala, Yewen Pu, Andrew Spielberg, Adriana Schulz, Daniela Rus, Armando Solar-Lezama, and Wojciech Matusik. InverseCSG : Automatic Conversion of 3D Models to CSG Trees . ACM Trans. Graph., 37 0 (6): 0 213:1--213:16, December 2018. ISSN 0730-0301. doi:10.1145/3272127.3275006. URL http://doi.acm.org/10.1145/3272127.3275006

  8. [8]

    Learning to Infer Graphics Programs from Hand-Drawn Images

    Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Joshua B. Tenenbaum. Learning to Infer Graphics Programs from Hand - Drawn Images . arXiv:1707.09627 [cs], July 2017. URL http://arxiv.org/abs/1707.09627. arXiv: 1707.09627

Show all 53 references
  1. [9]

    Learning to Infer Graphics Programs from Hand - Drawn Images

    Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to Infer Graphics Programs from Hand - Drawn Images . In Advances in Neural Information Processing Systems , volume 31. Curran Associates, Inc., 2018. URL https://papers.nips.cc/paper/2018/hash/678...

  2. [10]

    Computer- Aided Design as Language

    Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer- Aided Design as Language . In Advances in Neural Information Processing Systems , volume 34, pp.\ 5885--5897. Curran Associates, Inc., 2021. URL https://proceedings.neurips.cc/paper/2021/h...

  3. [11]

    Code llama: Open foundation models for code

    Wenhan Xiong Grattafiori, Alexandre D \'e fossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, and Gabriel Synnaeve. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023

  4. [12]

    ComplexGen : CAD reconstruction by B -rep chain complex generation

    Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. ComplexGen : CAD reconstruction by B -rep chain complex generation. ACM Transactions on Graphics, 41 0 (4): 0 129:1--129:18, July 2022. ISSN 0730-0301. doi:10.1145/3528223.3530078. URL https://dl.acm.org/d...

  5. [13]

    Pradeep Kumar Jayaraman, Joseph George Lambourne, Nishkrit Desai, Karl Willis, Aditya Sanghi, and Nigel J. W. Morris. SolidGen : An Autoregressive Model for Direct B -rep Synthesis . Transactions on Machine Learning Research, February 2023. ISSN 2835-8856. URL https://openrevi...

  6. [14]

    Kim, and Adriana Schulz

    Benjamin Jones, Dalton Hildreth, Duowen Chen, Ilya Baran, Vladimir G. Kim, and Adriana Schulz. AutoMate : a dataset and learning approach for automatic mating of CAD assemblies. ACM Transactions on Graphics, 40 0 (6): 0 227:1--227:18, December 2021. ISSN 0730-0301. doi:10.1145...

  7. [15]

    Shapeassembly: Learning to generate programs for 3d shape structure synthesis

    R Kenny Jones, Theresa Barton, Xianghao Xu, Kai Wang, Ellen Jiang, Paul Guerrero, Niloy J Mitra, and Daniel Ritchie. Shapeassembly: Learning to generate programs for 3d shape structure synthesis. ACM Transactions on Graphics (TOG), 39 0 (6): 0 1--20, 2020

  8. [16]

    Cad-signet: Cad language inference from point clouds using layer-wise sketch instance guided attention

    Mohammad Sadil Khan, Elona Dupont, Sk Aziz Ali, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-signet: Cad language inference from point clouds using layer-wise sketch instance guided attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

  9. [17]

    ABC : A Big CAD Model Dataset for Geometric Deep Learning

    Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. ABC : A Big CAD Model Dataset for Geometric Deep Learning . In 2019 IEEE / CVF Conference on Computer Vision and Pattern Recognition ...

  10. [18]

    Reconstructing editable prismatic CAD from rounded voxel models

    Joseph George Lambourne, Karl Willis, Pradeep Kumar Jayaraman, Longfei Zhang, Aditya Sanghi, and Kamal Rahimi Malekshan. Reconstructing editable prismatic CAD from rounded voxel models. In SIGGRAPH Asia 2022 Conference Papers , SA '22, pp.\ 1--9, New York, NY, USA, November 20...

  11. [19]

    Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J. Mitra. Free2CAD : parsing freehand drawings into CAD commands. ACM Transactions on Graphics, 41 0 (4): 0 1--16, July 2022. ISSN 0730-0301, 1557-7368. doi:10.1145/3528223.3530133. URL https://dl.acm.org/doi/10.1145/3528223.3530133

  12. [20]

    SECAD - Net : Self - Supervised CAD Reconstruction by Learning Sketch - Extrude Operations

    Pu Li, Jianwei Guo, Xiaopeng Zhang, and Dong-ming Yan. SECAD - Net : Self - Supervised CAD Reconstruction by Learning Sketch - Extrude Operations . 2023 a . doi:10.48550/ARXIV.2303.10613. URL https://arxiv.org/abs/2303.10613. Publisher: arXiv Version Number: 1

  13. [21]

    Sfmcad: Unsupervised cad reconstruction by learning sketch-based feature modeling operations

    Pu Li, Jianwei Guo, Huibin Li, Bedrich Benes, and Dong-Ming Yan. Sfmcad: Unsupervised cad reconstruction by learning sketch-based feature modeling operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 4671--4680, 2024

  14. [22]

    StarCoder : may the source be with you!, December 2023 b

    Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, Qian Liu, Evgenii Zheltonozhskii, Terry Yue Zhuo, Thomas Wang, Olivier Dehaene, Mishig Davaadorj, Joel Lamy-Poirier, João Monteiro, ...

  15. [23]

    Point2cad: Reverse engineering cad models from 3d point clouds

    Yujia Liu, Anton Obukhov, Jan Dirk Wegner, and Konrad Schindler. Point2cad: Reverse engineering cad models from 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3763--3772, 2024

  16. [24]

    StarCoder 2 and The Stack v2: The Next Generation , February 2024

    Anton Lozhkov, Raymond Li, Loubna Ben Allal, Federico Cassano, Joel Lamy-Poirier, Nouamane Tazi, Ao Tang, Dmytro Pykhtar, Jiawei Liu, Yuxiang Wei, Tianyang Liu, Max Tian, Denis Kocetkov, Arthur Zucker, Younes Belkada, Zijian Wang, Qian Liu, Dmitry Abulkhanov, Indraneil Paul, Z...

  17. [25]

    WizardCoder : Empowering Code Large Language Models with Evol - Instruct , June 2023

    Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. WizardCoder : Empowering Code Large Language Models with Evol - Instruct , June 2023. URL http://arxiv.org/abs/2306.08568. arXiv:2306.08568 [cs]

  18. [26]

    Draw step by step: Reconstructing cad construction sequences from point clouds via multimodal diffusion

    Weijian Ma, Shuaiqi Chen, Yunzhong Lou, Xueyang Li, and Xiangdong Zhou. Draw step by step: Reconstructing cad construction sequences from point clouds via multimodal diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 27154--27...

  19. [27]

    How can large language models help humans in design and manufacturing? arXiv preprint arXiv:2307.14377, 2023

    Liane Makatura, Michael Foshey, Bohan Wang, Felix H \"a hnLein, Pingchuan Ma, Bolei Deng, Megan Tjandrasuwita, Andrew Spielberg, Crystal Elaine Owens, Peter Yichen Chen, et al. How can large language models help humans in design and manufacturing? arXiv preprint arXiv:2307.14377, 2023

  20. [28]

    Toward programming languages for reasoning: Humans, symbolic systems, and ai agents

    Mark Marron. Toward programming languages for reasoning: Humans, symbolic systems, and ai agents. In Proceedings of the 2023 ACM SIGPLAN International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software, Onward! 2023, pp.\ 136–152, New York, NY, ...

  21. [29]

    Dag amendment for inverse control of parametric shapes

    Elie Michel and Tamy Boubekeur. Dag amendment for inverse control of parametric shapes. ACM Transactions on Graphics (TOG), 40 0 (4): 0 1--14, 2021

  22. [30]

    Synthesizing structured cad models with equality saturation and inverse transformations

    Chandrakana Nandi, Max Willsey, Adam Anderson, James R Wilcox, Eva Darulova, Dan Grossman, and Zachary Tatlock. Synthesizing structured cad models with equality saturation and inverse transformations. In Proceedings of the 41st ACM SIGPLAN Conference on Programming Language De...

  23. [31]

    Charlie Nash, Yaroslav Ganin, S. M. Ali Eslami, and Peter Battaglia. PolyGen : An Autoregressive Generative Model of 3D Meshes . In Proceedings of the 37th International Conference on Machine Learning , pp.\ 7220--7229. PMLR, November 2020. URL https://proceedings.mlr.press/v1...

  24. [32]

    OpenCascade , February 2021

    OCCT3D. OpenCascade , February 2021. URL https://occt3d.com/

  25. [33]

    Featurescript

    Onshape. Featurescript. https://cad.onshape.com/FsDoc/, 2024

  26. [34]

    W. Para, S. Bhat, Paul Guerrero, T. Kelly, N. Mitra, L. Guibas, and Peter Wonka. SketchGen : Generating Constrained CAD Sketches . June 2021 a . URL https://www.semanticscholar.org/paper/SketchGen

  27. [35]

    Sketchgen: Generating constrained cad sketches

    Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. Advances in Neural Information Processing Systems, 34: 0 5077--5088, 2021 b

  28. [36]

    Geocode: Interpretable shape programs

    Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A Yeh, and Rana Hanocka. Geocode: Interpretable shape programs. arXiv preprint arXiv:2212.11715, 2022

  29. [37]

    ExtrudeNet : Unsupervised Inverse Sketch -and- Extrude for Shape Parsing , September 2022

    Daxuan Ren, Jianmin Zheng, Jianfei Cai, Jiatong Li, and Junzhe Zhang. ExtrudeNet : Unsupervised Inverse Sketch -and- Extrude for Shape Parsing , September 2022. URL http://arxiv.org/abs/2209.15632. arXiv:2209.15632 [cs]

  30. [38]

    Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P. Adams. Vitruvion: A Generative Model of Parametric CAD Sketches . Technical Report arXiv:2109.14124, arXiv, April 2022. URL http://arxiv.org/abs/2109.14124. arXiv:2109.14124 [cs] type: article

  31. [39]

    Tenenbaum, Leslie Pack Kaelbling, and Michael Katz

    Tom Silver, Soham Dan, Kavitha Srinivas, Joshua B. Tenenbaum, Leslie Pack Kaelbling, and Michael Katz. Generalized Planning in PDDL Domains with Pretrained Large Language Models . 2023. doi:10.48550/ARXIV.2305.11014. URL https://arxiv.org/abs/2305.11014. Publisher: arXiv Versi...

  32. [40]

    SolveSpace , November 2022

    Jonathan Westhues. SolveSpace , November 2022. URL https://solvespace.com/

  33. [41]

    Karl D. D. Willis, Pradeep Kumar Jayaraman, Joseph G. Lambourne, Hang Chu, and Yewen Pu. Engineering Sketch Generation for Computer - Aided Design . In 2021 IEEE / CVF Conference on Computer Vision and Pattern Recognition Workshops ( CVPRW ) , pp.\ 2105--2114, Nashville, TN, U...

  34. [42]

    Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: a dataset and environment for programmatic CAD construction from human design sequences. ACM Transactions on Graphics, 40 0 (4): 0 ...

  35. [43]

    Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G

    Karl D.D. Willis, Pradeep Kumar Jayaraman, Hang Chu, Yunsheng Tian, Yifei Li, Daniele Grandi, Aditya Sanghi, Linh Tran, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. JoinABLe : Learning Bottom -up Assembly of Parametric CAD Joints . In 2022 IEEE / CVF Confer...

  36. [44]

    Deepcad: A deep generative network for computer-aided design models

    Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 6772--6782, 2021 a

  37. [45]

    DeepCAD : A Deep Generative Network for Computer - Aided Design Models

    Rundi Wu, Chang Xiao, and Changxi Zheng. DeepCAD : A Deep Generative Network for Computer - Aided Design Models . In 2021 IEEE / CVF International Conference on Computer Vision ( ICCV ) , pp.\ 6752--6762, Montreal, QC, Canada, October 2021 b . IEEE. ISBN 978-1-66542-812-5. doi...

  38. [46]

    WizardLM : Empowering Large Language Models to Follow Complex Instructions , June 2023

    Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. WizardLM : Empowering Large Language Models to Follow Complex Instructions , June 2023. URL http://arxiv.org/abs/2304.12244. arXiv:2304.12244 [cs]

  39. [47]

    Xiang Xu, Karl D. D. Willis, Joseph G. Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. SkexGen : Autoregressive Generation of CAD Construction Sequences with Disentangled Codebooks . In Proceedings of the 39th International Conference on Machine Learn...

  40. [48]

    Brepgen: A b-rep generative diffusion model with structured latent geometry

    Xiang Xu, Joseph Lambourne, Pradeep Jayaraman, Zhengqing Wang, Karl Willis, and Yasutaka Furukawa. Brepgen: A b-rep generative diffusion model with structured latent geometry. ACM Transactions on Graphics (TOG), 43 0 (4): 0 1--14, 2024

  41. [49]

    Capri-net: Learning compact cad shapes with adaptive primitive assembly

    Fenggen Yu, Zhiqin Chen, Manyi Li, Aditya Sanghi, Hooman Shayani, Ali Mahdavi-Amiri, and Hao Zhang. Capri-net: Learning compact cad shapes with adaptive primitive assembly. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11768--11778, 2022

  42. [50]

    Unifying the perspectives of NLP and software engineering: A survey on language models for code

    Ziyin Zhang, Chaoyu Chen, Bingchang Liu, Cong Liao, Zi Gong, Hang Yu, Jianguo Li, and Rui Wang. Unifying the perspectives of NLP and software engineering: A survey on language models for code. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openrev...

  43. [51]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  44. [52]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  45. [53]

    LLMs may be okay at getting syntax correct but they are especially bad at spatial reasoning

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.