Pith. sign in

REVIEW 4 major objections 43 references

Natural-language CAD assembly becomes production-ready when each mate is grounded in engineering axioms and realized by deterministic port-mate geometry, not free-form CAD code.

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

2026-07-11 08:28 UTC pith:VDHKBSWV

load-bearing objection Solid systems paper that actually separates assembly semantics from CAD code; headline APR gains are real-looking but partly judge- and constructibility-aligned. the 4 major comments →

arxiv 2607.05123 v1 pith:VDHKBSWV submitted 2026-07-06 cs.AI cs.CV

ASSEMCAD: Production-Ready CAD Assembly Generation from Natural Language

Yurui Dong , Shu Zou , Siqi Li , Nianchen Deng , Hongbin Zhou , Xuemeng Yang , Pinlong Cai , Licheng Wen
show 2 more authors
Xinyu Cai Botian Shi
This is my paper
classification cs.AI cs.CV
keywords Text-to-CADCAD assembly generationAssembly Specificationport-mate libraryengineering axiomsB-Rep verificationAssemBenchprogrammatic CAD
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.

Text-to-CAD has gotten good at single parts, but mechanical assemblies fail for a different reason: they need coordinated interfaces, assembly relations, engineering rules, and physical consistency that raw executable code does not enforce. AssemCAD therefore refuses to treat an assembly as one monolithic CAD program. It first builds an intermediate Assembly Specification of typed parts, geometry-backed ports, executable mates, and a curated set of engineering axioms that justify every connection. A port-and-mate CAD library then turns those symbolic relations into closed-form rigid transforms, checks that declared ports match real B-Rep geometry, synthesizes missing parametric component factories on demand, and verifies connectivity, clashes, degrees of freedom, and axiom compliance. On AssemBench, this pipeline lifts assembly-preservation rates far above strong code-centric baselines while remaining usable with different foundation-model backbones. The practical claim is that production-ready assemblies come from separating semantic engineering reasoning from deterministic geometric execution, not from better free-form CAD generation alone.

Core claim

Directly generating executable CAD code is insufficient for production-ready mechanical assemblies. AssemCAD shows that first constructing an axiomatic Assembly Specification—typed parts, geometry-backed ports, executable mates, and engineering axioms—and then realizing it through a deterministic port-and-mate library with B-Rep interface checks and multi-layer verification substantially improves assembly preservation and physical validity over code-centric baselines on AssemBench, and that this improvement generalizes across foundation-model backbones.

What carries the argument

The axiomatic Assembly Specification S = (P, M, A): typed parts with ports, typed mates, and engineering axioms that justify each mate. A port-and-mate CAD library executes mates via closed-form transforms, validates ports against concrete B-Rep evidence, and feeds failures back for bounded repair.

Load-bearing premise

The method assumes that AssemBench prompts and scores, plus a fixed set of 62 axioms and a small factory registry, are good enough stand-ins for real production assembly correctness and open-world mechanical coverage.

What would settle it

On a held-out set of real multi-part mechanisms outside Fusion360 Gallery style, measure whether AssemCAD’s Assembly Preservation Rate still stays near its Success Rate and whether independent engineers accept the STEP outputs and verification reports as production-ready; a large SR–APR gap or frequent rejections would falsify the central claim.

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

4 major / 0 minor

Summary. AssemCAD addresses production-ready mechanical CAD assembly generation from natural language by decomposing the task into (i) an axiomatic Assembly Specification S=(P,M,A) of typed parts, geometry-backed ports, executable mates, and engineering axioms, (ii) on-demand parametric factory synthesis over a port/mate CadQuery extension library, and (iii) deterministic mate transforms plus multi-layer geometric verification (port–geometry consistency, clash classification, connectivity, DOF). The paper reports large gains on AssemBench (120 Fusion360-derived assemblies) in Success Rate and Assembly Preservation Rate versus code-centric baselines and frontier models (Table 3), with backbone transfer (Table 4), and argues that Chamfer Distance is unsuitable for assembly evaluation (Appendix H).

Significance. If the empirical gains hold under stronger evaluation, the work is a meaningful step from single-part Text-to-CAD toward assembly-level design: the port–mate library, closed-form mate transforms (Eq. 2), SAFEUNION (Proposition 1), determinism under caching (Proposition 2), and B-Rep port-geometry checks (Definition 4) are concrete, reusable engineering contributions that separate semantic planning from deterministic geometric execution. The explicit axiom grounding and verification reports improve interpretability relative to monolithic CAD code generation. The formal critique of Chamfer Distance for assemblies (Appendix H) is also useful for the community. Significance is tempered by dependence on LLM/VLM judges and a constructibility-filtered benchmark for the headline APR claim.

major comments (4)
  1. §5.2 and Table 3: the central claim of substantially improved assembly preservation rests primarily on APR (LLM code-judge ≥4/5) and VLM-as-a-Judge scores. AssemCAD emits constrained YAML/library calls while baselines emit free CadQuery; an LLM judge may systematically favor the structured form. Please report (i) inter-rater agreement with human assembly-structure labels on a substantial subset, and/or (ii) independent structure metrics against Fusion360 GT (joint/contact recovery, mate-type accuracy, DOF match, part-count/graph edit distance). Without this, the SR–APR gap may overstate production-ready gains.
  2. §5.1 and Appendix F: AssemBench is curated for parametric constructibility from Fusion360 Gallery and annotated semi-automatically by a VLM. Combined with 13 built-in factories and 62 axioms aligned to the library (App. A–C), this risks circular coverage of the same component families. The open-world synthesis claim needs a holdout of assemblies outside the factory registry and axiom coverage, with failure rates and repair budgets reported separately for in-registry vs synthesized parts.
  3. §3–4 and Table 3: physical validity is a core claim, yet Table 3 does not break out the paper’s own deterministic checks (clash rates by Eq. 3, connectivity, DOF, port-geometry consistency failures) for AssemCAD vs baselines on the same instances. Please add these objective geometric metrics as primary columns so physical validity is not conflated with LLM/VLM preference scores.
  4. §3.2 / Appendix A: engineering-rule compliance is partly self-referential—axioms selected during decomposition are re-checked in the verification report. Clarify which verification steps are independent of the axiom set (B-Rep evidence, clash volume, connectivity BFS) versus axiom-referenced checks, and report how often axiom-justified mates still fail geometric verification.

Circularity Check

1 steps flagged

No load-bearing circular derivation: main gains are empirical; only mild self-reference is axiom attachment reused as engineering-rule compliance.

specific steps
  1. self definitional [§3.2–§4.2 and §4.4 Axiom Traceability; Def. 1; Alg. 1]
    "During decomposition, the LLM is required to explicitly justify every mate by selecting appropriate axioms. ... Beyond geometric verification, the axiom references attached to each mate during specification (Phase I) are propagated through assembly construction to the final verification report. ... a deterministic verification pipeline then checks interface validity, clash consistency, graph connectivity, degree-of-freedom constraints, and engineering-rule compliance"

    Engineering-rule compliance for mates is largely the presence and propagation of axiom IDs that generation was already required to attach from the fixed 62-axiom set. For admitted specifications, axiom grounding is therefore partly true by construction of the decomposition interface rather than an independent physical check of those principles (unlike clash volumes, port-geometry B-Rep evidence, or connectivity). This is a minor self-definitional loop on the engineering-semantics claim only; it does not force the SR/APR experimental results or the geometric validity gates.

full rationale

AssemCAD is a systems paper whose central claim is empirical improvement of SR/APR and physical validity on AssemBench versus code-centric baselines, not a first-principles prediction forced by fitted parameters or a uniqueness theorem. The realization chain (typed ports, closed-form mate transforms Eq. 2, B-Rep port-geometry checks Definition 4, clash classification Eq. 3, SAFEUNION, connectivity/DOF BFS) is deterministic geometric execution independent of the LLM prose that produced the specification. AssemBench is built from Fusion360 Gallery assemblies with semi-automatic NL briefs, not from AssemCAD outputs, so evaluation is not self-generated. Propositions 1–2 are design/correctness properties of the library (determinism under fixed LLM cache is by construction of content-addressed caching, not a disguised prediction). The only mild circularity is that mates must cite curated axioms at decomposition and those same references are later reported as engineering-rule compliance/traceability—so axiom presence is largely by construction for admitted specs—while independent geometric gates still do real work. No fitted-input-as-prediction, no load-bearing self-citation uniqueness chain, and no renaming of a known empirical law as a derived result. Score 2 for that minor self-definitional loop only.

Axiom & Free-Parameter Ledger

4 free parameters · 8 axioms · 4 invented entities

The central claim rests on engineering and systems assumptions more than fitted physical constants: a hand-curated axiom knowledge base, discrete port/mate type systems, geometric tolerances and clash thresholds, LLM agents with bounded repair, and a benchmark/judge stack that operationalizes “assembly preservation.” Invented intermediate objects (Assembly Specification, AssemBench protocol) are the paper’s main constructs; independent evidence for them is mostly internal experiments.

free parameters (4)
  • clash_volume_threshold_tau
    Interference is accepted or rejected relative to a predefined volume threshold τ (Def. 2 / Alg. 3); the paper does not report a data-driven calibration of τ across assemblies.
  • max_repair_rounds_K
    Bounded LLM repair loops for draft validation and factory synthesis (Alg. 1–2) depend on a chosen retry budget that affects success rates.
  • port_geometry_tolerances
    Face normal collinearity (~1e-3–1e-2 rad), sample offsets (e.g., ϕd/4, ±ϕL/2), and SAFEUNION micro-extension ε are fixed engineering thresholds that gate port validity and unions.
  • APR_score_threshold_4_of_5
    Assembly Preservation Rate counts LLM judge scores ≥4/5 as success; the cutoff is a free evaluation parameter that directly drives the headline APR claim.
axioms (8)
  • ad hoc to paper Production-ready assemblies are adequately captured by executability, connectivity, interface validity, and absence of unintended interference (author definition of production-ready).
    Stated in footnote/intro; excludes materials, tolerances, process planning, yet underwrites the paper’s central phrasing.
  • domain assumption A fixed set of 12 port types and 7 mate types with a compatibility matrix C_τ is sufficient to express the target mechanical assemblies.
    Def. 3 and Table 2; classical assembly modeling reduced to this discrete interface language.
  • domain assumption Closed-form SE(3) mate transforms (Eq. 2) plus type-specific checks realize mates without iterative constraint solvers.
    Phase III; assumes mates in scope are expressible as deterministic rigid placements with optional parameters.
  • domain assumption Declared ports are valid iff type-specific B-Rep classifiers find cavity/face/material evidence (Def. 4).
    Elevates ports from annotations to geometric interfaces; sampling-based checks approximate full feature recognition.
  • ad hoc to paper 62 MECE engineering axioms (41 textbook-distilled via LLM pipeline + 21 synthesized) correctly justify mates and cover library component families.
    §3.2 and App. A; extraction and synthesis are LLM-assisted and curated, not independently validated against a standard ontology.
  • domain assumption LLM semantic decomposition with bounded repair yields structurally valid, axiom-grounded AssemblyDrafts often enough for the pipeline to dominate code-only generation.
    Alg. 1 and system overview; load-bearing for end-to-end success.
  • standard math SAFEUNION either returns a single solid or errors; it never silently returns a disconnected compound (Prop. 1).
    Engineering lemma with proof sketch over OpenCascade union cases; assumed correct in implementation.
  • standard math With fixed description and LLM response cache, outputs are byte-identical (Prop. 2).
    Content-addressed caching argument; reproducibility claim conditional on cache, not open stochastic LLM calls.
invented entities (4)
  • Assembly Specification S=(P,M,A) no independent evidence
    purpose: Intermediate representation bridging NL intent and geometric realization with parts, mates, and axiom references.
    Core invented IR (Def. 1); evaluated only inside this system’s pipeline.
  • cad-assembly-ext port-mate library no independent evidence
    purpose: Deterministic execution layer for mates, port checks, SAFEUNION, and factory adapters.
    New software abstraction over CadQuery; no external independent adoption reported in the paper.
  • AssemBench + APR/VLM-as-a-Judge protocol no independent evidence
    purpose: Provide assembly-centric evaluation beyond single-part CAD metrics and Chamfer Distance.
    New benchmark/protocol built on Fusion360 Gallery; judges are model-based rather than human gold standards released with the paper.
  • 62-axiom engineering knowledge base no independent evidence
    purpose: Ground mates in engineering principles and enable traceable verification reports.
    Curated/synthesized artifact; provenance claimed to textbooks but full set is paper-specific.

pith-pipeline@v1.1.0-grok45 · 25564 in / 4355 out tokens · 39383 ms · 2026-07-11T08:28:09.877615+00:00 · methodology

0 comments
read the original abstract

Recent advances in large language models and programmatic CAD have significantly improved Text-to-CAD generation for individual parts. However, production-ready mechanical assembly generation remains largely unsolved. Unlike single-part modeling, assemblies require coordinated reasoning over multiple components, functional interfaces, assembly relations, engineering principles, and physical consistency. Consequently, directly generating executable CAD code is insufficient for constructing mechanically valid and reusable assemblies. We present AssemCAD, an axiom-grounded framework for production-ready CAD assembly generation from natural language. Instead of representing an assembly as monolithic CAD code, AssemCAD first constructs an axiomatic Assembly Specification consisting of typed parts, geometry-backed ports, executable mates, and engineering axioms. Each assembly relation is explicitly grounded in one or more engineering principles, making the resulting specification interpretable, reusable, and verifiable. To realize this specification, AssemCAD introduces a port- and mate-based CAD assembly library that executes symbolic assembly relations through deterministic mate transformations and validates declared interfaces using concrete B-Rep geometric evidence. Built on this representation and library, AssemCAD further supports on-demand synthesis of reusable parametric component factories for both standard and open-world geometries. Experiments on AssemBench show that AssemCAD substantially improves assembly preservation and physical validity over code-centric CAD generation baselines, while generalizing across different foundation-model backbones. By combining axiom-grounded assembly reasoning with deterministic geometric execution, AssemCAD extends Text-to-CAD from isolated part generation toward production-ready mechanical assembly design.

Figures

Figures reproduced from arXiv: 2607.05123 by Botian Shi, Hongbin Zhou, Licheng Wen, Nianchen Deng, Pinlong Cai, Shu Zou, Siqi Li, Xinyu Cai, Xuemeng Yang, Yurui Dong.

Figure 1
Figure 1. Figure 1: Representative production-ready CAD assemblies generated directly from natural language. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Motivation and overview of ASSEMCAD. Existing foundation models perform well on single-part CAD generation but struggle with production-ready assembly generation, where failures commonly arise from invalid components, incorrect assembly relations, and inconsistent assembly configurations. To address these challenges, ASSEMCAD decomposes assembly generation into three stages: structured assembly specificati… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of ASSEMCAD. The framework first transforms natural-language descriptions into a structured Assembly Specification, and then realizes the specification through an execution pipeline for geometry construction and engineering verification. Verification failures are fed back to the LLM agent for specification refinement, while successful executions produce production-ready CAD assemblies together wit… view at source ↗
Figure 4
Figure 4. Figure 4: The pipeline of constructing AssemBench engineering-aware annotator that jointly reasons over geometric appearance and assembly structure to generate standardized CAD briefs. During annotation, the VLM is instructed to preserve engineering semantics, accurately describe functional components and assembly relationships, reflect structural complexity, and avoid unnecessary implementation details. This proces… view at source ↗
Figure 5
Figure 5. Figure 5: Representative examples from AssemBench. Each row shows decomposed CAD compo￾nents and the corresponding assembled mechanical system. G Benchmark Visualization [PITH_FULL_IMAGE:figures/full_fig_p024_5.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

43 extracted references · 7 linked inside Pith

  1. [1]

    Comparing 3d cad models: uses, methods, tools and perspectives.Computer-Aided Design and Applications, 9(6):771–794, 2012

    Antoine Brière-Côté, Louis Rivest, and Roland Maranzana. Comparing 3d cad models: uses, methods, tools and perspectives.Computer-Aided Design and Applications, 9(6):771–794, 2012

    2012

  2. [2]

    Brepnet: A topological message passing system for solid models

    Joseph G Lambourne, Karl DD Willis, Pradeep Kumar Jayaraman, Aditya Sanghi, Peter Meltzer, and Hooman Shayani. Brepnet: A topological message passing system for solid models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12773–12782, 2021

    2021

  3. [3]

    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. InProceedings of the IEEE/CVF international conference on computer vision, pages 6772–6782, 2021

    2021

  4. [4]

    BPB Publications, 2018

    Gaurav Verma.Autodesk fusion 360 black book. BPB Publications, 2018

    2018

  5. [5]

    Cadquery/cadquery: Cadquery 2.4

    Jeremy Wright, Marcus Boyd, Hasan Yavuz ÖZDERY A, Bruno Agostini, Michael Greminger, Seth Fischer, Justin Buchanan, Miguel Sánchez de León Peque, Martin Budden, Peter Boin, et al. Cadquery/cadquery: Cadquery 2.4. 0.Zenodo, 2024

    2024

  6. [6]

    Parametric cad modeling for open source scientific hardware: Comparing openscad and freecad python scripts.Plos one, 14(12):e0225795, 2019

    Felipe Machado, Norberto Malpica, and Susana Borromeo. Parametric cad modeling for open source scientific hardware: Comparing openscad and freecad python scripts.Plos one, 14(12):e0225795, 2019

    2019

  7. [7]

    Text-to-cadquery: A new paradigm for cad generation with scalable large model capabilities.arXiv preprint arXiv:2505.06507, 2025

    Haoyang Xie and Feng Ju. Text-to-cadquery: A new paradigm for cad generation with scalable large model capabilities.arXiv preprint arXiv:2505.06507, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Cad translator: An effective drive for text to 3d parametric computer-aided design generative modeling

    Xueyang Li, Yu Song, Yunzhong Lou, and Xiangdong Zhou. Cad translator: An effective drive for text to 3d parametric computer-aided design generative modeling. InProceedings of the 32nd ACM International Conference on Multimedia, pages 8461–8470, 2024

    2024

  9. [9]

    Cadsmith: Multi-agent cad generation with programmatic geometric validation.arXiv preprint arXiv:2603.26512, 2026

    Jesse Barkley, Rumi Loghmani, and Amir Barati Farimani. Cadsmith: Multi-agent cad generation with programmatic geometric validation.arXiv preprint arXiv:2603.26512, 2026

    arXiv 2026

  10. [10]

    Cme-cad: Heterogeneous collaborative multi-expert reinforcement learning for cad code generation

    Ke Niu, Haiyang Yu, Zhuofan Chen, Zhengtao Yao, Weitao Jia, Xiaodong Ge, Jingqun Tang, Benlei Cui, Bin Li, and Xiangyang Xue. Cme-cad: Heterogeneous collaborative multi-expert reinforcement learning for cad code generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 39272–39281, 2026

    2026

  11. [11]

    Itercad: An iterative multimodal agent for visually-grounded cad generation and editing

    Tao Hu, Jiaxin Ai, Licheng Wen, Xueheng Li, Shu Zou, Siqi Li, Nianchen Deng, Xinyu Cai, Hongbin Zhou, Pinlong Cai, et al. Itercad: An iterative multimodal agent for visually-grounded cad generation and editing. arXiv preprint arXiv:2606.13368, 2026

    Pith/arXiv arXiv 2026

  12. [12]

    Comact: Reframing professional software manipulation via com-as-action paradigm

    Jiaxin Ai, Tao Hu, Xuemeng Yang, Shu Zou, Hairong Zhang, Daocheng Fu, Yu Yang, Hongbin Zhou, Nianchen Deng, Pinlong Cai, et al. Comact: Reframing professional software manipulation via com-as-action paradigm. arXiv preprint arXiv:2606.13239, 2026

    Pith/arXiv arXiv 2026

  13. [13]

    Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts.Advances in Neural Information Processing Systems, 37:7552–7579, 2024

    Mohammad S Khan, Sankalp Sinha, Talha U Sheikh, Didier Stricker, Sk A Ali, and Muhammad Z Afzal. Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts.Advances in Neural Information Processing Systems, 37:7552–7579, 2024

    2024

  14. [14]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3763–3772, 2024

    2024

  15. [15]

    cadrille: Multi-modal cad recon- struction with online reinforcement learning.arXiv preprint arXiv:2505.22914, 2025

    Maksim Kolodiazhnyi, Denis Tarasov, Dmitrii Zhemchuzhnikov, Alexander Nikulin, Ilya Zisman, Anna V orontsova, Anton Konushin, Vladislav Kurenkov, and Danila Rukhovich. cadrille: Multi-modal cad recon- struction with online reinforcement learning.arXiv preprint arXiv:2505.22914, 2025

    arXiv 2025

  16. [16]

    Claude code

    Anthropic. Claude code. https://docs.anthropic.com/en/docs/claude-code, 2025. Accessed: 2026-06-30

    2025

  17. [17]

    Introducing codex

    OpenAI. Introducing codex. https://openai.com/index/introducing-codex/, 2025. Accessed: 2026-06-30

    2025

  18. [18]

    Gemini 2.5: Pushing the frontier with advanced reasoning and multimodality.arXiv preprint arXiv:2507.06261, 2025

    George Comanici et al. Gemini 2.5: Pushing the frontier with advanced reasoning and multimodality.arXiv preprint arXiv:2507.06261, 2025. 13

    Pith/arXiv arXiv 2025

  19. [19]

    Glm-4.5: Agentic, reasoning, and coding (arc) foundation models.arXiv preprint arXiv:2508.06471, 2025

    GLM-4.5 Team. Glm-4.5: Agentic, reasoning, and coding (arc) foundation models.arXiv preprint arXiv:2508.06471, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Qwen3.7.https://chat.qwen.ai/, 2026

    Alibaba Qwen Team. Qwen3.7.https://chat.qwen.ai/, 2026

    2026

  21. [21]

    Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks

    Xiang Xu, Karl DD Willis, Joseph G Lambourne, Chin-Yi Cheng, Pradeep Kumar Jayaraman, and Yasutaka Furukawa. Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks. InInternational Conference on Machine Learning, pages 24698–24724. PMLR, 2022

    2022

  22. [22]

    Hierarchical neural coding for controllable cad model generation

    Xiang Xu, Pradeep Kumar Jayaraman, Joseph George Lambourne, Karl DD Willis, and Yasutaka Furukawa. Hierarchical neural coding for controllable cad model generation. InInternational Conference on Machine Learning, pages 38443–38461. PMLR, 2023

    2023

  23. [23]

    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 Trans. Graph., 40(4), July 2021

    2021

  24. [24]

    Changjian Li, Hao Pan, Adrien Bousseau, and Niloy J. Mitra. Free2cad: parsing freehand drawings into cad commands.ACM Trans. Graph., 41(4), July 2022

    2022

  25. [25]

    Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced multimodal llms

    Siyu Wang, Cailian Chen, Xinyi Le, Qimin Xu, Lei Xu, Yanzhou Zhang, and Jie Yang. Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced multimodal llms. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 7880–7888, 2025

    2025

  26. [26]

    Cad-mllm: Unifying multimodality-conditioned cad generation with mllm.arXiv preprint arXiv:2411.04954, 2024

    Jingwei Xu, Chenyu Wang, Zibo Zhao, Wen Liu, Yi Ma, and Shenghua Gao. Cad-mllm: Unifying multimodality-conditioned cad generation with mllm.arXiv preprint arXiv:2411.04954, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward.Advances in Neural Information Processing Systems, 38:59765–59789, 2026

    Yandong Guan, Xilin Wang, Ximing Xing, Jing Zhang, Dong Xu, and Qian Yu. Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward.Advances in Neural Information Processing Systems, 38:59765–59789, 2026

    2026

  28. [28]

    Cad-recode: Reverse engineering cad code from point clouds

    Danila Rukhovich, Elona Dupont, Dimitrios Mallis, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-recode: Reverse engineering cad code from point clouds. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9801–9811, 2025

    2025

  29. [29]

    Lambourne, Karl D

    Pradeep Kumar Jayaraman, Aditya Sanghi, Joseph G. Lambourne, Karl D. D. Willis, Thomas Davies, Hooman Shayani, and Nigel Morris. Uv-net: Learning from boundary representations. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11698–11707, 2021

    2021

  30. [30]

    Complexgen: Cad reconstruction by b-rep chain complex generation.ACM Transactions on Graphics (TOG), 41(4):1–18, 2022

    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 (TOG), 41(4):1–18, 2022

    2022

  31. [31]

    Lambourne, Nishkrit Desai, Karl D

    Pradeep Kumar Jayaraman, Joseph G. Lambourne, Nishkrit Desai, Karl D. D. Willis, Aditya Sanghi, and Nigel J. W. Morris. Solidgen: An autoregressive model for direct b-rep synthesis, 2023

    2023

  32. [32]

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

    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 Trans. Graph., 43(4), July 2024

    2024

  33. [33]

    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 Trans. Graph., 40(6), December 2021

    2021

  34. [34]

    Assembly modelling by geometric constraint satisfaction.Computer-Aided Design, 28(9):707–722, 1996

    Ram Anantha, Glenn A Kramer, and Richard H Crawford. Assembly modelling by geometric constraint satisfaction.Computer-Aided Design, 28(9):707–722, 1996

    1996

  35. [35]

    A review on geometric constraint solving.arXiv preprint arXiv:2202.13795, 2022

    Qiang Zou, Zhihong Tang, Hsi-Yung Feng, Shuming Gao, Chenchu Zhou, and Yusheng Liu. A review on geometric constraint solving.arXiv preprint arXiv:2202.13795, 2022

    Pith/arXiv arXiv 2022

  36. [36]

    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. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15828–15839, 2022

    2022

  37. [37]

    Jones, Karl Willis, Brian Curless, and Adriana Schulz

    James Noeckel, Benjamin T. Jones, Karl Willis, Brian Curless, and Adriana Schulz. Mates2motion: Learning how mechanical cad assemblies work, 2023. 14

    2023

  38. [38]

    Willis, and Wojciech Matusik

    Yunsheng Tian, Jie Xu, Yichen Li, Jieliang Luo, Shinjiro Sueda, Hui Li, Karl D. Willis, and Wojciech Matusik. Assemble them all: Physics-based planning for generalizable assembly by disassembly.ACM Transactions on Graphics, 41(6)

  39. [39]

    Yunsheng Tian, Karl D. D. Willis, Bassel Al Omari, Jieliang Luo, Pingchuan Ma, Yichen Li, Farhad Javid, Edward Gu, Joshua Jacob, Shinjiro Sueda, Hui Li, Sachin Chitta, and Wojciech Matusik. Asap: Automated sequence planning for complex robotic assembly with physical feasibility. In2024 IEEE International Conference on Robotics and Automation (ICRA), pages...

    2024

  40. [40]

    Cad-assistant: tool-augmented vllms as generic cad task solvers

    Dimitrios Mallis, Ahmet Serda Karadeniz, Sebastian Cavada, Danila Rukhovich, Niki Foteinopoulou, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad-assistant: tool-augmented vllms as generic cad task solvers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7284–7294, 2025

    2025

  41. [41]

    A solver-aided hierarchical language for llm-driven cad design

    Benjamin T Jones, Zihan Zhang, Felix Hähnlein, Wojciech Matusik, Maaz Ahmad, Vladimir Kim, and Adriana Schulz. A solver-aided hierarchical language for llm-driven cad design. InComputer Graphics Forum, volume 44, page e70250. Wiley Online Library, 2025. 15 Appendix Contents A Engineering Axiom Set 16 A.1 Axiom Extraction Pipeline . . . . . . . . . . . . ....

    2025

  42. [42]

    Coverage: every port type and component class in the CAD extension library (Appendix B) must be covered by at least one axiom

  43. [43]

    part_name:port_name

    DAG consistency: precedes/follows relationships among axioms must form a directed acyclic graph. Where the textbook corpus lacked coverage for component families present in the CAD libraries (e.g., certain bearing configurations),synthesized axiomsare created with explicit provenance markers and engineering justification. The disposition of the 139 raw ax...