Pith. sign in

REVIEW 2 major objections 4 minor 25 references

The paper claims CAD reverse engineering should proceed one operation at a time, choosing each step against the residual between the target mesh and the executed build — and that this loop beats single-pass systems by 12+ match-score points

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

2026-08-05 00:12 UTC pith:O7DPFUQI

load-bearing objection Serious empirical paper on stepwise CAD reverse engineering with honest reporting; the central mechanism claim is real but not yet isolated from the training corpus, so the missing same-corpus control should be resolved before acceptance. the 2 major comments →

arxiv 2608.00799 v1 pith:O7DPFUQI submitted 2026-08-01 cs.CV

CADENA: Stepwise CAD Reverse Engineering

Soslan Kabisov , Gennadiy Savrasov , Maksim Elistratov , Antonio Rodriguez , Daniil Ignatiev , Nikita Gavrilov , Rustam Uzdenov , Alexey I. Boyko
show 4 more authors
Igor Pasechnik Anton Konushin Andrey Kuznetsov Dmitrii Zhemchuzhnikov
This is my paper
classification cs.CV
keywords CAD reverse engineeringparametric CADvision-language modelsstepwise program generationgeometric feedbackreinforcement learningCADENA-Benchsurface-match metric
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.

Most AI systems that turn a 3D mesh into an editable CAD program emit the whole program in one pass and never look at what their code builds; the paper claims that is the wrong architecture, because early mistakes compound silently and choices follow the statistics of the training corpus rather than the shape in front of the model. CADENA is the alternative: a vision–language model adds one operation, executes it, renders the current build over the target (green versus red channel, eight viewpoints), and picks the next operation from that residual, with a hint point marking the region to modify — a comparison that costs nothing at test time, since the target mesh is the input to reverse engineering. Trained on 19.86 million generated steps and refined by reinforcement learning rewarded by the executed geometry, it reports a surface-match score (GMS, which requires surfaces to agree in position and normal direction) of 67.0 on CADENA-Bench, a new benchmark of 3396 real mechanical parts, against 54.8 for the best baseline, with invalid programs below 1.2% where baselines reach 31.7%. The paper also argues that the usual metrics (IoU, Chamfer distance) reward shape agreement rather than program recovery, and offers GMS as the grading instrument for reverse engineering. The stakes are practical: engineers edit parametric programs, not meshes, and the reported margin widens precisely on the real mechanical parts where single-pass systems fail.

Core claim

The paper's central claim is that conditioning each CAD operation on the residual between target and executed build — and never on the program text — is what drives reconstruction quality, and that this is practical because the target mesh is the input to reverse engineering, available at test time. CADENA formalises reconstruction as a sequential decision problem whose observation is a multi-view overlay image (target in green, build in red, depth as colour intensity) plus a text hint point marking the region to modify, with the executed build as the only history. A DSL whose every operation consumes and rebinds one variable makes any step a valid continuation of any prefix, so the rendered

What carries the argument

The load-bearing object is the observation function: one image tiling six orthographic and two isometric views, with the target mesh in the green channel and the current build in red, depth encoded as colour intensity. Yellow marks overlap, green marks missing geometry, red marks excess material, making the residual directly readable. A hint point — a 3D coordinate estimated at inference by clustering the target's discrepancy regions — tells the policy which region to modify. The companion mechanism is the DSL's single-variable design: every operation consumes and rebinds one variable r, so any step is a valid continuation of any prefix and the rendered image alone can drive the choice. Toge

Load-bearing premise

The method assumes that the eight-view rendered overlay of target and current build, plus one hint point, is enough information to pick the correct next operation — so any geometry hidden from every view (internal cavities, blind bores, undercuts) can never be recovered, a failure the paper itself documents in Appendix A.7.

What would settle it

Two experiments would settle it. First, run CADENA on parts whose internal cavities or blind bores project identically to solid regions in all eight views: if no step budget recovers them, the observation is insufficient — the paper's Appendix A.7 already predicts exactly this. Second, the control the paper says it has not run: train a single-pass model on the identical ~20-million-step corpus and reward; if it closes most of the measured gap, the gain belongs to data and reward engineering rather than to stepwise inference.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Single-pass generation becomes the wrong default for reconstruction: a decoder that never inspects the geometry its code produces inherits corpus statistics and compounds early errors, so stepwise loops with executed-geometry feedback are the route that scales to real mechanical parts.
  • Shape-agreement metrics (IoU, Chamfer distance) are the wrong grading instrument for reverse engineering, since a body built from the wrong primitives can satisfy them; surface-type agreement metrics such as GMS should replace them for this task.
  • Validity is a learnable property: with a reward that scores unbuildable programs as zero, the invalid rate falls from as high as 31.7% to below 1.2% on external sets and 1.6% on CADENA-Bench, with no post-hoc filtering.
  • Stepping is a targeted rescue rather than a uniform gain: the hardest quartile of parts gains 14–21 GMS points from the loop while the easiest quartile gains none, so the method raises the ceiling without touching the floor.
  • Because the target mesh is the input, comparing every executed prefix against it at test time is free, the policy extrapolates from a 10-step training horizon to a 20-step budget without degradation, and the paper notes that selection alone is worth little — the gain is the residual conditioning itself.

Where Pith is reading between the lines

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

  • The eight-view render is the method's ceiling, not a tunable detail: geometry hidden from every canonical viewpoint (internal cavities, blind bores, undercuts) is unrecoverable at any step budget — the paper's Appendix A.7 confirms the loop 'converges happily to a solid whose exterior matches' — so a point-cloud or cross-section input modality, which the paper names as motivation, is the natural n
  • The context-free DSL that makes each step executable in isolation also caps what can be expressed: features identified by construction history, and countable features like gear teeth whose residual barely changes when the count is wrong, are approximated rather than recovered — the per-family gap on gears (58.1 versus 61.3 for direct fitting) is the visible cost.
  • Because the geometric reward needs no reference programs, the same RL loop transfers to corpora like MCB that provide none; a family-conditioned or per-family policy could plausibly close the reported gears gap.
  • The paper itself states it has not yet run a single-pass model on its own training corpus (Section 6), so the share of the measured gain attributable to stepping rather than to the 20-million-step data remains unquantified.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper introduces CADENA, a stepwise CAD reverse-engineering system that reconstructs a target mesh by emitting one parametric operation at a time, executing it, rendering the target and current build as a multi-view overlay, and conditioning the next operation on this residual geometry rather than on the program text. Training combines supervised fine-tuning on a large procedural corpus (18M stepwise samples) with online reinforcement learning whose reward is the volumetric IoU between the executed partial build and the target mesh. The paper also introduces CADENA-Bench, a deduplicated benchmark of 3,396 real mechanical parts in six families, and GMS, a surface-type matching metric. The main experiments compare CADENA-SFT and CADENA-RL against released checkpoints of prior methods on DeepCAD, Fusion360, MCB, and CADENA-Bench, and additionally report a self-graded BenchCAD result.

Significance. If the central claim holds, CADENA is a substantial advance for mesh-to-CAD reconstruction: it demonstrates a practical closed-loop inference scheme in which the target mesh serves as an external, always-available objective, and it reports large margins on real mechanical parts, particularly on MCB and CADENA-Bench. The paper is unusually transparent about protocol choices: the fixed-frame evaluation and valid-only averaging are stated explicitly, their quantitative effect on a baseline is shown (cadrille MCB IoU moves 47.6 to 67.0 under the averaging convention, then to 66.2 under the fixed frame), and the failure modes in Appendix A.7 are honestly documented. The authors also release code, weights, and the benchmark. The central mechanistic claim, however, is not yet fully supported because no same-corpus single-pass control is provided, and the BenchCAD comparison is explicitly non-like-for-like.

major comments (2)
  1. The load-bearing attribution of the gains to stepwise execution feedback is not established. The k=1 ablation in Fig. 4 and Table 10 is not a single-pass model: it is the trained stepwise policy decoded with a budget of one operation, trained on stepwise snippets from an 18M-sample procedural corpus and refined with executed-geometry RL. Such a control cannot separate the inference loop from the training distribution, model scale, or reward design. The authors concede this in Section 6 ('we do not yet report a single-pass model trained on our own corpus'), but the abstract and contributions still assert the mechanism as the source of improvement. A same-corpus single-pass model—trained on the same generator, with the same backbone, and followed by the same RL recipe—is needed to support the claim that residual-geometry conditioning is what drives the reported margins. Until then, the cen
  2. The contribution bullet claims 'best reported results on all five datasets we evaluate,' including BenchCAD, but Table 3 is explicitly not like-for-like and its own note says it 'should not be read as a ranking.' CADENA receives the target mesh, renders its own views, and selects its returned prefix by IoU against the same ground truth the benchmark scores. The table is informative as an indicative comparison, but the contribution bullet overstates it. Either qualify the BenchCAD entry in the contributions or restrict the 'best reported results' claim to the controlled external datasets.
minor comments (4)
  1. Labeling the k=1 budget as 'a single-pass model' is misleading; it is a one-step decoding of the stepwise policy. Use 'policy decoded with budget 1' to avoid confusion with a genuinely single-pass trained model.
  2. The 70% human-agreement rate for family labels is acknowledged in the text but not in the table. Adding the caveat to the caption of Table 2 (and Tables 13–14) would prevent over-reading of per-family differences, especially for the tooling & gauges column with n=17.
  3. The protocol-change chain (47.6 → 67.0 → 66.2) is valuable. Consider also reporting the effect of the fixed-frame convention on CADENA itself, so the reader can see how much of the method's reported margin is protocol-dependent.
  4. The explicit omission list for non-comparable methods is exemplary and could be moved or summarized in the main text's related work, since silent omissions are a common reproducibility problem in this area.

Circularity Check

0 steps flagged

No circularity found: the stepwise reconstruction loop is driven by the external target mesh, and the acknowledged missing control is an attribution limit, not a definitional reduction.

full rationale

The paper's central derivation is self-contained with respect to its inputs. At each step the policy observes the target mesh and the current executed build, and the target is legitimately available at inference because the mesh is the input to reverse engineering, not a hidden label. The RL reward is volumetric IoU against that external mesh, and the paper explicitly shows that optimizing the evaluation metric GMS as a reward degrades geometry (Appendix A.3.4), so the reported GMS numbers are not fitted by construction. The claimed benefit of stepwise inference is supported by a within-model ablation (Appendix A.4) comparing operation budgets k=1 and k=20 on the same checkpoint; calling k=1 a 'single-pass model' in the Figure 4 caption is an overstatement, and the paper itself concedes in Section 6 that a same-corpus single-pass model is not yet reported. That is a missing control for attributing the gain to stepwise inference versus training data, but it is not circularity: no equation reduces to its inputs and no fitted parameter is renamed as a prediction. Self-citations to CADReasoner and CADEvolve are used only for the rendering protocol and data-generation approach, not to justify the central mechanism or to exclude alternatives; no uniqueness theorem is imported from the authors. The BenchCAD comparison is explicitly disclosed as not like-for-like because CADENA consumes the mesh and selects prefixes by IoU against that same mesh, and the paper states the gap should be attributed to the richer input and test-time selection rather than to the generator alone. The occluded-geometry failure mode (Appendix A.7) is an honest limitation bounding the method, not a circular validation. Overall, no step in the derivation chain is equivalent to its own inputs by definition.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The method carries five assumptions worth paying for explicitly: CAD kernel execution is a faithful oracle; the eight-view overlay plus hint point is a sufficient observation; the single-variable chain DSL can express the test parts; GMS surface agreement implies construction-tree correctness; and the synthetic generator's programs transfer to real parts. The free parameters are method hyperparameters (thresholds, tolerances, world scale, RL settings) rather than quantities fitted to game the benchmarks, but the GMS angular tolerance (25 degrees) materially shapes every reported GMS number.

free parameters (5)
  • discrepancy threshold tau = not stated numerically (relative to normalized frame)
    Defines which surface points count as missing geometry for hint-point sampling in training (interval (tau, 3tau)) and inference (walk termination at 2tau). Chosen by hand; central to the hint mechanism. Section 3.2.
  • GMS angular tolerance alpha_max = 25 degrees
    Upper bound of the angular tolerance integral in GMS. The paper's sensitivity analysis says the angular tolerance dominates the score, so this choice materially shapes every reported GMS number. Section 4.2.
  • GMS sample count N = 8192
    Sets the distance tolerance tau = N^(-1/3) ~ 0.05, chosen so tau equals mean sampling spacing. The tolerance is derived from N rather than free, but N itself is chosen. Section 4.2.
  • normalized world frame scale = 200 units; coordinates rounded to integers in [-100, 100]
    All DSL coordinates are absolute integers in this frame, making scale and placement errors count in evaluation. Chosen for tokenization; shapes the difficulty of the task. Section 3.3.
  • RL rollout hyperparameters = group size 32, T=1.0, keep 2 candidates per group, effective batch 48, 20 epochs, 10-op cap
    Standard GRPO-style settings chosen by the authors; not fitted to the benchmarks but set by hand and affecting the final policy. Section 3.4.
axioms (5)
  • domain assumption CadQuery/OpenCASCADE execution is a faithful oracle for the geometry of a generated program
    Every training, reward, and inference signal comes from executing DSL lines in CadQuery/OpenCASCADE. If execution semantics diverge from the intended solid, the feedback loop is corrupted. Sections 3.3, 3.4.
  • domain assumption The eight-view green/red overlay plus one hint point is a sufficient observation for selecting the next operation
    The policy conditions only on this image and hint, never on program text. The paper's own failure analysis shows occluded geometry is invisible to the observation, so the assumption is known to fail on parts with internal cavities or undercuts. Section 3.2, A.7.
  • domain assumption The single-variable chain DSL can express the constructions needed for the test parts
    The DSL forbids references to named features of earlier operations so that any step is a valid continuation of any prefix. Out-of-vocabulary features are approximated by stacks of extrusions; whether real engineering parts are adequately representable is assumed. Section 3.3, Section 6.
  • domain assumption Surface-type agreement (GMS) implies the construction tree is correct
    Stated explicitly in Section 4.2 as 'an assumption, but the weakest we know of that separates a recovered program from a recovered silhouette'.
  • domain assumption Synthetic procedural programs transfer to real mechanical parts
    Supervised training is entirely on a rule-based generator; the RL phase on about 4k real meshes (training splits only) partially mitigates domain shift. Transfer is assumed and only indirectly validated by held-out test performance. Sections 3.4, 5.
invented entities (1)
  • hint point no independent evidence
    purpose: A 3D coordinate fed to the policy as text that identifies which discrepancy region to edit next, biasing the model to extend existing geometry rather than start disconnected components.
    An algorithmic mechanism, not a physical entity. It has no external falsifiable handle; its value is validated only inside the pipeline, and an ablation removing it is not reported.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CADENA: Stepwise CAD Reverse Engineering." pith.science (2026). https://pith.science/paper/O7DPFUQI

@misc{pith2026260800799,
  author       = {Pith},
  title        = {Pith review of: CADENA: Stepwise CAD Reverse Engineering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O7DPFUQI}},
  note         = {Machine review of arXiv:2608.00799}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Computer-Aided Design (CAD) underpins modern engineering, yet converting existing shapes into editable models still demands substantial expert effort. Most AI systems emit the entire CAD program in a single pass, never inspecting the intermediate geometry. In contrast, human engineers build a part feature by feature, checking after each operation what remains to be modeled. We introduce CADENA (Spanish for "chain"), a model that reconstructs a 3D mesh as a parametric CAD program, growing its sequence of operations one at a time and comparing the target with the currently predicted geometry at every step. We also address the lack of benchmarks for evaluating reverse-engineering methods on mechanical parts, introducing CADENA-Bench, a benchmark that measures performance across categories of mechanical parts. CADENA outperforms prior methods on CADENA-Bench and on the DeepCAD, Fusion 360, and MCB datasets. Code is available at https://github.com/zhemdi/cadena, model weights at https://huggingface.co/kulibinai/cadena, and CADENA-Bench at https://huggingface.co/datasets/kulibinai/cadena-bench.

Figures

Figures reproduced from arXiv: 2608.00799 by Alexey I. Boyko, Andrey Kuznetsov, Antonio Rodriguez, Anton Konushin, Daniil Ignatiev, Dmitrii Zhemchuzhnikov, Gennadiy Savrasov, Igor Pasechnik, Maksim Elistratov, Nikita Gavrilov, Rustam Uzdenov, Soslan Kabisov.

Figure 1
Figure 1. Figure 1: CADENA overview. The target mesh (green channel) and the current partial prediction (red channel) are rendered into a single aligned multi-view image consisting of six orthographic views (with depth encoded as color intensity) and two isometric views. Predominantly green regions indicate missing geometry, while predominantly red regions indicate excess material. Given this image and a hint point specifying… view at source ↗
Figure 2
Figure 2. Figure 2: Stepwise reconstruction by CADENA. Each row follows one part from its first operation to the returned program. Columns are selected steps, annotated with the step index and the IoU of the build against the target at that point; the DSL line emitted at the step appears beneath each panel, and the rightmost column is the ground truth. The three parts are returned after 17, 4 and 16 operations. The model esta… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison across methods. Each row is one input part, labelled with the dataset it comes from; each column is one method, with the target in the rightmost column. The GMS of that single reconstruction is printed beneath each panel. Grey panels mark predictions that failed to build, and predictions that build without being watertight are annotated as such; both count toward the invalid rate and… view at source ↗
Figure 4
Figure 4. Figure 4: Inference-time ablations, all replayed from the stepwise records of the reported model. (a) Quality against the operation budget k: k = 1 is a single-pass model, and the curve saturates around eight operations, so the budget of 20 used throughout is not a tuned quantity. (b) The gain is concentrated in the hardest quartile of each dataset — parts whose first operation scores worst — and is essentially zero… view at source ↗
Figure 5
Figure 5. Figure 5: Characteristic failures. Five parts on which CADENA fails, with the target above and the returned reconstruction below. They illustrate three of the four modes discussed in this section: countable features are approximated rather than counted, so the impeller and the radial fan keep the right body and the wrong blades; helical geometry outside the DSL’s vocabulary is replaced by the nearest available primi… 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

25 extracted references · 10 canonical work pages · 4 internal anchors

  1. [2]

    HierCAD: Hierarchical Text-to-CAD Design via Structure Alignment and Parameter Grounding

    Jimin Xu, Tianbao Wang, Tao Jin, and Zhou Zhao. HierCAD: Hierarchical text-to-CAD design via structure alignment and parameter grounding.arXiv preprint arXiv:2607.11339,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [3]

    This appendix varies one thing only, the reward, and reports what each variant does to reconstruction quality. It carries the evidence behind two statements made there: that the KL penalty does not earn its place in the objective, and that GMS is used strictly for evaluation and never as a training signal. A.3.1 Setup All configurations below are trained ...

    work page arXiv

  3. [4]

    CADEvolve: Creating realistic CAD via program evolution.arXiv preprint arXiv:2602.16317,

    Maksim Elistratov, Marina Barannikov, Gregory Ivanov, Valentin Khrulkov, Anton Konushin, Andrey Kuznetsov, and Dmitrii Zhemchuzhnikov. CADEvolve: Creating realistic CAD via program evolution.arXiv preprint arXiv:2602.16317,

    arXiv

  4. [5]

    Pointer-CAD: Unifying B-Rep and command sequences via pointer-based edges and faces selection.arXiv preprint arXiv:2603.04337,

    Dacheng Qi, Chenyu Wang, Jingwei Xu, Tianzhe Chu, Zibo Zhao, Wen Liu, Wenrui Ding, Yi Ma, and Shenghua Gao. Pointer-CAD: Unifying B-Rep and command sequences via pointer-based edges and faces selection.arXiv preprint arXiv:2603.04337,

    Pith/arXiv arXiv

  5. [6]

    HistCAD: A constraint-aware parametric history-based CAD representation, dataset, and benchmark with industrial complexity.arXiv preprint arXiv:2602.19171, 2026a

    Xintong Dong, Chuanyang Li, Peng Zheng, Chuqi Han, Jiaxin Jing, Hailong Shen, Yanzhi Song, and Zhouwang Yang. HistCAD: A constraint-aware parametric history-based CAD representation, dataset, and benchmark with industrial complexity.arXiv preprint arXiv:2602.19171, 2026a. Haoxiang Guo, Shilin Liu, Hao Pan, Yang Liu, Xin Tong, and Baining Guo. ComplexGen: ...

    Pith/arXiv arXiv

  6. [7]

    B-Rep distance functions: How to represent a B-Rep model by volumetric distance functions?arXiv preprint arXiv:2511.14870,

    Fuyang Zhang, Pradeep Kumar Jayaraman, Xiang Xu, and Yasutaka Furukawa. B-Rep distance functions: How to represent a B-Rep model by volumetric distance functions?arXiv preprint arXiv:2511.14870,

    arXiv

  7. [9]

    cadrille: Multi-modal CAD reconstruction with reinforcement learning.arXiv preprint arXiv:2505.22914,

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

    arXiv

  8. [10]

    Doris, Md Ferdous Alam, Amin Heyrani Nobari, and Faez Ahmed

    Anna C. Doris, Md Ferdous Alam, Amin Heyrani Nobari, and Faez Ahmed. CAD-Coder: An open-source vision- language model for computer-aided design code generation.arXiv preprint arXiv:2505.14646,

    Pith/arXiv arXiv

  9. [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, Daocheng Fu, Yu Yang, Hairong Zhang, Botian Shi, and Xuemeng Yang. IterCAD: An iterative multimodal agent for visually-grounded CAD generation and editing.arXiv preprint arXiv:2606.13368,

    work page internal anchor Pith review Pith/arXiv arXiv

  10. [12]

    COSMO-Agent: Tool-Augmented Agent for Closed-loop Optimization,Simulation,and Modeling Orchestration

    Liyuan Deng, Shujian Deng, Yongkang Chen, Yongkang Dai, Zhihang Zhong, Linyang Li, Xiao Sun, Yilei Shi, and Huaxi Huang. COSMO-Agent: Tool-augmented agent for closed-loop optimization, simulation, and modeling orchestration.arXiv preprint arXiv:2604.05547,

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [13]

    CAD-Coder: Text-to-CAD generation with chain-of-thought and geometric reward.arXiv preprint arXiv:2505.19713,

    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.arXiv preprint arXiv:2505.19713,

    Pith/arXiv arXiv

  12. [14]

    CME-CAD: Heterogeneous collaborative multi-expert reinforcement learning for CAD code generation.arXiv preprint arXiv:2512.23333,

    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.arXiv preprint arXiv:2512.23333,

    arXiv

  13. [15]

    CAD-Judge: Toward efficient morphological grading and verification for text-to-CAD generation.arXiv preprint arXiv:2508.04002,

    Zheyuan Zhou, Jiayi Han, Liang Du, Naiyu Fang, Lemiao Qiu, and Shuyou Zhang. CAD-Judge: Toward efficient morphological grading and verification for text-to-CAD generation.arXiv preprint arXiv:2508.04002,

    Pith/arXiv arXiv

  14. [16]

    PR-CAD: Progressive refinement for unified controllable and faithful text-to-CAD generation with large language models.arXiv preprint arXiv:2604.19773,

    Jiyuan An, Jiachen Zhao, Fan Chen, Liner Yang, Zhenghao Liu, Hongyan Wang, Weihua An, Meishan Zhang, and Erhong Yang. PR-CAD: Progressive refinement for unified controllable and faithful text-to-CAD generation with large language models.arXiv preprint arXiv:2604.19773,

    Pith/arXiv arXiv

  15. [17]

    At 82% of that ceiling a tabulated number would confound their method with our reconstruction of their renderer, so we omit it

    Method CD 8k CD30k IoU GMS IR CD 8k CD30k IoU GMS IR CD 8k CD30k IoU GMS IR CAD-Recode 0.3931 0.3175 72.96 58.79 33.5 1.1215 0.9922 67.17 45.31 23.4 1.8759 1.7713 67.27 38.66 23.5 cadrille 0.3682 0.2645 70.82 62.060.91.4312 1.2847 66.03 46.311.02.1844 2.0251 75.48 41.790.0 CADReasoner 0.2964 0.1945 74.57 69.67 29.4 2.8767 2.7127 62.75 43.36 33.7 1.8975 1....

    work page arXiv 1945

  16. [18]

    Text2CAD-Bench: A benchmark for LLM-based text-to-parametric CAD generation.arXiv preprint arXiv:2605.18430,

    Liang Wang, Heng Meng, Zekai Xiang, Jin Liu, Pingyi Zhou, Litao Chen, and Yongqiang Tang. Text2CAD-Bench: A benchmark for LLM-based text-to-parametric CAD generation.arXiv preprint arXiv:2605.18430,

    Pith/arXiv arXiv

  17. [19]

    MUSE: Benchmarking manufacturable, functional, and assemblable text-to-CAD generation.arXiv preprint arXiv:2605.28579, 2026b

    Xiaoyu Dong, Zhi Li, and Xiao-Ming Wu. MUSE: Benchmarking manufacturable, functional, and assemblable text-to-CAD generation.arXiv preprint arXiv:2605.28579, 2026b. Haozhe Zhang, Kaichen Liu, Miaomiao Chen, Lei Li andShaojie Yang, Cheng Peng, and Hanjie Chen. Benchcad: A comprehensive, industry-standard benchmark for programmatic cad.arXiv preprint arXiv:...

    Pith/arXiv arXiv

  18. [20]

    16 CADENA: Stepwise CAD Reverse EngineeringA PREPRINT A Supplementary Material This appendix reports the reward-variant experiments behind the RL design (Section A.3), the greedy-versus-sampling comparison (Section 5.3), the surface grammar of the DSL (Section A.2), and the recurring failure modes of the method (Section A.7). A.1 Comparison with SOV-CAD S...

    work page 2026

  19. [22]

    RL (MPR, 0.9/0.1) was not evaluated on the full splits. deepcad_test_mesh fusion360_test_mesh mcb_test_mesh Model IoU↑CD 8k ↓GMS↑IR↓IoU↑CD 8k ↓GMS↑IR↓IoU↑CD 8k ↓GMS↑IR↓ SFT 0.9167 0.1581 0.9528 0.0263 0.8880 0.1422 0.9079 0.0377 0.7524 0.4316 0.6339 0.1204 RL (IoU)0.96100.1477 0.97000.00350.9410 0.1243 0.9334 0.01220.8825 0.2236 0.7373 0.0074 RL (IoU + KL...

    work page arXiv

  20. [23]

    four front-facing and four rear-facing

    and in nearly every individual class (Tables 13–14), with the sole exception of GMS on class Gears & bearings, where CADFit† is marginally higher despite its much higher IR on that class. At the same time, CADENA-RL maintains one of the lowest IR values among all models in every class, trailing only cadrille, which achieves a lower IR at the cost of subst...

    work page 2026

  21. [359]

    Method CD 8k CD30k IoU GMS IR CD 8k CD30k IoU GMS IR CD 8k CD30k IoU GMS IR CAD-Recode 1.0089 0.8402 71.63 53.75 20.1 1.2859 1.0795 62.34 39.09 35.9 1.1932 1.0762 73.50 56.08 34.3 cadrille 1.0506 0.9071 73.29 57.550.91.4828 1.2861 65.72 37.180.41.8209 1.7479 70.12 54.460.6 CADReasoner 2.0882 1.9453 69.41 55.55 21.0 2.5867 2.4641 65.69 46.90 74.2 2.1856 2....

    work page arXiv 1932

  22. [2022]

    Text-to-CAD Evaluation with CADTests

    Dimitrios Mallis, Marco Wang, Ahmet Serdar Karadeniz, Elisa Ricci, Anis Kacem, and Djamila Aouada. Text-to-CAD evaluation with CADTests.arXiv preprint arXiv:2605.07807,

    work page internal anchor Pith review Pith/arXiv arXiv

  23. [2023]

    Residual primitive fitting of 3d shapes with SuperFrusta.arXiv preprint arXiv:2512.09201,

    Aditya Ganeshan, Matheus Gadelha, Thibault Groueix, Zhiqin Chen, Siddhartha Chaudhuri, Vladimir Kim, Wang Yifan, and Daniel Ritchie. Residual primitive fitting of 3d shapes with SuperFrusta.arXiv preprint arXiv:2512.09201,

    arXiv

  24. [2025]

    CAD-MLLM: Unifying multimodality- conditioned CAD generation with MLLM.arXiv preprint arXiv:2411.04954, 2024a

    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, 2024a. Ruiyu Wang, Shizhao Sun, Weijian Ma, and Jiang Bian. CAD-Tokenizer: Towards text-based CAD prototyping via modality-specific tokenization.arXiv preprint arXiv:2509.21150,

    Pith/arXiv arXiv

  25. [2026]

    Zero-to-CAD: Agentic synthesis of interpretable CAD programs at million-scale without real data.arXiv preprint arXiv:2604.24479,

    Mohammadmehdi Ataei, Farzaneh Askari, Kamal Rahimi Malekshan, and Pradeep Kumar Jayaraman. Zero-to-CAD: Agentic synthesis of interpretable CAD programs at million-scale without real data.arXiv preprint arXiv:2604.24479,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.