pith. sign in

arxiv: 2606.11152 · v2 · pith:4R57AD37new · submitted 2026-06-09 · 💻 cs.CV

P3D-Bench: Benchmarking MLLMs for Parametric 3D Generation and Structural Reasoning

Yikang Yang , Zhanpeng Hu , Youtian Lin , Mengqi Zhou , Jingxi Xu , Feihu Zhang , Jiaheng Liu , Yao Yao This is my paper

Pith reviewed 2026-06-27 13:46 UTC · model grok-4.3

classification 💻 cs.CV
keywords parametric 3D generationMLLM evaluationtext-to-3Dimage-to-3Dassembly reasoninggeometric fidelitystructural consistency
0
0 comments X

The pith

P3D-Bench shows MLLMs recover global 3D shapes and semantics but fail to match precise parametric geometry and part structures.

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

The paper introduces P3D-Bench to evaluate how well multimodal large language models generate parametric 3D programs that must be geometrically precise, semantically aligned, and assembly-consistent from text, image, or assembly specifications. It applies a unified scoring protocol across Text-to-3D, Image-to-3D, and Assembly-3D tasks on hundreds of cases, measuring executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment, and part-level structure. The evaluation of frontier models finds that assemblies are hardest, global shape and identity are often recovered, yet exact parametric dimensions and correct part counts remain weak.

Core claim

P3D-Bench demonstrates that current MLLMs and text-only LLMs can often recover the global shape and semantic identity of target objects yet fail to reproduce the precise parametric geometry and part relations specified by the input, with part-level modeling especially weak on assemblies where neither individual part geometry nor the right number of parts is recovered.

What carries the argument

P3D-Bench benchmark protocol that scores parametric 3D program outputs for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment, and part-level structure across three task families.

If this is right

  • Assembly tasks expose the greatest gap in compositional structure modeling.
  • Global semantic recovery does not imply recovery of specified parametric values.
  • Part-level accuracy, including correct part counts and individual geometries, requires targeted improvement.

Where Pith is reading between the lines

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

  • The results point to a separation between semantic priors and the ability to enforce explicit geometric constraints in generated code.
  • The benchmark could be used to test whether fine-tuning on parametric constraint examples closes the observed gaps.
  • Similar evaluation protocols might apply to other domains where code must produce exact physical outputs rather than approximate appearances.

Load-bearing premise

The chosen metrics for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment and part-level structure are sufficient to measure whether a generated program recovers a design's structure.

What would settle it

A model that scores high on all P3D-Bench metrics but whose executed 3D output deviates in exact dimensions, topology, or part count from the input specification would show the metrics do not capture structural recovery.

Figures

Figures reproduced from arXiv: 2606.11152 by Feihu Zhang, Jiaheng Liu, Jingxi Xu, Mengqi Zhou, Yao Yao, Yikang Yang, Youtian Lin, Zhanpeng Hu.

Figure 1
Figure 1. Figure 1: Scores of different models across the three tasks in P3D-BENCH. The Score is the average of the four bucket scores (Geo, Topo, Judge, Part; §3.3), rescaled to 0–100; per-bucket results are reported in Tables 3a–3c. ABSTRACT Multimodal large language models can write code to produce complex programs as well as use programs to do 3D modeling, which opens up a new avenue for 3D generation powered by their pri… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the P3D-BENCH evaluation. Given a task input—a text specification (de￾scriptive or parametric), a single image, or an image with assembly-level and part-level annotations—a model (general-purpose LLM/MLLM or domain-specific) writes a program in one of four formats (JSON, OpenSCAD, CadQuery, Three.js). P3D-BENCH executes and renders each program, reports validity (Valid), and scores it along fou… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of our P3D-Dataset. The filter pipeline draws from the Text2CAD and Fusion 360 Gallery data sources, filters candidates with a review MLLM, removes near-duplicates, and samples 400 Text-to-3D and 400 Image-to-3D cases with category and complexity distributions. The annotation pipeline then annotates the filtered Text-to-3D data with an annotation MLLM into descriptive (desc) and parametric (param)… view at source ↗
Figure 4
Figure 4. Figure 4: The Text-to-3D and Image-to-3D 400-case filtered sets. visually redundant even when their source programs differ. Finally, complexity-balanced sampling draws the retained cases across semantic categories and easy, medium and hard tiers, yielding 400 Text-to-3D cases and 400 Image-to-3D cases without collapsing to only simple or extreme examples. Annotation Pipeline. Annotation converts retained geometry in… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the Judge bucket metrics. The Judge bucket combines two MLLM-based metrics. The QA metric builds a QA-S bank from the descriptive specification and a QA-P bank from the parametric specification, then scores the prediction by how many of those questions it answers correctly; the visual Judge (J-Sem / J-Geo / J-Aes) scores the prediction’s rendered views directly. drops assemblies that exceed the… view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the Assembly-3D Part bucket metrics. Part scoring runs in two stages. Assembly Decomposition: a fixed decomposition MLLM splits the predicted assembly into parts, and a fidelity gate keeps only those whose reassembled union still matches the original prediction. Part Metric: the retained parts are deduplicated, pose-aligned to the GT parts and matched one-to-one to compute the scores, where M i… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative OpenSCAD reconstructions across the three P3D-BENCH tasks for six representative models. Two cases per task group are shown (Text-to-3D split into Desc. and Param.; Image-to-3D; Assembly-3D). Each model cell prints the per case Geo and Judge scores (plus Part on the assembly tasks) above its rendered output. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-task cross-model bucket score distributions and GPT-5.5 cross-format bucket scores. (a) Per-task cross-model distribution of the bucket scores, averaged over output formats. Each panel is a violin plot with an inner box plot. (b) GPT-5.5 bucket scores across all tasks and output formats. Text-to-3D includes the descriptive (Desc.) and parametric (Param.) specifications. task gets harder: on the single-… view at source ↗
Figure 9
Figure 9. Figure 9: Judge bucket details. (a) GPT-5.5 Judge submetric scores across tasks and four output formats, normalized to [0, 1]. (b) Qualitative cases. (top) One descriptive and one parametric specification example, each with its QA results and reasons. (bottom) Two Image-to-3D cases with their visual Judge submetrics J-Geo/J-Aes/J-Sem. near 0.79–0.84, while J-Geo stays near 0.34–0.37. The strongest model can recogniz… view at source ↗
Figure 10
Figure 10. Figure 10: Part bucket details. (a) Part bucket submetrics for all models, averaged across CadQuery and OpenSCAD. (b) Part-match results for two assemblies. The ground-truth parts (top) are matched against the predicted parts (bottom), and each matched part is annotated with its per-part F-score. Colors indicate quality: green above 0.9, yellow for 0.7–0.9, and red below 0.7. 4.3 INVALID OUTPUT ANALYSIS The Valid me… view at source ↗
Figure 11
Figure 11. Figure 11: Execute-stage invalid breakdown across tasks, output formats and models. Each invalid output falls into one of four failure classes: Syntax (parser or import failures), Undefined Reference (missing names or attributes), Parameter (malformed arguments or schema fields), and Geometry (semantically legal programs that the geometry kernel cannot construct or export). For Text-to-3D we sum invalid cases over J… view at source ↗
Figure 12
Figure 12. Figure 12: Example annotations from the two data sources. The top row shows a retained Text2CAD [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Per-task quality score and cost to run the full task for each evaluated model. Each point is [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative OpenSCAD outputs from six representative models on the three [PITH_FULL_IMAGE:figures/full_fig_p031_14.png] view at source ↗
Figure 14
Figure 14. Figure 14: (Continued.) Qualitative OpenSCAD outputs for Image-to-3D, pairing each input render [PITH_FULL_IMAGE:figures/full_fig_p032_14.png] view at source ↗
Figure 14
Figure 14. Figure 14: (Continued.) Qualitative OpenSCAD outputs for Assembly-3D, pairing each assembly [PITH_FULL_IMAGE:figures/full_fig_p033_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative Text-to-3D outputs across the four (specification, format) combinations on five [PITH_FULL_IMAGE:figures/full_fig_p034_15.png] view at source ↗
Figure 15
Figure 15. Figure 15: (Continued.) Qualitative Text-to-3D outputs for descriptive specifications in OpenSCAD. [PITH_FULL_IMAGE:figures/full_fig_p035_15.png] view at source ↗
Figure 15
Figure 15. Figure 15: (Continued.) Qualitative Text-to-3D outputs for parametric specifications in JSON. The [PITH_FULL_IMAGE:figures/full_fig_p036_15.png] view at source ↗
Figure 15
Figure 15. Figure 15: (Continued.) Qualitative Text-to-3D outputs for parametric specifications in OpenSCAD. [PITH_FULL_IMAGE:figures/full_fig_p037_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative Image-to-3D outputs in CadQuery. The continued panels use the same [PITH_FULL_IMAGE:figures/full_fig_p038_16.png] view at source ↗
Figure 16
Figure 16. Figure 16: (Continued.) Qualitative Image-to-3D outputs in OpenSCAD on the same input cases and [PITH_FULL_IMAGE:figures/full_fig_p039_16.png] view at source ↗
Figure 16
Figure 16. Figure 16: (Continued.) Qualitative Image-to-3D outputs in Three.js on the same input cases and [PITH_FULL_IMAGE:figures/full_fig_p040_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Qualitative Assembly-3D outputs in CadQuery. The continued panel uses the same input [PITH_FULL_IMAGE:figures/full_fig_p041_17.png] view at source ↗
Figure 17
Figure 17. Figure 17: (Continued.) Qualitative Assembly-3D outputs in OpenSCAD on the same input cases [PITH_FULL_IMAGE:figures/full_fig_p042_17.png] view at source ↗
read the original abstract

Multimodal large language models can write code to produce complex programs as well as use programs to do 3D modeling, which opens up a new avenue for 3D generation powered by their priors, world knowledge and reasoning. Yet existing benchmarks rarely evaluate 3D modeling through code. Such modeling demands more than runnable code: from a text or visual specification, a model must generate a parametric 3D program that is geometrically precise, semantically aligned and assembly-consistent. We introduce P3D-Bench, a benchmark for parametric 3D generation. Unlike a 3D mesh, a parametric 3D program exposes explicit dimensions, construction operations and part relations, revealing whether a model recovers a design's structure, not just its appearance. Under a unified protocol, P3D-Bench covers three task families (Text-to-3D, Image-to-3D and Assembly-3D) and scores each output for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment and part-level structure. We evaluate frontier MLLMs and text-only LLMs on 400 text cases, 400 image cases and 203 annotated assemblies, with domain-specific models as reference points. Our extensive evaluation yields three findings. First, assemblies are the hardest setting, where models still fail to compose multiple parts into a coherent structure. Second, models can often recover the global shape and semantic identity of the target object, yet fail to reproduce the precise parametric geometry specified by the input. Third, part-level modeling remains weak on assemblies, where models recover neither the geometry of each part nor the right number of parts. These results position P3D-Bench as a benchmark for evaluating precise parametric geometry and part-level structure in parametric 3D generation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper introduces P3D-Bench, a benchmark for evaluating MLLMs on parametric 3D program generation from text, image, or assembly inputs across three task families. It scores model outputs using metrics for executability, geometric fidelity, topology, text-grounded constraints, multiview semantic alignment, and part-level structure on 400 text cases, 400 image cases, and 203 assemblies. The evaluation of frontier MLLMs and LLMs (with domain-specific references) yields three findings: assemblies are the hardest setting, models recover global shape and semantic identity but fail precise parametric geometry, and part-level modeling remains weak on assemblies.

Significance. If the metrics and case validations hold, P3D-Bench would provide a useful addition to 3D generation evaluation by focusing on parametric structure and part relations rather than appearance alone. The unified protocol across modalities and the scale of the evaluation (over 1000 cases) are positive aspects that could help identify specific gaps in current models' structural reasoning.

major comments (2)
  1. [Abstract] Abstract: the second finding claims models recover global shape/semantic identity yet fail to reproduce precise parametric geometry. However, the abstract lists the six metrics without definitions, formulas, or validation showing they isolate exact parameter mismatches independently of rendered mesh or global shape similarity (e.g., whether geometric fidelity compares program parameters directly). This is load-bearing for the finding.
  2. [Evaluation] Evaluation section: no details are supplied on metric definitions, how the 400+400+203 cases were validated or annotated (especially the 203 assemblies), or how scores were aggregated across metrics. Without this, support for all three findings cannot be verified.
minor comments (1)
  1. [Figures and Tables] Figure captions and tables should explicitly note the number of cases per task family and any aggregation method used for the reported scores.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments correctly identify that the current manuscript provides insufficient detail on metric definitions and case validation to fully substantiate the reported findings. We will revise the manuscript to address both points.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the second finding claims models recover global shape/semantic identity yet fail to reproduce precise parametric geometry. However, the abstract lists the six metrics without definitions, formulas, or validation showing they isolate exact parameter mismatches independently of rendered mesh or global shape similarity (e.g., whether geometric fidelity compares program parameters directly). This is load-bearing for the finding.

    Authors: We agree that the abstract is too concise and does not supply the requested definitions or validation. The geometric fidelity metric is intended to operate directly on program parameters (e.g., dimension values and construction operations) rather than on rendered meshes or global shape descriptors; the other five metrics address orthogonal aspects. To make this explicit and support the second finding, we will revise the abstract to include brief characterizations of the metrics and a pointer to their formal definitions and independence checks in the main text. revision: yes

  2. Referee: [Evaluation] Evaluation section: no details are supplied on metric definitions, how the 400+400+203 cases were validated or annotated (especially the 203 assemblies), or how scores were aggregated across metrics. Without this, support for all three findings cannot be verified.

    Authors: We concur that the Evaluation section currently lacks the necessary transparency. In the revised version we will expand it to provide: (i) complete definitions and formulas for all six metrics, including explicit demonstration that geometric fidelity measures parameter-level differences independently of mesh or semantic similarity; (ii) a description of the annotation and validation protocol used for the 400 text, 400 image, and 203 assembly cases (including how assemblies were constructed and cross-checked); and (iii) the precise aggregation procedure across metrics. These additions will enable verification of the three findings. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical benchmark with no derivations or fitted predictions

full rationale

The paper introduces P3D-Bench as an evaluation protocol for MLLMs on parametric 3D tasks, reporting empirical findings from model evaluations on fixed test cases. No mathematical derivations, parameter fitting, predictions from first principles, or self-citation chains appear in the abstract or described structure. The three findings are direct observations from scoring outputs on executability, fidelity, topology, and structure metrics; they do not reduce to inputs by construction. This matches the default case of a self-contained empirical benchmark with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that the listed scoring dimensions capture structural correctness of parametric programs; no free parameters or invented physical entities are described.

axioms (1)
  • domain assumption Parametric 3D programs can be meaningfully scored for geometric fidelity, topology, and part-level structure from text or image specifications.
    Invoked when the benchmark protocol is defined and when the three findings are stated.
invented entities (1)
  • P3D-Bench no independent evidence
    purpose: Unified evaluation protocol for parametric 3D generation across three task families
    New artifact introduced by the paper; no independent evidence outside the paper is provided in the abstract.

pith-pipeline@v0.9.1-grok · 5887 in / 1195 out tokens · 20141 ms · 2026-06-27T13:46:22.312590+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 9 linked inside Pith

  1. [1]

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

    Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    ACM Transactions on Graphics (TOG) , volume=

    Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences , author=. ACM Transactions on Graphics (TOG) , volume=. 2021 , publisher=

    2021

  4. [4]

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

    Deepcad: A deep generative network for computer-aided design models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  5. [5]

    2026 , howpublished =

    2026

  6. [6]

    2025 , eprint=

    CAD-Coder: An Open-Source Vision-Language Model for Computer-Aided Design Code Generation , author=. 2025 , eprint=

    2025

  7. [7]

    2026 , url=

    Prashant Govindarajan and Davide Baldelli and Jay Pathak and Quentin Fournier and Sarath Chandar , journal=. 2026 , url=

    2026

  8. [8]

    2026 , eprint=

    cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning , author=. 2026 , eprint=

    2026

  9. [9]

    arXiv preprint arXiv:2604.10992 , year=

    ArtiCAD: Articulated CAD Assembly Design via Multi-Agent Code Generation , author=. arXiv preprint arXiv:2604.10992 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2605.15187 , year=

    Articraft: An Agentic System for Scalable Articulated 3D Asset Generation , author=. arXiv preprint arXiv:2605.15187 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2508.08228 , year=

    Ll3m: Large language 3d modelers , author=. arXiv preprint arXiv:2508.08228 , year=

    arXiv

  12. [12]

    arXiv preprint arXiv:2601.11109 , year=

    Vision-as-inverse-graphics agent via interleaved multimodal reasoning , author=. arXiv preprint arXiv:2601.11109 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2207.04632 , year=

    Skexgen: Autoregressive generation of cad construction sequences with disentangled codebooks , author=. arXiv preprint arXiv:2207.04632 , year=

    arXiv

  14. [14]

    and Furukawa, Yasutaka , booktitle =

    Xu, Xiang and Jayaraman, Pradeep Kumar and Lambourne, Joseph George and Willis, Karl D.D. and Furukawa, Yasutaka , booktitle =. Hierarchical Neural Coding for Controllable. 2023 , editor =

    2023

  15. [15]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    Pith/arXiv arXiv

  16. [16]

    International Conference on Learning Representations , volume=

    Swe-bench: Can language models resolve real-world github issues? , author=. International Conference on Learning Representations , volume=

  17. [17]

    The Eleventh International Conference on Learning Representations , year=

    DreamFusion: Text-to-3D using 2D Diffusion , author=. The Eleventh International Conference on Learning Representations , year=

  18. [18]

    Communications of the ACM , volume=

    Nerf: Representing scenes as neural radiance fields for view synthesis , author=. Communications of the ACM , volume=. 2021 , publisher=

    2021

  19. [19]

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

    Deepsdf: Learning continuous signed distance functions for shape representation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    Direct3d-s2: Gigascale 3d generation made easy with spatial sparse attention , author=. Advances in Neural Information Processing Systems , volume=

  22. [22]

    International Conference on Learning Representations , volume=

    Lrm: Large reconstruction model for single image to 3d , author=. International Conference on Learning Representations , volume=

  23. [23]

    arXiv preprint arXiv:2310.02977 , year=

    T ^3 Bench: Benchmarking Current Progress in Text-to-3D Generation , author=. arXiv preprint arXiv:2310.02977 , year=

    arXiv

  24. [24]

    arXiv preprint arXiv:2503.21745 , year=

    3dgen-bench: comprehensive benchmark suite for 3d generative models , author=. arXiv preprint arXiv:2503.21745 , year=

    arXiv

  25. [25]

    arXiv preprint arXiv:2505.06507 , year=

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

    arXiv

  26. [26]

    arXiv preprint arXiv:2603.26512 , year=

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

    arXiv

  27. [27]

    arXiv preprint arXiv:2602.03045 , year=

    Clarify Before You Draw: Proactive Agents for Robust Text-to-CAD Generation , author=. arXiv preprint arXiv:2602.03045 , year=

    Pith/arXiv arXiv

  28. [28]

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

    Infinibench: Infinite benchmarking for visual spatial reasoning with customizable scene complexity , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  29. [29]

    The Fourteenth International Conference on Learning Representations , year=

    Theory of Space: Can Foundation Models Construct Spatial Beliefs through Active Exploration? , author=. The Fourteenth International Conference on Learning Representations , year=

  30. [30]

    arXiv preprint arXiv:2604.02580 , year=

    VoxelCodeBench: Benchmarking 3D World Modeling Through Code Generation , author=. arXiv preprint arXiv:2604.02580 , year=

    Pith/arXiv arXiv

  31. [31]

    2026 , eprint=

    3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code , author=. 2026 , eprint=

    2026

  32. [32]

    arXiv preprint arXiv:2605.18430 , year=

    Text2CAD-Bench: A Benchmark for LLM-based Text-to-Parametric CAD Generation , author=. arXiv preprint arXiv:2605.18430 , year=

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2605.10865 , year=

    BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD , author=. arXiv preprint arXiv:2605.10865 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2605.28579 , year=

    MUSE: Benchmarking Manufacturable, Functional, and Assemblable Text-to-CAD Generation , author=. arXiv preprint arXiv:2605.28579 , year=

    Pith/arXiv arXiv

  35. [35]

    2026 , eprint=

    UniCAD: A Unified Benchmark and Universal Model for Multi-Modal Multi-Task CAD , author=. 2026 , eprint=

    2026

  36. [36]

    2021 , howpublished =

    2021

  37. [37]

    Transactions on Machine Learning Research , issn=

    Maxime Oquab and Timoth. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024