Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

BrepLLM enables large language models to parse and reason over raw boundary representation (B-rep) CAD data, and reports state-of-the-art captioning and classification results with a compact 2.9B-parameter model.

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 · deepseek-v4-flash

2026-08-03 15:33 UTC pith:ZTKCHXX4

load-bearing objection New modality, sound architecture, but the SOTA numbers rest on a 200-sample split with no error bars and selection on the test set. the 3 major comments →

arxiv 2512.16413 v2 pith:ZTKCHXX4 submitted 2025-12-18 cs.CV

BrepLLM: Enabling Large Language Models to Understand Boundary Representations

Liyuan Deng , Hao Guo , Yongkang Dai , Yunpeng Bai , Yifan Zhu , Yuanyuan Gao , Huaxi Huang , Yilei Shi This is my paper
classification cs.CV
keywords Boundary RepresentationB-rep3D CAD understandingmultimodal LLMBrep2Textcontrastive learningMixture-of-Query Experts3D object captioning
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.

The paper claims that the exact parametric geometry and watertight topology of CAD boundary representations carry enough signal that a language model can be trained to read them directly, without converting to point clouds or command sequences. BrepLLM turns each B-rep into a face-edge graph via adaptive UV sampling, encodes it with a hierarchical BrepEncoder, aligns the result with text through a CLIP-style contrastive loss, and then fine-tunes a small LLM in three progressive stages. On the new Brep2Text benchmark of 269,444 auto-generated question-answer pairs, it reports higher captioning quality and classification accuracy than larger point-cloud-based 3D-LLMs. If true, this would give CAD and manufacturing an interface where models reason about the native geometric representation rather than a lossy proxy.

Core claim

The central discovery is that a boundary representation graph—faces as nodes with sampled UV geometry, edges as adjacency with length-adaptive point features—can be tokenized and fed into an LLM to produce fluent, dimensionally accurate CAD captions and category labels. The authors show that after contrastive pretraining of a hierarchical face/edge/topology encoder against CLIP text embeddings, and after a three-stage fine-tuning that first bridges geometry to the 2D vision-language space, then aligns through LoRA, and finally adds a residual mixture-of-query-experts, the resulting model outperforms point-cloud-based LLMs on both tasks. The result is attributed to the rich, exact geometric a

What carries the argument

The load-bearing mechanism is the hierarchical BrepEncoder: an adaptive UV sampling strategy discretizes each face into 10-dimensional attribute tensors (coordinates, normals, curvature, visibility, face type, area) and each edge into 8-dimensional tensors, building a face-edge adjacency graph. A PointTransformerV3 extracts per-face features, an NNConv edge-conditioned layer propagates boundary information, and EGAT-Conv layers aggregate global topology; concatenating these yields a 128-dimensional node token per face plus a global graph token. The global token is aligned to frozen CLIP text embeddings via InfoNCE, and the node tokens are progressively aligned to a Q-Former and Phi-2 LLM thr

Load-bearing premise

The whole evaluation rests on the assumption that 269,444 auto-generated question-answer pairs are faithful enough to human knowledge of CAD, and that a held-out set of 200 models is large enough to resolve the few-point gaps the paper reports.

What would settle it

Re-run the reported evaluation on a fresh random 200-model split (or a larger held-out set) and measure the distribution of the score gaps; if the gap to the best point-cloud baseline is consistently below one point or changes sign across splits, the claimed advantage of native B-rep reasoning does not hold.

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

If this is right

  • If correct, CAD question-answering and part captioning no longer require rendering point clouds or predicting command scripts; the raw B-rep can be the direct input modality.
  • A compact 2.9B model matching or beating 7B-13B point-cloud models suggests that representation fidelity matters more than scale for geometry-to-language tasks, potentially lowering hardware requirements.
  • The three-stage curriculum (CLIP alignment, vision-bridge MLP, then residual query experts) gives a reusable recipe for plugging other structured geometry modalities into frozen LLMs.
  • Brep2Text, with 269,444 QA pairs, becomes a benchmark for future B-rep language understanding and for comparing native-geometry approaches to point-cloud proxies.

Where Pith is reading between the lines

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

  • The same adaptive-UV graph encoding could be applied to other exact geometry formats, such as NURBS surfaces or trimmed patches, to let LLMs reason about free-form geometry rather than only CAD primitives.
  • If the B-rep signal is as rich as claimed, a testable extension is zero-shot or few-shot transfer to downstream tasks like constraint satisfaction, assemblability, or manufacturability analysis, where point-cloud representations are known to be lossy.
  • Because the dataset is synthesized by an LLM without reported human validation, a likely failure mode is that the model learns to imitate the synthetic captioning style rather than true geometric reasoning; this could be tested by probing with novel geometry attributes not present in the training distribution.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes BrepLLM, a framework that feeds raw boundary-representation (B-rep) CAD models directly into an LLM, bypassing command-sequence or point-cloud intermediaries. It introduces adaptive UV sampling to turn B-rep faces and edges into a graph, a hierarchical BrepEncoder that produces face, edge, and topology features, and a contrastive pre-training step that aligns a global Brep token with a frozen CLIP text encoder. The LLM (Phi-2) is then fine-tuned in three stages through an MLP projector, a Q-Former with LoRA, and a Mixture-of-Query Experts module. The paper also introduces Brep2Text, a dataset of 269,444 automatically generated B-rep–language QA pairs, and reports state-of-the-art (SOTA) results on captioning and generative classification tasks against point-cloud baselines on a 200-model test split.

Significance. If the SOTA claim holds, this is a significant contribution: it demonstrates that raw structured B-rep geometry alone can support LLM captioning and classification, and the proposed hierarchical encoder and dataset could become useful resources for CAD-focused multimodal research. The paper includes several strengths: a clear and plausible architecture, reproduction of point-cloud baselines on the same split, a large new dataset, and both automatic and human evaluation. However, the empirical evidence for the central claim is currently not strong enough. The headline margins over the best baseline are small (1.33–2.60 points) and are reported as point estimates on an n=200 split with no confidence intervals, significance tests, or repeated-seed variance. Moreover, the ablations that select the final configuration are computed on the same test split, so the final numbers are selected rather than independently validated. These issues are fixable, but they currently undermine the SOTA claim.

major comments (3)
  1. [§5.1, Tables 1–2] The headline SOTA margins are within sampling noise. With n=200, the classification accuracy difference 57.05% vs 54.90% has an approximate standard error of 5.0 percentage points, giving z≈0.4. No confidence intervals, p-values, or repeated-seed results are reported for any metric. The captioning scores in Table 1 also lack any variance measure. Since the reported deltas are the entire basis for the SOTA claim, the paper should report bootstrap CIs, significance tests, or results across multiple test splits/seeds.
  2. [§5.5, Tables 4–7] The ablations are conducted on the same 200-model test split used for the main results, and they are used to select the final configuration (stage composition, MQE placement, adaptive UV sampling). For example, Table 4 shows a final gain of +0.64 for adaptive sampling, Table 5 shows +2.42 for the hierarchical encoder, and Table 6 shows 57.05% for the full pipeline. Because the final configuration was chosen after inspecting these same test-set numbers, the reported SOTA is a selected maximum over several configurations and may overstate true performance. A separate validation split or a nested/selection-aware evaluation protocol is needed before the headline claim is credible.
  3. [§4 and §5.2] The training QA pairs in Brep2Text are generated by Qwen-Max (§4), and Qwen-Max is also used as the captioning judge (§5.2). Although the reference answers derive from human-written Text2CAD captions, using the same model family for training-data generation and evaluation risks rewarding stylistic mimicry of Qwen-Max-generated text rather than genuine geometric understanding. No human validation rate of the automatically generated questions/labels is reported. Please provide a validation sample with human agreement, use an additional independent LLM judge or human evaluation with inter-annotator statistics, and report the correlation between the Qwen-Max score and human precision.
minor comments (5)
  1. [Abstract and §3.3] The abstract and introduction describe a two-stage pipeline, while Section 3.3 and the conclusion describe a three-stage fine-tuning strategy. Please harmonize this terminology.
  2. [Table 3 caption] Typo: 'sxample' should be 'sample'. Please proofread the qualitative-results section.
  3. [Eq. (2)] Notation is inconsistent: MC is used in the text, while ℓC appears without a clear definition of C. Please define all symbols consistently.
  4. [References] Reference [1] has a malformed author list, and reference [23] is missing venue information. Please correct these entries.
  5. [Appendix] The text repeatedly refers to 'See Appendix' for training details, but no appendix is present in the submitted version. Please include it or clarify availability.

Circularity Check

0 steps flagged

No significant circularity: the reported SOTA numbers are external benchmark comparisons, not definitional reductions.

full rationale

BrepLLM's central claims are empirical comparisons against external point-cloud 3D-LLMs on a held-out 200-model split, with the model trained via standard CLIP contrastive loss (Eq. 7) and autoregressive language-modeling objectives. No equation-level identity connects a fitted parameter to a reported metric, and there is no self-citation chain or imported uniqueness theorem; components such as CLIP, PointTransformerV3, BLIP-2 Q-Former, and Phi-2 are external. The one self-referential element is that Qwen-Max generated the Brep2Text QA pairs and also served as a captioning judge; this is an evaluation-bias concern (possible style mimicry) rather than a circular derivation, and the captioning claim is also supported by frozen external metrics (Sentence-BERT, SimCSE) and human evaluation, while classification is judged against Text2CAD-derived ground-truth categories. Statistical concerns about the 200-model test split and ablations performed on the same split affect confidence in the SOTA delta but are correctness/risk issues, not circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The headline numbers rest on three things the paper does not independently establish: (1) the discretized face-adjacency graph preserves the exact geometry/topology that supposedly distinguishes B-reps from point clouds; (2) the Qwen-Max-generated QA labels are accurate ground truth; (3) a 200-sample test split, reused for configuration selection, has enough power for 1–3-point deltas. The free parameters are mostly standard recipe choices; the load-bearing ones are the sampling range [16,32], the 128-token truncation, and the unspecified MQE/LoRA settings.

free parameters (6)
  • Contrastive temperature τ (InfoNCE scale) = learned, value not reported
    Eqs. 5–7: temperature scales the shape-text similarity logits; a fitted hyperparameter the alignment quality depends on, value never given.
  • UV sampling counts N_face, M_edge = range [16,32] (Eqs. 1–2)
    Per-face area-adaptive and per-edge length-adaptive sample counts; caps the discretization resolution and directly bounds how much exact B-rep geometry the encoder sees.
  • Q-Former query count Q = 32
    Learnable cross-attention queries; a design choice, never ablated.
  • Node token length cap T_max = 128 (truncate/pad)
    Face tokens beyond 128 are dropped; for models with more faces, the 'direct' understanding claim silently loses information.
  • MQE expert count k and Top-G routing = not specified (appendix)
    Stage III adds k residual query experts with sparse routing; k, G, and expert dimensions are absent from the posted version.
  • LoRA rank/alpha (Q-Former, LLM) = not specified (appendix)
    Stage II LoRA settings are deferred to an appendix not present in v1.
axioms (6)
  • domain assumption UV-sampled face-adjacency graph faithfully represents B-rep geometry and topology for language grounding
    §3.2, Eqs. 1–2: the encoder never sees the exact parametric surface equations or trim curves; the 'direct B-rep understanding' claim rests on the adequacy of 16–32-point discretization per face/edge.
  • domain assumption Qwen-Max-generated Brep2Text QA pairs are valid ground truth for B-rep–language semantics
    §4: all 269,444 pairs are machine-generated from Text2CAD; no human-validation rate is reported, and both training and evaluation rest on this.
  • domain assumption Frozen CLIP (ViT-L/14) text space is a valid semantic anchor for CAD-related language
    §3.2: contrastive alignment assumes CLIP text embeddings are a good proxy for CAD semantics; no quantitative validation of the alignment is provided.
  • domain assumption The 200-model held-out split is representative and sufficient for 1–3-point metric deltas
    §5.1/§5.5: n=200 test set with no confidence intervals; the same split is reused for ablations, so its representativeness underpins the SOTA claim.
  • domain assumption Off-the-shelf components (PointTransformerV3, NNConv, EGATConv, Phi-2, Q-Former, 2D-VLM priors) behave as described in their source papers
    §3.2–3.3: these are invoked as pretrained tools; their correct behavior and suitability for B-rep inputs are assumed.
  • standard math Symmetric InfoNCE/CLIP contrastive loss yields a meaningful shared embedding under the dual-tower setup
    Eqs. 4–7: standard CLIP objective; validity assumed as in Radford et al. (2021).
invented entities (2)
  • Mixture-of-Query Experts (MQE) with residual query sets E_1..E_k no independent evidence
    purpose: Adds sparse-route residual experts over Q-Former queries so the final output is Q_base + Q_residual, to capture geometric diversity.
    Supported only by the paper's own ablation (Table 7); expert count, routing top-G, and dimensions are undeclared; no external falsifiable handle yet.
  • Brep2Text dataset (269,444 B-rep–text QA pairs) no independent evidence
    purpose: First large-scale instruction-tuning benchmark for B-rep-language understanding, built by Qwen-Max from Text2CAD.
    The dataset is not released or linked in the posted version and its labels have no reported human-validation rate, so it does not yet function as an independent benchmark others can test against.

pith-pipeline@v1.3.0-alltime-deepseek · 11390 in / 23685 out tokens · 212716 ms · 2026-08-03T15:33:04.495711+00:00 · methodology

0 comments
read the original abstract

Current token-sequence-based Large Language Models (LLMs) struggle to directly process 3D Boundary Representation (B-rep) models that contain complex geometric and topological information. To this end, we propose BrepLLM, the first multimodal framework that enables LLMs to directly parse and reason over raw B-rep data. BrepLLM adopts a two-stage training pipeline: cross-modal alignment pre-training and two-stage LLM fine-tuning. In the first stage, we design an adaptive UV sampling strategy to convert B-reps into graph representations that integrate geometric and topological information. Subsequently, we construct a hierarchical BrepEncoder to extract features from geometric elements (faces and edges) and topology, generating a global token and a sequence of node tokens. Then, via contrastive learning, we conduct an initial alignment between this global token and the text embeddings of a frozen CLIP text encoder (ViT-L/14). In the second stage, we integrate the pre-trained BrepEncoder into the LLM and employ a two-stage progressive strategy to align the sequence of node tokens: (1) training an MLP-based semantic mapping network that utilizes the prior knowledge of a 2D-VLM to align the B-rep representation to the 2D visual semantic space; (2) utilizing LoRA for parameter-efficient fine-tuning of the Q-Former and the LLM backbone network to achieve the final 3D-language generation capability. Furthermore, we construct the Brep2Text dataset, which contains 269,444 B-rep and text question-answer pairs. Experiments demonstrate that BrepLLM achieves SOTA performance on 3D object classification and captioning tasks. The project page is available at https://user-deng.github.io/BrepLLM/.

Figures

Figures reproduced from arXiv: 2512.16413 by Hao Guo, Huaxi Huang, Liyuan Deng, Yifan Zhu, Yilei Shi, Yongkang Dai, Yuanyuan Gao, Yunpeng Bai.

Figure 1
Figure 1. Figure 1: BrepLLM, trained to directly understand Brep data, en [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the BrepLLM architecture. The framework consists of two steps. Step 1 (Left): Cross-modal Alignment Pre [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The overview of BrepEncoder. (a) B-rep parameterization using area-adaptive UV sampling for faces and length-adaptive [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. BrepCoder: A Unified Multimodal Large Language Model for Multi-task B-rep Reasoning

    cs.LG 2026-02 conditional novelty 6.0

    BrepCoder trains one multimodal LLM to read B-rep CAD geometry and output CAD code, handling reverse engineering, completion, error correction, and CAD-QA.

Reference graph

Works this paper leans on

24 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Phi-2: The surprising power of small language models, 2023

    Marah Abdin, Aneja Jyoti, Bubeck Sebastien, C ´esar Teodoro Mendes Caio, Weizhu Chen, Allie Del Giorno, Ro- nen Eldan, Sivakanth Gopi, Suriya Gunasekar, Mojan Java- heripi, Piero Kauffmann, Tat Lee Yin, Yuanzhi Li, Anh Nguyen, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Michael Santacroce, Harkirat Singh Behl, Adam Tau- mann Kalai, Xin Wang,...

    2023

  2. [2]

    Query2cad: Generating cad models using natural language queries.arXiv preprint arXiv:2406.00144, 2024

    Akshay Badagabettu, Sai Sravan Yarlagadda, and Amir Barati Farimani. Query2cad: Generating cad models using natural language queries.arXiv preprint arXiv:2406.00144, 2024. 2

    Pith/arXiv arXiv 2024

  3. [3]

    Text2shape: Generating shapes from natural language by learning joint embeddings

    Kevin Chen, Christopher B Choy, Manolis Savva, An- gel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. InAsian conference on computer vision, pages 100–116. Springer, 2018. 3

    2018

  4. [4]

    An investigation on utilizing large language model for in- dustrial computer-aided design automation.Procedia CIRP, 128:221–226, 2024

    Haoxuan Deng, Samir Khan, and John Ahmet Erkoyuncu. An investigation on utilizing large language model for in- dustrial computer-aided design automation.Procedia CIRP, 128:221–226, 2024. 2, 3

    2024

  5. [5]

    A solver-aided hi- erarchical language for llm-driven cad design.arXiv preprint arXiv:2502.09819, 2025

    Benjamin T Jones, Felix H ¨ahnlein, Zihan Zhang, Maaz Ah- mad, Vladimir Kim, and Adriana Schulz. A solver-aided hi- erarchical language for llm-driven cad design.arXiv preprint arXiv:2502.09819, 2025. 3

    Pith/arXiv arXiv 2025

  6. [6]

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

    Mohammad Sadil Khan, Sankalp Sinha, Talha Uddin, Di- dier Stricker, Sk Aziz Ali, and Muhammad Zeshan Afzal. Text2cad: Generating sequential cad designs from beginner- to-expert level text prompts.Advances in Neural Information Processing Systems, 37:7552–7579, 2024. 2

    2024

  7. [7]

    Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models.arXiv preprint arXiv:2407.07895, 2024

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models.arXiv preprint arXiv:2407.07895, 2024. 3

    Pith/arXiv arXiv 2024

  8. [8]

    Llm4cad: Multi-modal large language models for 3d computer-aided design generation

    Xingang Li, Yuewan Sun, and Zhenghui Sha. Llm4cad: Multi-modal large language models for 3d computer-aided design generation. InInternational Design Engineering Technical Conferences and Computers and Information in Engineering Conference, page V006T06A015. American Society of Mechanical Engineers, 2024. 2

    2024

  9. [9]

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

    Dimitrios Mallis, Ahmet Serdar Karadeniz, Sebastian Cavada, Danila Rukhovich, Niki Foteinopoulou, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Cad- assistant: Tool-augmented vllms as generic cad task solvers. arXiv preprint arXiv:2412.13810, 2024. 2

    Pith/arXiv arXiv 2024

  10. [10]

    Shapellm: Universal 3d object understanding for embodied interaction,

    Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction,

  11. [11]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2, 5

    2021

  12. [12]

    Clip-forge: Towards zero-shot text-to-shape genera- tion

    Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malek- shan. Clip-forge: Towards zero-shot text-to-shape genera- tion. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 18603–18613,

  13. [13]

    Meshclip: Efficient cross-modal infor- mation processing for 3d mesh data in zero/few-shot learn- ing.Information Processing & Management, 60(6):103497,

    Yupeng Song, Naifu Liang, Qing Guo, Jicheng Dai, Junwei Bai, and Fazhi He. Meshclip: Efficient cross-modal infor- mation processing for 3d mesh data in zero/few-shot learn- ing.Information Processing & Management, 60(6):103497,

  14. [14]

    Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors, 2024

    Yuan Tang, Xu Han, Xianzhi Li, Qiao Yu, Yixue Hao, Long Hu, and Min Chen. Minigpt-3d: Efficiently aligning 3d point clouds with large language models using 2d priors, 2024. 6

    2024

  15. [15]

    Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced mul- timodal 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 mul- timodal llms. InProceedings of the AAAI Conference on Artificial Intelligence, pages 7880–7888, 2025. 2

    2025

  16. [16]

    Cad-llm: Large language model for cad generation

    Sifan Wu, Amir Khasahmadi, Mor Katz, Pradeep Kumar Ja- yaraman, Yewen Pu, Karl Willis, and Bang Liu. Cad-llm: Large language model for cad generation. InProceedings of the neural information processing systems conference. neurIPS, 2023. 2

    2023

  17. [17]

    Cad- vlm: Bridging language and vision in the generation of para- metric cad sketches

    Sifan Wu, Amir Hosein Khasahmadi, Mor Katz, Pradeep Ku- mar Jayaraman, Yewen Pu, Karl Willis, and Bang Liu. Cad- vlm: Bridging language and vision in the generation of para- metric cad sketches. InEuropean Conference on Computer Vision, pages 368–384. Springer, 2024. 2

    2024

  18. [18]

    Point transformer v3: Simpler, faster, stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Heng- shuang Zhao. Point transformer v3: Simpler, faster, stronger. 2024 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 4840–4851, 2023. 4

    2024

  19. [19]

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

    Jingwei Xu, Zibo Zhao, Chenyu Wang, Wen Liu, Yi Ma, and Shenghua Gao. Cad-mllm: Unifying multimodality- conditioned cad generation with mllm.arXiv preprint arXiv:2411.04954, 2024. 2

    Pith/arXiv arXiv 2024

  20. [20]

    Pointllm: Empowering large language models to understand point clouds, 2024

    Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiang- miao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds, 2024. 6

    2024

  21. [21]

    Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding

    Le Xue, Mingfei Gao, Chen Xing, Roberto Mart ´ın-Mart´ın, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1179–1189, 2023. 3

    2023

  22. [22]

    Cadtalk: An algorithm and benchmark for semantic commenting of cad programs

    Haocheng Yuan, Jing Xu, Hao Pan, Adrien Bousseau, Niloy J Mitra, and Changjian Li. Cadtalk: An algorithm and benchmark for semantic commenting of cad programs. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3753–3762, 2024. 2

    2024

  23. [23]

    Cad-editor: Text-based cad editing through adapting large language mod- els with synthetic data

    Yu Yuan, Shizhao Sun, Qi Liu, and Jiang Bian. Cad-editor: Text-based cad editing through adapting large language mod- els with synthetic data. 2

  24. [24]

    Pointclip: Point cloud understanding by clip

    Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu- peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552–8562, 2022. 3

    2022