Pith. sign in

REVIEW 4 major objections 5 minor 40 references

A fine-tuned language model can translate plain-English design descriptions into manufacture-ready STEP files, and this paper shows the steps that make it work.

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 09:42 UTC pith:EXTOWGBF

load-bearing objection A plausible first step at direct text-to-STEP generation, but the headline numbers are less clean than they look because of leakage risk and a self-referential RL reward. the 4 major comments →

arxiv 2601.12641 v1 pith:EXTOWGBF submitted 2026-01-19 cs.AI

STEP-LLM: Generating CAD STEP Models from Natural Language with Large Language Models

Xiangyu Shi , Junyang Ding , Xu Zhao , Sinong Zhan , Payal Mohapatra , Daniel Quispe , Kojo Welbeck , Jian Cao
show 3 more authors
Wei Chen Ping Guo Qi Zhu
This is my paper
classification cs.AI
keywords computer-aided designSTEP fileB-reptext-to-CADlarge language modelsretrieval-augmented generationreinforcement learningChamfer distance
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 tries to establish that large language models can generate STEP files directly from natural language, rather than relying on kernel-specific command sequences or script-based formats. It claims this is the first unified framework for direct STEP generation, and reports that on more than 2,000 held-out CAD descriptions it reaches a median scaled Chamfer distance of 0.53 versus 3.99 for the Text2CAD baseline, with renderability above 95%. The paper also claims each of its three design choices—DFS-based reserialization, retrieval-augmented fine-tuning, and reinforcement learning with a geometric reward—contributes to the gain. A sympathetic reader would care because STEP is the neutral, manufacturing-standard format, so direct generation could remove the conversion step between design intent and production.

Core claim

The central claim is that an LLM can be made to 'speak' STEP if the graph-structured file is first linearized in a way that preserves local reference locality, augmented with branch-level statistics, grounded with a retrieved similar example, and then refined with a reward that measures real geometric distance. The paper argues that the resulting model produces STEP files that are not only syntactically complete and renderable, but geometrically closer to the intended shape than a state-of-the-art script-based baseline, while also generating entity counts closer to human designs.

What carries the argument

The carrying mechanism is a depth-first-search reserialization of STEP's directed acyclic graph into a linear entity sequence with renumbered identifiers and normalized floating-point precision, plus chain-of-thought-style branch annotations that summarize child counts and depths. Around this sit retrieval-augmented generation during supervised fine-tuning and a piecewise-linear reward based on scale-normalized Chamfer distance used in group-relative policy optimization. The reserialization does the core work: it converts a format whose entities cross-reference each other non-locally into a token stream an autoregressive model can realistically produce.

Load-bearing premise

The retrieval database contains no near-duplicate of any test shape, so the retrieved STEP file serves only as a template and the measured fidelity is generated rather than copied.

What would settle it

For each test sample, compute the geometric similarity between the retrieved STEP file and the ground-truth file; if near-duplicate pairs are common, or if removing all near-duplicate retrievals from the index pushes the median Chamfer distance back up to the no-RAG level, the RAG gain is copying rather than generation.

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

If this is right

  • If correct, direct STEP generation lets non-experts produce manufacturing-ready files without an intermediate CAD kernel, simplifying the design-to-production pipeline.
  • The DFS linearization technique could make other graph-structured output formats, such as extended B-rep exchanges, tractable for autoregressive LLMs.
  • Reinforcement learning with a geometric reward pushes fidelity below what supervised fine-tuning alone achieves, suggesting that geometric signals can guide LLM training for CAD.
  • Entity-count fidelity indicates the model learns complexity calibration rather than over-generating, which matters for downstream CAM processing.
  • The framework's staged improvements (RAG, DFS, RL) give a recipe for adapting LLMs to other structured, cross-referenced engineering formats.

Where Pith is reading between the lines

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

  • An open question is whether the retrieval-augmented gain reflects generation or copying: the paper does not report deduplication between the retrieval database and test shapes, so a near-duplicate test shape could inflate the measured fidelity.
  • Because captions come from a vision-language model, the system likely inherits its biases toward common shapes; rare or expert-level geometry may be under-served.
  • The reward measures point-cloud distance, not manufacturability; adding constraints such as draft angle, wall thickness, or tool access could align the objective with real production requirements.
  • A stronger test would split the dataset by shape class rather than randomly, to see whether the model generalizes compositionally or only within familiar families.

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 / 5 minor

Summary. The paper proposes STEP-LLM, a fine-tuning framework for generating STEP files directly from natural-language captions. It contributes a curated caption–STEP dataset (filtered to <500-entity ABC files), a DFS-based reserialization with CoT-style branch annotations, RAG-augmented SFT, and GRPO RL with a scaled Chamfer distance reward. Experiments compare with Text2CAD on 2,056 held-out ABC samples using completion rate, renderability rate, median scaled Chamfer distance, and average entity count. The central claim is substantially higher geometric fidelity (MSCD 0.53 vs 3.99) while maintaining ~95% renderability, with further gains from RAG and RL. The paper also claims to be the first unified framework for direct STEP generation from natural language.

Significance. If the central comparison survives scrutiny, this would be a useful contribution: direct STEP generation bypasses kernel-dependent command/script formats and moves toward manufacturable B-rep output. The framework's modular design (reserialization, RAG, RL with geometric reward) is reasonable and the release of code and a dataset would help a young field. However, the quantitative advantage over Text2CAD currently rests on assumptions about train/test and retrieval overlap, and the RL improvement is measured with the same metric used as the reward. These issues must be resolved before the empirical claims can be considered established. The SFT-only result is less affected by reward circularity, but the leakage concern applies to all reported numbers.

major comments (4)
  1. [§3.1–3.2, Table 1] The headline MSCD advantage (0.53 vs 3.99) may be an artifact of dataset overlap. The SFT corpus, RAG database, and test set are all drawn from ABC without reported deduplication or geometric-similarity analysis. ABC contains many near-duplicate mechanical parts; if a near-copy of a test shape appears in training, the SFT model can memorize it, and the RAG module can retrieve it during inference because retrieval operates on captions from the same dataset. Text2CAD, trained on DeepCAD, would not benefit. The no-RAG row (MSCD 0.61) does not address training-set leakage. Please report train/test deduplication and the distribution of retrieval similarity; re-evaluate after removing near-duplicates or masking retrieved shapes whose similarity to the target exceeds a threshold.
  2. [§3.3, Eq. (3), §4.4, Table 4] The RL result is measured with exactly the quantity used as reward. Eq. (3) defines Rgeo as a scaled Chamfer distance and Table 4 reports median SCD, so the 0.53→0.098 improvement is not an independent validation of geometric fidelity; it mainly confirms that GRPO can optimize its own training objective. Please evaluate RL checkpoints with metrics not optimized during training (e.g., volumetric IoU, normal consistency, or a distance computed after a different alignment protocol), and report per-sample changes rather than only the aggregate median.
  3. [§4.3, Table 3] The RAG ablation is confounded by token count. The paper states that all models were trained for the same number of steps and that RAG prompts are longer, so the RAG condition consumed more tokens. The large differences (e.g., RR 0.13→0.95) could therefore be due to extra optimization compute rather than to retrieval grounding. Please provide a token-matched control (or a loss-matched control), and report average prompt lengths and optimizer steps for each condition.
  4. [§4.2, Tables 1–2] The central comparison lacks statistical quantification. Results are single runs on 2,056 samples with no confidence intervals, significance tests, or seeds. Because MSCD is a median, reporting the median alone hides the fraction of samples with degenerate or empty renderings. Please report bootstrap CIs for the median, the fraction of samples on which the distance is actually computed, and a paired comparison (e.g., Wilcoxon signed-rank) between STEP-LLM and Text2CAD on the same prompts. Also clarify whether Text2CAD is zero-shot; if so, a fine-tuned Text2CAD on the same captions would be a more informative baseline.
minor comments (5)
  1. [Abstract and §2.2] Typos and formatting: ‘manufacture models’ should be ‘manufacturable models’; ‘thatdirectly’ is missing a space; ‘F ACE’ appears in §2.2. Table 3 renders ‘0.970.95’ without spacing.
  2. [§1, Data Availability] The paper claims to release a curated dataset and metrics, but only a GitHub code link is given. Please provide a dataset link or state clearly where the data will be hosted.
  3. [§3.1] No example of a reserialized STEP file with CoT-style structural annotations is shown. A short concrete example would make the preprocessing concrete and reproducible.
  4. [§3.2] Please specify the split criterion for the held-out set (random file-level split? by shape?) and the number of retrieved examples per prompt in RAG. Also state whether the retrieval database can contain the target shape itself.
  5. [References] References [16] and [17] are both titled ‘CAD-Coder’ but are different works. Disambiguate these in the text, e.g., by author or venue.

Circularity Check

1 steps flagged

RL 'geometric fidelity' gain is self-referential because the RL reward is the evaluation metric; the core Text2CAD comparison is otherwise independent.

specific steps
  1. self definitional [Section 3.3 (Eq. 3), Section 4.1 metric definitions, Table 4]
    "To explicitly encode geometric fidelity into the training objective, we design a reward function based on the Scaled Chamfer Distance (SCD) ... MSCD quantifies geometric fidelity after alignment and scale normalization."

    The RL reward Rgeo is defined directly as a piecewise-linear function of SCD, and the RL result in Table 4 is reported as Median SCD (MSCD) (0.53 → 0.098). Thus the RL stage is trained to minimize exactly the quantity used to measure its success; the conclusion 'RL refinement leads to ... a significant reduction in median Chamfer Distance' restates the optimization objective rather than providing independent evidence of geometric fidelity.

full rationale

The SFT pipeline, DFS reserialization, RAG, and the baseline comparison are empirical and not constructionally circular: Table 1's Text2CAD comparison (MSCD 0.53 vs 3.99) is computed from generated STEP files against ground truth and does not reduce to the training inputs. The one genuine self-referential step is the RL reward/evaluation tie: Eq. (3) makes the reward a function of SCD, and Table 4 reports MSCD, so the RL improvement is by construction relative to that metric. The RAG near-duplicate concern and absence of train/test deduplication are data-leakage threats to external validity rather than definitional circularity, so they are not scored as circular steps here. Self-citations appear only in related work and are not load-bearing.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. Its contribution is an empirical system built on hand-chosen thresholds, a dataset filter, and assumptions about the validity of the baseline and the evaluation metric. The most consequential hidden assumption is that RAG retrieval does not leak near-duplicate training geometry into the prompt.

free parameters (5)
  • Entity-count filter threshold = 500 entities
    Training corpus restricted to STEP files with <500 entities; test set therefore max 477; no sensitivity analysis or evaluation on larger files. Section 3.2.
  • SCD reward thresholds (delta_low, delta_high) = 0.01, 0.5
    Piecewise-linear reward in Eq. 3; thresholds chosen by hand, no sweep; directly shapes RL results. Section 4.4.
  • GRPO hyperparameters = KL 0.02, entropy 0.005, lr 3e-6, batch 8, 80 steps
    Chosen by hand for RL; no sensitivity analysis. Section 4.4.
  • Number of retrieved examples = not reported
    RAG retrieval count is never specified; affects prompt length and leakage risk. Section 3.2.
  • SFT hyperparameters = lr 2e-4, batch 16, 10 epochs, LoRA
    Standard choices; not the main target but still hand-selected. Section 4.1.
axioms (6)
  • domain assumption STEP entity graphs can be losslessly linearized by DFS + pruning with renumbered ids and normalized floats
    Central preprocessing; pruning may drop references; no proof that pruned serialization preserves all topology. Section 3.1.
  • domain assumption CoT-style statistical annotations (child counts, branch depth) improve LLM global coherence
    Used as guidance tokens; no isolated ablation of annotations vs DFS. Section 3.1.
  • ad hoc to paper Retrieved STEP examples ground generation without revealing the target
    No deduplication or nearest-neighbor overlap analysis between RAG DB and test set; potential leakage. Section 3.2.
  • ad hoc to paper Text2CAD is a fair open-source baseline
    Used zero-shot without fine-tuning on authors' caption distribution; export pipeline gives it RR advantage. Section 4.2.
  • domain assumption Scaled Chamfer Distance is a valid geometric fidelity/manufacturability proxy
    Used for both RL reward and evaluation; no topology or manufacturing constraints. Sections 3.3, 4.1.
  • ad hoc to paper Training on <500-entity files generalizes to more complex cases
    Claim in Section 3.2; test set max is 477 entities, so >500 is untested.

pith-pipeline@v1.3.0-alltime-deepseek · 10253 in / 11185 out tokens · 114356 ms · 2026-08-03T09:42:44.712249+00:00 · methodology

0 comments
read the original abstract

Computer-aided design (CAD) is vital to modern manufacturing, yet model creation remains labor-intensive and expertise-heavy. To enable non-experts to translate intuitive design intent into manufacturable artifacts, recent large language models-based text-to-CAD efforts focus on command sequences or script-based formats like CadQuery. However, these formats are kernel-dependent and lack universality for manufacturing. In contrast, the Standard for the Exchange of Product Data (STEP, ISO 10303) file is a widely adopted, neutral boundary representation (B-rep) format directly compatible with manufacturing, but its graph-structured, cross-referenced nature poses unique challenges for auto-regressive LLMs. To address this, we curate a dataset of ~40K STEP-caption pairs and introduce novel preprocessing tailored for the graph-structured format of STEP, including a depth-first search-based reserialization that linearizes cross-references while preserving locality and chain-of-thought(CoT)-style structural annotations that guide global coherence. We integrate retrieval-augmented generation to ground predictions in relevant examples for supervised fine-tuning, and refine generation quality through reinforcement learning with a specific Chamfer Distance-based geometric reward. Experiments demonstrate consistent gains of our STEP-LLM in geometric fidelity over the Text2CAD baseline, with improvements arising from multiple stages of our framework: the RAG module substantially enhances completeness and renderability, the DFS-based reserialization strengthens overall accuracy, and the RL further reduces geometric discrepancy. Both metrics and visual comparisons confirm that STEP-LLM generates shapes with higher fidelity than Text2CAD. These results show the feasibility of LLM-driven STEP model generation from natural language, showing its potential to democratize CAD design for manufacturing.

Figures

Figures reproduced from arXiv: 2601.12641 by Daniel Quispe, Jian Cao, Junyang Ding, Kojo Welbeck, Payal Mohapatra, Ping Guo, Qi Zhu, Sinong Zhan, Wei Chen, Xiangyu Shi, Xu Zhao.

Figure 1
Figure 1. Figure 1: The typical workflow of CAD from design to manufacturing. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Framework of STEP-LLM. The top panel illustrates data preprocessing: raw STEP files are rendered [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Entity count distribution of generated STEP files vs. ground truth. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The visual comparison between Text2CAD, STEP-LLM-noRAG and STEP-LLM. [PITH_FULL_IMAGE:figures/full_fig_p009_4.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

40 extracted references · 16 linked inside Pith

  1. [1]

    Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts,

    M. S. Khan, S. Sinha, T. Uddin, D. Stricker, S. A. Ali, and M. Z. Afzal, “Text2cad: Generating sequential cad designs from beginner-to-expert level text prompts,”Advances in Neural Information Processing Systems, vol. 37, pp. 7552–7579, 2024

    2024

  2. [2]

    Text-to-cadquery: A new paradigm for cad generation with scalable large model capabilities,

    H. Xie and F. 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

  3. [3]

    Shop-r1: Rewarding llms to simulate human behavior in online shopping via reinforcement learning,

    Y. Zhang, T. Wang, J. Gesi, Z. Wang, Y. Lu, J. Lin, S. Zhan, V. Gao, R. Jiao, J. Liuet al., “Shop-r1: Rewarding llms to simulate human behavior in online shopping via reinforcement learning,”arXiv preprint arXiv:2507.17842, 2025

    arXiv 2025

  4. [4]

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

    X. Xu, J. Lambourne, P. Jayaraman, Z. Wang, K. Willis, and Y. Furukawa, “Brepgen: A b-rep generative diffusion model with structured latent geometry,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–14, 2024

    2024

  5. [5]

    Solidgen: An autoregressive model for direct b-rep synthesis,

    P. K. Jayaraman, J. G. Lambourne, N. Desai, K. D. Willis, A. Sanghi, and N. J. Morris, “Solidgen: An autoregressive model for direct b-rep synthesis,”arXiv preprint arXiv:2203.13944, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    Approaching step file analysis as a language processing task: A robust and scale-invariant solution for machining feature recognition,

    V. Miles, S. Giani, and O. Vogt, “Approaching step file analysis as a language processing task: A robust and scale-invariant solution for machining feature recognition,”Journal of Computational and Applied Mathematics, vol. 427, p. 115166, 2023. 10

    2023

  7. [7]

    Recursive encoder network for the automatic analysis of step files,

    M. Victoria, G. Stefano, and V. Oliver, “Recursive encoder network for the automatic analysis of step files,”Journal of Intelligent Manufacturing, vol. 34, no. 1, pp. 181–196, 2023

    2023

  8. [8]

    Graph-kv: Breaking sequence via injecting structural biases into large language models,

    H. Wang, P. Wang, M. Li, S. Liu, S. Miao, Z. Wang, and P. Li, “Graph-kv: Breaking sequence via injecting structural biases into large language models,”arXiv preprint arXiv:2506.07334, 2025

    arXiv 2025

  9. [9]

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

    S. Wang, C. Chen, X. Le, Q. Xu, L. Xu, Y. Zhang, and J. Yang, “Cad-gpt: Synthesising cad construction sequence with spatial reasoning-enhanced multimodal llms,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 8, p. 7880–7888, Apr. 2025. [Online]. Available: http://dx.doi.org/10.1609/aaai.v39i8.32849

    doi:10.1609/aaai.v39i8.32849 2025

  10. [10]

    Flexcad: Unified and versatile controllable cad generation with fine-tuned large language models,

    Z. Zhang, S. Sun, W. Wang, D. Cai, and J. Bian, “Flexcad: Unified and versatile controllable cad generation with fine-tuned large language models,”arXiv preprint arXiv:2411.05823, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Cad-mllm: Unifying multimodality-conditioned cad generation with mllm,

    J. Xu, C. Wang, Z. Zhao, W. Liu, Y. Ma, and S. Gao, “Cad-mllm: Unifying multimodality-conditioned cad generation with mllm,”arXiv preprint arXiv:2411.04954, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Cad-llama: leveraging large language models for computer-aided design parametric 3d model generation,

    J. Li, W. Ma, X. Li, Y. Lou, G. Zhou, and X. Zhou, “Cad-llama: leveraging large language models for computer-aided design parametric 3d model generation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 18 563–18 573

    2025

  13. [13]

    Img2cad: Reverse engineering 3d cad models from images through vlm-assisted conditional factorization,

    Y. You, M. A. Uy, J. Han, R. Thomas, H. Zhang, S. You, and L. Guibas, “Img2cad: Reverse engineering 3d cad models from images through vlm-assisted conditional factorization,”arXiv preprint arXiv:2408.01437, 2024

    arXiv 2024

  14. [14]

    Cadvlm: Bridging language and vision in the generation of parametric cad sketches,

    S. Wu, A. H. Khasahmadi, M. Katz, P. K. Jayaraman, Y. Pu, K. Willis, and B. Liu, “Cadvlm: Bridging language and vision in the generation of parametric cad sketches,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 368–384

    2024

  15. [15]

    Openecad: An efficient visual language model for editable 3d-cad design,

    Z. Yuan, J. Shi, and Y. Huang, “Openecad: An efficient visual language model for editable 3d-cad design,” Computers & Graphics, vol. 124, p. 104048, 2024

    2024

  16. [16]

    Cad-coder: An open-source vision-language model for computer-aided design code generation,

    A. C. Doris, M. F. Alam, A. H. Nobari, and F. Ahmed, “Cad-coder: An open-source vision-language model for computer-aided design code generation,”arXiv preprint arXiv:2505.14646, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward,

    Y. Guan, X. Wang, X. Ming, J. Zhang, D. Xu, and Q. Yu, “Cad-coder: Text-to-cad generation with chain-of-thought and geometric reward,”arXiv preprint arXiv:2505.19713, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Llm4cad: Multimodal large language models for three-dimensional computer- aided design generation,

    X. Li, Y. Sun, and Z. Sha, “Llm4cad: Multimodal large language models for three-dimensional computer- aided design generation,”Journal of Computing and Information Science in Engineering, vol. 25, no. 2, p. 021005, 2025

    2025

  19. [19]

    Query2cad: Generating cad models using natural language queries,

    A. Badagabettu, S. S. Yarlagadda, and A. B. Farimani, “Query2cad: Generating cad models using natural language queries,” 2024. [Online]. Available: https://arxiv.org/abs/2406.00144

    Pith/arXiv arXiv 2024

  20. [20]

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

    R. Wu, C. Xiao, and C. Zheng, “Deepcad: A deep generative network for computer-aided design models,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 6772–6782

    2021

  21. [21]

    Complexgen: Cad reconstruction by b-rep chain complex generation,

    H. Guo, S. Liu, H. Pan, Y. Liu, X. Tong, and B. Guo, “Complexgen: Cad reconstruction by b-rep chain complex generation,”ACM Transactions on Graphics (TOG), vol. 41, no. 4, pp. 1–18, 2022

    2022

  22. [22]

    Point2cad: Reverse engineering cad models from 3d point clouds,

    Y. Liu, A. Obukhov, J. D. Wegner, and K. Schindler, “Point2cad: Reverse engineering cad models from 3d point clouds,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3763–3772. 11

    2024

  23. [23]

    Abc: A big cad model dataset for geometric deep learning,

    S. Koch, A. Matveev, Z. Jiang, F. Williams, A. Artemov, E. Burnaev, M. Alexa, D. Zorin, and D. Panozzo, “Abc: A big cad model dataset for geometric deep learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9601–9611

    2019

  24. [24]

    Empowering au- tonomous driving with large language models: A safety perspective,

    Y. Wang, R. Jiao, S. S. Zhan, C. Lang, C. Huang, Z. Wang, Z. Yang, and Q. Zhu, “Empowering au- tonomous driving with large language models: A safety perspective,”arXiv preprint arXiv:2312.00812, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Shedding light on vln robustness: A black-box framework for indoor lighting-based adversarial attack,

    C. Li, W. Tang, Y. Huang, S. S. Zhan, M. Hu, X. Jia, and Y. Liu, “Shedding light on vln robustness: A black-box framework for indoor lighting-based adversarial attack,”arXiv preprint arXiv:2511.13132, 2025

    arXiv 2025

  26. [26]

    Sentinel: A multi-level formal framework for safety evaluation of llm-based embodied agents,

    S. S. Zhan, Y. Liu, P. Wang, Z. Wang, Q. Wang, Z. Ruan, X. Shi, X. Cao, F. Yang, K. Wanget al., “Sentinel: A multi-level formal framework for safety evaluation of llm-based embodied agents,”arXiv preprint arXiv:2510.12985, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    A survey on large language models for code generation,

    J. Jiang, F. Wang, J. Shen, S. Kim, and S. Kim, “A survey on large language models for code generation,” arXiv preprint arXiv:2406.00515, 2024

    Pith/arXiv arXiv 2024

  28. [28]

    Fine-tuning large language models for domain adaptation: Ex- ploration of training strategies, scaling, model merging and synergistic capabilities,

    W. Lu, R. K. Luu, and M. J. Buehler, “Fine-tuning large language models for domain adaptation: Ex- ploration of training strategies, scaling, model merging and synergistic capabilities,”npj Computational Materials, vol. 11, no. 1, p. 84, 2025

    2025

  29. [29]

    Can LLMs understand unvoiced speech? exploring EMG-to-text conversion with LLMs,

    P. Mohapatra, A. Pandey, X. Zhang, and Q. Zhu, “Can LLMs understand unvoiced speech? exploring EMG-to-text conversion with LLMs,” inProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). Association for Computational Linguistics, 2025, pp. 703–712

    2025

  30. [30]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,”arXiv preprint arXiv:1908.10084, 2019

    Pith/arXiv arXiv 1908

  31. [31]

    The faiss library,

    M. Douze, A. Guzhva, C. Deng, J. Johnson, G. Szilvasy, P.-E. Mazar´ e, M. Lomeli, L. Hosseini, and H. J´ egou, “The faiss library,”arXiv preprint arXiv:2401.08281, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Rayet al., “Training language models to follow instructions with human feedback,”Advances in neural information processing systems, vol. 35, pp. 27 730–27 744, 2022

    2022

  33. [33]

    Token buncher: Shielding llms from harmful reinforcement learning fine-tuning,

    W. Feng, L. Wang, T. Wei, J. Zhang, C. Gao, S. Zhan, P. Lv, and W. Dong, “Token buncher: Shielding llms from harmful reinforcement learning fine-tuning,”arXiv preprint arXiv:2508.20697, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    The lessons of developing process reward models in mathematical reasoning,

    Z. Zhang, C. Zheng, Y. Wu, B. Zhang, R. Lin, B. Yu, D. Liu, J. Zhou, and J. Lin, “The lessons of developing process reward models in mathematical reasoning,”arXiv preprint arXiv:2501.07301, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Model-based reward shaping for adversarial inverse reinforcement learning in stochastic environments,

    S. S. Zhan, P. Wang, Q. Wu, Y. Wang, R. Jiao, C. Huang, and Q. Zhu, “Model-based reward shaping for adversarial inverse reinforcement learning in stochastic environments,”arXiv preprint arXiv:2410.03847, 2024

    arXiv 2024

  36. [36]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  37. [37]

    Unsloth,

    M. H. Daniel Han and U. team, “Unsloth,” 2023. [Online]. Available: http://github.com/unslothai/uns loth 12

    2023

  38. [38]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.”ICLR, vol. 1, no. 2, p. 3, 2022

    2022

  39. [39]

    pyocct – python bindings for opencascade via pybind11,

    T. Laughlin, “pyocct – python bindings for opencascade via pybind11,” 2020, https://github.com/trelau/pyOCCT

    2020

  40. [40]

    Cad-judge: Toward efficient morphological grading and verification for text-to-cad generation,

    Z. Zhou, J. Han, L. Du, N. Fang, L. Qiu, and S. Zhang, “Cad-judge: Toward efficient morphological grading and verification for text-to-cad generation,”arXiv preprint arXiv:2508.04002, 2025. 13

    Pith/arXiv arXiv 2025