Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

LLM-to-Phy3D: Physically Conform Online 3D Object Generation with LLMs

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An iterative prompt-refinement loop can make frozen text-to-3D models produce physically conforming, novel car designs, improving a combined physics-and-domain score by up to 106.7% without retraining.

desk verdict Promising prompt-refinement recipe for physics-aware text-to-3D, but Equation (3) has a sign error that contradicts the reported results and needs to be resolved before the paper is publishable. read the letter →

arxiv 2506.11148 v1 pith:FLLCTSLU submitted 2025-06-11 cs.CV cs.LG

Melvin Wong , Yueming Lyu , Thiago Rios , Stefan Menzel , Yew-Soon Ong This is my paper
classification cs.CVcs.LG
keywords text-to-3Dgenerationlargelanguagemodelsblack-boxoptimizationphysicalAIaerodynamicvehicledesigncomputationalfluiddynamicsgeometricnoveltyonlinerefinementloop
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

LLM-to-Phy3D claims that existing text-to-3D models can be made to produce physically usable 3D objects without retraining by wrapping them in an online black-box refinement loop. The loop has a large language model write prompts, a text-to-3D model turn them into car shapes, and an evaluator score each shape on aerodynamic drag, visual match to the car domain, and geometric novelty against reference cars. The best-scoring prompts are fed back as exemplars, and the LLM iteratively learns which words produce better designs. In vehicle-design experiments, this raises the domain-and-physics alignment rating (DPAR) by 4.46% to 106.74% over the same LLM-to-3D pipelines used without refinement, across four LLMs and two generators. If the claim holds, physical design constraints can be imposed on frozen generative models through prompt search alone.

What carries the argument

The load-bearing mechanism is the three-term black-box objective of Equation (3), minimized over prompts: $f_{\mathrm{physical}} + f_{\mathrm{domain}} - \beta F_{\mathrm{novelty}}$. Here $f_{\mathrm{physical}}$ is the drag coefficient from an OpenFOAM CFD simulation, normalized to the range [0,1]; $f_{\mathrm{domain}}$ is a temperature-softened BLIP-2 cosine similarity between orthographic renders and the concept "car"; and $F_{\mathrm{novelty}}$ is the minimum over reference cars of pixel-level plus EfficientNet feature-map differences on masked orthographic renders, with weight $\beta = e^{-\hat{\mu}/\hat{\sigma}}$ derived from the reference set's drag statistics. The framework alternates sampling $N$ prompts, generating and scoring artifacts, and selecting the $N$ best prompt-score pairs as exemplars, so memory stays constant while the LLM receives in-context directional feedback. Orthographic physical light simulation (via OPTIX) is the shared visual representation that makes both domain recognition and novelty comparison distortion-free; switching to perspective projection degrades DPAR gains by 3.38% to 52.21%.

What would settle it

Re-run the same 40-step loop on the car task, then re-evaluate all final candidates with a high-fidelity CFD mesh or a different drag code; if the prompts picked by the proxy do not yield lower drag than the unrefined baseline, the physical score is gameable. A second decisive test is to remove $F_{\mathrm{novelty}}$ from Equation (3) while keeping the drag and domain terms; if DPAR still improves or malformed cars appear among the exemplars, the claimed need for the full three-term objective is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that physical conformance can be injected into frozen text-to-3D pipelines by steering the prompt distribution rather than retraining the generator. LLM-to-Phy3D runs an online loop: an LLM proposes prompts of the form "A Car in the shape of ...", the text-to-3D model (Shap-E or Trellis) turns each into a mesh, and each mesh is scored by Equation (3), combining normalized CFD drag, a BLIP-2 domain-alignment score, and a geometric-novelty score computed on orthographic renders against a reference set. The prompts that produced the best-scoring cars are kept as exemplars for the next round, so the LLM learns which of its own wordings translate into physically better designs. On the DPAR metric, the expected ratio of domain alignment to drag, the refined pipeline beats the same LLM-to-3D model without refinement for all four LLMs and both generators, with improvements from 4.46% to 106.74%. The ablations support the three-term design: drag alone rewards malformed and fragmented cars, domain alone yields near-copies, and novelty alone favors impractical shapes.

Load-bearing premise

The load-bearing premise is that the three-term score in Equation (3), combining drag, domain match, and geometric novelty, correctly identifies cars that are genuinely physically conforming rather than shapes that merely game the proxy; the paper's own ablation shows drag alone can be gamed by fragmented cars, so the full objective must keep excluding them.

Editorial extensions

If this is right

  • Applying the loop to GPT-4o Mini, GPT-3.5, Gemini 2.0 Flash Lite, and Mistral 3.1 Small improves DPAR over the unrefined baseline with both Shap-E and Trellis, from +4.46% to +106.74%.
  • Using orthographic instead of perspective camera projection improves DPAR by +3.38% to +52.21%, showing that geometric distortion in the visual evaluation materially changes which designs are found.
  • The full three-term objective is necessary: drag-only search favors malformed or fragmented cars, domain-only search produces near-copies of reference cars, and novelty-only search favors impractical shapes.
  • Prompt exemplars give the LLM enough directional feedback that GPT-4o Mini and GPT-3.5 converge in under 10 iterations, so the approach works with constant memory and a fixed number of evaluations per step.
  • The inherited weaknesses of the base models remain: outputs can be fragmented, non-watertight, or exhibit Janus-face artifacts, and such cases are handled by regeneration and bounded scores.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An untested extension: replace the drag simulator with another physical evaluator, such as lift, structural stress, or energy dissipation, and the same loop should push the LLM toward that target because only the scorer changes.
  • Because the loop touches the generator only through prompts and rendered images, it should also work with proprietary or closed text-to-3D services, provided the evaluators can view the outputs.
  • The reported margins come from a single automotive setup with one reference set and no repeated runs; a stress test with different seeds, reference cars, and target domains would show how stable the gains are.
  • The novelty term is anchored to the reference set through $\beta$ and the minimum over references, so the kind and radicalness of discovered designs will depend on that set; changing it would change what counts as novel.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes LLM-to-Phy3D, an online black-box refinement loop that iteratively prompts an LLM to generate text-to-3D objects, evaluates them with a CFD drag simulator, a BLIP2-based domain alignment measure, and a geometric novelty metric over orthographic renders, and feeds the best-scoring prompts back as exemplars. The central claim is that this loop improves physical conformity and novelty over conventional LLM-to-3D pipelines, with reported DPAR improvements of 4.46% to 106.74% across four LLMs and two text-to-3D generators. Ablation studies show that each objective term alone is insufficient, and the paper argues that the combined objective is necessary.

Significance. If the reported results hold, the framework offers a practical way to inject physics and domain knowledge into frozen text-to-3D models, which is a relevant contribution for engineering design. The use of external non-differentiable simulators (OpenFOAM) and a fixed domain VLM avoids the circularity of fitting to a self-generated target; the orthographic rendering pipeline is a reasonable design choice; and the authors are explicit about inherited failure modes such as non-watertight meshes and Janus artifacts. However, the formal objective as written is internally inconsistent with the reported behavior, and the empirical evidence rests on single runs without error bars, so the central claim is not yet established.

major comments (4)
  1. [Section 3.1, Eq. (3); Section 3.2(b), Eq. (5)] The second term of the objective is written as +fdomain(g(x), S), and Eq. (5) defines fdomain as the softmax of the BLIP2 similarity to S over the similarity to ¬S, so fdomain is high for in-domain objects. Minimizing Eq. (3) therefore rewards low fdomain, i.e., artifacts that do not resemble the target domain. This contradicts the sentence immediately after Eq. (3) that artifacts less resembling S are penalized, and it is inconsistent with the ablation in Figure 5(b), where optimizing 'fdomain only' produces cars that 'visually highly resemble the reference cars.' Every DPAR improvement in Table 1 is attributed to this objective, so the formal method, as written, cannot produce the reported behavior. The authors should correct the sign (e.g., use -fdomain or +(1-fdomain)) and revise the surrounding text to state the intended direction consistently.
  2. [Section 4.1 and Table 1; Appendix A.3] All results come from a single run with fixed seeds (Appendix A.3: 'the same random seeds are used throughout the iterations'), with no standard deviation, confidence intervals, or repeated trials. The headline improvements (4.46% to 106.74%) and the exact 1.0000 DPAR entries for GPT-4o Mini are therefore not established as robust. Please report multiple independent runs with different seeds and error bars, and explain why the GPT-4o Mini DPAR saturates exactly at 1.0000 for both Shap-E and Trellis.
  3. [Section 4.1, DPAR definition] DPAR is defined as E_{x∈D}[ fdomain(g(x),S) / fphysical(x) ], a ratio of two normalized scores that are both in [0,1]. The rating is not bounded above, so a single object with fphysical near its lower bound can dominate the expectation, and the metric conflates improvements in physical performance with improvements in domain alignment. For a fair and interpretable comparison, please report the two components separately (or justify the ratio with a bounded and physically meaningful formulation) and clarify the behavior when fphysical approaches zero.
  4. [Section 4.1; Table 2] The related work and Table 2 acknowledge prior black-box engineering design optimization methods (Rios et al. [2023], Wong et al. [2024], Xu et al. [2025]), but the experimental section compares only against plain LLM-to-3D baselines, not against these existing optimizers. Without such comparisons, the specific contribution of the selection-pressure mechanism and the combined visual/physics surrogate relative to prior black-box approaches cannot be assessed. Please add comparisons to these methods or explain why they are not applicable in this setting.
minor comments (5)
  1. [Section 3.2(b), Eq. (5)] The temperature hyperparameter Γ is mentioned in the main text as set to 0.01, but its role in the softmax is not explained until the supplementary material; please define it clearly in the main text.
  2. [Section 3.2(c), Eqs. (6)-(7)] The notation switches from lowercase fnovelty(x,y) in Eq. (6) to uppercase Fnovelty(x,Y) in Eq. (7); please make the distinction between per-reference and set-level novelty explicit and consistent.
  3. [Section 4.1] The normalization bounds a=-1.0 and b=1.0 are introduced for fphysical, but it is not stated whether these correspond to the realistic range of drag coefficients for the test scenario; please clarify the mapping between the raw CFD output and the normalized score.
  4. [Appendix A.2] The heading 'Failure Scenarios and Mitgation Approaches' contains a typo; it should read 'Mitigation.'
  5. [Figures 1, 4, and 11-18] The rendered images are small and the drag values are not annotated on the figures; consider labeling the physical performance directly on the images so the claimed improvement is visually verifiable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DPAR gains are measured against external CFD, VLM, and rendering components, not derived from the paper's own definitions or fitted parameters.

full rationale

No circular step is exhibited. The optimizer in Eq. (3) combines fphysical from OpenFOAM CFD, fdomain from a BLIP2 VLM, and Fnovelty from EfficientNet features and OPTIX orthographic renderings; these components are external to the LLM and to the reported DPAR metric, and none is defined in terms of the claimed result. The hyperparameter beta is fitted from the reference set, but it only weights the novelty term, which is not part of the DPAR evaluation, so the Table 1 improvements are not a fitted quantity renamed as a prediction. The paper does cite the authors' prior PREDO line (Rios et al. 2023, Wong et al. 2024) as motivation and in Table 2, but the central empirical comparison is against non-optimizing LLM-to-3D baselines and does not depend on those citations being true, so this is a positioning weakness rather than load-bearing self-citation. One non-circular correctness concern should be noted separately: Eq. (3) as written minimizes +fdomain, while fdomain in Eq. (5) is high for target-domain objects; the surrounding text and the Figure 5 ablations describe the opposite behavior. This is an internal sign inconsistency that would invalidate the formal objective as stated, but it is a mathematical correctness issue, not a reduction of the claimed prediction to its inputs by construction.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on imported black-box surrogates (CFD, BLIP2, EfficientNet, OptiX rendering) rather than on new physical entities or derivations. The only hand-fit quantities are the objective weights and normalization bounds. No new particles, forces, or conserved quantities are introduced.

free parameters (3)
  • beta (novelty weight) = beta = exp(-mu/sigma) from reference set
    Set from mean and standard deviation of drag performance of reference objects Y in Section 4.1; controls trade-off between physical score and novelty in Equation (3).
  • Gamma (domain temperature) = 0.01
    Hand-set in Section 3.2 for Equation (5); controls sharpness of VLM domain-alignment softmax; no sensitivity analysis reported.
  • a and b (normalization bounds) = a=-1.0, b=1.0
    Chosen in Section 4.1 to clip normalized drag scores in Equation (4); affects scale of fphysical and hence DPAR.
assumptions (6)
  • standard math Navier-Stokes and continuity equations govern the incompressible flow used to compute drag.
    Section A.1.1; used by OpenFOAM and not re-derived in the paper.
  • domain assumption OpenFOAM CFD simulation is a valid surrogate for aerodynamic drag of generated 3D vehicles.
    Used in Equation (4) and Section A.1.1; all physical performance scores and the objective in Equation (3) depend on this.
  • domain assumption BLIP2 cosine similarity on rendered views measures target-domain alignment.
    Used in Equation (5) and Equation (21); domain scores drive exemplar selection.
  • domain assumption Pixel and EfficientNet feature differences on orthographic renders measure meaningful geometric novelty.
    Used in Equation (6); the novelty term Fnovelty and the claim of greater geometric novelty rest on this.
  • domain assumption In-context exemplars of prompt-score pairs improve LLM prompt generation.
    Algorithm 1 feeds selected prompt-score pairs back into the LLM; the convergence plots in Figure 3 are the only evidence, with no theoretical guarantee.
  • domain assumption Orthographic rendering preserves surface topology relevant for novelty comparisons.
    Section 3.2 and Section A.1.2 motivate orthographic projection; the paper provides only a DPAR comparison (orthographic vs perspective), not a direct validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-to-Phy3D: Physically Conform Online 3D Object Generation with LLMs." pith.science (2026). https://pith.science/paper/FLLCTSLU

@misc{pith2026250611148,
  author       = {Pith},
  title        = {Pith review of: LLM-to-Phy3D: Physically Conform Online 3D Object Generation with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLLCTSLU}},
  note         = {Machine review of arXiv:2506.11148}
}
read the original abstract

The emergence of generative artificial intelligence (GenAI) and large language models (LLMs) has revolutionized the landscape of digital content creation in different modalities. However, its potential use in Physical AI for engineering design, where the production of physically viable artifacts is paramount, remains vastly underexplored. The absence of physical knowledge in existing LLM-to-3D models often results in outputs detached from real-world physical constraints. To address this gap, we introduce LLM-to-Phy3D, a physically conform online 3D object generation that enables existing LLM-to-3D models to produce physically conforming 3D objects on the fly. LLM-to-Phy3D introduces a novel online black-box refinement loop that empowers large language models (LLMs) through synergistic visual and physics-based evaluations. By delivering directional feedback in an iterative refinement process, LLM-to-Phy3D actively drives the discovery of prompts that yield 3D artifacts with enhanced physical performance and greater geometric novelty relative to reference objects, marking a substantial contribution to AI-driven generative design. Systematic evaluations of LLM-to-Phy3D, supported by ablation studies in vehicle design optimization, reveal various LLM improvements gained by 4.5% to 106.7% in producing physically conform target domain 3D designs over conventional LLM-to-3D models. The encouraging results suggest the potential general use of LLM-to-Phy3D in Physical AI for scientific and engineering applications.

Figures

Figures reproduced from arXiv: 2506.11148 by the authors.

Figure 1
Figure 1. Most representative vehicles generated with various conventional LLM-to-3D models using [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Aerodynamic Visualization of two vehicles generated (a) without considering physical [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Search optimization score (Equation 3) of LLM-to-Phy3D with different LLM and text-to [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (14 more)
Figure 5
Figure 5. Figure 5: Ablation Studies on the effectiveness of different terms in the objective function on [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Example of geometric novelty between a reference car (in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Illustration of virtual wind tunnel for aerodynamics physics simulation. In a simplified [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Example of objects rendered under different camera projections and overlay projected [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Example of meta prompt designed for aerodynamic design optimization task. The instruc [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Similarity of 3D generated artifacts between and in various LLMs in the last iteration. [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Examples of 3D cars generated with existing GPT-4o-Mini and Trellis text-to-3D genera [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Examples of 3D cars generated with existing GPT-4o-Mini and Shap-E text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]
Figure 13
Figure 13. Figure 13: Examples of 3D cars generated with existing GPT-3.5-Turbo and Trellis text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p023_13.png]
Figure 14
Figure 14. Figure 14: Examples of 3D cars generated with existing GPT-3.5-Turbo and Shap-E text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: Examples of 3D cars generated with existing Gemini-2.0-Lite and Trellis text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: Examples of 3D cars generated with existing Gemini-2.0-Lite and Shap-E text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p026_16.png]
Figure 17
Figure 17. Figure 17: Examples of 3D cars generated with existing Mistral-3.1-Small and Trellis text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p027_17.png]
Figure 18
Figure 18. Figure 18: Examples of 3D cars generated with existing Mistral-3.1-Small and Shap-E text-to-3D [PITH_FULL_IMAGE:figures/full_fig_p028_18.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. GReFEM: Multimodal LLMs as Zero-Shot Semantic Assistants for Physics-Guided 3D Mesh Refinement

    cs.GR 2026-07 conditional novelty 6.5 of 10

    GReFEM shows MLLMs zero-shot isolate load-activated geometric features for volumetric mesh refinement with higher precision than matched-budget geometric heuristics.

Reference graph

Works this paper leans on

18 extracted references · 9 canonical work pages · cited by 1 Pith paper

  1. [2024]

    Jun and A

    H. Jun and A. Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463,

    arXiv

  2. [2025]

    doi: 10.1109/TPAMI.2024.3522305. T. Bolkart, T. Li, and M. J. Black. Instant multi-view head capture through learnable registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 768–779, June

    arXiv 2024

  3. [1]

    A comprehensive list of Large Language Model knowledge cut off dates - ALLMO: Boost Your Brand’s Visibility in AI Search — allmo.ai

    ALLMO.ai. A comprehensive list of Large Language Model knowledge cut off dates - ALLMO: Boost Your Brand’s Visibility in AI Search — allmo.ai. https://www.allmo.ai/articles/ list-of-large-language-model-cut-off-dates . [Accessed 14-05-2025]. C. Banerjee, K. Nguyen, C. Fookes, and K. George. Physics-informed computer vision: A review and perspectives. ACM ...

    work page 2025

  4. [5]

    URL https://api.semanticscholar.org/ CorpusID:277150679. 10 N. Fabian, K. Moreland, D. Thompson, A. C. Bauer, P. Marion, B. Gevecik, M. Rasquin, and K. E. Jansen. The paraview coprocessing library: A scalable, general purpose in situ visualization library. In 2011 IEEE symposium on large data analysis and visualization, pages 89–96. IEEE,

    work page 2011

  5. [7]

    J. Li, D. Li, S. Savarese, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning , pages 19730–19742. PMLR, 2023a. M. Li, Y . Duan, J. Zhou, and J. Lu. Diffusion-sdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF conference...

    arXiv

  6. [8]

    semanticscholar.org/CorpusID:267334657

    URL https://api. semanticscholar.org/CorpusID:267334657. D. Liu, J. Zhang, A.-D. Dinh, E. Park, S. Zhang, A. Mian, M. Shah, and C. Xu. Generative physical ai in vision: A survey. arXiv preprint arXiv:2501.10928,

    arXiv

  7. [12]

    Z. Wang, J. Lorraine, Y . Wang, H. Su, J. Zhu, S. Fidler, and X. Zeng. Llama-mesh: Unifying 3d mesh generation with language models. arXiv preprint arXiv:2411.09595,

    arXiv

  8. [14]

    Xiang, Z

    J. Xiang, Z. Lv, S. Xu, Y . Deng, R. Wang, B. Zhang, D. Chen, X. Tong, and J. Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506,

    arXiv

Show all 18 references
  1. [15]

    URL https: //ojs.aaai.org/index.php/AAAI/article/view/20188

    doi: 10.1609/aaai.v36i3.20188. URL https: //ojs.aaai.org/index.php/AAAI/article/view/20188. Q. Xu, J. Liu, M. Wong, C. Chen, and Y . Ong. Looks great, functions better: Physics compliance text-to-3d shape generation. In International Joint Conference on Neural Networks (IJCNN). IEEE,

    doi

  2. [16]

    Zhang, L

    Y . Zhang, L. Wang, C. Zou, T. Wu, and R. Ma. Diff3ds: Generating view-consistent 3d sketch via differentiable curve rendering. ArXiv, abs/2405.15305,

    arXiv

  3. [17]

    semanticscholar.org/CorpusID:270045697

    URL https://api. semanticscholar.org/CorpusID:270045697. 12 A Technical Appendices and Supplementary Material A.1 Background Much interest in computer vision, computer graphics, and continuum dynamics is on the boundary between mediums. In physics, particularly in solid or flu...

    2001

  4. [18]

    The instruc- tions highlighted in green provide contextual information on the role, the search task the LLM needs to solve, and the operating constraints it must follow

    ✗ ✓ LLM-to-Phy3D (Ours) ✓ ✓ 17 Figure 9: Example of meta prompt designed for aerodynamic design optimization task. The instruc- tions highlighted in green provide contextual information on the role, the search task the LLM needs to solve, and the operating constraints it must ...

    2019

  5. [1997]

    Shimabucoro, S

    L. Shimabucoro, S. Ruder, J. Kreutzer, M. Fadaee, and S. Hooker. Llm see, llm do: Guiding data generation to target non-differentiable objectives. arXiv preprint arXiv:2407.01490,

    arXiv

  6. [1998]

    M. Wong, T. Rios, S. Menzel, and Y . S. Ong. Prompt evolutionary design optimization with generative shape and vision-language models. In 2024 IEEE Congress on Evolutionary Computation (CEC), pages 1–8. IEEE,

    2024

  7. [2000]

    T. Rios, S. Menzel, and B. Sendhoff. Large language and text-to-3d models for engineering design optimization. In 2023 IEEE Symposium Series on Computational Intelligence (SSCI) , pages 1704–1711. IEEE,

    2023

  8. [2001]

    Nie, C.-A

    A. Nie, C.-A. Cheng, A. Kolobov, and A. Swaminathan. Importance of directional feedback for llm-based optimizers. In NeurIPS 2023 Foundation Models for Decision Making Workshop,

    2023

  9. [2018]

    Dutta, M

    A. Dutta, M. Zheng, Z. Gao, B. Planche, A. Choudhuri, T. Chen, A. K. Roy-Chowdhury, and Z. Wu. Chrome: Clothed human reconstruction with occlusion-resilience and multiview-consistency from a single image. ArXiv, abs/2503.15671,

  10. [2022]

    doi: 10.1145/3554727

    ISSN 0360-0300. doi: 10.1145/3554727. URL https://doi.org/10.1145/3554727. P. Dutre, P. Bekaert, and K. Bala. Advanced global illumination. AK Peters/CRC Press,

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.