acceptodds
Under review as a conference paper at ICLR 2027

OAT3D: A Generative Model for 3D Shape- and Resolution-Free Structural Topology Optimization

Abstract

Topology optimization determines how material should be distributed within a design domain to optimize structural performance under prescribed loads, supports, and constraints. The problems are highly nonconvex, and conventional solvers require repeated finite element analyses, making high-resolution 3D topology optimization computationally expensive. Data-driven approaches face an additional bottleneck: generating training datasets requires solving large numbers of these costly problems, so existing work in data-driven topology optimization primarily focuses on 2D problems. Large-scale level 3D datasets remain scarce, and existing models remain limited in problem scale and boundary-condition diversity. Here, we introduce Optimize Any Topology-3D (OAT3D), a generative model for high-resolution 3D structural topology optimization. OAT3D introduces a novel sparse problem encoding to represent load and fixity conditions and combines a sparse volumetric variational autoencoder with a conditional diffusion transformer. We construct a dataset of 501,865 optimized topologies, generated using 51,800 GPU-hours and encompassing structures with up to one million elements and substantially more complex loading and support configurations than any existing structural topology optimization dataset. To our knowledge, this is the largest dataset for 3D topology optimization structures. Trained on this dataset, OAT3D generates structures at 512 resolution and achieves a 45-fold speedup over conventional topology optimization. These contributions expand the scale and diversity of data-driven topology optimization for three-dimensional structural design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Topology optimization, Latent diffusion models, 3D generative modeling, Foundation models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_L3f1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a large-scale dataset for 3D topology optimization under a linear-elastic minimum-compliance formulation and proposes OAT3D, a conditional generative model for structural topology design. OAT3D outperforms the evaluated NITO-3D baselines on the proposed benchmark, with further improvements achieved through best-of-N sampling and post-optimization.

Critical strengths
  1. The paper addresses a practically important problem in 3D structural topology optimization and introduces a large-scale dataset covering diverse domain dimensions, loading conditions, supports, and material budgets.

  2. The proposed framework is well suited to low-volume-fraction structures: its two-stage representation concentrates fine-scale modeling on occupied regions, while spatial encodings of loads and supports guide generation according to the prescribed mechanical conditions.

Critical weaknesses
  1. Limited external validation. The main performance evaluation is conducted on the proposed dataset. Although the paper discusses differences from NITO-3D and evaluates a partially overlapping subset, evaluation on the NITO-3D benchmark would provide stronger evidence of generalization. If direct evaluation is not feasible, the authors should clarify the limitations and any adaptations required for a meaningful comparison.

  2. Insufficient ablation of mechanical performance. The ablations primarily assess geometric overlap through IoU, which does not necessarily reflect structural performance. In particular, the benefits of BNG conditioning should be evaluated using end-to-end compliance error, volume constraint violations, and failure rates, to establish whether it improves mechanically valid generation rather than only geometric similarity.

Questions for authors
  1. Could the authors analyze the main failure modes beyond the volume-fraction breakdown in Appendix I?
  2. How well does OAT3D generalize to unseen loading and support configurations, particularly those outside the training distribution?
  3. How representative are the randomly sampled loads of practical structural design problems, and what checks ensure that the resulting problems are physically well posed?
3 comments
AuthorsOpenReview ↗

These comments address reviewer's comments on critical weaknesses

We genuinely appreciate the reviewer’s valuable feedback.

Reviewer:

Limited external validation. The main performance evaluation is conducted on the proposed dataset. Although the paper discusses differences from NITO-3D and evaluates a partially overlapping subset, evaluation on the NITO-3D benchmark would provide stronger evidence of generalization. If direct evaluation is not feasible, the authors should clarify the limitations and any adaptations required for a meaningful comparison.

Authors: We evaluated OAT3D generation on 2,000 problems from the NITO-3D test set and reported the results in Appendix M.2, together with the one adaptation we found necessary because NITO-3D domains are smaller than any domain in the OAT3D training set. The model never saw NITO-3D data during training or fine-tuning, and we used checkpoints from the initial draft submission. With the adaptation, OAT3D achieves lower compliance error on the NITO-3D test set than on the OAT3D test set. After 5 post-optimization updates, the median compliance error of OAT3D-generated topologies on the NITO-3D test set is 0.89%.

We also present common topology optimization problems at three resolutions in Appendix M.1, including the MBB beam, the middle- and bottom-loaded cantilever beams, and a bridge.

Reviewer:

Insufficient ablation of mechanical performance. The ablations primarily assess geometric overlap through IoU, which does not necessarily reflect structural performance. In particular, the benefits of BNG conditioning should be evaluated using end-to-end compliance error, volume constraint violations, and failure rates, to establish whether it improves mechanically valid generation rather than only geometric similarity.

Authors: We ran an additional best-of-N experiment without the conditioning, and the results are appended in Table 7. Across the experiments, the BNG conditioning does lead to better structural performance.

Best-of-N selection from OAT3D, comparison between with BNG and without BNG.

MethodNPO StepsCE mean (%) ↓CE median (%) ↓Mean VFE (%)Failure (%) ↓
Best-of-N selection from OAT3D (with BNG)
OAT3D2-30.8722.80−2.1834.80
OAT3D4-26.2518.27−0.7929.70
OAT3D8-22.7415.990.0624.80
OAT3D16-20.2314.301.0622.15
OAT3D32-17.9211.972.1719.85
Best-of-N selection from OAT3D without BNG
OAT3D (no BNG)2-33.5426.47−2.8342.10
OAT3D (no BNG)4-28.7621.55−1.8237.25
OAT3D (no BNG)8-25.2718.92−0.8133.20
OAT3D (no BNG)16-22.4716.72−0.1629.70
OAT3D (no BNG)32-20.3315.040.8426.20
AuthorsOpenReview ↗

These comments address reviewer's questions to authors

Reviewer:

Could the authors analyze the main failure modes beyond the volume-fraction breakdown in Appendix I?

Authors: We provide an additional study in Appendix J, where we (1) break down the overall failure rate, (2) investigate the mechanical causes, and (3) examine additional problem attributes like number of loaded nodes, number of loaded regions, number of fixed nodes, and number of elements (Table 18 in revised manuscript). We also added a figure to show some sample generations of all the four models side-by-side, demonstrating OAT3D’s advantage in generating structures with intact load paths, which the table below (Table 17) quantifies.

Failed single-generation predictions on the OAT test set by structural cause, as a percentage of the problems in each volume-fraction range. The five cause columns are exclusive and sum to the overall failure rate. Numbers in parentheses are shares of compliance failures in each class.

Volume fractionProblemsMethodOverall failure (%) ↓Solve failed (%)Compliance failure: incomplete load path (%)Compliance failure: partial load path (%)Compliance failure: intact load path (%)Volume failure only (%)
0.05 ≤ v* < 0.10312NITO-3D (NITO-trained)1001.9243.3 (44.4)48.4 (49.7)5.77 (5.9)0.64
NITO-3D (OAT-trained, 23M)97.40.0046.8 (48.0)36.2 (37.2)14.4 (14.8)0.00
NITO-3D (OAT-trained, 811M)96.50.0040.7 (42.2)40.1 (41.5)15.7 (16.3)0.00
OAT3D71.80.000.96 (1.4)6.41 (8.9)64.4 (89.7)0.00
0.10 ≤ v* < 0.20793NITO-3D (NITO-trained)99.42.4041.4 (42.8)50.7 (52.4)4.67 (4.8)0.25
NITO-3D (OAT-trained, 23M)93.80.0017.2 (18.3)44.4 (47.3)32.3 (34.4)0.00
NITO-3D (OAT-trained, 811M)91.60.0013.4 (14.6)38.7 (42.3)39.5 (43.1)0.00
OAT3D50.30.000.50 (1.0)2.14 (4.3)47.5 (94.7)0.13
0.20 ≤ v* ≤ 0.30895NITO-3D (NITO-trained)99.42.4621.5 (22.4)65.1 (67.9)9.27 (9.7)1.12
NITO-3D (OAT-trained, 23M)81.00.003.58 (4.4)27.9 (34.5)49.5 (61.1)0.00
NITO-3D (OAT-trained, 811M)79.70.003.46 (4.4)21.0 (26.4)55.1 (69.2)0.11
OAT3D24.80.000.22 (0.9)1.34 (5.4)23.2 (93.7)0.00

Reviewer:

How well does OAT3D generalize to unseen loading and support configurations, particularly those outside the training distribution?

Authors: We added Appendix M.1, which presents common topology optimization problems unseen during training, including the MBB beam, the middle- and bottom-loaded cantilever beams, and a bridge, across three resolutions. Without post-optimization, OAT3D achieves an average compliance error of 23.84% on these problems. For configurations outside the training distribution, we evaluated OAT3D on 2,000 problems from the NITO-3D test set, which differs from our training data in domain size, volume-fraction range, and boundary-condition sampling (Appendix M.2). With a factor-of-two grid refinement to match our training domain sizes, OAT3D achieves lower compliance error on the NITO-3D test set than on the OAT3D test set, and after 5 post-optimization updates the median compliance error is 0.89%.

Reviewer:

How representative are the randomly sampled loads of practical structural design problems, and what checks ensure that the resulting problems are physically well posed?

Authors: It was our intention to sample as randomly as possible for the OAT3D dataset. We did check for physical well-posedness through our dataset filtering scheme described in Appendix B.3. We ensured that after thresholding (quantization), the topology did not exhibit rigid body motion. Mechanically, we also ensured that the loaded degrees of freedom do not coincide with the fixed degrees of freedom.

Additional Remark

We hope these clarifications address the reviewer’s concerns and improve the presentation of this work.

Thank you for the thoughtful clarifications and additional experiments, particularly the mechanical ablations and failure analysis, which substantially address my concerns. The external evaluation is also helpful, but it does not yet establish a consistent advantage of OAT3D over the baselines beyond the proposed benchmark. I therefore remain somewhat uncertain about the settings in which OAT3D is most beneficial and the limits of its applicability. Thus, my overall assessment remains Weak Accept.

Reviewer_YYydOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a generative model for high-resolution 3D structural topology optimization, called Optimize Any Topology-3D (OAT3D). This involves a sparse problem encoding to represent load and fixity conditions and combines a sparse volumetric variational autoencoder with a conditional diffusion transformer. The authors build a dataset of 501,865 optimized topologies, generated using 51,800 GPU-hours that include structures with up to one million elements and signifiacntly complex loading and support configurations.

Critical strengths
  1. The scale and diversity of the 3D training dataset is quite impressive and is a significant contribution to the Design and Manufacturing (specifically, Topology Optimization) research community.

  2. The adaptation of sparse volumetric generative models to continuous material-density fields, and the proposed mechanics-informed conditioning strategy are interesting approaches.

Critical weaknesses
  1. The "First shape- and resolution-free generative topology optimization" claim is too strong. The paper involves topology generation for rectangular domains and resolutions that fit within the model's grid. So, the claim should be more like a high-resolution, conditional generative framework for diverse 3D minimum-compliance problems.

  2. The paper seems to be missing a few important references. For example, Giannone and Ahmed (2023), Diffusing the Optimal Topology: A Generative Optimization Approach, ASME IDETC/CIE 2023. This paper introduced TopoDiff-FF that seems to have conceptual similarity with BNG conditioning.

Questions for authors
  1. To establish the "shape- and resolution-free" claim, can the authors present some generalization studies varying aspect ratio, native mesh resolution and boundary-condition complexity, including out-of-distribution testing.

  2. Present at least a discussion comparing the BNG approach with the simpler Green's-function based idea in the TopoDiff-FF paper by Giannone and Ahmed (2023).

3 comments
AuthorsOpenReview ↗

These comments address reviewer's comments on critical weaknesses

We thank the reviewer for the thoughtful feedback.

Reviewer:

The "First shape- and resolution-free generative topology optimization" claim is too strong. The paper involves topology generation for rectangular domains and resolutions that fit within the model's grid. So, the claim should be more like a high-resolution, conditional generative framework for diverse 3D minimum-compliance problems.

Authors: We agree that our previous characterization of the method as "shape- and resolution-free" was too broad and not sufficiently supported by the presented experiments. We have therefore revised the corresponding claims in the manuscript to more accurately describe our contribution as a high-resolution, conditional generative framework for diverse 3D minimum-compliance problems. We have also added an experiment in Appendix N.3 to investigate the possibility of extending generation beyond the original training grid through guided denoising, as discussed later in our response.

Reviewer:

The paper seems to be missing a few important references. For example, Giannone and Ahmed (2023), Diffusing the Optimal Topology: A Generative Optimization Approach, ASME IDETC/CIE 2023. This paper introduced TopoDiff-FF that seems to have conceptual similarity with BNG conditioning.

Authors: We sincerely thank the reviewer for bringing this important prior work to our attention, and we apologize for the omission. We agree that TopoDiff-FF by Giannone and Ahmed (2023) is highly relevant to our work, particularly in its use of Green's-function-inspired conditioning to incorporate physical information into generative topology optimization. We have added the corresponding citation and expanded the discussion in Appendix C.1 to more clearly acknowledge this connection and distinguish our BNG approach from the previous work.

AuthorsOpenReview ↗

These comments address reviewer's questions to authors

Reviewer:

To establish the "shape- and resolution-free" claim, can the authors present some generalization studies varying aspect ratio, native mesh resolution and boundary-condition complexity, including out-of-distribution testing.

Authors: We agree that evaluating the generalization of our method across different problem configurations is important for understanding its capabilities and limitations.

To further investigate this, we have added additional experiments. First, in Appendix M.1, we evaluate the method on several common engineering topology optimization problems, including the MBB beam, middle- and bottom-loaded cantilever beams, and bridge configurations, across varying resolutions. Additionally, to explore generation beyond the native training grid, in Appendix N.3, we demonstrate a guided joint denoising approach that combines six overlapping blocks to generate a 703 × 60 × 60 structure containing more than 2.5 million voxels. This experiment demonstrates the potential to extend our framework to larger design domains while maintaining structural connectivity between blocks. We acknowledge that these additional experiments do not fully establish generalization to arbitrary geometries, resolutions, or boundary condition configurations. We have therefore revised our claims to more accurately reflect the demonstrated capabilities of our framework.

In addition, we evaluated OAT3D generation on some common engineering problems and 2,000 problems from the NITO-3D test set. The results are reported in Appendices M.1 and M.2, respectively. The model never saw NITO-3D data during training or fine-tuning.

Reviewer:

Present at least a discussion comparing the BNG approach with the simpler Green's-function based idea in the TopoDiff-FF paper by Giannone and Ahmed (2023).

Authors: We agree that a more explicit comparison with the Green's-function-based conditioning approach introduced in TopoDiff-FF would help clarify the motivation and distinction of our proposed BNG approach. To address this, we have added a discussion and quantitative comparison in Appendix C.1. While TopoDiff-FF employs kernel relaxation inspired by Green's functions to represent loads and boundary conditions, our BNG approach estimates support reactions that satisfy global force and moment balance, and subsequently constructs an approximation of the strain-energy density field. To further investigate the differences between these conditioning representations, we compared the raw BNG field against a 3D extension of the publicly released TopoDiff-FF kernels, using the strain-energy density obtained from finite element analysis of the fully solid domain as a reference. Across 200 held-out test problems, we evaluated the intersection-over-union (IoU) between the top 10% of elements identified by each conditioning field and the corresponding FEM energy mask. The resulting mean IoU was 0.453 for BNG and 0.172 for the extended TopoDiff-FF kernels.

Additional Remark

We hope these clarifications address the reviewer’s concerns and improve the presentation of this work.

Reviewer_YYydOpenReview ↗

Thanks for addressing my concerns

Thanks for addressing my concerns. I will maintain my positive score.

Reviewer_1o9DOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

OAT3D is a 3D generative model to solve high-resolution structural optimization problems. The contributions include a large 3D dataset for optimized topology covering a wide range of constraints and grid resolutions. It gains substantial speedup over conventional topology optimization with its two-stage sparse VAE.

Critical strengths
  • The dataset is the large and diverse: 4x larger than NITO-3D’s and covers diverse boundary conditions and aspect ratios,
  • The generative model gains 45.9x speedup over the conventional TO.
  • Using density instead of SDF is well justified
  • The BNG conditioning field offers a cheap approximation of strain energy that needs no FE solve, although the measure effect is small (Table 4).
Critical weaknesses
  • From the statistics (Table 6), it reveals that sampling alone is not enough for a reliable design, and the model heavily depends on best-of-N and post optimization.
  • This paper claims 45.9x speedup over traditional TO, however, it only counts generating a single raw candidate. From Table 6, it suggests post-optimization is essential to performance of the final topology. Therefore, the cost of post-optimizations should be counted as well.
  • All the evaluations were done on in-distribution problems. It would be more convincing if the authors can also include some out-of-distribution problems especially the practical ones in engineering, such as a cantilever, a 3D MBB beam, a bridge, etc.
Questions for authors
  • The conventional topology optimization uses PGD with conjugate directions, whereas post-optimization of OAT3D uses OC. It is worth comparing them with the same optimizer.
  • Iteration counts are not reported. Although the paper gives TO wall-clock time, it highly depends on computing power. Iteration count can be a good indicator of computational efficiency.
  • More explanation is needed for filtering criteria when creating the dataset. If hard problems are dropped, both training and test sets are biased toward easier case.
  • In order to demonstrate scalability of the model, ther should be a breakdown of speedup by problem size, ie, number of elements, number of boundary conditions.
3 comments
AuthorsOpenReview ↗

These comments address reviewer's critical weaknesses

We sincerely thank the reviewer for this valuable suggestion.

Reviewer:

From the statistics (Table 6), it reveals that sampling alone is not enough for a reliable design, and the model heavily depends on best-of-N and post optimization.

Authors: We agree that, when evaluated in terms of compliance, our method still relies on best-of-N selection and post-optimization to achieve performance comparable to traditional topology optimization. We acknowledge this limitation and have revised the manuscript to more clearly distinguish between the efficiency of raw structure generation and the overall computational cost required to achieve comparable compliance. In particular, we have moderated our speedup claims to provide a more accurate and balanced representation of our method's performance. We have also conducted additional timing experiments to account for the computational cost of post-optimization, as detailed in our response to the next comment and inserted the table below to the manuscript. We sincerely appreciate the reviewer for highlighting this important distinction. Please note that the original Table 6 is now Table 7 in the revised manuscript.

Mean and median per-problem generation and compliance evaluation time on an H200 GPU for raw generation, best-of-N selection, and post-optimization of the best-of-2 selection, profiled over the 2,000-problem OAT3D test set.

MethodGeneration Time (s)Compliance Evaluation Time (s)Total Time (s)Compliance EvaluationsSpeedup w/ EvaluationSpeedup w/o Evaluation
Mean
Direct Optimization54.9-54.9124--
Raw Generation1.221.672.90119.0×44.9×
Best-of-22.453.355.7929.48×22.4×
Best-of-2 + 5 post-optimization2.458.4510.975.04×22.4×
Best-of-2 + 10 post-optimization2.4512.615.0123.65×22.4×
Best-of-2 + 15 post-optimization2.4516.719.2172.86×22.4×
Median
Direct Optimization37.0-37.089--
Raw Generation1.221.252.47115.0×30.3×
Best-of-22.442.504.9427.49×15.2×
Best-of-2 + 5 post-optimization2.446.859.2973.98×15.2×
Best-of-2 + 10 post-optimization2.4410.112.6122.94×15.2×
Best-of-2 + 15 post-optimization2.4413.516.0172.31×15.2×

Reviewer:

This paper claims 45.9x speedup over traditional TO, however, it only counts generating a single raw candidate. From Table 6, it suggests post-optimization is essential to performance of the final topology. Therefore, the cost of post-optimizations should be counted as well.

Authors: We fully agree that accounting for the computational cost of post-optimization is essential for a fair comparison with traditional topology optimization. To address this concern, we have conducted additional timing experiments and inserted the table above to provide a more comprehensive comparison, distinguishing between raw structure generation, best-of-N evaluation, and post-optimization. For completeness, we have also updated the timing measurements using the full test set of 2,000 problems on H200 GPUs, rather than the subset of 100 problems used in the original evaluation. We agree with the reviewer that comparing computational efficiency at comparable compliance levels provides a more meaningful assessment of the practical performance of our method. Specifically, our experiments show that achieving a compliance error below 5% requires 10 steps of post-optimization, resulting in an average speedup of 3.65x when accounting for post-optimization time on H200. We have accordingly revised the manuscript to place greater emphasis on this more representative speedup and to clearly distinguish it from the raw generation speedup.

Reviewer:

All the evaluations were done on in-distribution problems. It would be more convincing if the authors can also include some out-of-distribution problems especially the practical ones in engineering, such as a cantilever, a 3D MBB beam, a bridge, etc.

Authors: We thank the reviewer for pointing out the missing link to practical engineering problems. We added Appendix M.1, which presents common topology optimization problems across three resolutions, including the MBB beam, the middle- and bottom-loaded cantilever beams, and a bridge.

In addition, we evaluated OAT3D generation on 2,000 problems from the NITO-3D test set and reported the results in Appendix M.2. We did this evaluation with the NITO dataset unseen during training or fine tuning. OAT3D achieves lower compliance error on the NITO-3D dataset than on the OAT3D test set. After 5 post-optimization updates, the median compliance error on the NITO-3D test set is 0.89%.

AuthorsOpenReview ↗

These comments address reviewer's questions to authors

Reviewer:

The conventional topology optimization uses PGD with conjugate directions, whereas post-optimization of OAT3D uses OC. It is worth comparing them with the same optimizer.

Authors: We added an ablation in Appendix K, comparing PGD and OC as post-optimizers. OC achieves a lower compliance error. We attribute this to how PGD scales its updates. PGD normalizes its step globally first by the largest sensitivity magnitude and then by spectral estimates that take several iterations to calibrate. This is acceptable over a full optimization run, but not within a budget of a few updates.

Reviewer:

Iteration counts are not reported. Although the paper gives TO wall-clock time, it highly depends on computing power. Iteration count can be a good indicator of computational efficiency.

Authors: We agree that iteration count is an important indicator of computational efficiency, and we have added a detailed breakdown of iteration counts and inference times for the dense and sparse stages across different problem sizes and boundary condition configurations. These additional results are provided in the new Appendix L. From the OAT3D generation timing results, we observe that the raw generation time varies only slightly across problem configurations, differing by at most 0.13 s between bins (Tables 20-21). In contrast, for traditional TO, both the number of iterations and the computational cost per iteration depend on the problem size and the extent of the domain constrained by boundary conditions. While iteration count provides a valuable complementary perspective, we note that it may not fully capture these differences in per-iteration computational cost, particularly when comparing methods with different computational characteristics. We therefore believe that reporting both iteration counts and wall-clock times provides a more comprehensive assessment of computational efficiency.

Reviewer:

More explanation is needed for filtering criteria when creating the dataset. If hard problems are dropped, both training and test sets are biased toward easier case.

Authors: We added a list of the criteria and the reason for filtering in Appendix B.3 (Table 8) of the revised paper. None of the criteria was to bias the dataset to achieve an easier distribution. Because problems are sampled at random, a fraction of them are ill-posed or yield degenerate solutions that are not meaningful for training. We filter out topologies that have 10 or more disconnected islands, topologies that are not constrained against rigid-body motion, optimization runs that did not converge within the 500-iteration cap, and topologies with any side length greater than 512 elements. Early in data generation, prior to training the OAT3D model, we discovered an error in the code that resulted in a mismatch between the sampled input and the input fed into the optimizer, and we excluded the affected runs.

Reviewer:

In order to demonstrate scalability of the model, there should be a breakdown of speedup by problem size, ie, number of elements, number of boundary conditions.

Authors: We agree that understanding how computational efficiency scales with problem size is important for evaluating the practical applicability of our method. To provide further insight, we have included additional timing analyses across different problem sizes and boundary condition configurations in the new Appendix L. As shown in the results, the generation time of OAT3D remains relatively stable across different problem sizes, whereas conventional topology optimization is more sensitive to problem size, as both the number of iterations and the computational cost per iteration can increase with problem complexity. For a more comprehensive speedup comparison, we included two more tables. Table 23 reports the speedup based solely on raw structure generation, without accounting for compliance evaluation or post-optimization. Table 24 presents a more comprehensive comparison that considers best-of-2 sampling, where the candidate with better inference-stage compliance is selected and subsequently refined through 10 steps of post-optimization. We believe the latter provides a more representative assessment of computational efficiency when considering the compliance performance of the resulting structures.

Additional Remark

While we highlight the speedup compared to direct optimization, we consider the pre-trained model equally valuable, since it can be adapted by the community for other downstream tasks as demonstrated at the end of the main body of this work (Section 4.4).

We hope these clarifications address the reviewer’s concerns and improve the presentation of this work.

Reviewer_1o9DOpenReview ↗

Revisions read

Thank you for your clarification. I have updated the score accordingly.

Comments on the paper

1 comment

Dear authors and reviewers,

The author-reviewer discussion period closes on November 18 (AoE). Please engage/continue to engage during this period so that we have our issues discussed thoroughly.

AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.