acceptodds
Under review as a conference paper at ICLR 2027

Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion

Abstract

Generating crystals has recently attracted significant interest due to their broad applications in materials science. However, existing generative models struggle to produce complete crystallographic specifications, limiting their ability to capture full crystal symmetries and structural dependencies. In particular, current state-of-the-art approaches generate crystals only up to site symmetries and rely on sampling space groups from empirical distributions during inference. Inspired by spontaneous symmetry breaking in physics, where crystals break symmetries of the underlying system, we propose a novel diffusion-based framework that generates full crystallographic specifications by reversing from the lowest-symmetry priors. Our method leverages a Markovian jump-diffusion process to model these symmetry-breaking dynamics, enabling it to traverse different space groups in a physically motivated manner. Our model, dubbed Symmetry-breaking Crystal Diffusion (SbCD), introduces a principled approach to explicitly incorporate inter-space-group transitions into the generative process. In de novo crystal generation experiments, SbCD substantially outperforms its symmetry-preserving counterpart and achieves state-of-the-art thermodynamic stability on MPTS-52, offering a promising perspective for generative modeling of crystalline materials.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
de novo crystal generation, jump diffusion models, symmetry breaking, space group transition dynamics

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

The authors of this paper introduce a diffusion process for symmetry breaking, allowing to jointly generate all aspects of crystal structures without sampling from empirical distributions.

Critical strengths

Interesting new approach, clear conceptually new idea compared to previous work. Some ablations and overall results look promising.

Critical weaknesses

The evidence that the approach has a clear advantage over other (simpler) approaches is not strong. The ablations in the last lines of Table 2 are good, but the manuscript would be stronger if other ablations are performed that actually demonstrate the advantage of the new approach over an empirical non-learned marginal distribution over G, as assumed in other models.

Comparison and benchmarking, especially on MPTS-52 (Table 3): Latest SOTA methods are missing. Please add and discuss the latest baseline models. While some metrics improve, others are worse than SOTA, which is overrated in the abstract.

According to P4.2, it seems like only symmetry-breaking effects from any space group directly to P1 are modelled, which is not well justified (see question below).

Questions for authors

Please better explain the ablations you introduced in Section 5. All SbCD variants are described in only one sentence, which makes it very hard to follow.

Where do you model the symmetry hierarchy of space groups? Does the process only allow jumps to P1 directly? This seems to severely limit what was promised throughout the paper.

3 comments

We thank the Reviewer for their constructive and detailed feedback. We address the raised weaknesses (W) and questions (Q) as follows.

W1+Q1. Details of the SbCD variants and an ablation using an empirical marginal distribution over GG

We have clarified the descriptions of the SbCD variants in the revised manuscript. Below, we explain the design and purpose of each variant.

(1) SbCD♯\text{SbCD}^{\sharp} learns site symmetries using the 15×1315 \times 13 binary matrix representation introduced by Levy et al. [1]. This representation specifies the generators of the site-symmetry group along 15 possible crystallographic axes, including body and face diagonals, with up to 13 distinct symmetry operations per axis (Lemma 3.2). Figure 3 and the second ablation study (lines 491–503) show that SbCD, which uses a simple one-hot encoding of oriented site-symmetry symbols (Section 4.4), is more robust than SbCD♯\text{SbCD}^{\sharp}. This result suggests that our simpler representation helps address the challenges of modeling complex site symmetries.

(2) SbCD♢\text{SbCD}^{\diamondsuit} is an SbCD variant with a smaller architecture comparable in size to DiffCSP and DiffCSP++. It generates less stable structures but achieves better S.U.N. results, suggesting slightly less overfitting on MP-20.

(3) SbCD♢\text{SbCD}^{\diamondsuit} (w/o SB) is an ablation of SbCD♢\text{SbCD}^{\diamondsuit} that removes symmetry breaking and explicitly conditions on space groups drawn from the empirical marginal distribution during both training and inference. We used the smaller architecture for this ablation to speed up training and inference. Under the same experimental setup, SbCD♢\text{SbCD}^{\diamondsuit} outperforms SbCD♢\text{SbCD}^{\diamondsuit} (w/o SB) across multiple evaluation metrics. This ablation demonstrates the benefit of explicitly modeling crystalline symmetry breaking relative to sampling space groups from the empirical distribution.

W2. Missing recent baselines on MPTS-52 and overstatement in the abstract

In Table 3, we add two recent baselines for MPTS-52, MCFlow (Seong et al., ICLM 2026) [2] and SymADiT (Karmush et al., NeurIPS 2026) [3]. Both baselines are transformer-based and have significantly more parameters than SbCD. Few existing baselines can scale to a dataset as large as MPTS-52. MCFlow did not report thermodynamic stability results, and SymADiT reported results based on 100 DFT calculations. Neither method has released its code or model weights, which makes it difficult to evaluate them using a more feasible MLIP-based protocol. Because of this limitation, we have revised the abstract to state that our thermodynamic stability results are competitive with those of baselines for which results are available.

We also include additional metastability results in Tables 4 and 9, as requested by Reviewer 6eUs, to further support our claims about SbCD’s performance. In both the original and revised versions of the manuscript, our claim of strong performance on MPTS-52 is restricted to thermodynamic stability.

W3+Q2. SbCD's approach to modeling symmetry breaking

In the manuscript, we present the first proof of concept for modeling symmetry breaking in crystal generative models. SbCD breaks the symmetry of crystal structures by transitioning directly to P1, the space group with the lowest symmetry. During generation, SbCD reverses this process starting from the P1 prior (Algorithm 1, page 28). We acknowledge that this approach to modeling symmetry breaking is simplified. As discussed in the conclusion and future work (lines 803–809), the framework of Proposition 4.2 could also incorporate more physically realistic transition pathways.

We have revised the manuscript accordingly. We sincerely thank the Reviewer for their valuable time.

References

[1] (Levy et al.) SymmCD: Symmetry-preserving crystal generation with diffusion models. In ICLR 2025.

[2] (Seong et al.) Multimodal crystal flow: Any-to- any modality generation for unified crystal modeling. In ICLM 2026.

[3] (Karmush et al.) Symadit: Symmetry-aware crystal generation with latent flow matching. In NeurIPS 2026.

Thank you very much for the answers. I am sure this will make the manuscript more accessible.

Dear Reviewer sgvf,

We have updated our revised manuscript with additional results. Particularly, we would like to highlight our recent benchmark using multiple MLIPs.

We evaluate SbCD on MP-20 using LetMat GenBench [1], a recent, highly standardized benchmark for crystal generation. To evaluate thermodynamic stability, Betala et al. [1] compute the energy above the convex hull, EhullE_{\text{hull}}, using an ensemble of advanced MLIPs comprising MACE-MP [3], UMA [4], and Orb-v3 [5]. They also use LetMat-Bulk [2], which contains approximately 55 million structures and provides the broadest coverage of competing phases relaxed under consistent DFT settings. The same database serves as a reference for accurately assessing the novelty of generated structures. As recommended by Betala et al. [1], we generated 25002500 structures and pre-relaxed them to assess their thermodynamic viability better. We report the performance across three runs alongside leaderboard results [1] for the baselines listed in Table 2.

Stable(\mathrm{Stable} (\\%)S.U.N(\mathrm{S.U.N} (\\%) [count]MetaStable(\mathrm{MetaStable} (\\%)M.S.U.N(\mathrm{M.S.U.N} (\\%)
MatterGen0.50.08[2]3615.8
ADiT0.10.00[0]37.11.
DiffCSP1.10.04[1]32.29.3
DiffCSP++0.70.16[4]27.65.3
SymmCD0.50.04[1]19.42.5
SbCD (run1)1.880.36[9]61.486.64
SbCD (run2)1.920.16[4]59.065.96
SbCD (run3)2.080.24[6]62.246.52

We observe that SbCD outperforms all the baselines on multiple metrics. SbCD generates significantly higher numbers of stable and metastable structures and also yields more S.U.N. structures. Note that novelty is assessed against 55 million structures from LetMat-Bulk, so sampling a S.U.N. structure is particularly meaningful. Given the promising results on LetMat GenBench and the availability of our open-source code and model weights, we hope that SbCD, with its principled approach to modeling symmetry breaking, can make meaningful contributions toward addressing real-world materials discovery problems. We have released the CIF files for each run via the anonymous link (https://drive.google.com/file/d/1UIB-MONgBlZsj8CQ-6gNHML5spcagueN/view?usp=sharing).

We sincerely thank the Reviewer again for their valuable feedback, which has improved the quality of our manuscript significantly.

References

[1] (Betala et al.) Lemat-genbench: A unified evaluation framework for crystal generative models. In arXiv:2512.04562

[2] (Siron et al.) Lemat-bulk: aggregating, and de-duplicating quantum chemistry materials databases. In arXiv:2511.05178

[3] (Batatia et al.) A foundation model for atomistic materials chemistry. In arXiv:2401.00096

[4] (Wood et al.) UMA: A family of universal models for atoms. In NeurIPS 2025.

[5] (Rhodes et al.) Orb-v3: atomistic simulation at scale. In arXiv:2504.06231

Reviewer_6eUsOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors addressed an important problem in material design: symmetry-aware crystal generation. Existing models on symmetry-aware crystal generation like SlayerGen, SGEquiDiff, SymmCD, and DiffCSP++ pick the space group first, usually by sampling it from the training-data frequencies, and then generate the structure within that fixed group. So during the whole sampling/diffusion process from noise to crystal, the space group never changes. However, the authors rightly pointed out, space group of a noisy structure, intermediate structure and final target structure cant be same. Model should learn the space group that none of prior work addresses.

Hence, the authors solved that issue by a Markovian Jump diffusion approach, where they make the space group part of the diffusion process. The forward (noising) process lowers a crystal’s symmetry by jumping to the lowest-symmetry group, P1, at a random time. During sampling, it starts from P1 and jumps to a learned space group while denoising the rest of the structure.

Critical strengths
  1. Overall, the paper is very well written and the problem statement is well motivated.

  2. The idea of Markovian Jump diffusion and learning the space group as part of generation is novel and interesting. Authors proposes clear theoretical proof of their work and gives one training objective covering all crystal parts, with closed-form noising and denoising steps.

  3. Authors provide extensive empirical results evaluating their model against popular benchmarks and using popular datasets like MP-20 and MPTS-52

Critical weaknesses
  1. Some of the recent baselines on de novo crystal material generation are missing (not cited and not evaluated). These models include MatterGen, ADiT, UniMat, CrysLLMGen, and LLM-based models such as FlowLLM. Several of these report strong S.U.N. and stability results. It would be good if the authors incorporated their results as well in their de novo crystal material generation baselines.

  2. All stability and S.U.N. results rely on a single MLIP i.e., CHGNet. The authors did not evaluate on DFT or any other MLIP. Also, many prior works report meta-stability as well, which is not included.

  3. Table 8 varies only the number of sampling runs, not independently trained models. Baselines are single runs, and the margin over SymmCD on S.U.N. is small (7.15% vs. 6.86%). Results over multiple training seeds with confidence intervals are needed.

Questions for authors
  1. SbCD’s structural validity (89.80%) is well below DiffCSP (97.43%), DiffCSP++ (99.44%), and SGFM (99.87%), and the ablation without symmetry breaking drops to 78.22%. What causes this? An analysis of which invalid structures arise (by space group or Wyckoff position) would help.

  2. Recent models such as CrysLLMGen report higher S.U.N. on MP-20 than SbCD’s 7.15%. Could the authors include these models and discuss why SbCD does not lead on S.U.N.?

  3. The authors did not provide the code. I would love to look at the code and try to reproduce the results.

5 comments

We thank the Reviewer for their thoughtful and detailed feedback. We address the raised weaknesses (W) and questions (Q) as follows.

W1+Q2: Missing baselines, SbCD versus CrysLLMGen on thermodynamic stability.

We thank the reviewer for suggesting these baselines. Following Khastagir et al. [1], we have included the additional baselines in Table 2 and discussed them in the revised manuscript. One plausible explanation for the performance of CrysLLMGen (Khastagir et al., NeurIPS 2025) [1] is that it combines the strengths of LLMs and diffusion models. The LLM handles discrete atom types and generates initial structures, while the diffusion model refines the continuous lattice parameters and atomic coordinates (Khastagir et al.) [1]. In addition, large-scale pretraining may also capture chemical priors beyond those available in the MP-20 fine-tuning set, making LLM-generated structures a more informative starting point for refinement than pure noise. We therefore consider it a natural next step to use LLM-generated structures as inputs to SbCD and test whether our symmetry-breaking approach further improves crystal validity and stability. We have included this discussion in the revised manuscript.

W2+Q3: DFT, other MLIPs, meta-stability, and code.

Due to our lab constraints, we can not obtain a VASP license and perform expensive DFT calculations. We were also unable to complete evaluations with other advanced MLIPs within the rebuttal period. We acknowledge this limitation in the conclusion.

Nevertheless, we believe that CHGNET provides a reliable estimate of the relative performance among different methods. Specifically, Levy et al. [2] (Table 2) show that DiffCSP outperforms FlowMM and CDVAE (DiffCSP > FlowMM > CDVAE) according to the CHGNET-based S.U.N. metric. Meanwhile, Miller et al. [3] (Table 2) report a consistent performance ranking, i.e., DiffCSP > FlowMM > CDVAE, when evaluated using DFT-based calculations.

We also include additional metastability results in Tables 4 and 9 to further support our claims about SbCD’s performance.

We have therefore released our code and the full model weights for our best checkpoints to date to support community adoption, benchmarking, and further development. Link: https://anonymous.4open.science/r/submission_iclr27-F340/README.md

W3: Independently trained models with different seeds.

We have updated Table 9 to report results for three models trained independently with different random seeds. We performed three sampling runs for each trained model. Overall, performance is consistent across sampling runs for each model. The model trained with seed 1000 yields slightly better results and outperforms the state-of-the-art CrysLLMGen (Khastagir et al., NeurIPS 2025) [1] on both metastability (64.427 for SbCD vs. 62.02 for CrysLLMGen) and M.S.U.N. (36.290 for SbCD vs. 35.94 for CrysLLMGen).

Q1. Structural validity compared with DiffCSP, DiffCSP++, and SGFM; comparison with the SbCD variant without symmetry breaking; and performance by space group.

We hypothesize that these baselines achieve higher structural validity because they use unit-cell fractional coordinates, as indicated in Table 2, which allow them to model interatomic distances directly. In contrast, SbCD, as well as SymmCD, employs fractional coordinates from the smallest asymmetric unit. This representation is memory efficient, but SbCD variants do not directly model all interatomic distances in the full unit cell. When SbCD variants expand the asymmetric unit into the full unit cell using learned site symmetries, they tend to generate denser atomic structures, which may affect structural validity checks. We discuss this effect in the ablation study on the symmetry-breaking time window (lines 475–489).

Under the same experimental setup, SbCD♢\text{SbCD}^{\diamondsuit} achieves higher structural validity than its variant without symmetry breaking, SbCD♢\text{SbCD}^{\diamondsuit} (w/o SB). The variant explicitly conditions on empirical space groups during both training and inference. This ablation demonstrates the benefit of modeling symmetry breaking. In Appendix E.4, we analyze the learned space-group distributions and report performance for the 10 most representative space groups among crystals generated on MP-20.

We have revised the manuscript accordingly. We sincerely thank the Reviewer for their valuable time.

References

[1] (Khastagir et al.) LLM meets diffusion: A hybrid framework for crystal material generation. In NeurIPS 2025.

[2] (Levy et al.) SymmCD: Symmetry-preserving crystal generation with diffusion models. In ICLR 2025.

[3] (Miller et al.) FlowMM: Generating materials with Riemannian flow matching. In ICML 2024.

Thank you for the detailed response, for adding the baselines, and for sharing the codebase.

W1/Q2 (baselines). In the revised Table 2, CrysLLMGen outperforms SbCD on both stability and S.U.N. For completeness, I would also suggest including ADiT, a strong recent de novo generation baseline.

Regarding the explanation for CrysLLMGen’s performance, I agree that combining the strengths of LLMs (for discrete atom types) and diffusion models (for continuous lattice parameters and coordinates) is a plausible factor. However, the argument about large-scale pretraining should be stated more carefully. To my understanding, CrysLLMGen trains its crystal-generation components on MP-20 alone, with no additional materials data. Any chemical prior beyond MP-20 would therefore come only from the general-purpose pretraining of the base LLM, not from materials-specific pretraining. I would suggest the authors clarify this point in the revised discussion.

W2 (stability evaluation). Thank you for adding the metastability results. However, these are currently reported only for MPTS-52. Since MP-20 is the main benchmark and most prior works report metastability on it, I would suggest also including MP-20 metastability in Table 2.

I understand the constraints on DFT calculations. However, I am not fully convinced by the ranking argument. The agreement between CHGNet and DFT holds for DiffCSP, FlowMM and CDVAE, which differ by large margins, whereas SbCD’s advantage over SymmCD on MP-20 S.U.N. is small (7.15% vs. 6.86%). At such margins, the energy errors of a single MLIP could change the ranking. I would encourage the authors to evaluate their models with LeMat-GenBench (Betala et al., 2025), which estimates stability using an ensemble of universal MLIPs and is therefore less sensitive to the errors of any single potential. Even a comparison of SbCD and SymmCD on around 2,500 generated structures each (or fewer, given time constraints), reported with confidence intervals, would considerably strengthen the stability claims.

We thank the Reviewer for the prompt responses.

W1/Q2 (baselines)

We initially used the baselines from CrysLLMGen, which achieved the best results so far on MP20. We will add ADiT (with the MP-20-trained-only version for a fair comparison) in the revised manuscript as requested by the Reviewer.

In our comparison to CrysLLMGen, we meant that the general-purpose pretraining of LLMs themselves can capture chemical priors; importantly, we can not verify which sources of chemical knowledge these LLMs were pretrained on. We did not mean the pretraining of CrysLLMGen, and we agree that CrysLLMGen was only trained on the MP-20 dataset. We will make this explicit in the revised manuscript.

W2 (stability evaluation)

We updated the MP-20 meta-stability results and compared them with the state-of-the-art baseline CrysLLMGen in Table 9. We will update Table 2 for baselines with known metastability results in the revised manuscript.

We thank the Reviewer for the suggested benchmark, and we agree that this will make a fairer and more accurate evaluation. In the rest of the rebuttal, we will try to implement the LeMat-GenBench benchmark (Betala et al., 2025) and update the results.

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

The authors present SbCD, which is a crystal generative model which enables joint diffusion of crystal attributes that have directional constraints (space groups, Wyckoff positions, and atom positions) via jump diffusion. They derive discrete noising processes specifically for space group and Wyckoff position transitions and correspondingly adapt the continuous noising processes over crystal lattices. They show competitive performance compared to previous methods and ablations demonstrating the advantage of symmetry breaking during diffusion.

Critical strengths
  • The authors resolve a seeming dichotomy in the crystal generative modeling literature between enforcing space group constraints and jointly generating all crystal attributes
Critical weaknesses
  • Empirical performance is only state-of-the-art on 1 metric on MP20, while MPTS-52 experiments have only 2 baselines so it's not clear whether it is state-of-the-art on that dataset
  • Stability claims are weak given that they are done with CHGNet, which does not have gold standard accuracy amongst machine learning interatomic potentials anymore. Using a newer MLIP (and ideally multiple models) would strengthen claims.
Questions for authors
  • How are fractional coordinates generated if not with dynamic Wyckoff symmetry constraints?
4 comments

We thank the Reviewer for their constructive feedback. We address the raised weaknesses (W) and questions (Q) as follows.

W1: Missing baselines on MPTS-52.

In Table 3, we add two new baselines for MPTS-52, MCFlow (Seong et al., ICLM 2026) [1] and SymADiT (Karmush et al., NeurIPS 2026) [2]. Note that both baselines are transformer-based and have significantly larger model sizes than SbCD. We note that few existing baselines can scale to a dataset as large as MPTS-52. MCFlow did not report thermodynamic stability results, and SymADiT reported results for 100 DFT calculations. Neither method has released its code or model weights, which makes it difficult to benchmark these methods using a more feasible MLIP-based protocol. Because of this limitation, we have revised the abstract to state only that our results are competitive with those of baselines for which results are available.

We also include additional metastability results in Tables 4 and 9, as requested by Reviewer 6eUs, to further support our claims about SbCD’s performance.

W2: DFT & multiple MLIP evaluations.

Due to our lab constraints, we can not obtain a VASP license and perform expensive DFT calculations. We were also unable to complete evaluations with other advanced MLIPs within the rebuttal period. We acknowledge this limitation in the conclusion.

Nevertheless, we believe that CHGNET provides a reliable estimate of the relative performance among different methods. Specifically, Levy et al. [3] (Table 2) show that DiffCSP outperforms FlowMM and CDVAE (DiffCSP > FlowMM > CDVAE) according to the CHGNET-based S.U.N. metric. Meanwhile, Miller et al. [4] (Table 2) report a consistent performance ranking, i.e., DiffCSP > FlowMM > CDVAE, when evaluated using DFT-based calculations.

We have therefore released our code and the full model weights for our best checkpoints to date to support community adoption, benchmarking, and further development. Link: https://anonymous.4open.science/r/submission_iclr27-F340/README.md

Q1: Fractional coordinate generation without symmetry constraints.

In Algorithm 1 (page 28, line 1509) and Algorithm 2 (page 31, line 1777), we provide the sampling and training of fractional coordinates, which rely on DiffCSP (Jiao et al.) [5]. During training, we use a score-matching objective with a wrapped normal distribution to account for the periodicity of fractional coordinates. During sampling, we draw the initial fractional coordinates from a periodic translation-invariant uniform distribution U(0,1)\mathcal{U}(0,1) and use the same predictor-corrector sampling framework as DiffCSP.

We have revised the manuscript accordingly. We sincerely thank the Reviewer for their valuable time.

References

[1] (Seong et al.) Multimodal crystal flow: Any-to- any modality generation for unified crystal modeling. In ICLM 2026.

[2] (Karmush et al.) Symadit: Symmetry-aware crystal generation with latent flow matching. In NeurIPS 2026.

[3] (Levy et al.) SymmCD: Symmetry-preserving crystal generation with diffusion models. In ICLR 2025.

[4] (Miller et al.) FlowMM: Generating materials with Riemannian flow matching. In ICML 2024.

[5] (Jiao et al.) Crystal Structure Prediction by Joint Equivariant Diffusion. In NeurIPS 2023.

AuthorsOpenReview ↗

Benchmarking with multiple MLIPs on LetMat GenBench

Dear Reviewer PgKR,

We have finalized our benchmark using multiple MLIPs to address W2. We evaluate SbCD on MP-20 using LetMat GenBench [1], a recent, highly standardized benchmark for crystal generation. To evaluate thermodynamic stability, Betala et al. [1] compute the energy above the convex hull, EhullE_{\text{hull}}, using an ensemble of advanced MLIPs comprising MACE-MP [3], UMA [4], and Orb-v3 [5]. They also use LetMat-Bulk [2], which contains approximately 55 million structures and provides the broadest coverage of competing phases relaxed under consistent DFT settings. The same database serves as a reference for accurately assessing the novelty of generated structures. As recommended by Betala et al. [1], we generated 25002500 structures and pre-relaxed them to assess their thermodynamic viability better. We report the performance across three runs alongside leaderboard results [1] for the baselines listed in Table 2.

Stable(\mathrm{Stable} (\\%)S.U.N(\mathrm{S.U.N} (\\%) [count]MetaStable(\mathrm{MetaStable} (\\%)M.S.U.N(\mathrm{M.S.U.N} (\\%)
MatterGen0.50.08[2]3615.8
ADiT0.10.00[0]37.11.
DiffCSP1.10.04[1]32.29.3
DiffCSP++0.70.16[4]27.65.3
SymmCD0.50.04[1]19.42.5
SbCD (run1)1.880.36[9]61.486.64
SbCD (run2)1.920.16[4]59.065.96
SbCD (run3)2.080.24[6]62.246.52

We observe that SbCD outperforms all the baselines on multiple metrics. SbCD generates significantly higher numbers of stable and metastable structures and also yields more S.U.N. structures. Note that novelty is assessed against 55 million structures from LetMat-Bulk, so sampling a S.U.N. structure is particularly meaningful. Given the promising results on LetMat GenBench and the availability of our open-source code and model weights, we hope that SbCD, with its principled approach to modeling symmetry breaking, can make meaningful contributions toward addressing real-world materials discovery problems.

We have released the CIF files for each run via the anonymous link (https://drive.google.com/file/d/1UIB-MONgBlZsj8CQ-6gNHML5spcagueN/view?usp=sharing). We have also updated our revised manuscript. We sincerely thank the reviewer again for their helpful feedback, which contributed to significant improvements in our manuscript.

References

[1] (Betala et al.) Lemat-genbench: A unified evaluation framework for crystal generative models. In arXiv:2512.04562

[2] (Siron et al.) Lemat-bulk: aggregating, and de-duplicating quantum chemistry materials databases. In arXiv:2511.05178

[3] (Batatia et al.) A foundation model for atomistic materials chemistry. In arXiv:2401.00096

[4] (Wood et al.) UMA: A family of universal models for atoms. In NeurIPS 2025.

[5] (Rhodes et al.) Orb-v3: atomistic simulation at scale. In arXiv:2504.06231

I thank the authors for the response. Given that the two new baselines on MPTS-52 were not re-run (so most of their metrics are empty) and other models like DiffCSP(++) were not benchmarked, I maintain my concern about the MPTS-52 claims. I appreciate conducting the additional experiments with an ensemble of MLIPs, which adds robustness of the results. However, the claim that SbCD outperforms all baselines seems overstated; Stable and Metastable rate are not as meaningful as SUN and MSUN. SbCD underperforms at MSUN, and the SUN rates are reported based on a very small number of stable structures (fewer than 10 out of 2500) without any error bars, which casts large doubt about the statistical significance.

Overall I still think the paper is interesting and novel from a methodological point of view. I maintain my score.

We thank the Reviewer for the prompt response.

In our rebuttal, we acknowledge that neither of the two new baselines has released code; therefore, we could not benchmark against them. We tried to include all baselines that can scale to MPTS-52; however, DiffCSP++ was not trained on this dataset, and no available paper reports its performance on MPTS-52.

On the new benchmark LetMat GenBench, novelty is assessed over ~5 million structures from LetMat-Bulk [1], a unified dataset combining Materials Project, OQMD, and Alexandria. Thus, we think that generating a single S.U.N. structure is particularly meaningful. On the leaderboard ([2], Table 1), only a single model, PLaID++, a state-of-the-art RL property-guided model and different from our setting (no guidance/unconditional), can generate more than 10 S.U.N. structures (22 S.U.N. structures). Thus, we think LetMat GenBench provides a meaningful evaluation benchmark for real-world material discovery problems. We also reported three runs to evaluate our model's robustness.

We hope these points address the Reviewer's remaining concerns. We thank the reviewer again for their helpful and timely feedback.

References

[1] (Siron et al.) Lemat-bulk: aggregating, and de-duplicating quantum chemistry materials databases. In arXiv:2511.05178

[2] (Betala et al.) Lemat-genbench: A unified evaluation framework for crystal generative models. In arXiv:2512.04562

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.