acceptodds
Under review as a conference paper at ICLR 2027

CrystalAR: Autoregressive Modeling for Symmetric Crystal Generation

11 views

Abstract

Deep generative modeling has achieved remarkable progress in crystal generation. Given the rich symmetries inherent in crystalline structures, recent work has increasingly sought to explicitly incorporate space-group symmetry into generative models, enabling more effective exploration of novel symmetry templates. Yet introducing symmetry templates substantially increases the combinatorial complexity of stoichiometric relationships between atomic species and crystallographic sites, while creating a fundamental tension between flexible exploration of discrete crystallographic templates and faithful realization of these templates as continuous three-dimensional structures. To address these challenges, we propose **CrystalAR**, a unified autoregressive framework that models Wyckoff symmetry templates and complete crystal structures within a single generative hierarchy. Specifically, CrystalAR reformulates continuous coordinate generation as a symmetry-conditioned hierarchy of discrete spatial localization decisions, autoregressively refining the positions of all symmetry-inequivalent sites across resolution levels while jointly modeling their spatial configuration within each level. This discrete hierarchical formulation further enables constraint-aware search over both crystallographic templates and geometric realizations. As a result, CrystalAR achieves competitive retrieval-free crystal structure prediction and de novo generation performance comparable to state-of-the-art models, while retaining strong symmetry-template novelty. To our knowledge, CrystalAR is the first symmetry-native approach to simultaneously demonstrate these capabilities, providing a promising alternative paradigm for symmetry-preserving crystal generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystal Generation, Autoregressive Modeling, Symmetry

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Qgx5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce an autoregressive framework that hierarchically generates crystals from discrete symmetry configurations to three-dimensional geometry. The main contribution is an iterative coordinate refinement, achieving very competitive results in crystal generation benchmarks.

Critical strengths

Very promising benchmarking results. Interesting and highly relevant new contributions regarding the iterative refinement.

Critical weaknesses

The discussion regarding what is actually new here should be improved (representation, model, refinement, ...).

The experiments section requires improvement and clarification:

  • Most important point: The results of the SV model on LeMat-GenBench should also be included in Table 1. The non-SV version gets a really strong prior with the information on (G, M, N, K). The tuple restricts the possible templates (sometimes only one is possible), which makes the non-SV Version not completely template-free.
  • The metrics in Tables 1 and 2 are only explained in the appendix (which is not mentioned in the main text).
  • Novelty metrics should also be compared with respect to the training data, not only LeMat-Bulk
  • The bold and underline labels in Table 1 are misleading (bold value should be the best in the whole column)
  • The comparison in the CSP task is not fair, as some models (including theirs) get the space group as input, which is an advantage, and others don't
  • CrysBFN should be included in the CSP, as they have better results than the mentioned models in CSP on MP-20
Questions for authors

Please add a more detailed conceptual comparison to other autoregressive models, e.g. WyFormer and CrystalFormer, to discuss the novel contributions. The approach seems to be quite expensive. Would it be possible to discuss a runtime comparison with other models? What is the performance for CSP and DNG with ancestral sampling (no beam search)? What is the performance for CSP and DNG without the geometric consistency term (lambda = 0 in Eq. 7), and additionally without pruning at d_prune? What are the values for d_min and d_prune, and where do they come from?

6 comments
AuthorsOpenReview ↗

Replying to Reviewer Qgx5

Thank you for your positive feedback and constructive comments. We believe that all the concerns raised can be fully addressed. Please find our detailed responses below:

W1 SV-variant Model Issue

Note that the results in this table are obtained using ancestral sampling without any search.

VariantAttemptsValid (%)Unique (%)Novel (%)E_f (eV/atom)E_hull (eV/atom)RMSD (Å)Stable (%)S.U.N.Meta (%)MSUN (%)
CrystalAR2,50064.4062.2418.76−0.451 ± 1.3860.457 ± 1.1370.382 ± 0.6805.44 (136)536.68 (917)2.12 (53)
CrystalAR (only M)2,50064.0862.6019.16−0.423 ± 1.3560.475 ± 1.0890.428 ± 0.9854.88 (122)436.12 (903)1.92 (48)

To investigate whether our non-SV variant benefits from excessive size priors, we conducted an additional experiment in which CrystalAR is conditioned only on the number of symmetry-inequivalent sites (M), without explicitly specifying the other size variables. As shown above, the results remain highly consistent with our original setting across validity (64.08% vs. 64.40%), uniqueness (62.60% vs. 62.24%), novelty (19.16% vs. 18.76%), and stability (4.88% vs. 5.44%). This demonstrates that the additional size priors are not essential to CrystalAR's generation performance.

Importantly, conditioning on the number of sites is conceptually analogous to conditioning on the total atom count (N), as commonly practiced in generic crystal generative models. In symmetry-native generation, (M) specifies only the number of symmetry-inequivalent sites, not their Wyckoff assignments, elemental identities, or geometric configurations, all of which must still be generated by the model. Therefore, providing (M) does not amount to supplying a crystallographic template, and CrystalAR remains template-free in the sense that no specific Wyckoff configuration is given or retrieved.

Overall, these additional results alleviate the concern that CrystalAR's performance relies heavily on informative size priors and further support the effectiveness of our symmetry-native generative formulation.

W2: Clarification of Evaluation Metrics

Thank you for pointing out these issues. Due to space constraints, we originally provided detailed explanations of the evaluation metrics in the appendix. To improve clarity and readability, we have revised the manuscript to include an additional paragraph introducing the key evaluation metrics in the main text. We have also expanded the table captions to clarify the relevant metric definitions, making the results easier to interpret without referring to the appendix.

W3 Novelty Evaluation Against the Training Data

MethodNovel vs. MP-20 (%)Novel vs. LeMat-Bulk (%)S.U.N. Count ↑S.U.N. Rate (%) ↑
CrystalAR (Ours)60.042.8321.28
FlowMM84.265.0311.24
DiffCSP86.766.2251.00
CrystalAR-SV (Ours)66.948.9220.88
DiffCSP++86.662.0160.64
SymmCD64.447.290.36
WyFormer-DiffCSP++ (1,000)86.165.280.80
CrystalFormer (1,000)51.031.770.70
SGEquiDiff78.258.030.12

The results reveal several important observations. First, the relative trends in structural novelty remain consistent across the two reference datasets, indicating that our original evaluation against LeMat-Bulk provides a meaningful assessment of structural novelty. Second, while some baseline methods achieve higher raw novelty percentages, higher novelty does not necessarily translate into better stable material discovery. Notably, CrystalAR achieves the highest S.U.N. discovery rate against the MP-20 training set (32 structures, 1.28%), outperforming DiffCSP (25, 1.00%) and FlowMM (31, 1.24%), despite their higher raw novelty rates. This highlights the importance of jointly considering stability, uniqueness, and novelty rather than optimizing structural novelty in isolation.

Importantly, CrystalAR does not aim to maximize novelty at the expense of physical plausibility. Instead, our symmetry-native formulation seeks to generate crystallographically consistent structures while maintaining competitive discovery performance. The additional results demonstrate that CrystalAR achieves a favorable balance between structural exploration and stable material discovery, rather than merely producing a larger fraction of structures distant from the reference dataset. Note that here reported S.U.N. results are evaluated with respect to the MP-20 training set, which serves as the reference dataset for assessing structural novelty.

AuthorsOpenReview ↗

Further Replies to Reviewer Qgx5

W5 Fair Comparison and Space-Group Consistency in CSP

(1) Clarification of CSP settings We thank the reviewer for raising this important question regarding the fairness of CSP comparisons. Previous symmetry-native generative models either cannot directly perform CSP due to limitations in their probabilistic factorization or rely on external template retrieval methods, such as CSPML, to obtain the required crystallographic information. In the latter case, retrieving a crystal template effectively provides both the space group GG and the Wyckoff configuration WW, along with additional structural priors. In contrast, CrystalAR addresses a more challenging template-free, space-group-conditioned CSP setting, where only the chemical composition and space group GG are provided, while the Wyckoff configuration WW and crystal geometry must be predicted without external template retrieval. We consider conditioning on GG a well-motivated setting, as space-group information is a fundamental conditioning variable commonly adopted by symmetry-native generative models. In our original comparison, symmetry-native approaches were highlighted in red to distinguish them from generic crystal generation models.

(2) Fair comparison under unified conditions & space-group consistency evaluation. To provide a more comprehensive and fair evaluation, we have conducted additional experiments in which generic CSP baselines are retrained with space-group conditioning, following the soft-conditioning strategy introduced by MatterGen. This establishes a more unified comparison in which all evaluated models receive the same chemical composition and ground-truth space group, allowing us to directly assess their ability to exploit crystallographic symmetry information. Furthermore, we introduce Space-Group Accuracy (SG Acc.), which measures the proportion of generated structures whose crystallographic space groups agree with those of the ground-truth structures, as identified by symmetry analysis. Related space-group consistency evaluations have also been adopted in previous studies, including CrystaLLM~\cite[1].

Importantly, SG Acc. and the conventional CSP Match Rate evaluate two distinct and complementary aspects of structure prediction. Match Rate measures geometric agreement between predicted and reference structures under predefined tolerances, whereas SG Acc. evaluates whether the predicted structures preserve the target crystallographic symmetry. The relatively permissive tolerances commonly used in CSP benchmarks may identify two geometrically similar structures as matching even when they belong to different space groups (See Page 32 Appendix Figure 10 in the newly revised manuscript). This discrepancy can persist even when relatively generous symmetry-detection tolerances are applied. Such differences are physically meaningful: symmetry breaking can fundamentally alter the allowed physical properties of a material~\cite[2], as exemplified by structural phase transitions associated with the emergence of ferroelectric polarization. Therefore, a high Match Rate alone does not necessarily guarantee the correct recovery of crystallographic symmetry, motivating the inclusion of SG Acc. as an independent evaluation metric.

The updated CSP results demonstrate a distinctive advantage of CrystalAR: our model achieves nearly perfect space-group accuracy through its symmetry-native autoregressive generation mechanism, while maintaining a Match Rate comparable to state-of-the-art CSP models. In contrast, the competing methods do not simultaneously achieve these two properties, even when explicitly conditioned on the same ground-truth space groups.

Composition & SG

MethodSC-MR, 0.1 ÅSG Acc, 0.1 Å
CrystalAR62.4798.12
DiffCSP-SG-CFG54.1163.30
DiffCSP-SG53.1662.99
FlowMM-SG34.5041.43
CrystalFormer27.5487.78
CDVAE-SG15.0220.10
WyFormer (no geometry for 7,611/9,046 targets in slot 0)7.6315.68
OMatG-SG1.353.55
CrystalFlow-SG1.233.59

SG-MR measures the proportion of predicted structures that both match the ground-truth structures under the MR criterion and preserve their ground-truth space-group symmetry.

Composition Only

MethodSC-MR, 0.1 ÅSG Acc, 0.1 Å
Uni-3DAR (released checkpoint, rescored)52.1459.22
CrystalAR, 1 of 20 by a frozen rule51.4665.11
CrystalAR, first candidate49.2064.66
DiffCSP41.8349.03
FlowMM34.5740.50
CDVAE10.0713.94
CrystalFlow1.373.53
OMatG1.283.35
CrysBFN35.6741.66

We have updated CSP section in the manuscript accordingly.

[1] Crystal structure generation with autoregressive large language modeling. Nature Communications 2024

[2] Physical Properties of Crystals: Their Representation by Tensors and Matrices 1957

AuthorsOpenReview ↗

Further Replies to Reviewer Qgx5

W6 CrysBFN

We have included CrysBFN as an additional baseline in the revised manuscript. Please refer to Table 3 for the corresponding experimental results.

Q1 Conceptual Comparison to Prior AR approaches

We have added an additional page to the revised manuscript (page 9-10) to provide a detailed comparison of our contributions with prior autoregressive approaches. We believe this analysis clarifies the fundamental differences between CrystalAR and existing methods, highlighting its key methodological innovations and advantages. We appreciate the reviewer for raising this important point and will further refine the discussion in the camera-ready version.

Q2 Runtime Comparison

Method / SettingWall Time (s/attempt)Wall Time (s/generated structure)Generated / AttemptsBatch Size
CrystalAR, Ancestral Sampling (No Search)0.0690.0692,500 / 2,50032
CrystalAR, SGBS Q = 64 (Including Ancestral Stage)0.4040.768202 / 38432
DiffCSP, 1,000 Timesteps (2,000 Network Evaluations)0.6790.6791,500 / 1,500500
SymmCD, 1,000 Steps0.6180.6191,497 / 1,500500
CrystalFormer, K = 1, T = 10.0360.0362,400 / 2,400100

As shown in the table, CrystalAR is computationally efficient and does not necessarily incur substantial inference overhead. With pure ancestral sampling, CrystalAR requires only 0.069 seconds per structure, substantially faster than DiffCSP (0.679 s) and SymmCD (0.619 s). Even with our simulation-guided beam search (SGBS, Q = 64), CrystalAR requires only 0.404 seconds per attempt, with an average of 0.768 seconds per successfully generated structure, which remains comparable to diffusion-based approaches. These results demonstrate that CrystalAR offers a flexible trade-off between generation efficiency and structural quality, allowing users to choose between fast ancestral sampling and more extensive geometry-guided search depending on their computational budget.

Q3 CSP and DNG Performance with Ancestral Sampling/ Q4 Without geometric consistency & d_prune

We thank the reviewer for this important question. We have conducted additional experiments to evaluate CrystalAR under pure ancestral sampling without beam search, together with detailed ablations of our search strategy.

As shown below, CrystalAR achieves 54.12% MR@1 with pure ancestral sampling, demonstrating that the autoregressive model itself possesses strong structure prediction capability without relying on search. Incorporating our hierarchical search strategy further improves MR@1 to 62.92%, yielding a substantial 8.80 percentage-point improvement (p=7.2×10−127p = 7.2 \times 10^{-127}).

Importantly, our search mechanism goes beyond simply increasing the beam size. When coordinate candidates are ranked solely by autoregressive likelihood (λ=0\lambda=0), MR@1 decreases to 60.92%, demonstrating the additional value of geometry-aware ranking. Meanwhile, reducing the coordinate beam size to 1 or 4 produces relatively small performance differences, suggesting that the improvement is not merely attributable to larger coordinate beams.

These results demonstrate that hierarchical search and geometry-aware candidate selection provide complementary benefits to our autoregressive modeling, enabling more effective exploration of plausible crystal configurations rather than relying solely on ancestral sampling.

SettingMR@1 (%)Δ vs. Defaultp-value
Default (Ours)62.92——
Ancestral sampling (no search)54.12−8.807.2e-127
Likelihood-only ranking (λ=0\lambda=0)60.92−2.004.1e-23
Without 0.5 Å contact mask62.96+0.030.508
Coordinate beam size = 162.61−0.310.010
Coordinate beam size = 462.79−0.130.201
ArmMR@1 %Δ vs default95% CI (paired)p
Default62.92———
Ancestral (no search anywhere)54.12−8.80—7.2e-127
λ = 0 (coordinates ranked by log p only)60.92−2.00—4.1e-23
λ = 0 and no d_prune (mask and cap removed)59.40−3.53[−3.98, −3.07]1.9e-56
No d_prune mask only (cap kept)62.96+0.03—0.508
No clearance cap only (d_min counterpart removed)62.68−0.24[−0.53, +0.05]0.117
Coordinate beam 162.61−0.31—0.010
Coordinate beam 462.79−0.13—0.201
Sampling strategyValid (%)Unique (%)Novel (%)E_f (eV/atom)E_hull (eV/atom)RMSD (Å)Stable (% (n))S.U.N. (% (n))Meta (% (n))M.S.U.N. (% (n))
Ancestral sampling (no search)64.4062.2418.76−0.45 ± 1.390.46 ± 1.140.38 ± 0.685.44 (136)0.20 (5)36.68 (917)2.12 (53)
SGBS (default)91.4489.0842.76−0.23 ± 1.270.71 ± 1.080.77 ± 1.256.00 (150)0.28 (7)39.24 (981)2.72 (68)
Reviewer_idsHOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CrystalAR is a single decoder-only Transformer that generates a crystal autoregressively: first the space group, then an element-decorated Wyckoff template, then the lattice, and finally the free Wyckoff coordinates. The coordinates are discretized coarse-to-fine by repeatedly bisecting each Wyckoff manifold, and all inequivalent sites are refined in parallel within each level. Because the generation is discrete, the authors can add a composition-constrained template search for template-free crystal structure prediction (CSP) and a beam search over coordinates. They evaluate de novo generation on LeMat-GenBench, symmetry and template novelty on MP-20, and CSP on MP-20 and MPTS-52. The paper claim strong template generation together with competitive de novo and template-free CSP performance.

Critical strengths

Improvement over prior autoregressive symmetry-native models. Under identical inputs, CrystalAR more than doubles the single-hypothesis CSP match rate of CrystalFormer and WyFormer on MP-20, and achieves a substantially higher stability rate than CrystalFormer in de novo generation, even without beam search, advancing within this AR model family.

Critical weaknesses

1. The methodological novelty is limited relative to prior work on crystal generation.

  • Wyckoff-based autoregressive generation follows CrystalFormer (Cao et al., 2024) and WyFormer (Kazeev et al., 2025). Hierarchical, octree-based autoregression for crystals has already been explored in Uni-3DAR (Lu et al., 2025; arXiv:2503.16278), which also addresses de novo generation and CSP but is not cited.
  • The contribution therefore appears to lie mainly in combining these ideas. The paper does not clearly state what is new beyond this combination.

2. The comparisons are not matched, so model quality and evaluation setup are hard to separate.

  • In CSP, CrystalAR is given the ground-truth space group, while the diffusion and flow baselines receive only the composition. Its competitive results may partly come from this extra information.
  • The de novo results rely on a beam search that filters candidates by interatomic distance, which the main text does not mention. Without it, validity and novelty drop considerably. The baselines receive no comparable search, and inference cost is not reported.

3. The claim of competitive de novo performance is hard to support due to missing baselines.

  • The paper claims competitiveness beyond symmetry-native methods, but the baselines are mostly earlier models. Recent models such as OMatG (Hoellmer et al., 2025), MCFlow (Seong et al., 2026), and Crystalite (Veljković et al., 2026), are not compared. In addition, CrystalAR's higher stability comes with lower novelty than several baselines.
Questions for authors
  1. Beyond combining Wyckoff-based autoregression with hierarchical coordinate generation, what do you consider the key novel component?
  2. Could you discuss Uni-3DAR in the related work and compare against it?
  3. Could you report CSP results with a predicted space group, and separate the baselines by their inputs?
  4. What is the inference cost of the beam search? How do the baselines perform with a comparable filter and compute budget?
  5. Could you show the trade-off between template novelty and stability across configurations, together with the baselines, to clarify whether CrystalAR is Pareto-efficient?
8 comments
AuthorsOpenReview ↗

Replying to Reviewer idsH

We sincerely appreciate your encouraging feedback and insightful suggestions. We believe that the concerns raised can be satisfactorily addressed, and we provide our point-by-point responses below:

W1Q1 Novelty Issue & W1Q2 Uni-3DAR Discussion

We added a new Section 6, "Comparisons to Prior AR Approaches", which writes the generative factorizations of all four models in a common notation (Eqs. 8–11) and summarizes the differences in Table 4. Those factorizations show that CrystalAR is not a combination of existing pieces; it differs from each prior model in specific terms of the factorization. We see three key novel components:

1. Symmetry-native hierarchical localization, as opposed to Cartesian octree autoregression (vs. Uni-3DAR). Uni-3DAR's factorization (Eq. 11) contains neither the space group GG nor the Wyckoff template TT. It partitions the Cartesian unit cell over all NN atoms, so symmetry can only emerge implicitly and sequence length grows as O(NL)\mathcal{O}(NL). CrystalAR instead performs coarse-to-fine refinement within each Wyckoff manifold fi=ϕwi(ui)\mathbf{f}_i=\phi_{w_i}(\mathbf{u}_i): each site is refined in its own did_i-dimensional free-parameter space with 2di≤82^{d_i}\le 8 children per level, and only the M≤NM\le N symmetry-inequivalent sites are refined. As a result:

  • every generated structure is exactly invariant under GG by construction;
  • the sequence is shorter;
  • the template is an explicit variable that can be generated, conditioned on, or searched.

The effect shows in symmetry fidelity. CrystalAR reaches 98.12% space-group accuracy on MP-20, while Uni-3DAR reaches 59.22%. Uni-3DAR does not support space-group conditioning, so its number is evaluated given composition only.

2. Level-wise joint localization of all sites, as opposed to site-by-site continuous coordinates (vs. CrystalFormer). In CrystalFormer (Eq. 9), each site's coordinates ui\mathbf{u}_i are drawn from a continuous density, interleaved with the template tokens and conditioned only on earlier sites s<i\mathbf{s}_{\lt i}. A site's precise position is therefore fixed before later sites exist, and the lattice is predicted last. CrystalAR (Eq. 8) first completes the template and the lattice. Then, at each resolution level, it predicts the refinement decisions R(ℓ)\mathbf{R}^{(\ell)} of all inequivalent sites jointly, conditioned on the coarser regions R(<ℓ)\mathbf{R}^{(\lt \ell)} of every site, using block-causal attention. Interatomic geometry is thus resolved globally from coarse to fine, under the metric L\mathbf{L} that determines distances. On MP-20 this improves template-free CSP over CrystalFormer as follows:

  • match rate: 29.62% → 62.92% with 1 hypothesis, and 64.56% → 74.55% with 20 hypotheses (Table 3);
  • symmetry-consistent match rate: 27.54% → 62.47%;
  • space-group accuracy: 87.78% → 98.12% (Table 4).

3. One end-to-end likelihood over template, lattice, and geometry, which supports template-free CSP and joint search (vs. WyFormer). WyFormer (Eq. 10) learns only the template term pθ(T∣G)p_\theta(T\mid G) and delegates geometry to a separately trained generator or relaxation qextq_{\mathrm{ext}}. Its template model therefore receives no signal about whether a template can be realized geometrically. In CrystalAR, TT, L\mathbf{L}, and R(1:L)\mathbf{R}^{(1:L)} share one likelihood and one Transformer. Because every variable, including geometry, is discrete, a single decoding procedure can search jointly over templates (constraint-guided template search) and over their geometric realizations (hierarchical exploration; Section 4.3). This makes template-free CSP possible without any retrieved Wyckoff template. The comparison with WyFormer:

  • match rate 62.92% vs. 10.25%;
  • symmetry-consistent match rate 62.47% vs. 7.63%;
  • WyFormer produces no geometry for 7,611 of the 9,046 test targets.

Supporting contribution: symmetry-aware CSP evaluation. We show that the standard match rate cannot tell symmetry-preserving predictions apart from predictions that match approximately but break symmetry. We therefore report a symmetry-consistent match rate (SC-MR) and space-group accuracy (SG Acc), defined in Appendix A.5. Figure 10 in Appendix A.11 gives examples in which every prediction from the space-group-conditioned baselines counts as a match under the standard match rate, yet only CrystalAR recovers the ground-truth space group.

In summary, the novelty is not stacking Wyckoff tokens on top of an octree. It is a symmetry-native discrete parameterization of geometry: each site is refined within its own Wyckoff manifold, all sites are localized jointly at every level, and template and geometry are modeled together in one likelihood. This parameterization yields exact symmetry by construction, a template that can be searched, and native template-free CSP. As Eqs. 8–11 show, none of CrystalFormer, WyFormer, or Uni-3DAR provides these properties.

AuthorsOpenReview ↗

Further Replies to Reviewer idsH

W2Q3 Fair Comparison and Space-Group Consistency in CSP

(1) Clarification of CSP settings We thank the reviewer for raising this important question regarding the fairness of CSP comparisons. Previous symmetry-native generative models either cannot directly perform CSP due to limitations in their probabilistic factorization or rely on external template retrieval methods, such as CSPML, to obtain the required crystallographic information. In the latter case, retrieving a crystal template effectively provides both the space group GG and the Wyckoff configuration WW, along with additional structural priors. In contrast, CrystalAR addresses a more challenging template-free, space-group-conditioned CSP setting, where only the chemical composition and space group GG are provided, while the Wyckoff configuration WW and crystal geometry must be predicted without external template retrieval. We consider conditioning on GG a well-motivated setting, as space-group information is a fundamental conditioning variable commonly adopted by symmetry-native generative models. In our original comparison, symmetry-native approaches were highlighted in red to distinguish them from generic crystal generation models.

(2) Fair comparison under unified conditions & space-group consistency evaluation. To provide a more comprehensive and fair evaluation, we have conducted additional experiments in which generic CSP baselines are retrained with space-group conditioning, following the soft-conditioning strategy introduced by MatterGen. This establishes a more unified comparison in which all evaluated models receive the same chemical composition and ground-truth space group, allowing us to directly assess their ability to exploit crystallographic symmetry information. Furthermore, we introduce Space-Group Accuracy (SG Acc.), which measures the proportion of generated structures whose crystallographic space groups agree with those of the ground-truth structures, as identified by symmetry analysis. Related space-group consistency evaluations have also been adopted in previous studies, including CrystaLLM~\cite[1].

Importantly, SG Acc. and the conventional CSP Match Rate evaluate two distinct and complementary aspects of structure prediction. Match Rate measures geometric agreement between predicted and reference structures under predefined tolerances, whereas SG Acc. evaluates whether the predicted structures preserve the target crystallographic symmetry. The relatively permissive tolerances commonly used in CSP benchmarks may identify two geometrically similar structures as matching even when they belong to different space groups (See Page 32 Appendix Figure 10 in the newly revised manuscript). This discrepancy can persist even when relatively generous symmetry-detection tolerances are applied. Such differences are physically meaningful: symmetry breaking can fundamentally alter the allowed physical properties of a material~\cite[2], as exemplified by structural phase transitions associated with the emergence of ferroelectric polarization. Therefore, a high Match Rate alone does not necessarily guarantee the correct recovery of crystallographic symmetry, motivating the inclusion of SG Acc. as an independent evaluation metric.

The updated CSP results demonstrate a distinctive advantage of CrystalAR: our model achieves nearly perfect space-group accuracy through its symmetry-native autoregressive generation mechanism, while maintaining a Match Rate comparable to state-of-the-art CSP models. In contrast, the competing methods do not simultaneously achieve these two properties, even when explicitly conditioned on the same ground-truth space groups.

Composition & SG

MethodSC-MR, 0.1 ÅSG Acc, 0.1 Å
CrystalAR62.4798.12
DiffCSP-SG-CFG54.1163.30
DiffCSP-SG53.1662.99
FlowMM-SG34.5041.43
CrystalFormer27.5487.78
CDVAE-SG15.0220.10
WyFormer (no geometry for 7,611/9,046 targets in slot 0)7.6315.68
OMatG-SG1.353.55
CrystalFlow-SG1.233.59

SG-MR measures the proportion of predicted structures that both match the ground-truth structures under the MR criterion and preserve their ground-truth space-group symmetry.

Composition Only

MethodSC-MR, 0.1 ÅSG Acc, 0.1 Å
Uni-3DAR (released checkpoint, rescored)52.1459.22
CrystalAR, 1 of 20 by a frozen rule51.4665.11
CrystalAR, first candidate49.2064.66
DiffCSP41.8349.03
FlowMM34.5740.50
CDVAE10.0713.94
CrystalFlow1.373.53
OMatG1.283.35
CrysBFN35.6741.66

We have updated CSP section in the manuscript accordingly.

[1] Crystal structure generation with autoregressive large language modeling. Nature Communications 2024

[2] Physical Properties of Crystals: Their Representation by Tensors and Matrices 1957

AuthorsOpenReview ↗

Further Replies to Reviewer idsH

W2Q4 Inference Cost

Method / SettingWall Time (s/attempt)Wall Time (s/generated structure)Generated / AttemptsBatch Size
CrystalAR, Ancestral Sampling (No Search)0.0690.0692,500 / 2,50032
CrystalAR, SGBS Q = 64 (Including Ancestral Stage)0.4040.768202 / 38432
DiffCSP, 1,000 Timesteps (2,000 Network Evaluations)0.6790.6791,500 / 1,500500
SymmCD, 1,000 Steps0.6180.6191,497 / 1,500500
CrystalFormer, K = 1, T = 10.0360.0362,400 / 2,400100

As shown in the table, CrystalAR is computationally efficient and does not necessarily incur substantial inference overhead. With pure ancestral sampling, CrystalAR requires only 0.069 seconds per structure, substantially faster than DiffCSP (0.679 s) and SymmCD (0.619 s). Even with our simulation-guided beam search (SGBS, Q = 64), CrystalAR requires only 0.404 seconds per attempt, with an average of 0.768 seconds per successfully generated structure, which remains comparable to diffusion-based approaches. These results demonstrate that CrystalAR offers a flexible trade-off between generation efficiency and structural quality, allowing users to choose between fast ancestral sampling and more extensive geometry-guided search depending on their computational budget.

W3 Comparison with recent baselines and the stability–novelty trade-off

Our baseline selection follows two principles: (i) we compare with methods that have publicly available implementations or checkpoints, so that every result can be reproduced under a common evaluation protocol; and (ii) we follow the ICLR 2027 policy on recent and non-peer-reviewed work.

OMatG (Hoellmer et al., 2025). OMatG was already included in our CSP comparison (Table 3: 63.75% MR / 0.0720 RMSE on MP-20, 25.15% / 0.1931 on MPTS-52). In the revision, we have also added OMatG to the de novo generation benchmark (Table 1).

MCFlow (Seong et al., 2026). At the time of our evaluation, we could not find a public implementation or released checkpoints for MCFlow. Without them, we cannot generate samples or evaluate it under the same benchmark setting and validity checks as the other methods. We therefore omit it from the main comparison (see Appendix A.6), consistent with our principle of comparing only with reproducible baselines.

Crystalite (Veljković et al., 2026). To our knowledge, Crystalite is available only on arXiv and has not yet been peer-reviewed. The ICLR 2027 Reviewer Guidelines state that arXiv is not a peer-reviewed venue and that authors are not required to compare with papers available solely on arXiv; reviewers may point out such work, but the lack of such comparisons "cannot be a basis for rejection."

Stability vs. novelty. We agree that CrystalAR trades some structural novelty for stability. We view this as a well-documented, inherent tension in crystal generation and as a choice of operating point, not a flaw of the framework:

  • The trade-off is intrinsic and well documented. Park & Walsh (2026, Nature Machine Intelligence) identify a trade-off between novelty and stability that constrains conventional generative models. Likelihood-based training concentrates samples in well-represented, energetically favorable regions of the training distribution, whereas novel compounds lie in underexplored regions. Addressing this requires dedicated mechanisms such as their RL-based guidance. A lower novelty rate accompanied by higher stability is therefore the expected behavior of a well-calibrated likelihood model, not a deficiency specific to CrystalAR.

  • Joint discovery metrics favor CrystalAR. Novelty alone also counts unstable or implausible structures. In Table 1, CrystalAR achieves the highest stability rate in both settings (13.7% with and 6.0% without pre-relaxation) and the best S.U.N. and metastability rates without pre-relaxation (0.3% and 39.2%). With pre-relaxation, its S.U.N. (0.9%) is second only to PLaID++ (1.0%), which uses a 7B LLM with RL post-training, compared with CrystalAR's 33M parameters.

  • Structural novelty is not the same as crystallographic novelty. Structure-matching novelty can largely come from elemental substitutions of known prototypes (Negishi et al., 2026). On the more stringent template-level novelty (Table 2), symmetry-native generation explores substantially more new Wyckoff configurations: CrystalAR (SV) discovers 268 novel unique templates, versus 48 for DiffCSP and 5 for DiffCSP++, while keeping a realistic P1 ratio (1.94% vs. 2.23% reference).

  • The trade-off is controllable. Sampling atom and site counts from the empirical prior (default CrystalAR) favors stability and S.U.N. (0.28% S.S.U.N.), while explicitly predicting both counts (CrystalAR (SV)) broadens the explored size space and increases template novelty (Table 2).

Reviewer_6euaOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the challenge of combining both flexible exploration of discrete crystallographic templates with the generation of continuous 3D coordinates in crystalline materials. It proposes CrystalAR, a unified autoregressive model that generates element-decorated Wyckoff templates and lattice parameters before refining the positions of the sites across spatial resolutions. They used a discrete hierarchical formulation that jointly predicts refinement decisions for all symmetry-inequivalent sites within each level and recursively narrows their feasible regions with successive levels. The authors evaluate their method on de novo crystal generation based on LeMat-GenBench, and crystal structure prediction (CSP) on MP-20 and MPTS-52, with chemical composition and space group provided for the CSP tasks.

Critical strengths
  1. Hierarchical coordinates in CrystalAR can improve structural validity in de novo generation. In the coordinate representation ablation (Table 7), they show that the hierarchical learned-STOP reference can achieve 90.12% validity compared to 49.12% for full-cell AR, and 68.16% for flat 512-bin coordinates. So these may support refining coordinates within their allowed Wyckoff regions can produce more valid structures. However, this experiment shows a benefit for geometric validity, but does not establish better energetic stability or quality.

  2. CSP results show that feasible-set supervision improves structure recovery. The authors show that feasible set supervision can raise structure match rate from 58.57% (standard cross entropy) to 61.35%, following same template-ranking and coordinate-search settings. Their control experiments also provide evidence that training on choices compatible with target composition can help CSP, beyond effects of extra training or trainable params.

Critical weaknesses
  1. The contribution beyond existing symmetry-aware autoregressive models is insufficiently established. Other methods such as CrystalFormer already generate complete crystals using Wyckoff representations and coordinates. So a key method distinction in CrystalAR is the level-wise coarse-to-fine coordinate refinement with interaction between the symmetry-inequivalent sites. While their Table 7 compares the hierarchy with flat, continuous-coordinate, and full-cell alternatives, there is no matched site-by-site hierarchical baseline. These results support the chosen representation but do not establish the added value of the proposed interaction mechanism, and so this limits the evidence of the paper's claimed contributions.

  2. The main symmetry related benefits are insufficiently established. For CSP, the reported match rates do not isolate the benefit of supplying the target space group, since several other baselines in the CSP comparisons receive only composition (which is the original CSP task). Furthermore, although CrystalAR uses Wyckoff constraints and symmetry expansion, the paper does not report whether the same generated candidate both matches the target structure (within some tolerance of the StructureMatcher) and recovers the correct space group symmetry. It would be useful to see whether CrystalAR leads to better symmetry fidelity given that conditioning. Additionally, in de novo generation, S.S.U.N. requires non-P1 symmetry but does not require an unseen Wyckoff template. The released evaluator likewise does not show the intersection of stable discoveries with unseen templates. The connection central to the paper's motivation is missing, the current results do not establish that exploring new templates lead to more stable discoveries.

Questions for authors
  1. Contribution of the level-wise site interaction. Could the authors compare their level-wise refinement with a site-wise hierarchical baseline with same search budget? This would help isolate the benefits of interactions between sites from the hierarchical discretizations.

  2. Stable and novel crystalline materials discovery. In Table 7, the hierarchical learned-STOP reference gives 4 S.U.N structures compared to 0 for flat 512-bin coordinates, but has a lower overall stable rate across the 2500 attempts (3.52% vs 4.24%). How consistent are these results across independent runs? And among the generated S.U.N structures, how many have Wyckoff templates that are unseen in the reference data, before and after relaxation? And how many retain their supplied space group symmetry after relaxation?

  3. Implementations of the proposed architecture. Which implementation in the released code and checkpoints produced the reported results in the paper?

  4. Fair comparison on CSP task and important evaluations of space group agreement. Appendix A.6 states that CrystalFormer and WyFormer support composition- and space-group-conditioned CSP. Which exact conditioning inputs and decoding settings were used for their results in Table 3? Importantly, can the authors also report the agreement between the supplied space group and the detected space group of the generated structures (up till a tolerance)? It would also be useful to report how often the same candidate both matches the target structure and recovers its space group. This can help assess whether structural recovery in CSP also corresponds to correct space group symmetry.

9 comments
AuthorsOpenReview ↗

Replying to Reviewer 6eua

Thank you for your positive feedback and valuable suggestions. We believe that all the issues raised can be effectively addressed. Our detailed point-by-point responses are provided below.

Question 2

ArmRunValidUniqueNovelE_f (eV/atom)E_hull (eV/atom)RMSD (Å)StableS.U.N.MetaMSUN
Learned STOP reference (Table 7)1 (in Table 7)90.1288.6448.92-0.09 ± 1.290.85 ± 1.180.87 ± 1.1688 (3.52%)433.7665 (2.60%)
290.0888.3648.44-0.05 ± 1.450.85 ± 1.370.84 ± 1.08105 (4.20%)733.2068 (2.72%)
390.3288.6848.96-0.16 ± 1.280.82 ± 1.100.84 ± 1.06114 (4.56%)132.3654 (2.16%)
all 3 runs90.17 ± 0.1388.56 ± 0.1748.77 ± 0.29-0.10 ± 1.340.84 ± 1.220.85 ± 1.10307/7500 (4.09%)1233.11 ± 0.70187/7500 (2.49%)
Flat 512-bin coordinates (Table 7)1 (in Table 7)68.1666.4829.12-0.22 ± 1.380.67 ± 1.200.58 ± 0.93106 (4.24%)029.9261 (2.44%)
267.7266.5227.40-0.27 ± 1.380.61 ± 1.200.55 ± 0.99114 (4.56%)330.9658 (2.32%)
368.7667.6429.52-0.23 ± 1.400.68 ± 1.190.59 ± 0.99126 (5.04%)230.3250 (2.00%)
all 3 runs68.21 ± 0.5266.88 ± 0.6628.68 ± 1.13-0.24 ± 1.390.65 ± 1.190.58 ± 0.97346/7500 (4.61%)530.40 ± 0.52169/7500 (2.25%)

Result Consistency To assess the robustness of our results, we conducted three independent runs for both hierarchical refinement (Learned STOP) and flat 512-bin coordinate generation, with 2,500 samples per run. The results demonstrate strong consistency across independent runs, particularly in validity, uniqueness, and novelty.

Model / SettingS.U.N.S.U.N. with templates unseen in MP-20 training data
CrystalAR (raw)70
CrystalAR (Orb pre-relaxed)220
CrystalAR (SV variant, raw)40
DiffCSP (raw)30
DiffCSP++ (raw)40
SymmCD (raw)10
SGEquiDiff (raw)10
FlowMM (raw)90
CrystalFormer (public 1,000, raw)0—
WyFormer-DiffCSP++ (public 1,000, raw)0—

S.U.N under Novel Templates. We have conducted additional experiments to evaluate S.U.N. (Stable, Unique, and Novel) structures generated under previously unseen crystallographic templates, as summarized in the table above. The results show that none of the evaluated models successfully discovers S.U.N. structures under novel templates, highlighting a common challenge faced by current crystal generative models. We have also explicitly acknowledged this limitation in the Limitations section of our paper. Nevertheless, we emphasize that the ability to generate novel templates remains scientifically important, even when it does not immediately lead to stable structures, for three key reasons:

(1). Expanding the crystallographic search space. Novel templates enable the exploration of previously unseen Wyckoff configurations beyond existing structural templates, overcoming a fundamental limitation of template-retrieval-based approaches.

(2). Decoupling structural novelty from stability. Thermodynamic stability depends jointly on chemical composition, lattice geometry, and atomic coordinates. Therefore, the instability of a generated candidate does not necessarily imply that no stable configurations exist within the same template family.

(3). Enabling subsequent structure optimization. Novel templates provide new starting points for energy-guided optimization, structural relaxation, and iterative search, potentially facilitating the discovery of stable or metastable polymorphs.

These considerations are consistent with the broader motivation of symmetry-native generative modeling, which aims to explore crystallographically meaningful configurations beyond those represented in existing datasets. Importantly, template novelty and S.U.N. measure complementary aspects of crystal discovery: the former characterizes the exploration of new crystallographic configurations, while the latter evaluates the discovery of stable, unique, and novel structures.

symprec 0.01 Åsymprec 0.05 Åsymprec 0.1 Å
relax19/2218/2217/22
no relax22/2222/2221/22

Space-group symmetry preservation after relaxation. As shown above, 22/22 structures preserve their target space groups before relaxation at a symmetry tolerance of 0.05 Å, while 18/22 (81.82%) retain their space groups after relaxation. Even under the stricter tolerance of 0.01 Å, 19/22 (86.36%) relaxed structures preserve their original space groups. These results demonstrate that CrystalAR generates structures with high crystallographic symmetry consistency, which is largely preserved after geometric relaxation.

AuthorsOpenReview ↗

Further Replies to Reviewer 6eua

W2 Q4 Fair Comparison and Space-Group Consistency in CSP

(1) Clarification of CSP settings We thank the reviewer for raising this important question regarding the fairness of CSP comparisons. Previous symmetry-native generative models either cannot directly perform CSP due to limitations in their probabilistic factorization or rely on external template retrieval methods, such as CSPML, to obtain the required crystallographic information. In the latter case, retrieving a crystal template effectively provides both the space group GG and the Wyckoff configuration WW, along with additional structural priors. In contrast, CrystalAR addresses a more challenging template-free, space-group-conditioned CSP setting, where only the chemical composition and space group GG are provided, while the Wyckoff configuration WW and crystal geometry must be predicted without external template retrieval. We consider conditioning on GG a well-motivated setting, as space-group information is a fundamental conditioning variable commonly adopted by symmetry-native generative models. In our original comparison, symmetry-native approaches were highlighted in red to distinguish them from generic crystal generation models.

(2) Fair comparison under unified conditions & space-group consistency evaluation. To provide a more comprehensive and fair evaluation, we have conducted additional experiments in which generic CSP baselines are retrained with space-group conditioning, following the soft-conditioning strategy introduced by MatterGen. This establishes a more unified comparison in which all evaluated models receive the same chemical composition and ground-truth space group, allowing us to directly assess their ability to exploit crystallographic symmetry information. Furthermore, we introduce Space-Group Accuracy (SG Acc.), which measures the proportion of generated structures whose crystallographic space groups agree with those of the ground-truth structures, as identified by symmetry analysis. Related space-group consistency evaluations have also been adopted in previous studies, including CrystaLLM~\cite[1].

Importantly, SG Acc. and the conventional CSP Match Rate evaluate two distinct and complementary aspects of structure prediction. Match Rate measures geometric agreement between predicted and reference structures under predefined tolerances, whereas SG Acc. evaluates whether the predicted structures preserve the target crystallographic symmetry. The relatively permissive tolerances commonly used in CSP benchmarks may identify two geometrically similar structures as matching even when they belong to different space groups (See Page 32 Appendix Figure 10 in the newly revised manuscript). This discrepancy can persist even when relatively generous symmetry-detection tolerances are applied. Such differences are physically meaningful: symmetry breaking can fundamentally alter the allowed physical properties of a material~\cite[2], as exemplified by structural phase transitions associated with the emergence of ferroelectric polarization. Therefore, a high Match Rate alone does not necessarily guarantee the correct recovery of crystallographic symmetry, motivating the inclusion of SG Acc. as an independent evaluation metric.

The updated CSP results demonstrate a distinctive advantage of CrystalAR: our model achieves nearly perfect space-group accuracy through its symmetry-native autoregressive generation mechanism, while maintaining a Match Rate comparable to state-of-the-art CSP models. In contrast, the competing methods do not simultaneously achieve these two properties, even when explicitly conditioned on the same ground-truth space groups.

Composition & SG

MethodSC-MR, 0.1 ÅSG Acc, 0.1 Å
CrystalAR62.4798.12
DiffCSP-SG-CFG54.1163.30
DiffCSP-SG53.1662.99
FlowMM-SG34.5041.43
CrystalFormer27.5487.78
CDVAE-SG15.0220.10
WyFormer (no geometry for 7,611/9,046 targets in slot 0)7.6315.68
OMatG-SG1.353.55
CrystalFlow-SG1.233.59

SG-MR measures the proportion of predicted structures that both match the ground-truth structures under the MR criterion and preserve their ground-truth space-group symmetry.

Composition Only

MethodSC-MR, 0.1 ÅSG Acc, 0.1 Å
Uni-3DAR (released checkpoint, rescored)52.1459.22
CrystalAR, 1 of 20 by a frozen rule51.4665.11
CrystalAR, first candidate49.2064.66
DiffCSP41.8349.03
FlowMM34.5740.50
CDVAE10.0713.94
CrystalFlow1.373.53
OMatG1.283.35
CrysBFN35.6741.66

We have updated CSP section in the manuscript accordingly.

[1] Crystal structure generation with autoregressive large language modeling. Nature Communications 2024

[2] Physical Properties of Crystals: Their Representation by Tensors and Matrices 1957

AuthorsOpenReview ↗

Further Replies to Reviewer 6eua

W1 & Q1: Novelty & Level-wise Prediction Ablation

Level-wise joint localization of all sites, as opposed to site-by-site continuous coordinates (vs. CrystalFormer).

In CrystalFormer (Eq. 9), each site's coordinates ui\mathbf{u}_i are drawn from a continuous density, interleaved with the template tokens and conditioned only on earlier sites s<i\mathbf{s}_{\lt i}. A site's precise position is therefore fixed before later sites exist, and the lattice is predicted last. CrystalAR (Eq. 8) first completes the template and the lattice. Then, at each resolution level, it predicts the refinement decisions R(ℓ)\mathbf{R}^{(\ell)} of all inequivalent sites jointly, conditioned on the coarser regions R(<ℓ)\mathbf{R}^{(\lt \ell)} of every site, using block-causal attention. Interatomic geometry is thus resolved globally from coarse to fine, under the metric L\mathbf{L} that determines distances. On MP-20 this improves template-free CSP over CrystalFormer as follows:

  • match rate: 29.62% → 62.92% with 1 hypothesis, and 64.56% → 74.55% with 20 hypotheses (Table 3);
  • symmetry-consistent match rate: 27.54% → 62.47%;
  • space-group accuracy: 87.78% → 98.12% (Table 4).

Contribution of Level-Wise Inter-Site Interactions.

We thank the reviewer for this valuable suggestion. To directly address this concern, we have conducted an additional controlled experiment comparing CrystalAR with a site-wise hierarchical variant (CrystalAR-site) under the same search budget. Both variants employ the same hierarchical coordinate discretization, allowing us to isolate the contribution of level-wise inter-site interactions.

As shown below, CrystalAR achieves 62.47% SC-MR, outperforming CrystalAR-site (58.36%) by 4.11 percentage points, while maintaining nearly identical space-group accuracy (98.12% vs. 97.87%). These results demonstrate that the performance improvement is not solely attributable to hierarchical discretization, but also benefits from the proposed level-wise inter-site interactions. By allowing all symmetry-inequivalent sites to attend to one another at each refinement level, CrystalAR enables more effective geometric coordination than conventional site-wise autoregressive decoding.

MethodSC-MR, 0.1 Å (%) ↑SG Acc, 0.1 Å (%) ↑
CrystalAR (Ours)62.4798.12
CrystalAR-site58.3697.87

We have incorporated this additional ablation into the revised manuscript. This controlled comparison directly addresses the reviewer's concern and provides empirical evidence for the effectiveness of our level-wise interaction mechanism beyond hierarchical discretization alone.

Note that we added a new Section 6, "Comparisons to Prior AR Approaches", which writes the generative factorizations of all four models in a common notation (Eqs. 8–11) and summarizes the differences in Table 4. Those factorizations show that CrystalAR is not a combination of existing pieces; it differs from each prior model in specific terms of the factorization.

Question 3

We have already included a link to our anonymous GitHub repository in the appendix, where the implementation code is publicly available. Specifically, the text “This repository” contains a hyperlink to the anonymous repository. We apologize for any confusion caused by the link not being immediately visible. To improve clarity, we have updated the appendix to explicitly display the full anonymous repository URL.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Discussion and Thanks to Reviewers

Dear Reviewers and AC,

As the discussion period draws to a close, we would like to sincerely thank all reviewers for their active engagement and briefly summarize the current status.

To Reviewer Qgx5: We are delighted that our responses have fully addressed your concerns, and we have been deeply impressed by your expertise and insightful feedback. We sincerely invite you to continue supporting our work in the upcoming reviewer discussion. Some of the remaining concerns from other reviewers are similar to those you raised earlier, so we believe your expert perspective would be invaluable in facilitating an effective discussion and helping resolve them.

To Reviewer 6eua: We are very glad that our rebuttal has addressed most of your concerns and are grateful that you raised your score. We also sincerely appreciate that you raised your remaining concerns in such a specific and timely manner, which gave us sufficient time to address them carefully. We are confident that our latest response resolves your two remaining concerns, and we would be grateful if you could reassess our work in light of the new evidence. Your careful and responsible engagement throughout the process has been evident, and we trust you will reach a well-informed judgment.

To Reviewer idsH: We sincerely thank you for responding before the deadline and clearly stating your two remaining concerns. Although your score has not yet changed, we believe our latest response addresses both concerns, and we would be grateful if you could reassess our work based on the evidence provided.

We understand that novelty is your main concern. The other two reviewers initially raised similar concerns; after we added a full page of discussion to the revised manuscript (Section 6, Eqs. (8)–(11), Table 4), neither has remaining novelty concerns. We believe yours can be resolved in the same way. With respect, we believe that viewing CrystalAR as a combination of existing approaches overlooks key differences in representation, symmetry enforcement, factorization and search, which we detail in our latest response to you.

We would also like to highlight the significance of the work. CrystalAR is the only method that achieves all three of the following simultaneously:

  • CSP: When only the space group is given (without the Wyckoff template), it achieves exact symmetry consistency together with a competitive match rate (Table 3, Figure 10 in Appendix A.11).
  • De novo generation: It lies on the stability–novelty Pareto frontier, with no baseline dominating it (Figure 6, Appendix A.9).
  • Template generation: It explicitly generates Wyckoff templates, including novel ones, with a tractable likelihood that allows them to be scored, ranked and searched (Table 2, Section 4.3).

Individual prior methods offer some of these properties, but none combines all three. More broadly, CrystalAR shows that autoregressive models have distinctive advantages for materials generation. Compared with the mainstream diffusion and flow-matching approaches, it offers a new perspective and paradigm and an additional path for the community. We also see autoregressive and diffusion models as complementary rather than competing.

Finally, we have done our very best to address every concern raised. We once again thank all reviewers and the AC for their time, effort and active participation.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.