acceptodds
Under review as a conference paper at ICLR 2027

CrystalDLM: Feedback Learning of Periodic Diffusion Language Models for Crystal Generation

Abstract

Crystal generation requires compatible periodic arrangements of atoms, yet geometric validity alone does not guarantee physical quality. We introduce CrystalDLM, a framework built on pretrained masked diffusion language models whose conditional prediction supports both crystal construction and reference-conditioned reconstruction. A lattice-conditioned periodic head couples coordinate candidates in a tractable joint distribution, making periodic compatibility an intrinsic part of draft construction. After diffusion refinement, structured records linking reference structures, executed modifications, and their physical outcomes train a reconstruction diffusion language model to propose targeted replacements and a relative verifier to assess candidates against their references. This feedback stage turns physical evaluations into reusable supervision for subsequent structural decisions. Experiments on MP-20, Perov-5, and MPTS-52 demonstrate higher draft structural validity than existing language-model constructors. On MP-20, the complete framework achieves a stable, unique, and novel (S.U.N.) rate of 10.50%, with component comparisons showing that periodic construction improves structural validity under a common diffusion refiner while feedback learning further increases discovery yield.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Material Generation, Crystal Generation, Diffusion Language Models, Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UYCVOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CrystalDLM, a framework that combines diffusion language models with periodic coordinate modeling and physical-feedback learning for crystal generation. The method uses a lattice-conditioned periodic module to improve the geometric consistency of generated structures, followed by continuous diffusion refinement. It further introduces a feedback-learning stage that uses physically evaluated structural modifications to train a reconstruction model and a relative verifier for proposing and selecting improved candidates. Experiments on MP-20, Perov-5, and MPTS-52 show improvements in structural validity and several generation metrics. On MP-20, the complete framework achieves a S.U.N. rate of 10.50%, outperforming the evaluated baselines. Overall, the paper aims to improve crystal generation by incorporating periodic structural constraints and learning from physical evaluation outcomes.

Critical strengths
  1. The paper explores an interesting direction by incorporating periodic structural constraints and physical feedback into diffusion language models for crystal generation. The proposed framework provides a reasonable attempt to improve both geometric validity and physical quality.

  2. The paper conducts experiments on three crystal generation benchmarks, including MP-20, Perov-5, and MPTS-52. The results demonstrate improvements in structural validity and stable, unique, and novel crystal discovery compared with several existing baselines.

Critical weaknesses
  1. Unclear advantages of diffusion language models for crystal generation. The paper does not convincingly demonstrate the unique advantages of using diffusion language models (DLMs) for crystal generation. The proposed framework combines masked prediction, periodic interactions, and continuous diffusion refinement, but these techniques are not inherently specific to DLMs. Although the periodic DLM improves structural validity, its contribution to stable and novel crystal discovery is relatively limited. More importantly, the paper lacks controlled comparisons with alternative generative backbones under the same refinement and feedback pipeline. It therefore remains unclear whether DLMs offer meaningful benefits over existing autoregressive or diffusion-based crystal generative models.

  2. Insufficient validation of the physical-feedback learning mechanism. The physical-feedback stage appears to be the primary source of improvement in stable and novel crystal discovery. However, the proposed approach mainly relies on reward-weighted supervised fine-tuning and a learned verifier for candidate selection, rather than introducing a substantially new feedback-learning algorithm. More importantly, the current ablation studies do not distinguish the contributions of feedback-trained reconstruction, candidate sampling, and verifier-based selection. For example, it remains unclear whether comparable improvements could be obtained by simply generating more candidates and selecting them using the same verifier, without updating the generative model. Without these essential comparisons, the actual contribution of feedback learning remains insufficiently established.

  3. Limited performance improvements relative to framework complexity. The complete framework involves a pretrained 8B diffusion language model, a periodic interaction module, continuous diffusion refinement, and an additional feedback-learning pipeline with a reconstruction model and verifier. Despite this considerable complexity, the reported improvements over existing methods are relatively modest. For example, the S.U.N. rate on MP-20 increases from 9.42% for CrysLLMGen to 10.50% for CrystalDLM. Furthermore, the paper does not provide a comprehensive analysis of training costs, inference overhead, or the computational resources required for physical-feedback collection. Without a clear cost-performance comparison, it is difficult to justify the complexity of the proposed framework over simpler alternatives.

  4. Lack of standardized evaluation for stable and novel crystal discovery. The paper evaluates S.U.N. and M.S.U.N. on MP-20 using its own evaluation protocol rather than the standardized framework provided by LeMat-GenBench. Although both evaluations can be performed on MP-20, differences in relaxation procedures, stability assessment, and novelty criteria may substantially affect the reported results. Consequently, the reported S.U.N. improvements may not be directly comparable with results obtained under other evaluation settings. Given the relatively modest performance gains, evaluating CrystalDLM and its baselines under the same LeMat-GenBench protocol would provide more reliable evidence of its effectiveness in stable and novel materials discovery.

Questions for authors
  1. What are the unique advantages of using a diffusion language model (DLM) for crystal generation? Could the authors clarify which capabilities are specifically enabled by the DLM architecture and whether similar results could be achieved using alternative generative backbones under the same refinement and feedback pipeline?

  2. How much of the improvement comes from feedback learning itself rather than candidate sampling and verifier-based selection? Could the authors provide additional ablation studies comparing feedback-trained reconstruction with a simpler generate-and-rank baseline that uses the same candidate budget and verifier without updating the generative model?

3 comments

We sincerely thank the reviewer for the careful evaluation of our paper and for the detailed comments and questions.

W1 & Q1. Why use a diffusion language model?

A masked DLM provides a common conditional-prediction interface for crystal construction and local revision. By holding specified composition, atom count, and visible geometry fixed as observed tokens while predicting masked fields, it supports both lattice-conditioned periodic coordination of unresolved positions and local re-masking of selected fields within the surrounding structure (§4.2.2; Appendix D.3).

Table R3-1 compares Periodic DLM with crystal-adapted LLaMA-3-8B. In the MPTS-52 backbone comparison, both receive the same composition and atom count and share the continuous refiner, reconstruction model, and relative verifier.

Table R3-1. Discovery results (%).

DatasetConstructorBefore feedback: S.U.N. / M.S.U.N.After feedback: S.U.N. / M.S.U.N.
MP-20LLaMA-3-8B4.00 / 22.104.96 / 24.96
MP-20Periodic DLM8.48 / 47.2410.50 / 51.24
MPTS-52LLaMA-3-8B1.14 / 3.521.52 / 4.86
MPTS-52Periodic DLM4.00 / 22.965.72 / 32.20

1. Satisfying the specified chemical condition is a key difficulty for the tested LLaMA-3-8B constructor. On MP-20, 99.90% of LLaMA-3-8B outputs are parseable, 53.42% match the specified composition and atom count, and 44.38% additionally pass structural validity; on MPTS-52, these rates drop to 99.62%, 10.58%, and 8.76%. Conditioning geometry prediction directly on observed composition and atom-count tokens avoids this failure mode and contributes to Periodic DLM's higher discovery rates in Table R3-1.

2. Periodic coordination becomes more valuable as cells grow. Periodic DLM achieves 100.00% structural validity on MP-20 after continuous refinement. Table R3-2 shows its MPTS-52 construction results stratified by unit-cell size, reaching 97.14% overall validity with local re-masking.

Table R3-2. MPTS-52 structural validity by cell size (%).

Atoms per cellBase DLMPeriodic DLMBase DLM + local re-maskingPeriodic DLM + local re-masking
All86.2689.6496.0697.14
1–2096.2297.8299.6099.92
21–3278.4683.7694.8697.40
33–5251.6660.3481.0083.20

Periodic construction improves single-pass validity by 1.60, 5.30, and 8.68 percentage points across the three size groups, and local re-masking further raises 33–52-atom validity from 60.34% to 83.20%. A separate fixed-sampler control likewise shows gains of 2.96 points on MP-20 and 2.84 points on MPTS-52 from learned periodic interactions. Together, these results support combining DLM conditional prediction with explicit periodic coordination as unit-cell geometry grows more complex.

W2 & Q2. Why train proposals with physical feedback when a verifier can rank generated candidates?

Physical-feedback training improves the proposals supplied to the same verifier, including at a smaller candidate budget. Whereas geometric pretraining recovers clean fields from corrupted training crystals, feedback training learns where to edit, which fields to change, and what replacements to propose from the measured outcomes of edits to generated references (§4.3.2, Eqs. 12–13). Appendix C.1 connects these executed edits and their relative utility to conditional prediction targets, while forward-KL regularization retains the starting model's geometric prior.

Holding the final encoder, relative verifier, references, and candidate budget (up to eight candidates) fixed, Table R3-3 compares geometrically pretrained and feedback-trained reconstruction across three runs before bounded single-field revision.

Table R3-3. Proposal learning with the same verifier (mean ± sample SD over three runs).

Proposal modelS.U.N. (%)M.S.U.N. (%)
Before physical-feedback training9.60 ± 0.2448.27 ± 0.13
After physical-feedback training10.54 ± 0.2649.05 ± 0.71

Under the identical verifier, proposal training improves both discovery metrics in all three runs. To verify that proposal quality improves independently of learned verifier ranking, Table R3-4 replaces the verifier with random selection in Run 1.

Table R3-4. Proposal learning with random selection, Run 1 (%).

Proposal modelS.U.N.M.S.U.N.
Geometrically pretrained reconstruction9.5041.50
Feedback-trained reconstruction9.9043.50

Furthermore, when paired with the verifier, feedback-trained reconstruction using up to four candidates reaches 10.02% S.U.N. / 48.44% M.S.U.N., exceeding geometrically pretrained reconstruction using up to eight candidates (9.86% / 48.24%). Proposal training thus makes useful edits accessible within a smaller candidate budget, while relative verification selects among them.

W3. What benefits do the additional components and computation provide?

The added computation improves periodic geometry, learns reusable edits, and increases effective conditional-generation throughput. On MP-20, CrystalDLM achieves 10.50% S.U.N. / 51.24% M.S.U.N., compared with CrysLLMGen's 9.42% / 44.78%, gains of 1.08 and 6.46 percentage points.

1. Module capacity. Both the periodic head and the relative verifier are lightweight modules built on the backbone representations.

Table R3-5. Parameter counts of the periodic head and relative verifier.

ComponentParameters
Lattice-conditioned periodic head0.537M
Relative verifier1.049M

The lattice-conditioned periodic head adds 0.537M parameters (0.0067% of the backbone) and improves draft structural validity by 8.68 percentage points on 33–52-atom MPTS-52 cells.

2. One-time physical-feedback cost. Appendix A and Table 5 give the upstream models' training settings; Table R3-6 reports physical-supervision collection and feedback training alongside the LLaMA-3-8B baseline training cost.

Table R3-6. Physical-feedback preparation and baseline training costs.

ComputationA100 GPU-hours
Physical-supervision collection5.56
Reconstruction and verifier training0.65
Supervision collection + feedback training6.21
LLaMA-3-8B baseline training6.71

This 6.21 A100 GPU-hour investment (versus 6.71 A100 GPU-hours for LLaMA-3-8B baseline training) learns proposal and selection abilities that are reused on new structures without adding physical relaxations during candidate ranking; the reported relaxation workload includes reference and final-output assessment.

3. Conditional construction throughput. Table R3-7 compares prediction modes under the same chemical conditions on the same physical A100 GPU with batch size one. Wall-clock time includes failed attempts and construction recovery, and valid drafts must satisfy both the specified chemical condition and structural validity.

Table R3-7. Conditional construction efficiency.

Here, condition matching means that the generated structure has exactly the specified chemical composition and total number of atoms.

DatasetConstructor and prediction modeCondition & Struc. Valid (%)Seconds per condition-matched valid draft
MP-20LLaMA-3-8B44.3818.644
MP-20Periodic DLM, sequential fields99.806.025
MP-20Periodic DLM, joint coordinate axis71.524.061
MPTS-52LLaMA-3-8B8.76228.282
MPTS-52Periodic DLM, sequential fields92.9614.901
MPTS-52Periodic DLM, joint coordinate axis92.102.463

Sequential-field Periodic DLM produces condition-matched valid drafts 3.1× faster on MP-20 and 15.3× faster on MPTS-52 than LLaMA-3-8B. Joint-axis prediction reaches 2.463 seconds per valid draft with 92.10% validity on MPTS-52, and 4.061 seconds with 71.52% validity on MP-20, offering a direct speed–yield trade-off.

For CrysLLMGen's native free-generation task, batch-16 generation of 5,000 outputs takes 36.46 minutes, with 89.94% validity and 0.4865 seconds per valid draft. Common continuous refinement takes 12.31 minutes, yielding 94.78% validity. This native task is reported separately from conditional construction.

W4. Evaluation with LeMat-GenBench

Physical-feedback gains persist under a shared LeMat-GenBench configuration. All methods use CHGNet 0.3 relaxation, the Materials Project phase reference (2026-04-13), the MP-20 novelty reference, and common validity criteria. Novelty and uniqueness are evaluated on relaxed structures using StructureMatcher, with lattice, site, and angular tolerances of 0.1, 0.3, and 5°, respectively.

Table R3-8. MP-20 results under the shared LeMat configuration (%).

MethodNoveltyS.U.N.M.S.U.N.
CrystalDLM before feedback90.986.3839.90
CrystalDLM92.896.9642.96
CrysLLMGen81.595.8240.70
ADiT37.243.7021.92
MatterGen89.736.9045.40

Novelty is measured over valid relaxed structures; M.S.U.N. combines LeMat's stable and metastable discovery categories.

Feedback raises novelty from 90.98% to 92.89%, S.U.N. from 6.38% to 6.96%, and M.S.U.N. from 39.90% to 42.96%. CrystalDLM achieves the highest observed novelty and S.U.N. among the compared methods; its M.S.U.N. exceeds CrysLLMGen and ADiT, while MatterGen reaches 45.40%. The feedback improvement therefore extends to evaluation on relaxed structures.

We thank the reviewer for the detailed assessment. We hope these results clarify our design choices and their empirical benefits, and we would welcome further discussion of any remaining questions.

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

This paper proposes CrystalDLM, a novel framework for crystal generation that combines periodic geometric modeling with physical feedback learning. The authors introduce lattice-conditioned periodic interactions into a diffusion language model (DLM) to improve the geometric validity of generated crystal structures. They further train a Reconstruction DLM and a relative Verifier using physical evaluation feedback to propose and select structural modifications. Experiments on MP-20, Perov-5, and MPTS-52 demonstrate improved structural validity and competitive generation performance. On MP-20, CrystalDLM achieves a S.U.N. rate of 10.50% and an M.S.U.N. rate of 51.24%.

Critical strengths
  1. Clear motivation and intuitive methodology. The paper addresses two meaningful challenges in crystal generation: periodic geometric compatibility and physical stability. The proposed periodic interactions and feedback learning framework are well motivated, and the overall methodology is intuitive.

  2. Solid experimental results and ablation studies. The proposed framework achieves competitive performance across three crystal generation benchmarks, particularly in structural validity and stable, unique, and novel crystal discovery. The controlled ablation studies in Table 4 provide evidence for the effectiveness of both periodic construction and physical feedback learning.

Critical weaknesses
  1. Mixed results on property distribution matching with limited analysis. Although CrystalDLM improves structural validity and coverage recall, its performance on property distribution matching is inconsistent. For example, in Table 1, the density Wasserstein distance on MP-20 increases from 0.39 (CrysLLMGen) to 0.79 (CrystalDLM). While the authors acknowledge this discrepancy, further analysis is needed to explain its causes and understand the trade-offs.

  2. The necessity of the learned feedback components is not sufficiently justified. The proposed Reconstruction DLM and Verifier are trained using physical evaluation results from CHGNet. However, it remains unclear whether learning these additional components is preferable to directly using the physical oracle for optimization or candidate selection. Prior studies, such as Symbolic Music Generation with Non-Differentiable Rule Guided Diffusion (2024) and Chemistry-Inspired Diffusion with Non-Differentiable Guidance (2025), demonstrate alternative approaches for guiding generative models using non-differentiable rules or physical oracles. Although these methods address different domains, they motivate comparisons with direct oracle-guided optimization. A CHGNet-guided optimization or candidate-selection baseline, ideally under comparable computational budgets, would help clarify the advantages of the learned feedback framework. Additionally, separate ablations of the Reconstruction DLM and Verifier would better establish their individual contributions.

  3. Limited ablation studies across datasets. The ablation studies in Table 4 are conducted only on MP-20, despite the evaluation covering three datasets. It remains unclear whether the contributions of periodic construction and physical feedback learning generalize to Perov-5 and MPTS-52, particularly for larger crystal structures. Additional ablation studies on these datasets would strengthen the experimental evidence.

  4. Insufficient justification for the binary tree topology. The periodic head uses a fixed binary tree based on site indices to model pairwise coordinate interactions. Although the tree structure enables efficient exact inference, the paper provides limited intuition for choosing this particular topology. Why is a fixed binary tree preferred over other tree structures, especially those informed by the geometric relationships between atoms? Experiments comparing different tree topologies would help establish whether this design choice is important for the proposed method.

Questions for authors

See weakness.

3 comments

We sincerely thank the reviewer for the positive assessment of our work and for the thoughtful comments and constructive suggestions.

W1. Explaining the density distribution difference

CrystalDLM's lower density is driven by lighter compositions, with smaller volumes per atom partly offsetting the shift. To explain the MP-20 density Wasserstein distances of 0.79 for CrystalDLM and 0.39 for CrysLLMGen, Table R2-1 compares the property-evaluation distributions of mean mass per atom mˉ\bar m, median volume per atom V/NV/N, and mass density ρ=1.66054,mˉ/(V/N)\rho=1.66054\\,\bar m/(V/N).

Table R2-1. Composition, volume, and density on MP-20.

MethodMean mass/atom (amu)Median volume/atom (ų)Mean density (g/cm³)Density W1 ↓
TEST79.5018.376.710.0000
CrysLLMGen86.9819.207.080.3944
CrystalDLM67.0717.205.920.7889

Decomposing the mean log-density shift relative to TEST isolates the two terms:

ΔE[log⁡ρ]=ΔE[log⁡mˉ]−ΔE[log⁡(V/N)]=−0.1719=−0.2712−(−0.0993).\Delta\mathbb{E}[\log\rho] = \Delta\mathbb{E}[\log\bar m] - \Delta\mathbb{E}[\log(V/N)] = -0.1719 = -0.2712 - (-0.0993).

Here, the negative mass term (−0.2712-0.2712) accounts for the downward shift in mean log-density, while the smaller volume per atom (−0.0993-0.0993) partly compensates in the opposite direction; Wasserstein-1 additionally reflects differences in distribution shape and tails.

Training and generation difficulty provide a possible explanation for the sampled composition mix. Because continuous refinement and physical feedback preserve species and atom count (§§4.2.3–4.3), the evaluated mass distribution is determined by sampled chemical conditions and admission to property evaluation. In our representation of 7+4N7+4N fields (including 6+3N6+3N geometric fields), smaller atom counts require fewer predictions and fewer positions to coordinate. Although the training loss is averaged within each crystal (Appendix A, Eq. A.2), condition frequency and geometric difficulty can affect which structures are learned, generated successfully, and admitted to evaluation, thereby reweighting the evaluated composition mix. Periodic coordination and feedback improve geometry and discovery, while the mass–volume decomposition identifies the composition-related shift in density.

W2. Learning reusable structural edits from physical evaluation

Physical-feedback training turns evaluated edits into reusable proposal and selection abilities. Whereas direct oracle selection must evaluate each new candidate batch, the trained reconstruction model and relative verifier reuse their learned decisions on new structures without relaxing every candidate.

Following your suggestion (Huang et al., 2024; Shen et al., 2025), Table R2-2 compares learned feedback against single-point energy selection (K=8K=8) and full-relaxation oracle selection (K=1,4,8K=1,4,8).

Table R2-2. Candidate-selection results (costs projected).

Selection strategyS.U.N. (%)M.S.U.N. (%)Relaxation inputsA100 relaxation cost (GPU-hours)
CrystalDLM without physical feedback8.4847.245,000 (default evaluation)5.07
CrystalDLM (ours)10.5051.247,6057.71
Single-point energy (K=8)8.4847.525,1625.23
Full-relaxation oracle (K=1)10.0050.589,5389.67
Full-relaxation oracle (K=4)12.0058.1823,08623.40
Full-relaxation oracle (K=8)12.8664.3840,50541.05

Single-point energy selection (K=8K=8) yields 8.48% S.U.N. / 47.52% M.S.U.N., close to the 8.48% / 47.24% baseline without feedback. By contrast, CrystalDLM reaches 10.50% / 51.24%, exceeding the K=1K=1 full-relaxation oracle (10.00% / 50.58%) while requiring a smaller full relaxation workload (7,605 vs. 9,538 relaxation inputs, or 7.71 vs. 9.67 projected A100 GPU-hours, including reference and selected-output validation). Scaling the full-relaxation oracle to K=8K=8 reaches 12.86% / 64.38%, but requires 40,505 relaxation inputs (41.05 projected A100 GPU-hours), or 5.3 times CrystalDLM's full relaxation workload.

Separate controls identify the contributions of both learned components before bounded single-field revision. Holding the verifier and candidate budget (K≤8K\leq8) fixed, proposal training raises three-run mean S.U.N. from 9.60% to 10.54% and M.S.U.N. from 48.27% to 49.05%. Holding the feedback-trained candidate pool fixed in Run 1, relative verification improves S.U.N. from 9.90% to 10.50% and M.S.U.N. from 43.50% to 49.86% over random selection. Physical supervision thus improves both the proposed edits and the decision of which edit to accept.

W3. Component effects on larger crystal structures

MPTS-52 demonstrates the contributions of periodic coordination and physical feedback in cells with up to 52 atoms. In the simpler five-atom Perov-5 setting, by contrast, the base and periodic constructors already achieve the same near-saturated structural validity.

1. Periodic coordination and local re-masking. The periodic head scores candidate compatibility under the lattice (§4.2.2, Eqs. 6–7), while construction-stage local re-masking re-predicts selected fields conditioned on the surrounding geometry (Appendix D.3). In a controlled comparison with the constructor and joint sampler held fixed, learned periodic interactions raise MPTS-52 structural validity from 86.10% to 88.94% in single-pass construction, and from 95.80% to 97.24% with local re-masking. To examine how this effect scales with unit-cell size, Table R2-3 further compares the base and periodic constructors across 5,500 conditions stratified by atom count.

Table R2-3. MPTS-52 structural validity by cell size (%).

Atoms per cellBase DLMPeriodic DLMBase DLM + local re-maskingPeriodic DLM + local re-masking
All86.2689.6496.0697.14
1–2096.2297.8299.6099.92
21–3278.4683.7694.8697.40
33–5251.6660.3481.0083.20

In single-pass construction, the gain from periodic coordination grows across the three size groups from 1.60 to 5.30 and 8.68 percentage points. For 33–52-atom cells, local re-masking further raises Periodic DLM validity from 60.34% to 83.20% (and overall validity to 97.14%). The larger gains in bigger unit cells support coordinating unresolved positions under periodic constraints as atom count increases.

2. Physical feedback. In a separate fixed-composition experiment on MPTS-52, we track the same continuously refined references before and after the feedback stage.

Table R2-4. Physical-feedback comparison on MPTS-52 (%).

StageS.U.N.M.S.U.N.
Before feedback4.0022.96
After feedback5.7232.20

Physical feedback improves MPTS-52 S.U.N. by 1.72 percentage points and M.S.U.N. by 9.24 points, extending the discovery gains to larger crystal structures.

W4. Choice of the fixed binary tree

A fixed binary tree provides a usable dependency structure before all coordinates are generated, while learned edge potentials supply lattice-dependent geometry. Its logarithmic depth and exact conditional message passing make it a natural topology for coordinating masked positions (§4.2.2, Eqs. 6–7; Appendix B). Table R2-5 compares binary, geometry-informed, and chain trees under the same conditions on MP-20 before physical feedback.

Table R2-5. Tree-topology comparison on MP-20.

TreeConstruction validity (%)Validity with local re-masking (%)Backbone forward callsS.U.N. (%)M.S.U.N. (%)
Binary95.6299.5242,4378.9049.60
Geometry-informed94.7699.5242,7848.5048.60
Chain94.2899.0442,7358.8048.60

S.U.N. and M.S.U.N. are evaluated after the common continuous refinement, before physical feedback.

The binary tree achieves the highest single-pass construction validity and pre-feedback discovery rates, ties the geometry-informed tree after local re-masking, and uses the fewest backbone forward calls, supporting it as a simple and effective default topology.

We appreciate the reviewer's insightful suggestions on the mechanisms and experiments. We would welcome the opportunity to discuss any points that would benefit from further clarification.

Reviewer_Pu8POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper has proposed CrystalDLM, a framework for crystal generation using pretrained masked diffusion language models (DLMs). In this work, the authors used a hybrid framework that combines a discrete masked diffusion language model to jointly predict lattice parameters and atomic coordinates with continuous geometric diffusion refinement to refine numerical coordinates and lattice geometry. The authors proposed two main contributions: the first is the lattice conditioned periodic head, which they added to the DLM (LLaDA-8B constructor), and the second is the feedback stage applied after diffusion refinement that uses previously evaluated structural modifications to guide the reconstruction and candidate selection. It first fixes composition and count using Planner (Llama-3-8B), then passes to a periodic DLM (LlaDA-8B) that couples coordinate prediction conditioned on the lattice, and finally, using diffusion refinement, continuously refines fractional coordinates (X) and lattice vectors (L). Then, using offline feedback records, it reconstructs X and L and, finally, using the relative verifier, selects the best candidate or keeps the reference, after which the final crystal verifier is applied. For validation, they have used the MP-20, Perov-5, and MPTS-52 datasets against various baselines.

Critical strengths

This paper has outperformed other methodologies such as CrysLLMGen(draft) and Llama-2 (7B) on the larger MPTS-52 dataset. These results show the framework can generate structurally valid crystal drafts for larger unit cells as well. The condition on the lattice periodic head is another strength of this paper, as it addresses an important limitation of independently predicting atomic coordinates. Although the structured dependencies in DLMs have been explored previously, their adaptation to periodic crystal geometry is a relevant contribution.

Critical weaknesses
  1. Limited ablation studies: Table 4 shows that the feedback stage improves S.U.N. from 8.48% to 10.50% and M.S.U.N. from 47.24% to 51.24%. However, it is unclear how much each component contributes to these improvements. A more detailed ablation study would help clarify their individual roles.

  2. Limited baseline comparisons: The paper has been compared with several existing methods, but some recent methods, such as All-atom Diffusion Transformers (ADiT), are missing. It would be helpful if the authors could include one or two additional baselines, where possible, to provide a better comparison with recent work.

Questions for authors
  1. Contribution of the feedback components: Table 4 shows that the feedback stage improves S.U.N. from 8.48% to 10.50% and M.S.U.N. from 47.24% to 51.24%. However, the contributions of the individual components are not evaluated separately. It would be helpful if the authors could include a more detailed ablation study to show how much each component contributes to these improvements.

  2. Independent stability evaluation: The feedback stage uses CHGNet to evaluate candidate structures, and the S.U.N. and M.S.U.N. scores are also based on CHGNet. Could the authors explain whether using the same model for both feedback and evaluation might affect the reported results?

  3. Variability across runs: Tables 2 and 4 did not report standard deviations. Since the improvements over CrysLLMGen in S.U.N. (1.08 percentage points) and MatterGen in M.S.U.N. (1.80 percentage points) are relatively small, could the authors report the mean and standard deviation across multiple random seeds? This would have helped in assessing how consistent the improvements are.

  4. Comparison with All-atom Diffusion Transformers (ADiT): ADiT (Joshi et al., 2025) is another diffusion-based method evaluated on MP-20. Could the authors clarify why it was not included in the comparisons? A comparison using the same evaluation protocol would have been helpful.

2 comments

We sincerely thank the reviewer for carefully reading our paper and for the thoughtful comments and suggestions for strengthening the work.

W1 & Q1. Individual contributions of the physical-feedback components

Physical feedback improves crystal discovery through three decisions: learning which edits to propose, selecting edits relative to the reference, and applying a bounded single-field revision. Submitted Table 4 reports increases from 8.48% to 10.50% S.U.N. and 47.24% to 51.24% M.S.U.N. on MP-20. Under the same CHGNet evaluation protocol, Table R1-1 tracks M.S.U.N. across the three stages of our supplementary runs: continuous refinement → proposals and relative selection → bounded single-field revision.

Table R1-1. M.S.U.N. across feedback stages (three runs, mean ± sample SD, %).

Continuously refined referenceAfter proposals and relative selection, before revisionFinal feedback output, after bounded revision
47.57 ± 1.7749.05 ± 0.7151.45 ± 0.41

1. Proposal learning. Geometric pretraining recovers clean fields from corrupted training crystals, whereas physical-feedback training uses the outcomes of edits to generated references to supervise modification scope, sites, and replacement values (§4.3.2, Eqs. 12–13; Appendix C.1). Holding the final encoder, relative verifier, references, and 80-call budget (up to eight candidates) fixed, we compare the two proposal models before bounded single-field revision.

Table R1-2. Proposal training with the same verifier, before bounded revision (three runs, mean ± sample SD, %).

Geometrically pretrained: S.U.N.Feedback-trained: S.U.N.Geometrically pretrained: M.S.U.N.Feedback-trained: M.S.U.N.
9.60 ± 0.2410.54 ± 0.2648.27 ± 0.1349.05 ± 0.71

With the same verifier, physical-feedback training improves both discovery metrics across the three runs, showing the contribution of learning which structural edits to propose.

2. Relative verification. The verifier predicts each candidate's quality relative to retaining its reference (§4.3.1, Eq. 11; Appendix C.2). Fixing the Run 1 candidate pools before bounded single-field revision, we compare learned relative verification against random selection, with reference retention available under both rules.

Table R1-3. Selection-rule comparison on fixed candidate pools, before bounded revision.

Proposal modelSelection ruleS.U.N. (%)M.S.U.N. (%)
Unchanged referencesRetain reference8.9049.60
Geometrically pretrained reconstructionRandom selection9.5041.50
Geometrically pretrained reconstructionRelative verifier9.8648.24
Feedback-trained reconstructionRandom selection9.9043.50
Feedback-trained reconstructionRelative verifier10.5049.86

Relative verification improves M.S.U.N. over random selection by 6.74 percentage points for pretrained proposals and 6.36 points for feedback-trained proposals; paired with feedback-trained proposals, it reaches 49.86% M.S.U.N., exceeding the reference's 49.60%. This isolates the contribution of learning which proposed edits to accept.

3. Bounded single-field revision. Within the shared 80-call budget, the model can revise one field of a selected single-site edit using the risk-guided conditional distribution in Appendix D.2. Fixing the trained weights and verifier-selected structures, we compare outputs with bounded revision disabled versus enabled.

Table R1-4. From selected structures to final feedback outputs: bounded revision with trained weights and starting structures fixed.

Paired comparisonRevision off: S.U.N. (%)Revision on: S.U.N. (%)Revision off: M.S.U.N. (%)Revision on: M.S.U.N. (%)M.S.U.N. change (pp)
Run 210.3010.3048.5451.00+2.46
Run 310.8210.8248.7451.54+2.80

Bounded single-field revision preserves S.U.N. while adding 2.46 and 2.80 percentage points of M.S.U.N., yielding the final outputs in Table R1-1. Together, the three controls connect the overall feedback gain to improved proposals, learned acceptance, and subsequent local revision.

W2 & Q4. Comparison with ADiT

We have added ADiT (Joshi et al., 2025) to the revised main comparison (Table 2). Using its official fixed collection of 5,000 generated MP-20 structures under our shared CHGNet relaxation and Materials Project competing-phase hull protocol, ADiT achieves 5.18% S.U.N. and 31.96% M.S.U.N., compared with 10.50% S.U.N. and 51.24% M.S.U.N. for CrystalDLM.

Q2. Independent stability evaluation

Physical-feedback gains persist under independent MACE evaluation, and the same editing framework can learn from MACE supervision. Using 1,000 sampled structures per evaluation, we cross the feedback-supervision potential with the evaluation potential in Table R1-5.

Table R1-5. Reference results and crossed supervision–evaluation comparison.

Feedback supervisionEvaluation potentialS.U.N. (%)M.S.U.N. (%)
Before feedbackCHGNet8.9049.60
Before feedbackMACE2.8041.80
CHGNetCHGNet10.5051.80
CHGNetMACE3.1046.10
MACECHGNet10.7049.40
MACEMACE3.4047.80

Under MACE evaluation, CHGNet-supervised feedback raises S.U.N. from 2.80% to 3.10% and M.S.U.N. from 41.80% to 46.10%. Supervising the feedback stage directly with MACE further reaches 3.40% S.U.N. and 47.80% M.S.U.N. under MACE evaluation (and 10.70% S.U.N. and 49.40% M.S.U.N. under CHGNet evaluation).

We also compare CrystalDLM (CHGNet-supervised) against baseline generators under MACE evaluation.

Table R1-6. Method comparison under MACE evaluation.

MethodS.U.N. (%)M.S.U.N. (%)
CrystalDLM3.1046.10
CrysLLMGen2.7043.00
ADiT2.9030.30
MatterGen3.3045.20

Under MACE evaluation, CrystalDLM achieves the highest M.S.U.N. (46.10%) among the compared methods and exceeds CrysLLMGen and ADiT on both metrics, while MatterGen attains a higher S.U.N. (3.30% vs. 3.10%).

Q3. Variability across random seeds

Physical feedback improves both discovery metrics in all three independent sampling runs on MP-20. Holding the trained models and chemical conditions fixed, we generate 5,000 structures per seed and evaluate each run independently before and after feedback.

Table R1-7. CrystalDLM results across three seeds: mean ± sample SD.

Stage (rates: %; gains: pp)S.U.N., Runs 1 / 2 / 3Mean ± SDM.S.U.N., Runs 1 / 2 / 3Mean ± SD
Before feedback8.90 / 8.60 / 8.708.73 ± 0.1549.60 / 46.70 / 46.4047.57 ± 1.77
After feedback10.50 / 10.30 / 10.8210.54 ± 0.2651.80 / 51.00 / 51.5451.45 ± 0.41
Gain+1.60 / +1.70 / +2.12+1.81 ± 0.28+2.20 / +4.30 / +5.14+3.88 ± 1.51

Mean paired gains are 1.81 ± 0.28 percentage points in S.U.N. and 3.88 ± 1.51 points in M.S.U.N. The three-seed means (10.54% ± 0.26 pp and 51.45% ± 0.41 pp) are close to the submitted 10.50% and 51.24%, and the observed M.S.U.N. sample standard deviation decreases from 1.77 pp to 0.41 pp.

For reference, we also report three-seed statistics for two baselines under their respective protocols: MatterGen-MP20 uses 2,000 native samples per seed evaluated against the 2026-04 phase-reference archive, and crystal-adapted LLaMA-3-8B is evaluated under fixed composition and atom count after continuous refinement (5,000 requests per seed).

Table R1-8. Baseline results across three seeds: mean ± sample SD (%).

MethodS.U.N., Runs 1 / 2 / 3Mean ± SDM.S.U.N., Runs 1 / 2 / 3Mean ± SD
MatterGen-MP208.90 / 9.55 / 8.859.10 ± 0.3954.35 / 55.15 / 53.9554.48 ± 0.61
LLaMA-3-8B4.00 / 4.38 / 3.804.06 ± 0.2922.10 / 21.90 / 23.5222.51 ± 0.88

We thank the reviewer again for the constructive feedback. These results clarify the feedback components and broaden the evaluation. We would be happy to discuss any further questions or suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.