acceptodds
Under review as a conference paper at ICLR 2027

CrystalReasoner: Reasoning and RL for Property-Conditioned Crystal Structure Generation

Abstract

Generative modeling has emerged as a promising approach for crystal structure discovery. However, existing LLM-based generative models struggle with low-level atomic precision, while diffusion-based methods fall short in integrating high-level scientific knowledge. As a result, generated structures are often invalid, unstable, or do not possess desirable properties. To address this gap, we propose CrystalReasoner (CrysReas), an end-to-end LLM framework that generates crystal structures from natural language instructions through reasoning and alignment. Before emitting atomic coordinates, CrysReas writes a thinking trace that states the crystallographic symmetry, the local coordination environments and the physical properties of the target structure, so that the textual knowledge a language model already carries is turned into constraints on the coordinates it is about to write. CrysReas then employs reinforcement learning (RL) with a multi-objective, dense reward function to align generation with downstream desired properties such as physical validity, chemical consistency, and thermodynamic stability. For property-conditioned tasks, we design task-specific reward functions and train specialized models for discrete constraints (e.g., space group) and continuous properties (e.g., elasticity, thermal expansion). On the MP-20 benchmark, CrysReas raises the share of generated structures that are stable, unique and novel from to over the strongest prior language-model baseline re-implemented in the same pipeline, and each of three property specialists improves adherence to its target constraint over the unconditioned model. Our work demonstrates the potential of leveraging thinking traces and RL for generating valid, stable, and property-conditioned crystal structures. Code is available at: https://anonymous.4open.science/r/CrystalReasoner/

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Crystal structure generation, Large language models, Generative models, Reinforcement learning, Scientific reasoning, Property-conditioned generation, Materials discovery

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

This paper presents CrystalReasoner, an LLM-based crystal structure generation framework combining explicit reasoning traces with GRPO reinforcement learning. The method incorporates structure-aware textual descriptions and physical property supervision for condition-guided generation and is evaluated on MP-20 for validity, stability, and condition consistency.

Critical strengths
  1. Explicit thinking traces with crystallographic information provide a promising direction for enabling LLMs to incorporate structural knowledge during crystal generation.

  2. The method combines thinking-trace supervision with reinforcement learning and extends to generation conditioned on different physical properties. Comparisons among Base, Thinking, RL, and the full model provide insights into the contributions of reasoning traces and GRPO.

Critical weaknesses

The evaluation protocol is not presented clearly enough to interpret performance across generation settings or identify the sources of improvement. Section 4.1 refers to “textual specifications” (Line 259) and evaluates 1,024 test compositions (Lines 292-295), but does not explicitly describe the evaluated input conditions. Lines 962-968 describe prompts that always provide the formula and select additional conditions from formation energy, band gap, energy above the hull, and space group, with 20% containing only the formula. Together with Lines 989-992, this suggests that Table 1 uses these mixed conditions. However, the actual distribution of evaluation conditions and results for each setting are not reported, preventing separate assessment of formula-only, space-group-conditioned, and energy/property-conditioned generation. Moreover, the interpretation of Chemical Validity, Novelty, Uniqueness, S.U.N., and S.S.U.N. under formula-conditioned evaluation requires clarification. Chemical Validity is largely determined by the supplied composition when the model follows the formula. Generating one structure per composition per seed and computing Uniqueness only within reduced-formula groups (Lines 1159-1169) does not adequately measure structural diversity for a given composition. Novelty also depends substantially on the overlap between test and training compositions. Consequently, it remains unclear how much S.U.N. and S.S.U.N. reflect actual structure exploration rather than the supplied compositions and evaluation protocol.

Questions for authors
  1. What specifically constitutes the “atomic precision” problem? Lines 14-15 and 198-200 identify insufficient low-level atomic precision as a key motivation. Which existing methods exhibit this limitation, and does it concern coordinate errors, unreasonable interatomic distances, symmetry inconsistencies, or other issues? Could the authors provide direct quantitative evidence and clarify how reasoning traces and GRPO individually improve atomic precision?
  2. What information is encoded in the thinking traces, and how does it constrain generation? Lines 161-184 state that the traces contain Wyckoff letters and multiplicities, whereas the case study in Appendix D (Lines 886-889) does not appear to include Wyckoff position information. Please reconcile these descriptions and explain how the traces implement the “progressively shrinking” crystallographic solution space described in Section 3.2.
  3. Could results be reported separately by input setting? Is the condition selection strategy in Lines 962-968 used in Table 1? Please separate formula-only, formula with space-group conditions, and formula with additional property conditions, and clarify the interpretation of Chemical Validity, Novelty, Uniqueness, S.U.N., and S.S.U.N. in these settings.
  4. Why is PLAID++’s DPO stage omitted? Table 1 uses only its Wyckoff representation without DPO (Lines 218-219), whereas the proposed method includes RL alignment. Could the authors include complete PLAID++ or clarify which conclusions about superiority over existing methods the current comparison supports?
  5. How were the data filtered, and how representative is the evaluation subset? The standard MP-20 train/validation/test split contains 27,136 / 9,047 / 9,046 structures, but the paper reports 24,231 training and 8,141 test structures (Lines 704-707), then selects 1,024 test compositions (Lines 755-759). Please explain both reductions and the resulting subset’s representativeness. Evaluation on the full test set would strengthen the reliability of the reported improvements.
4 comments
AuthorsOpenReview ↗

Author Rebuttal to Reviewer jDwD

We thank the reviewer for recognizing the value of crystallographic thinking traces and the CrysReas-Base/CrysReas-Thinking/CrysReas-RL/CrysReas ablations. The additional experiments address the missing DPO baseline, evaluation by input setting, structural diversity for a fixed prompt, and coverage of the full test set. We also clarify the trace contents, the meaning of "atomic precision" and the filtering from the original MP-20 splits.

Q1. Definition of “atomic precision” and quantitative evidence from the thinking and RL ablations.

We agree that the term "atomic precision" was insufficiently specific. Our intended meaning is the ability to generate coordinates that yield physically plausible structures and satisfy the requested constraints, including reasonable interatomic distances and consistent symmetry. The number of decimal places follows the training representation; the quality of the resulting structures is assessed through structural validity, symmetry consistency, energy, and property-target satisfaction. We therefore correct it to "structural validity and target-property satisfaction" to state the objective in terms of the quantities directly evaluated.

Q2. Wyckoff information in the thinking traces and the “progressively shrinking” crystallographic solution space.

We agree that the construction program computes Wyckoff letters for grouping and ordering sites, but does not explicitly print those letters in the trace. We correct the wording to match these actual outputs in the revision.

Three parts of thinking traces describe symmetries, local coordinations, and physical properties. We add an experiment to prove that they progressively shrink the crystal solution space. We select 32 distinct formula-only prompts from the held-out generation split, testing the generated crystal when the thinking traces are truncated at different positions. The observed cluster count decreases at every prefix stage, with the largest drop after the symmetry and site-count segment.

Trace prefixMean distinct clusters per prompt (total)Eligible outputs (%)Space-group match to trace (%)
c0c_0: prompt only15.94 (510)62.11.5
c1c_1: symmetry and sites11.66 (373)95.955.5
c2c_2: plus coordination10.47 (335)98.073.7
c3c_3: complete trace9.34 (299)97.474.6
AuthorsOpenReview ↗

Author Rebuttal to Reviewer jDwD (Part 2)

W & Q3. Formula-only, space-group-conditioned, and property-conditioned results, and interpretation of validity, novelty, and diversity.

We evaluate the models on the full CDVAE test set now. The full evaluation uses the mixed-condition prompt protocol, with the following fixed population shared by all models:

Input settingPrompts
Formula only1,660
Formula + space group416
Formula + properties2,408
Formula + space group + properties3,657
Total8,141

Each prompt has five candidates. Table 1 evaluates structural validity, chemical validity, and composition consistency on the 1,660 formula-only prompts; space-group consistency and S.S.U.N. on the 4,073 prompts explicitly requesting a space group; and stability, energy, uniqueness, and S.U.N. on all 8,141 prompts. Condition-specific results are also available. Failed attempts remain in rate denominators; mean energy uses finite hull distances.

We agree that Chemical Validity is a composition-based check and is largely determined by the supplied formula when the model follows it. Every evaluation prompt specifies a formula, so we interpret this metric together with Composition Consistency and Structural Validity. Novelty also depends on the reference set: 88.5% of test compositions are not in the training set. We therefore repeat novelty matching against the larger MP+Alexandria reference set. The table reports the resulting rates on the same 8,141 prompts.

Reference setMetric (%)Wyckoff+DPOWyckoff+GRPOCrysReas
SFT training structuresNovelty78.5081.8882.32
SFT training structuresS.U.N.3.333.494.58
MP+AlexandriaNovelty45.1446.2041.38
MP+AlexandriaS.U.N.0.140.160.21

The larger reference set substantially lowers novelty and S.U.N. for every model. CrysReas has lower standalone novelty against MP+Alexandria than either Wyckoff baseline, while retaining the highest S.U.N. under that reference. This means RL of CrysReas is able to improve structure quality (hence improve S.U.N.), but since the reward does not contain novelty, the novelty over MP+Alexandria is the lowest.

We report the uniqueness over the same formula. We select 256 distinct formula-only prompts from the full test set and generate 32 outputs per prompt. We find this number is sensitive to sample count, indicating a reasonable evaluation requires more prompts, making it hard to evaluate for multiple conditions. It is also unfair for us to compare our uniqueness (always condition on formula) and prior works' uniqueness (unconditional), hence we retain uniqueness in Table 1 as the uniqueness over 8,141 test prompts, averaged by 5 seeds.

Outputs per promptMean within-formula Uniqueness (%)
1100.00
853.20
1644.99
3238.07

Q4. PLAID++'s DPO stage and comparison with post-trained Wyckoff baselines.

We include Wyckoff+DPO, incorporating PLAID++-style offline preference optimization, and Wyckoff+GRPO. Both start from the same Wyckoff SFT checkpoint, use the same reward definition and 8,000 training prompts, and generate eight candidates per prompt. DPO uses these candidates to construct offline preference pairs, while GRPO samples candidates online during training. All models are evaluated on the same 8,141 test prompts with five candidates per prompt.

MetricWyckoff+DPOWyckoff+GRPOCrysReas
Structural validity (%)90.4094.7795.58
Space-group consistency (%)61.5562.7068.51
Stability (%)3.663.925.13
Mean hull distance (eV/atom)0.5270.4800.434
S.U.N. (%)3.333.494.58

CrysReas retains an advantage over both post-trained Wyckoff baselines on these metrics. Thus, the comparison remains favorable after including the preference-optimization stage omitted from the original evaluation, and when the Wyckoff generator also receives GRPO. Within the shared Wyckoff representation and SFT initialization, GRPO also achieves higher performance than DPO on these metrics.

AuthorsOpenReview ↗

Author Rebuttal to Reviewer jDwD (Part 3)

Q5. MP-20 filtering, the 1,024-prompt subset, and evaluation on the full test set.

The difference in dataset size comes from two filtering steps: we remove MP-20 entries unavailable in the Materials Project data used for this study, then exclude structures containing elements outside MatterGen's fixed 76-element vocabulary. The following counts account for the reductions in all three splits:

SplitOriginal MP-20Unavailable MP entriesExcluded by element vocabularyRetained
Train27,1361,0981,80724,231
Validation9,0473705828,095
Test9,0463455608,141

Following the reviewer's suggestion, the main evaluation now covers all 8,141 retained test prompts, with five candidates per prompt, compared with the earlier 1,024-prompt subset. Every model is evaluated on the same prompts and candidate count. The updated main results therefore cover the entire filtered test set, removing dependence on the representativeness of the earlier subset.

Dear Reviewer jDwD, We posted a response to your concerns on Nov 15, and we wanted to check whether our clarifications address the issues you raised. If any point remains unresolved, we are happy to provide further evidence. Thank you.

Reviewer_RaqXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CrystalReasoner (CrysReas), an end-to-end LLM framework that generates crystal structures from natural language instructions. Before emitting atomic coordinates, the model writes a thinking trace stating the crystallographic symmetry, local coordination environments and physical properties, synthesized by a deterministic crystallographic program from the ground-truth structure. The model is then aligned with online RL (GRPO) using a multi-objective, dense reward for physical validity, chemical validity and thermodynamic stability (Ehull from MatterSim), and three property specialists are trained with property-specific rewards. On MP-20, CrysReas produces the largest share of stable, unique and novel (S.U.N.) structures compared with CrystalTextLLM and the Wyckoff representation of PLAID++, both re-implemented in the same pipeline.

Critical strengths

Thinking traces measurably improve generation. With everything else fixed, adding the trace before the CIF improves structural validity, composition and space group consistency, stability and S.U.N. over SFT without traces. Replacing the symmetry or coordination segment with that of another sample degrades the generated structure, so the decoder does condition on the trace, and RL improves hull distance only when traces are present. Synthesizing such traces from structure records is a cheap and reproducible way to add an intermediate structural description, which other structured scientific generation tasks could adopt.

Critical weaknesses

Limited methodological novelty. Fine-tuning an LLM with SFT followed by RL against MLIP-based stability and property rewards has already been explored for crystal generation, e.g., CrysTune (Govindarajan et al., ICML 2026, OpenReview Oe5iihLiiV), which applies GRPO with stability and crystal-system rewards, and CRYSTAL (Yao et al., ICML 2026, OpenReview u1r1C1EJBh); neither is cited or discussed in the paper. The new component is therefore the thinking trace, but the trace is a templated summary of the ground-truth structure, acts as a prior rather than a specification, and its property segment has little effect.

Questions for authors

Clarifications

  1. What does CrysReas add beyond the thinking trace relative to CrysTune and CRYSTAL? Section 5 states that prior RL works do not support property-conditioned generation, whereas CrysTune supports it with GRPO.
  2. Which DFT settings (functional, U values, corrections) were used in Fig. 3?

Suggestions

  1. Please cite and discuss CrysTune and CRYSTAL in Section 5.
  2. Scoring the generated structures with LeMat-GenBench (Betala et al., arXiv 2512.04562), as CrysTune does, would give an evaluation independent of the reward model and comparable to recent LLM generators.
3 comments
AuthorsOpenReview ↗

Author Rebuttal to Reviewer RaqX

We thank the reviewer for recognizing the improvements from thinking traces and the evidence from the trace-replacement experiments. Our responses clarify the contributions relative to CrysTune and CRYSTAL, provide the DFT settings used in Figure 3, and report property evaluation with two different potentials, MACE-MP-0 and SevenNet-Omni.

W, Clarification Q1 & Suggestion 1. Contributions beyond CrysTune and CRYSTAL, and the effect of the property segment.

We agree that the earlier statement about prior RL methods lacking property-conditioned generation was too broad. We recognize CrysTune and CRYSTAL as prior RL-based crystal generators, including the conditional optimization supported by CrysTune, and make this distinction explicit in the related-work discussion. Our contribution is structured intermediate reasoning about symmetry, coordination, and physical properties before coordinate generation. We additionally evaluate conditional optimization for continuous physical targets, including elastic moduli and thermal expansion.

As the reviewer mentioned, the trace intervention experiment (Table 2) provides evidence that the generated structures depend on the thinking traces. On the 256 prompts used in this experiment, replacing the symmetry/formula segment or the local-coordination segment with another sample's description reduces space-group consistency by 57.91 and 48.05 percentage points, respectively. The experiment therefore provides the strongest evidence for the role of symmetry and coordination descriptions in guiding generation. 

To further demonstrate the effectiveness of thinking traces, we additionally trained a Wyckoff+GRPO baseline to compare CrysReas with a Wyckoff generator that also receives online RL. On the same 8,141 test prompts with five candidates per prompt, CrysReas achieves higher space-group consistency (68.51% vs. 62.70%), stability (5.13% vs. 3.92%), and S.U.N. (4.58% vs. 3.49%), together with a lower mean hull distance (0.434 vs. 0.480 eV/atom). This comparison shows that CrysReas retains an advantage when the Wyckoff baseline also benefits from GRPO, strengthening the evidence beyond the original comparison with a Wyckoff SFT model.

MetricWyckoff+GRPOCrysReas
Structural validity (%)94.7795.58
Space-group consistency (%)62.7068.51
Stability (%)3.925.13
Mean hull distance (eV/atom)0.4800.434
S.U.N. (%)3.494.58

Clarification Q2. DFT functional, Hubbard-U values, and energy corrections used in Figure 3.

The calculations use VASP 6.4.3 with PBE and PAW_PBE potentials, a plane-wave cutoff of 700 eV, and a Γ-centered k-point mesh with KSPACING = 0.2 Å⁻¹. Calculations are spin-polarized. Variable-cell relaxation uses ISIF = 3, a maximum of 99 ionic steps, and a force threshold of 0.02 eV/Å. The electronic convergence threshold is 10⁻⁵ eV for most relaxations, with tighter thresholds in a small number of runs. Where applicable, Hubbard-U values are 3.32 eV for Co, 5.30 eV for Fe, 3.25 eV for V, 3.90 eV for Mn, and 3.70 eV for Cr. Hull distances use the MP/Alexandria/WBM reference phase diagram with MP2020 compatibility corrections.

Suggestion 2. LeMat-GenBench and evaluation independent of the MatterSim reward potential.

We evaluated the generalist and property specialists with two independent potentials, MACE-MP-0 and SevenNet-Omni, on 512 elastic-property prompts and 256 thermal-expansion prompts. The metrics are the joint bulk/shear-modulus target-hit rate and the thermal-expansion target-hit rate at 300 K. Failed calculations remain in the denominator.

TaskPotentialGeneralist (%)Specialist (%)Gain (percentage points)
ElasticityMACE-MP-041.4142.97+1.56
ElasticitySevenNet-Omni48.8352.93+4.10
Thermal expansionMACE-MP-021.8824.22+2.34
Thermal expansionSevenNet-Omni33.2034.77+1.56

Both specialists retain higher target-hit rates under both independent potentials, supporting some transfer beyond MatterSim, the reward potential used during training. Absolute hit rates are lower than under MatterSim. These experiments address evaluation independent of the reward model.

We have not used LeMat-GenBench, so we can not provide a direct comparison with CrysTune's published scores.

Thank you for the response and the additional Wyckoff+GRPO comparison. However, my main concern still remains the same. The SFT + GRPO training pipeline with MLIP-based rewards has been explored in prior work, the main new component is the thinking trace, and its property segment appear to have a limited effect on generation. Therefore, I will keep my score.

AuthorsOpenReview ↗

Reply to RaqX

We thank the reviewer for the reply. As RaqX said, one of our main contribution is the thinking traces (and the other is conditioning on continuous physical properties that require DFT or MLIP computation, which does not appear in prior works such as CrysTune or CRYSTAL). Although the property segment has a little effect, the other two parts are effective, as structural validity will decrease by ~10% and space group consistency decrease by ~50% (Table 2). We also do an experiment to show that thinking traces are shrinking the solution space, reducing the probability of generating invalid crystals (Table 3, number of clusters is decreasing). Moreover, thinking traces provide better results for structural validity and stability than Wyckoff representation (Table 1).

Scaling up thinking traces and RL is a very effective approach for large language models, and we believe this method can be utilized in material discovery, as there are large-scale internet data for materials, such as Materials Project, Robocrystallgrapher (what we use). Though the experiments may not resolve the reviewer's concern, the idea is still valuable for future works.

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

CrystalReasoner trains an LLM to reason about crystal symmetry, local coordination, and physical properties before generating atomic coordinates. The reasoning traces are constructed programmatically from known structures, followed by GRPO training with validity and stability rewards. The method also supports property-conditioned generation through specialized RL rewards using multi-objective RL. Experiments on MP-20 show improvements over LLM-based baselines, with ablations studying the contributions of reasoning and RL. The authors further evaluate trace consistency and perform DFT checks on generated structures.

Critical strengths

The main strength is bringing reasoning and RL together for crystal generation. The reasoning traces introduce structural knowledge before coordinate generation, while RL provides physical feedback to refine the outputs. The ablations suggest that these two components complement each other, improving validity, instruction following, and stability. I also appreciate the trace-swapping experiment, which provides some evidence that the intermediate reasoning actually affects the generated structures.

Critical weaknesses
  1. The comparison with prior methods is incomplete. PLAID++ uses its Wyckoff representation without its preference optimization stage, while CrystalReasoner benefits from additional RL. It is unclear whether the gain comes primarily from reasoning or simply stronger post-training. A comparison with a Wyckoff generator trained using the same RL setup would be more convincing.

  2. The materials discovery results are less convincing. Novelty is primarily evaluated against the training split, and the improvement becomes less conclusive when considering the larger materials database. Also, the property-conditioned results rely on the same surrogate potential used during training.

Questions for authors
  1. How does the Wyckoff baseline perform with the same RL rewards and training budget?

  2. Do the property-conditioned improvements hold under an independent potential or DFT?

  3. Can the authors clarify the discrepancy between the energy results in Table 1 and Figure 3, particularly for the RL-only model?

3 comments
AuthorsOpenReview ↗

Author Rebuttal to Reviewer pcrW

We thank the reviewer for recognizing the complementary contributions of reasoning and RL. The additional experiments address the omission of PLAID++'s RL stage, broader novelty references, and property-conditioned evaluation under independent potentials.

W1 & Q1. A Wyckoff generator trained using the same RL setup.

We trained Wyckoff+DPO, incorporating PLAID++-style offline preference optimization, and Wyckoff+GRPO. They use the same Wyckoff SFT initialization and reward definition. Both methods use the same 8,000 training prompts and generate eight candidates per prompt. DPO uses these candidates to construct offline preference pairs, while GRPO samples candidates online during training. All models are evaluated on the full filtered CDVAE test set (8,141 test prompts) with five candidates per prompt.

MetricWyckoff+DPOWyckoff+GRPOCrysReas
Structural validity (%)90.4094.7795.58
Space-group consistency (%)61.5562.7068.51
Stability (%)3.663.925.13
Mean hull distance (eV/atom)0.5270.4800.434
S.U.N. (%)3.333.494.58

CrysReas retains an advantage over both post-trained Wyckoff baselines. The Wyckoff comparison also shows higher performance with GRPO than DPO on these metrics.

W2 & Q2. Property evaluation with independent MLIPs and novelty against a larger database.

We evaluated the generalist and property specialists with MACE-MP-0 and SevenNet-Omni on 512 elastic-property prompts and 256 thermal-expansion prompts. Each model produces one candidate per prompt, and failed calculations remain in the denominator. The metrics are the joint bulk/shear-modulus target-hit rate and the thermal-expansion target-hit rate at 300 K.

TaskPotentialGeneralist (%)Specialist (%)Gain (percentage points)
ElasticityMACE-MP-041.4142.97+1.56
ElasticitySevenNet-Omni48.8352.93+4.10
Thermal expansionMACE-MP-021.8824.22+2.34
Thermal expansionSevenNet-Omni33.2034.77+1.56

Both specialists retain higher hit rates under both independent potentials, supporting some cross-potential transfer. Absolute hit rates are lower than under MatterSim, the potential used for reward calculation.

We also repeated structural novelty matching against MP+Alexandria. The table below reports novelty and S.U.N. against both reference sets, using the same stability threshold of E_hull ≤ 0 eV/atom for S.U.N.

Reference databaseMetric (%)Wyckoff+DPOWyckoff+GRPOCrysReas
SFT training structuresNovelty78.5081.8882.32
SFT training structuresS.U.N.3.3293.4914.579
MP+AlexandriaNovelty45.1446.2041.38
MP+AlexandriaS.U.N.0.1350.1570.209

The larger database lowers novelty and S.U.N. for all three models. CrysReas retains the highest S.U.N., indicating a higher rate of structures that jointly satisfy stability, uniqueness, and novelty under the broader reference. Its standalone novelty rate against MP+Alexandria is lower than those of the Wyckoff baselines, indicating CrysReas is able to improve structure quality (hence improve S.U.N.) but unable to improve the novelty, as the reward does not contain novelty.

AuthorsOpenReview ↗

Author Rebuttal to Reviewer pcrW (Part 2)

Q3. Why do hull distances and RL gains differ between Table 1 and Figure 3, especially for CrysReas-RL?

The larger RL improvement in Figure 3 comes from sample selection. The DFT evaluation prioritizes cases with larger improvements under MatterSim, using the score E_hull(CrysReas-Base) + E_hull(CrysReas-Thinking) − E_hull(CrysReas-RL) − E_hull(CrysReas). This selection makes the CrysReas-Base–CrysReas-RL difference more pronounced in Figure 3. Table 1 averages over the full test set, including cases with smaller improvements.

The difference in absolute hull distance comes changing the energy model and further relaxing the structures with DFT. VASP first evaluates the MatterSim-relaxed structures using DFT, then optimizes their atomic positions and lattice parameters on the DFT energy surface. For CrysReas-RL, a paired comparison gives mean hull distances of 0.354 eV/atom under MatterSim, 0.837 eV/atom under VASP static evaluation, and 0.150 eV/atom after VASP relaxation. The static-to-relaxed decrease shows the substantial effect of further structural optimization, explaining why Figure 3's relaxed DFT energies differ from the MatterSim estimates in Table 1.

Thank you for addressing my questions. My concerns are addressed and I will keep my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Rebuttal

We thank the reviewers for their thoughtful feedback and for recognizing several strengths of CrysReas:

  • Crystallographic thinking before coordinate generation. pcrW and jDwD recognize the value of incorporating structural knowledge through explicit thinking traces.
  • Measured improvements from thinking traces. RaqX highlights the gains in validity, composition and space-group consistency, stability, and S.U.N. under controlled comparisons.
  • Evidence that the generated structures depend on the traces. pcrW and RaqX recognize the trace-replacement experiments showing the effects of symmetry and coordination descriptions.
  • Complementary contributions of thinking and RL. pcrW and jDwD highlight the ablations separating these contributions and the extension to property-conditioned generation.

In response to the main concerns, we present the following additional evaluations and strengthen the evidence for CrysReas:

  1. A shared evaluation population and explicit metric scopes. All models are evaluated on the 8,141 retained MP-20 test prompts, with five candidates per prompt. We report the four input-condition groups, identify the applicable population for each metric, and explain the filtering from the original MP-20 split.
  2. Comparable post-training baselines. We include Wyckoff+DPO, incorporating PLAID++-style preference optimization, and Wyckoff+GRPO, which receives online RL. Both use the same Wyckoff SFT initialization, RL settings, and RL training budgets.
  3. Stable, unique, and novel structures under a broader reference. We evaluate novelty and S.U.N. against both the SFT training structures and MP+Alexandria. Under MP+Alexandria, CrysReas achieves a S.U.N. yield of 0.21%, compared with 0.14% for Wyckoff+DPO and 0.16% for Wyckoff+GRPO. This indicates that CrysReas's RL optimization increases the share of stable crystals, while the RL rewards contain no novelty component.
  4. Property-conditioning improvements under independent potentials and DFT settings explaination. We evaluate the effectiveness of property-conditioned specialists over two different potentials, MACE-MP-0 and SevenNet-Omni, and find they achieve higher target-hit rates than the generalist. We provide detailed DFT settings.
  5. Crystallographic reasoning and its measured effects. We specify the symmetry, site, coordination, and physical-property descriptions emitted in the traces and report a trace-prefix experiment measuring their effects on structural diversity and symmetry. We clarify the structural meaning of “atomic precision” and discuss CrysTune and CRYSTAL alongside our contribution of explicit intermediate crystallographic reasoning and property-conditioned generation.

Together, these evaluations show that CrysReas generates more valid, stable, symmetry-consistent structures and a higher stable-unique-novel yield than the post-trained Wyckoff baselines on the same conditional benchmark. The independent-potential and trace experiments further substantiate the property-conditioning and crystallographic-reasoning contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.