acceptodds
Under review as a conference paper at ICLR 2027

Reinforcement Learning on the Discrete Composition Channel of a Crystal Generator: Validated Gains and Reward Hacking

Abstract

Inverse materials design is a long-standing goal of computational materials discovery. Generative models for crystalline materials are typically trained to match the distribution of a structure database, while nothing in their training objective points them at specific design goals such as targeted properties. We use group-relative policy optimization (GRPO) to align a generative model based on stochastic interpolants and discrete flow matching with general black-box reward functions through reinforcement learning. Atom types are generated by a discrete flow and the policy gradient of our generalization of GRPO directly acts on the likelihoods of the atom-type transitions, which differentiates our work from previous reinforcement-learning approaches for diffusion and flow-based generative models of crystalline materials. We introduce a reward function that raises the yield of metastable, unique and novel structures (mSUN) from 13.4% for the pretrained model to 45.5% for the reinforced model, as evaluated by a community benchmark. Our reward also improves the performance of a reinforcement learning framework for crystalline materials based on latent denoising diffusion models. At the same time, we find that directly reinforcing atom-type transition likelihoods enables reward exploitation that has to be prevented with explicit guards. The same analysis also exposes a gap in the community metric. Single-element structures in distinct packings are counted as metastable, unique and novel materials and inflate mSUN without yielding any new compounds. A stability claim is only as good as its reference hull. We report every result split by the number of reference phases behind it and argue that benchmarks should do the same.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
reinforcement learning, GRPO, generative models, materials discovery, crystal structure generation, discrete flow matching, stochastic interpolants, reward hacking

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 5 reviews, above 27% of the others.
Reviewer_V7baOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper main focus is to show the reward hacking by RL models on LeMat-GenBench: the oversampling of unary materials and materials for which the reference hull has fewer than 12 structures on or near it. They design a reward function which prevents said reward hacking, and show that it works with OMatG and Chemeleon2.

Methodology-wise, they apply GRPO to the discrete flow-matching composition channel of a crystal generator, using exact per-event likelihoods so that the elements of a structure are actions of the policy in their native discrete.

Critical strengths

The papers addresses an important problem in the field. There is an overabundance of crystal generation models regurgitating DFT datasets and not so many real discoveries made. The proxy metrics used for benchmarking the generators are indeed deficient, and RL indeed hacks them.

On the low-level the paper appears sound - it reasonably measures what it claims to measure. Can’t be sure without the code though.

Critical weaknesses

I What, exactly, is wrong with single-element structures?

A. “face-centeredcubic Mg” in Figure 1 is an interesting example, as it both supports and undermines the authors’ point. An fcc Mg phase has been synthesised by Ji and Jona - so it is possible that the particular model-generated structure is not novel and is just absent from the DFT database. On the other hand, it shows that thermodynamically metastable single-element phases can be valuable enough to warrant experimental synthesis.

On top of that, LeMat-Bulk does contain fcc Mg, e. g. agm002159588 & mp-1056702. Could the authors please:

  1. Supply the CIF file for the fcc MG structure depicted in Figure 1
  2. Confirm that the Figure 1 fcc MG has been relaxed with UMA prior to novelty-scoring
  3. Use pymatgen.core.StructureMatcher to compare this supposedly novel phase against all the Mg entries in LeMat-Bulk

The reason I’m asking is because a novel fcc Mg can be registered in a few ways, it’s important to distinguish between them:

  1. A novel local energy minimum for Mg system
  2. BAWL fingerprint bug
  3. A configuration of Mg atoms which is novel but is not at the local energy minimum

B. Multi-element compounds also have a lot of metastable local energy minima

https://www.science.org/doi/10.1126/sciadv.1600225

https://doi.org/10.1063/1.3079326

C. The burden is on authors to show why discovering new single-element metastable local minima is “meaningless”, but discovering mutli-element metastable local minima is not. As the authors correctly write in the Related Work section, RL can produce many metastable generations which are duplicates or substitution variants of training structures (Negishi & Walsh, 2026), and, quoting (Negishi et.al), “While it is natural that the sample distribution becomes biased toward a specific area of material space after RL, the generation of 900 or 611 novel, stable structures for the same composition is unrealistic.”

II Sparse hull impact

On a hull with only a few references a value near zero says only that the structure is no worse than the few phases it was compared against, and every benchmark that scores stability against a reference set inherits this limit.

The number of structures on the hull depends on the chemical system (it doesn’t have to be the same for all), not only on the arbitrary database construction procedure.

On top of this, there is a purely combinatoric aspect. The 12 threshold for near-hull reference phases is the same for all arities - practically impossible to satisfy for unary systems, and easily satisfiable for quarterly (4 unary + 6 binary already give 10).

The authors show a correlation and a direction, but they don’t prove the magnitude of the effect. Had the true hull been known, how many of the structures scored as metastable would turn out to actually not be metastable? I see two possible ways to address this:

  1. Take some chemical systems with sparse hulls, generate a large number of structures in those chemical systems, relax with a MLIP, add them to the MLIP-based hull. MatterGen and WyFormer can do this out of the box. // Just in case, LeMat used MLIPs to compute single point energy of DFT-relaxed structures - it's likely that MLIP-relaxed structures will have slightly lower energy, this needs to be accounted for
  2. Take some chemical systems with dense hulls, exclude some structures on the hull, check how it affects the stability verdicts. This is cheap but random removal won’t settle the question

Table 1 contains the SUN counts, 8-72 out of 2500 generated structures. This doesn't conclusively proof anything, but intuitively one would expect that if the database hull was somehow deficient, the models would have identified a large number of structures below it as well.

III DFT

This is not critical and doesn’t have a big impact on Accept/Reject, but you still might want to fix.

DFT settings are wrong, MP used MPGGADoubleRelaxStaticMaker, not a single MPRelaxSet+MPStaticSet. The impact is likely not catastrophic: Miller et al. (2024) had the same mistake, see Table 1 and Section H.1, in Kazeev et al. (2025).

DFT spot checks only include the structures with MLIP E_hull < 0.1 eV/atom, missing errors where MLIP overestimates the energy.

IV Minor bits

They are not critical and don’t have an impact on Accept/Reject, but you still might want to fix.

“known elements” is a weird phrase. What is the alternative, ununennium-119?

Figure 3 caption has weird commas: “a,”

Questions for authors

See Critical Weaknesses

11 comments

We thank the reviewer for this expert assessment. It has led to the most extensive changes in the manuscript, which are described below. It also made the paper scientifically more sound, because the revision now discusses what a new material means and states that an improvement in a benchmark does not always translate into new materials. Section numbers refer to the general response. All values are the re-scored values (Section 1).

I. Regarding single-element structures

We agree that single-element structures are not meaningless. The wording of the submission was unfair, and we take responsibility for it. We see this work as a step toward the inverse design of bulk materials with properties of interest for applications. Such properties arise in compounds more often than in elemental structures. The single-element guard is therefore a restriction of the task to compounds and not a statement about the value of elemental phases.

There is also a chemical reason. For an element, the convex hull consists of the ground state alone, so EhullE_\mathrm{hull} measures only the energy difference to that one phase. For many metals a large family of stacking variants and distorted cells of the close-packed structures lies within 0.1 eV/atom of it. The metastability criterion is therefore less selective for elements than for compounds, which also have to compete with phases at other compositions.

We observed that in the MP-20 training set, 0.8% of the structures are single-element (226 of 27,136). In the training run without the guard, 52.9% of the mSUN structures are single-element (499 of 944). With the guard, the share is 0.4% (5 of 1,158). Without the guard the policy moves far from the training distribution, toward the structures for which the reward is easiest to obtain.

Change in the manuscript. The words "meaningless" and "repackings" are removed. The guard is described as a restriction of the task to compounds. Changes are in the abstract, the caption of Fig. 1, Sec. 3 and Sec. 4.4.

I.A. The fcc Mg example

Response. The assessment of the reviewer is correct. We thank the reviewer for identifying this error, and we apologize for it. The example has been removed.

The three requests.

  1. The CIF file is given in full at the end of this comment.
  2. The structure was relaxed with UMA (FIRE with a cell filter, force tolerance 0.05 eV/Å, at most 500 steps). The log of this relaxation was not retained, so we cannot give the number of steps. The recomputed maximum force is 0.020 eV/Å. The structure came from our internal evaluation, which tests novelty against the MP-20 training set only. The benchmark never scored it. Presenting it next to benchmark values was our error.
  3. We compared the structure with all 38 Mg entries of LeMat-Bulk using StructureMatcher, with the default settings of pymatgen and with the settings of the benchmark (ltol=0.1). Under both it matches two entries, mp-1056702 and agm002159588, which are the two that the reviewer names. We also applied the novelty test of the benchmark to the file. It classifies the structure as not novel.

Further checks. The structure has space group Fm3ˉmFm\bar{3}m at a symmetry tolerance of 0.05 Å and a lattice constant of 4.494 Å. The benchmark set of the same model contains a corresponding fcc Mg structure, and the benchmark classifies it as not novel as well. In DFT the structure of Fig. 1 remains fcc and lies 12 meV/atom above hcp Mg at the densest k-point mesh.

CIF file of the fcc Mg structure of Fig. 1 of the submission, as relaxed with UMA (request 1 of point I.A). The cell is the generated 10-atom cell. Its space group is Fm3ˉmFm\bar{3}m at a symmetry tolerance of 0.05 Å.

# generated using pymatgen
data_Mg
_symmetry_space_group_name_H-M   'P 1'
_cell_length_a   4.49781729
_cell_length_b   9.53368363
_cell_length_c   6.35267752
_cell_angle_alpha   120.07390771
_cell_angle_beta   90.05783947
_cell_angle_gamma   76.40423149
_symmetry_Int_Tables_number   1
_chemical_formula_structural   Mg
_chemical_formula_sum   Mg10
_cell_volume   226.91099134
_cell_formula_units_Z   10
loop_
 _symmetry_equiv_pos_site_id
 _symmetry_equiv_pos_as_xyz
  1  'x, y, z'
loop_
 _atom_site_type_symbol
 _atom_site_label
 _atom_site_symmetry_multiplicity
 _atom_site_fract_x
 _atom_site_fract_y
 _atom_site_fract_z
 _atom_site_occupancy
  Mg  Mg0  1  -0.15832017  0.58505558  0.82277210  1
  Mg  Mg1  1  0.64138158  0.98535408  0.62434489  1
  Mg  Mg2  1  0.44168075  0.38503127  0.42222510  1
  Mg  Mg3  1  0.24158093  0.78489656  0.22312820  1
  Mg  Mg4  1  0.04208248  0.18469505  0.02118603  1
  Mg  Mg5  1  0.24169081  0.78482200  0.72301490  1
  Mg  Mg6  1  0.84181604  0.58495946  0.32265772  1
  Mg  Mg7  1  0.44174168  0.38496994  0.92210234  1
  Mg  Mg8  1  0.64154264  0.98523574  0.12417282  1
  Mg  Mg9  1  0.04217532  0.18455470  0.52099794  1

I.A, continued. The three ways in which a single-element structure can be registered as novel

The reviewer lists three ways in which a novel fcc Mg can be registered. For the structure of Fig. 1 the answer is the possibility the reviewer raised first, that it is not novel. It is the known fcc phase and was novel only with respect to the MP-20 training set, which our internal evaluation uses and not in LeMat-GenBench. We should have stated this in the submission.

To answer the question for the structures that the benchmark does count, we examined every single-element structure that it classifies as novel, in five structure sets (1,531 structures). Each was relaxed with UMA to a tight tolerance (0.005 eV/Å on atomic and cell forces), then perturbed three times and relaxed again. A structure is a known phase if it matches a LeMat-Bulk entry of its element. It is not a local minimum if it leaves its geometry or lowers its energy by more than 5 meV/atom. The rest are candidate new minima. These criteria were fixed before the calculations.

The table refers to the training run without the single-element guard (left side of Fig. 1), which has 499 single-element structures in mSUN.

ClassStructures
Known phase (matches a LeMat-Bulk entry)140 (28%)
Not a local energy minimum, case (3) of the reviewer231 (46%)
Candidate new local energy minimum, case (1) of the reviewer128 (26%)

Case (2) does not occur. No structure that the benchmark classifies as novel matches a LeMat-Bulk entry under the settings of the benchmark. Our evaluation uses StructureMatcher and does not use BAWL. For the candidates, DFT calculations on ten of them retain the geometry of all ten, at 3 to 138 meV/atom above the ground state of the element. The perturbation test is not a phonon calculation, which is why we call them candidates. With the guard, OMatGRPO has 8 single-element structures that the benchmark classifies as novel (5 known phases, 2 not local minima, 1 candidate).

Fig. 1 now shows a Sn structure of the run without the guard. The benchmark counted it in mSUN, and after tight relaxation it matches the known body-centered tetragonal Sn entry mp-55 (I4/mmmI4/mmm). The caption names the training run and the evaluation. A new appendix table reports the classification for the five sets.

I.B and I.C. Metastable minima of compounds, and substitution variants

We agree that compounds also have many metastable local minima. Metastable configurations of compounds can carry properties of interest. One example is the superconductivity created in WB$_2$ by pressure-induced metastable planar defects (Lim et al., Nat. Commun. 13, 7901 (2022), https://doi.org/10.1038/s41467-022-35191-8). We focus on compounds for this reason.

We applied the criterion that the reviewer quotes, the number of novel and stable structures of one composition.

Number of mSUN structures
With the guard, largest number of structures sharing one formula8 (1,153 compounds over 1,071 formulas)
Without the guard, largest number of structures of one element73 (Sn)

Of the 73 Sn structures, 13 are known phases, 41 are not local minima and 19 are candidates. Compounds are less concentrated, although the chemical system with the most mSUN structures of OMatGRPO, Hg-K, has 38.

We also classified the mSUN structures against the MP-20 training set with the released code of Negishi and Walsh.

ModelSubstitution variants of training structures
Pretrained model48.2%
OMatGRPO68.0%
Chemeleon280.3%

Two thirds of the mSUN structures of OMatGRPO are substitution variants by this definition. The classification follows their pipeline, which relaxes each structure with MACE before the comparison. The revision will reports both tables in the appendix and does not present mSUN as a number of discoveries.

Regarding the single-element structures, I think it would be useful to view the issue at two levels.

Material discovery I and the authors appear to be in agreement that both single-element and multi-element (meta)stable structures can be valuable novel materials - but also that formal MSUN wrt a DFT database doesn't gurantee that.

There is also a chemical reason. For an element, the convex hull consists of the ground state alone, so measures only the energy difference to that one phase. For many metals a large family of stacking variants and distorted cells of the close-packed structures lies within 0.1 eV/atom of it. The metastability criterion is therefore less selective for elements than for compounds, which also have to compete with phases at other compositions.

This qualitive assertion requires evidence and loops us back to "what counts as a new material". Graphite and diamond, martensitic and austenitic steels, α\alpha, γ\gamma, δ\delta -iron share chemical composition and are genuinely different real materials. On the other hand, ordered realizaitons of the same disordered material can have different chemical compostions.

ML model proxy metrics Now, we don't have to precisely solve what counts as discovery to advance machine learning.

We observed that in the MP-20 training set, 0.8% of the structures are single-element (226 of 27,136). In the training run without the guard, 52.9% of the mSUN structures are single-element (499 of 944). With the guard, the share is 0.4% (5 of 1,158). Without the guard the policy moves far from the training distribution, toward the structures for which the reward is easiest to obtain.

I agree with the authors here. A dramatic collapse in the diversity of chemical compositions is a bad trait for a generative model. Also, the observation that the reward is easiest to obtain on single-element structures is interesting; qualitively, one might expect that uniqueness & novelty would suffer.

I.A. The fcc Mg example

Good, I'm glad that it's settled.

Reviewer_SeBmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper takes the pretrained OMatG model and applies GRPO on top of it, extending OMatG-IRL (which reinforced only positions and lattice) by also reinforcing the discrete atom-type channel directly through the masked discrete flow-matching commit probabilities, rather than through a latent space or with a fixed composition. The reward is largely assembled from previous works(Chemeleon2, MatInvent) plus new guards (sparse-hull abstention, single-element penalty, below-hull) that penalize meaningless crystals. Evaluation is on LeMat-GenBench, where mSUN rises from 13.4% for the prior model to 45.5%.

Critical strengths
  • Investigated the quality of crystals generated under naive RL rewards and identified meaningless ones (single-element repackings, sparse-hull artifacts).
  • DFT validation of MLIP stability claims.
  • Composition is directly influenced by the reward in this work.
  • Reports the number of reference phases behind every generated structure and tiers results by it (Sec. 4.5).
  • Addresses reward hacking, one of the common failure cases of RL training, through domain-driven penalties, and successfully reduces meaningless and single-element crystals.
Critical weaknesses
  • Incremental contribution- OMatG-IRL already applied RL to OMatG; the addition here is the species channel and a reward built mostly from Chemeleon2/MatInvent components plus guards.
  • Missing several baselines- No comparison to MatterGen, DiffCSP, FlowMM or many other generators. Only comparisons and ablations and Chemeleon2 are included.
  • The reward uses UMA and the LeMat-GenBench’s 2-of-3 MLIP ensemble includes UMA. An appendix table with UMA removed/replaced would be better.
  • Significant drop in validity. Quite concerning that the validity drops significantly (90+% → ~70%) compared to the baselines despite the reward choices. Authors have not fully acknowledged or addressed. Besides, most percentages (in LeMat-GenBench) are w.r.t. the valid crystals. This is acknowledged only in the captions. A table in the appendix with 2500 as the denominator in percentages for all metrics could be reported.
  • No RL training curves (individual and total reward terms, loss, or mSUN vs. iteration). At present there are no analyses/plots to study the learning behavior of the model during RL training.
  • Stability–novelty trade-off is not studied, unlike Chemeleon2. Besides the explanation provided in the paper, it would be great if the authors do a deeper analysis to see how the additional guards in the reward affect novelty and diversity (since we’re restricting the exploration space) .
  • Not all metrics provided by LeMat-Genbench are mentioned in the main part: e.g. diversity, relaxation RMSD. From table 8 , it looks like Chemeleon2 has better Eform and RMSD than the OMat-GRPO.
  • Benchmark results are only models pre-trained on MP-20 and the authors have not shown how the benefit of design choices and the proposed reward function scale with pre-trained data (e.g. Alex-MP20) - Appendix C only contains an ablation study starting with OMatG pre-trained on larger data.
  • Every rollout is relaxed before scoring, so the policy is not directly rewarded for raw geometries. There is no ablation of this in their own pipeline (no Eq. 5 + OMatG run without relaxation), and the swap runs in Table 7 show relaxation helping in one pipeline and hurting in the other, so the choice is not justified yet.
  • Clarity. It should be made clearer early on that the paper improves on OMatG(-IRL) and Chemeleon2. The labels “prior” and “discovery” are defined only in table/figure captions. Additionally, several sections could benefit from clearer writing and presentation.
Questions for authors

How does the novelty increase after GRPO especially when the novelty reference used in the training is a much smaller set compared to LeMat-GenBench?

4 comments

Response to Reviewer SeBm (1 of 3)

We thank the reviewer for the detailed list of weaknesses, and we ran every experiment that the review requests. "Section N" refers to our general response and "Sec. N" to the paper. All values are from the corrected benchmark version (Section 1).

Weakness 1. Incremental contribution

We agree that the work builds on OMatG-IRL, which applied RL to the positions and the lattice of OMatG, and that the reward combines components of Chemeleon2 and MatInvent with our guards. The addition of the paper is RL on the discrete atom-type channel, with the exact likelihood of every atom-type transition. A matched control in which RL on the atom-type channel is switched off remains at the 357 mSUN structures (14.3%) of the pretrained model, against 1,158 (46.3%) for OMatGRPO (Section 2 in general comment). In our pipeline the gain is therefore obtained largely through RL on the atom types. The revised introduction states at its beginning which parts come from the three earlier works and which part is new.

Weakness 2. Baselines

We evaluated the released MP-20 checkpoints of MatterGen and DiffCSP with the same protocol as all other models (2,500 generated structures, relaxation with UMA, same benchmark version).

ModelValidmSUN
Pretrained OMatG model94.1%14.3% (357)
DiffCSP96.7%17.9% (447)
MatterGen96.1%23.4% (584)
Chemeleon294.9%37.4% (936)
OMatGRPO, five training runs66.8% to 70.8%34.1% to 46.3%
OMatGRPO with charge-neutrality penalty, three training runs96.6% to 98.6%53.2% to 66.8%

MatterGen and DiffCSP are pretrained generators without RL, so they correspond to the pretrained OMatG model. The charge-neutrality penalty is described in Section 5 of general comment. FlowMM is not included because its authors have not released a checkpoint. The revision adds both baselines in a new appendix table.

Weakness 3. UMA in the reward and in the benchmark

We evaluated the same structure sets again with the Orb and MACE potentials only.

ModelmSUN, three potentialsmSUN, Orb and MACE only
Pretrained model357360
Chemeleon2936946
OMatGRPO1,1581,171

mSUN changes by at most 13 structures, and the ordering of the models is unchanged. The benchmark installation contains no fourth potential, so UMA was removed and not replaced. All structure sets were relaxed with UMA before the benchmark, and this test does not remove that dependence. The description of the stability criterion as agreement of two of three potentials came from our submission and was inaccurate, because the benchmark applies the threshold to the mean EhullE_\mathrm{hull} over the potentials (Section 1). The revision adds the table to the appendix and corrects Sec. 4.1.

Weakness 4. Validity

We agree that the submission did not address the decrease in validity. The decrease comes from the charge-neutrality check of the benchmark. OMatGRPO generates anion-rich oxides, halides and chalcogenides that fail this check, while more than 99% of its structures pass the two geometry checks. A penalty for such structures in the reward restores the validity to between 96.6% and 98.6% and raises mSUN to between 53.2% and 66.8% in three training runs (Section 5).

The requested table with 2,500 as the denominator for all metrics is added to the appendix for every model, together with the three validity checks. Four of its rows are given here.

Fraction of the 2,500 generated structuresValidUniqueNovelMetastablemSUN
Pretrained model94.1%93.6%77.2%28.3%14.3%
Chemeleon294.9%86.0%64.6%66.5%37.4%
OMatGRPO69.8%68.9%61.6%54.2%46.3%
OMatGRPO with charge-neutrality penalty96.6%95.2%89.4%60.1%53.2%

The revision explains the decrease in Sec. 4.2 and states in the main text, and not only in the captions, which values are fractions of the valid structures.

Response to Reviewer SeBm (2 of 3)

Weakness 5. Training curves

We evaluated the benchmark at four intermediate checkpoints of the OMatGRPO training and of the training with the charge-neutrality penalty. Each row is one training run, and iteration 0 is the pretrained model.

Iteration0150300450600750
OMatGRPO, mSUN14.3%18.8%31.1%31.9%37.2%46.3%
OMatGRPO, valid94.1%86.8%80.4%75.1%66.1%69.8%
With charge-neutrality penalty, mSUN14.3%22.4%28.2%34.5%41.2%53.2%
With charge-neutrality penalty, valid94.1%96.4%97.8%96.9%98.8%96.6%

mSUN rises throughout both trainings. The validity of OMatGRPO falls as mSUN rises, whereas with the penalty it stays above 96%.

The revision adds a figure with the training curves of OMatGRPO in the appendix. It shows the total reward, the creativity term, the KL term of the loss and mSUN against the training iteration. The mean total reward rises from 0.87 in the first 25 iterations to 1.06 in the last 25. The stability term and the MMD bonus were not logged separately during training, so the figure shows the creativity term as the only individual reward term.

Weakness 6. Stability and novelty

The revision adds the plot that corresponds to Fig. 4c of Chemeleon2, with one point per model. Both columns of the table below are fractions of the valid structures.

ModelNovelEhull≤0.1E_\mathrm{hull} \leq 0.1 eV/atom
Pretrained model82.1%31.4%
DiffCSP68.1%44.3%
MatterGen74.6%45.3%
Chemeleon268.1%73.3%
Our reward in the Chemeleon2 pipeline (A′)71.7%74.4%
OMatGRPO88.3%78.1%
OMatGRPO with charge-neutrality penalty92.5%62.8%

In our pipeline the fraction of structures within 0.1 eV/atom of the hull rises from 31.4% to 78.1% during training, while the novel fraction stays between 82% and 88%. This is a comparison across models. We did not vary a reward weight, so it is not a trade-off curve for one model. The meaning of benchmark novelty is discussed in Section 6.

For the effect of the guards, we compare the training without the single-element guard with OMatGRPO.

TrainingNovelUniqueElement diversity
Without the single-element guard73.9%84.1%0.669
With the single-element guard (OMatGRPO)88.3%98.7%0.661

The guard does not lower novelty or uniqueness, and the element diversity of the benchmark is nearly the same. Without the guard, 52.9% of the mSUN structures are single-element structures, and both novelty and uniqueness are lower than with the guard.

Weakness 7. Further metrics, EformE_\mathrm{form} and RMSD

The reviewer is right that Chemeleon2 has the more favorable values for both quantities. The revision adds a table with all benchmark metrics to the appendix.

ModelMean EformE_\mathrm{form} (eV/atom)RMSD of the benchmark (Å)Displacement under our relaxation (Å)
Pretrained model-0.740.460.95
MatterGen-0.790.250.14
DiffCSP-0.770.320.27
Chemeleon2-0.810.180.24
A′-0.840.150.18
OMatGRPO-0.060.310.99

Formation energy. The value near zero has a chemical cause. OMatGRPO moves toward alloys of chemically similar metals, and 96.3% of its mSUN structures contain metallic elements only, against 44.0% for the pretrained model. Such systems have a shallow hull, so an ordered arrangement on a close-packed lattice lies within 0.1 eV/atom of it almost by construction. We regard this as the same kind of exploit as the single-element structures, one step further, and the revision reports it as such in Sec. 4.3 (Section 6 and our response to Reviewer P1HN). The same reward in the Chemeleon2 pipeline gives -0.84 eV/atom, so the shift belongs to the combination of our reward with our pipeline.

RMSD. The benchmark measures its RMSD after our own relaxation, so it is a residual. The last column is the mean displacement of the atoms between the generated structure and the structure after our relaxation with UMA, without the strain of the cell. It is about 1 Å in our pipeline before and after RL, against 0.1 to 0.3 Å for the other generators. RL therefore does not improve the unrelaxed geometry of OMatG. With the native sampler of OMatG (710 steps in place of the 64 steps used in training and evaluation), the pretrained model gives 0.49 Å on 1,000 structures, so about half of the displacement comes from the short sampler. The revision states this limitation next to the metric table. The difference to Chemeleon2 is therefore already present in the pretrained models and is not created by RL. It reflects the generator and its sampler, and we have not separated the two.

Response to Reviewer SeBm (3 of 3)

Weakness 8. Pretraining on Alex-MP-20

We trained OMatGRPO with the same settings from the OMatG model pretrained on Alex-MP-20. The novelty reference and the composition reference of the reward were changed to the Alex-MP-20 training set.

ModelValidmSUN
Model pretrained on Alex-MP-2092.3%30.2% (756)
After RL91.8%56.7% (1,417)

mSUN rises by a factor of 1.9, compared with 2.4 to 3.2 for the five training runs on MP-20, and the validity does not decrease. This is one training run. The revision reports it in a new appendix table and in Sec. 4.2.

Weakness 9. Relaxation before scoring

We ran the requested ablation, which is the reward of Eq. 5 in our pipeline with all terms computed on the generated structure without relaxation.

TrainingValidmSUN
Without relaxation before the reward99.6%21.2% (531)
With relaxation (OMatGRPO, five training runs)66.8% to 70.8%34.1% to 46.3%

Without relaxation, the creativity term equals 1 for nearly all structures, because an unrelaxed structure rarely matches a known one, and the term then gives no learning signal. The generated compositions also concentrate on a few alkali elements, with 27 distinct elements in the 2,500 structures against 79 for OMatGRPO.

We agree that the policy is not rewarded for the unrelaxed geometry. The generated structures of OMatGRPO remain about 1 Å from their relaxed geometry (Weakness 7). This distance is also consistent with the opposite effect of relaxation in the two pipelines of Table 7. The structures of the Chemeleon2 pipeline are 0.2 Å from their relaxed geometry, so that scoring them without relaxation loses little. When the Chemeleon2 pipeline is trained with relaxation before the reward, this distance grows to 0.83 Å and mSUN falls from 44.3% to 25.3%. The revision reports the ablation in App. F and states the limitation.

Weakness 10. Clarity

We agree. The revised introduction states early that the paper builds on OMatG-IRL, Chemeleon2 and MatInvent. The labels "prior" and "discovery" and the names of the controls are now defined once in Sec. 4.1 and no longer only in the captions. The revision also clarifies the wording of Sec. 3 and Sec. 4.

Question. Increase in novelty

The benchmark compares a generated structure only with reference entries of the same reduced formula. We therefore sorted the valid structures into three groups.

Fraction of the valid structuresPretrained modelOMatGRPO
Formula absent from LeMat-Bulk39.6%40.9%
Formula present in LeMat-Bulk, structure not matched42.5%47.4%
Structure matched with a known entry17.9%11.7%

The first two groups are novel for the benchmark. The fraction with a formula that is absent from LeMat-Bulk is nearly the same before and after RL. The increase in novelty comes from structures of known formulas that match no entry of that formula, and from fewer matches with known entries.

The smaller reference of the reward is not the reason for the increase. A training run without the creativity term, and therefore without any novelty test in the reward, still reaches 40.7% mSUN (1,017 structures). A structure whose formula is absent from the reference is counted as novel without a comparison of its structure, and many of these structures are alloy compositions on a known lattice. For this reason the revision no longer presents novelty in the benchmark as evidence of new materials (Section 6). The revision adds the table above to the appendix and updates Sec. 4.3.

Robustness and composition of the mSUN gain, and implications for benchmarks of generative materials models

The experiments that the reviewer requested support the main result. The gain of RL on the atom-type channel over the pretrained model is present in all five training runs (factor 2.4 to 3.2), without UMA in the benchmark, without the creativity term in the reward, and for a model pretrained on Alex-MP-20 (factor 1.9). The decrease in validity is removed by a penalty that raises mSUN further. The closer analysis also showed that a large part of the gain consists of alloy compositions on the lattice of a known phase, which the benchmark counts as novel. This is not specific to OMatGRPO, a known phase of the same chemical system has the same lattice for 66.2% of the mSUN structures of OMatGRPO, 70.5% for Chemeleon2, 49.2% for DiffCSP and 48.1% for MatterGen (Section 6). We consider both results useful for the community. The first is a method that applies RL directly to the composition of a crystal generator, with the exact likelihood of every atom-type transition. The second is a limit of the current metric, which counts a new composition on a known lattice as a new material. Methods that optimize mSUN directly are appearing quickly, and the checks of this study show how much of such a gain remains when known lattices and alloy classes are taken into account.

Reviewer_P1HNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper applies GRPO to OMatG, including its masked discrete atom-type channel through exact categorical likelihood ratios at unmasking events. Its reward combines relaxed hull distances, novelty, uniqueness and composition diversity with guards against sparse-reference hulls and single-element outputs. Reported LeMat-GenBench mSUN increases from 13.4% to 45.5%. Transferring the reward to Chemeleon2 raises its mSUN from 36.6% to 43.7%. The paper also documents reward exploitation and provides DFT spot-checks.

Critical strengths
  • Eqs. 4/10 and the species KL (Eq. 11) are correct under the stated transition model. Reinforcing the species channel directly is a concrete difference from crystal RL work that freezes, relaxes or decodes compositions. Policy gradients for discrete flows already exist (DoMinO, dFlowGRPO, Graph-GRPO, cited in §2), so the contribution lies mainly in the application and the analysis.
  • The gain over the prior is large: from 336 to 1,138 of 2,500 nominal draws. The nominal denominator already charges invalid outputs. Without the creativity term the run still reaches 982, about 80% of the gain.
  • The exploit catalog, the reward swap and the single-element finding should help others who design rewards for crystal generators or report mSUN. The appendices disclose single-seed runs, the unseeded Chemeleon2 sampler, the DFT exclusions and per-structure DFT values.
  • The Chemeleon2 baseline is the released checkpoint scored under the same protocol, and I see no sign that it is under-performing. The authors are fairly honest about the baseline setup.
Critical weaknesses
  1. The chemical substance of the gains needs to be established. The single-element guard is motivated by the observation that a repacked element relaxes near its own hull and counts as novel without being a relevant new material (lines 126–129). The same argument applies to ordered dilute alloys (an elemental packing with a few substituted, chemically similar atoms) and to orderings of mutually soluble rare earths. Their formation energies are near zero, so Ehull is small almost by construction. A guard that counts distinct elements cannot exclude them. The paper's own data suggest the policy moved in this direction:

    • Table 8 gives mean formation energies on the valid set of −0.733 (prior), −0.804 (Chemeleon2), −0.837 (A′) and −0.055 eV/atom (OMatGRPO). Eform − Ehull estimates the mean hull depth at the generated compositions: about −0.93, −0.88, −0.91 and −0.13 eV/atom, respectively. The fixed 0.1 eV/atom metastability window is therefore comparable to the hull depth for OMatGRPO, but only about a tenth of it for the other methods.
    • In Table 11, 19 of the 22 trusted OMatGRPO spot checks are all-metal and 11 have one element at ≥70% (e.g., HgK19, CdMgTl16, Cu3Zn17, ErMg9Y, Li5Zn, Nd8TmY). Seven consist only of rare earths/Sc/Y (AcLaNdTm, ErNd7SmTm, Er2Sc8SmTm, HoNdPrTb5, LuPr2Sm5, LuSmY3, Nd8TmY). Six of the seven sparse spot checks are dilute metallic structures, while none of the four prior trusted spot checks are. The two “DFT-confirmed compounds” shown in Fig. 1 (ErMg9Y, HoMgTmY9) belong to this class.
    • Several of these are trusted on the UMA count but have only 4–8 near-hull phases on the DFT hull (LuPr2Sm5, LuSmY3, AcLaNdTm, Nd8TmY). Near-degenerate orderings may inflate the 1 meV/atom count in mutually soluble systems.

    The DFT sample is stratified by element count and is not an unbiased sample of the mSUN set, and I am not arguing that such structures are worthless. Agreement between UMA and DFT validates the energies, but near-hull energies are expected for dilute alloys, so the agreement does not show that these are new compounds. Please report, for the prior, Chemeleon2, A′ and OMatGRPO mSUN sets, the all-metal fraction, the distribution of the largest element fraction, the rare-earth/Sc/Y-only fraction and the Eform distribution. Please also report mSUN after excluding these classes, or the substitution-derived fraction under the workflow of Negishi & Walsh (2026), which the paper cites. If the margin over the prior and Chemeleon2 survives, the main claim stands. Otherwise, I would lean toward the view that the way the paper counts its results has a loophole, and the credibility and usefulness of the results would be in question.

  2. The controls do not isolate composition-channel RL, and “best overall” depends on the metric.

    • The frozen-composition control (Appendix F, lines 1378–1386) changes several things at once. Compositions are shared within a group, the reward keeps only the stability term, relaxation moves positions only, and the channel weights change from 0.1 to 1. The reasons are disclosed, but 45.5% versus 9.8% compares two recipes. OMatG-IRL (Appendix K, Table 6) shows only a small DNG gain with frozen species (MSUN from 16.9% to 18.1–18.2%). This supports the need to change the composition distribution. It does not show that reinforcing the discrete channel directly is required, since A′ (latent diffusion, frozen decoder) reaches a comparable level. The paper's claims may therefore be overstated.
    • Best-of-N ranks by stability alone (lines 1388–1395), whereas mSUN requires novelty, so a selection dominated by rediscoveries (11.5% novel) is expected. The 47,795 samples roughly match the RL rollout budget (750 × 64). Re-ranking them with the full reward (stability, creativity and in-set uniqueness) requires no new generation and is the budget-matched control that lines 113–115 need.
    • mSUN in the pinned benchmark excludes on-hull SUN structures (Fig. 2a notes “+72 SUN, on-hull” for best-of-N). Adding the SUN counts of Table 7 gives 1,145 for both OMatGRPO (1,138 + 7) and A′ (1,093 + 52). On trusted hulls (mSUN × (1 − sparse share), Tables 7–8, approximate because the sparse shares are rounded), OMatGRPO reaches ≈34.8%, released Chemeleon2 ≈33.7% and A′ ≈42.8%. By the paper's own standard that sparse-tier mSUN marks candidates (lines 483–484), A′ is ahead.
    • The transfer result depends on the relaxation convention. With the reward as defined in Sec. 3.3, which relaxes before scoring, Chemeleon2's pipeline reaches 23.5% (A), below the released 36.6%. Only without relaxation does it reach 43.7% (A′; lines 1125–1127). A′ is single-seed, uses an unseeded sampler and ran the full horizon without early stopping (lines 1118–1122). The Table 2 caption discloses the convention, but the abstract's statement that the reward improves Chemeleon2 should carry this qualification.

    Please either add a matched ablation (full reward, species channel frozen, independent per-rollout compositions, same relaxation and channel weights) together with a full-reward best-of-N, or narrow the claims to the tested recipes and drop or qualify “best overall performance” (lines 119–121).

  3. The benchmark implementation does not match the described protocol. I checked the pinned commit 58e6eae3. The preset sets MACE to model_type: mp with the mace_mp hull, and MACECalculator calls mace_mp(device=...) without a model, so mace-torch 0.3.13 loads medium-mpa-0 by default. The maintainers' README now states that the mace_mp reference hull was computed with medium-0b3 and that all earlier mace_mp-hull-based SUN/MSUN results “mixed two energy scales”. Fix fbf1ba48 (2026-09-15) pins medium-0b3. In addition, SUNMetric and SUNNewMetric threshold e_above_hull_mean, the arithmetic mean over available MLIPs. min_mlips_required = 2 counts available models, not agreeing votes, so the “two of three must agree” description (lines 312–314; Table 3) is inaccurate. All methods share the evaluator, but OMatGRPO's set is chemically distinct from the others (Weakness 1), so a chemistry-dependent MACE offset, one third of which enters the mean, could change the ranking. Please re-score the saved structures with the fixed benchmark and correct the description. No retraining is needed.

Minor comments

  • The DFT tier and agreement definitions are explicit and the counts are consistent. Although 35/67 UMA-trusted cases have fewer than 12 DFT near-hull references, applying that additional filter still gives 30/32 agreement, and 12/13 for the final model. This is a useful sensitivity summary, not evidence that the original tier labels are wrong. Use “threshold classifications agree” consistently in the main text, since the disclosed Co10Ho case is negative under both methods.
  • Please identify the DFT hull file revision and the provenance of its energy column. Appendix E explains the decision not to add MP2020 corrections, and the LeMat-Bulk data card describes its energy field as uncorrected VASP energy and documents compatibility filtering. These facts do not establish a correction mismatch. The reported shift from correcting candidates alone is a sensitivity check, not a measured error in the actual hull. Additional DFT calculations are not required merely because correction metadata need clarification.
  • Excluding elemental polymorphs is a legitimate restriction to compound discovery, but they are not intrinsically scientifically meaningless. Likewise, lying below an existing reference hull does not by itself prove a potential failure. Present these guards as conservative task-specific rules.
  • Explaining the validity decline from 94.1% to 69.8% would be useful. The nominal mSUN denominator already accounts for it.
Questions for authors

(1) What checkpoint/hull files and aggregation rule produced the benchmark tables? If compatibility needs correction, are the main gains preserved? (2) Can the species-update attribution be supported by a matched control or qualified to the tested recipes? (3) How were Fig. 2's intervals calculated? (4) Can the DFT hull revision and energy-column provenance be provided?

  1. The chemical substance of the gains needs to be established. The single-element guard is motivated by the observation that a repacked element relaxes near its own hull and counts as novel without being a relevant new material (lines 126–129). The same argument applies to ordered dilute alloys (an elemental packing with a few substituted, chemically similar atoms) and to orderings of mutually soluble rare earths. Their formation energies are near zero, so Ehull is small almost by construction. A guard that counts distinct elements cannot exclude them. The paper's own data suggest the policy moved in this direction:

    • Table 8 gives mean formation energies on the valid set of −0.733 (prior), −0.804 (Chemeleon2), −0.837 (A′) and −0.055 eV/atom (OMatGRPO). Eform − Ehull estimates the mean hull depth at the generated compositions: about −0.93, −0.88, −0.91 and −0.13 eV/atom, respectively. The fixed 0.1 eV/atom metastability window is therefore comparable to the hull depth for OMatGRPO, but only about a tenth of it for the other methods.
    • In Table 11, 19 of the 22 trusted OMatGRPO spot checks are all-metal and 11 have one element at ≥70% (e.g., HgK19, CdMgTl16, Cu3Zn17, ErMg9Y, Li5Zn, Nd8TmY). Seven consist only of rare earths/Sc/Y (AcLaNdTm, ErNd7SmTm, Er2Sc8SmTm, HoNdPrTb5, LuPr2Sm5, LuSmY3, Nd8TmY). Six of the seven sparse spot checks are dilute metallic structures, while none of the four prior trusted spot checks are. The two “DFT-confirmed compounds” shown in Fig. 1 (ErMg9Y, HoMgTmY9) belong to this class.
    • Several of these are trusted on the UMA count but have only 4–8 near-hull phases on the DFT hull (LuPr2Sm5, LuSmY3, AcLaNdTm, Nd8TmY). Near-degenerate orderings may inflate the 1 meV/atom count in mutually soluble systems.

    The DFT sample is stratified by element count and is not an unbiased sample of the mSUN set, and I am not arguing that such structures are worthless. Agreement between UMA and DFT validates the energies, but near-hull energies are expected for dilute alloys, so the agreement does not show that these are new compounds. Please report, for the prior, Chemeleon2, A′ and OMatGRPO mSUN sets, the all-metal fraction, the distribution of the largest element fraction, the rare-earth/Sc/Y-only fraction and the Eform distribution. Please also report mSUN after excluding these classes, or the substitution-derived fraction under the workflow of Negishi & Walsh (2026), which the paper cites. If the margin over the prior and Chemeleon2 survives, the main claim stands. Otherwise, I would lean toward the view that the way the paper counts its results has a loophole, and the credibility and usefulness of the results would be in question.

  2. The controls do not isolate composition-channel RL, and “best overall” depends on the metric.

    • The frozen-composition control (Appendix F, lines 1378–1386) changes several things at once. Compositions are shared within a group, the reward keeps only the stability term, relaxation moves positions only, and the channel weights change from 0.1 to 1. The reasons are disclosed, but 45.5% versus 9.8% compares two recipes. OMatG-IRL (Appendix K, Table 6) shows only a small DNG gain with frozen species (MSUN from 16.9% to 18.1–18.2%). This supports the need to change the composition distribution. It does not show that reinforcing the discrete channel directly is required, since A′ (latent diffusion, frozen decoder) reaches a comparable level. The paper's claims may therefore be overstated.
    • Best-of-N ranks by stability alone (lines 1388–1395), whereas mSUN requires novelty, so a selection dominated by rediscoveries (11.5% novel) is expected. The 47,795 samples roughly match the RL rollout budget (750 × 64). Re-ranking them with the full reward (stability, creativity and in-set uniqueness) requires no new generation and is the budget-matched control that lines 113–115 need.
    • mSUN in the pinned benchmark excludes on-hull SUN structures (Fig. 2a notes “+72 SUN, on-hull” for best-of-N). Adding the SUN counts of Table 7 gives 1,145 for both OMatGRPO (1,138 + 7) and A′ (1,093 + 52). On trusted hulls (mSUN × (1 − sparse share), Tables 7–8, approximate because the sparse shares are rounded), OMatGRPO reaches ≈34.8%, released Chemeleon2 ≈33.7% and A′ ≈42.8%. By the paper's own standard that sparse-tier mSUN marks candidates (lines 483–484), A′ is ahead.
    • The transfer result depends on the relaxation convention. With the reward as defined in Sec. 3.3, which relaxes before scoring, Chemeleon2's pipeline reaches 23.5% (A), below the released 36.6%. Only without relaxation does it reach 43.7% (A′; lines 1125–1127). A′ is single-seed, uses an unseeded sampler and ran the full horizon without early stopping (lines 1118–1122). The Table 2 caption discloses the convention, but the abstract's statement that the reward improves Chemeleon2 should carry this qualification.

    Please either add a matched ablation (full reward, species channel frozen, independent per-rollout compositions, same relaxation and channel weights) together with a full-reward best-of-N, or narrow the claims to the tested recipes and drop or qualify “best overall performance” (lines 119–121).

The benchmark implementation does not match the described protocol. I checked the pinned commit 58e6eae3. The preset sets MACE to model_type: mp with the mace_mp hull, and MACECalculator calls mace_mp(device=...) without a model, so mace-torch 0.3.13 loads medium-mpa-0 by default. The maintainers' README now states that the mace_mp reference hull was computed with medium-0b3 and that all earlier mace_mp-hull-based SUN/MSUN results “mixed two energy scales”. Fix fbf1ba48 (2026-09-15) pins medium-0b3. In addition, SUNMetric and SUNNewMetric threshold e_above_hull_mean, the arithmetic mean over available MLIPs. min_mlips_required = 2 counts available models, not agreeing votes, so the “two of three must agree” description (lines 312–314; Table 3) is inaccurate. All methods share the evaluator, but OMatGRPO's set is chemically distinct from the others (Weakness 1), so a chemistry-dependent MACE offset, one third of which enters the mean, could change the ranking. Please re-score the saved structures with the fixed benchmark and correct the description. No retraining is needed.

6 comments

Response to Reviewer P1HN (1 of 3)

We thank the reviewer for the detailed examination of the paper and of the benchmark code. "Section N" refers to our general response and "Sec. N" to the paper. All values are from the corrected benchmark version (Section 1). mSUN is the number of the 2,500 generated structures that are metastable, unique and novel.

Weakness 1. Chemical substance of the gains

The reviewer is right that OMatGRPO moved toward alloys with formation energies near zero, and that a guard which counts elements cannot exclude them. We computed the requested quantities for the mSUN structures of the four models. A′ is Chemeleon2 trained with our reward.

mSUN structuresPretrained modelChemeleon2A′OMatGRPO
Total3579361,1071,158
Metallic elements only44.0%43.7%48.2%96.3%
Rare-earth elements, Sc and Y only1.7%0.1%0.3%29.4%
One element at 70% or more26.3%4.6%2.4%52.7%
Median largest element fraction0.600.500.500.70
Median EformE_\mathrm{form} (eV/atom)-0.38-0.59-0.69+0.05

The median formation energy confirms the estimate of the reviewer. The window of 0.1 eV/atom is comparable to the hull depth for OMatGRPO and is a small fraction of it for the other models.

The second table gives mSUN after the exclusions that the reviewer proposed. In the first row we exclude single-element structures, structures of rare-earth elements, Sc and Y only, and structures of metallic elements only in which one element reaches 70%.

mSUN structures that remainPretrained modelChemeleon2A′OMatGRPO
After excluding the classes above2908971,080301
With at least one nonmetal20052757343
Neither a duplicate nor a substitution variant of a training structure (Negishi and Walsh)167163103366
Without a known parent lattice (Section 6)218276210391

The answer to the question of the reviewer depends on the criterion. After the exclusion of the alloy classes, OMatGRPO retains about as many structures as the pretrained model, and Chemeleon2 and A′ are far above it. Under the workflow of Negishi and Walsh, OMatGRPO has the most structures that are neither duplicates nor substitution variants of training structures, and it also has the most structures without a known parent lattice. The margin over the pretrained model and Chemeleon2 therefore survives the substitution criterion and does not survive the exclusion of the alloy classes.

We agree that this is a loophole in how mSUN counts. Alloys of chemically similar metals lie within 0.1 eV/atom of the hull almost by construction, and the novelty test passes them because their formula is new. This is the same kind of exploit as the single-element structures that the paper documents, and our guard closed only that one. Training with a charge-neutrality penalty (Section 5) moves the model partly back toward compounds, with 540 structures after the exclusion of the alloy classes and 144 with a nonmetal.

We also agree on the DFT sample. The agreement between DFT and UMA validates the energies and does not show that the structures are new compounds, and near-degenerate orderings can raise the number of reference entries in mutually soluble systems.

We thank the reviewer for this analysis, which changes how we present the result. The revision discusses these alloys in Sec. 4.3 together with the other exploits and gives both tables for all models in the appendix. It no longer presents the gain in mSUN as a gain in new compounds, and it manuscript will acknowledges that the two DFT-confirmed structures of Fig. 1 belong to class of alloys.

Response to Reviewer P1HN (2 of 3)

Weakness 2. Controls and "best overall"

We agree with all four points and have run both experiments that the reviewer requested.

Matched ablation. In the matched control, RL on the atom-type channel is switched off and every other setting is that of OMatGRPO. It uses the full reward, independent compositions for every generated structure, the same relaxation and the same channel weights. It reaches 357 mSUN structures, the value of the pretrained model, against 1,158 for OMatGRPO (table in Section 2). Its composition stays that of the pretrained model, with a Jensen-Shannon divergence of the element frequencies of 0.003 against 0.179 for OMatGRPO. The control removes the reward gradient on the atom-type channel and does not strictly freeze the composition, because the three output heads share one network.

We also agree with the conclusion that the reviewer draws from A′ and from OMatG-IRL. The control shows that, in our pipeline, the gain appears only when the reward acts on the composition. It does not show that the discrete channel must be reinforced directly, since A′ changes the composition through a latent space and reaches a comparable level. The revision removes the word "require" from the introduction and makes only this narrower statement.

Best-of-N with the full reward. We re-ranked the same 47,795 samples of the pretrained model as the reviewer proposed, keeping the highest-ranked samples that are stable and also novel and unique. This selection reaches 1,441 mSUN structures (57.6%) and exceeds OMatGRPO (Section 2). Our statement that oversampling cannot reach the gain of RL was inaccurate and has been removed. The selection obtains its 1,441 structures from 47,795 generated ones (3.0%), whereas OMatGRPO obtains 1,158 from 2,500 (46.3%). The revision will also report this best-of-N row with the full reward in Table 1.

"Best overall". The counts of the reviewer are confirmed after re-scoring. mSUN plus SUN is 1,159 for OMatGRPO (1,158 + 1) and 1,156 for A′ (1,107 + 49). On hulls with at least 12 reference entries, the "trusted" tier of the submission, the exact values are 35.5% for OMatGRPO, 34.6% for Chemeleon2 and 43.3% for A′, so A′ is ahead. As discussed in Section 4, however, the threshold of 12 reference entries is a coarse heuristic. The count includes the entries of all subsystems, so structures with more elements reach it more easily, and this tier favors a model that generates structures with more elements. Four additional training runs of OMatGRPO with different random seeds give between 34.1% and 46.3% mSUN, and Chemeleon2 and A′ lie inside this range (Section 3). The revision removes the claim of "best overall performance" and does not rank the three methods.

Relaxation convention. The reviewer is right that the transfer result depends on it. With relaxation before scoring, as the reward is defined in Sec. 3.3, the Chemeleon2 pipeline reaches 25.3% (632 structures), below the released Chemeleon2 model (37.4%). Only without relaxation does it reach 44.3% (A′). We will add this qualification to the abstract as the reviewer asks, so that it states that the reward improves Chemeleon2 when it is applied without relaxation.

Intervals of Fig. 2 (Question 3). They are Wilson score intervals with z = 1.96 on the mSUN count out of 2,500. The sampling unit is one generated structure from one trained model, so they contain no variation between training runs. The revised caption states this.

Response to Reviewer P1HN (3 of 3)

Weakness 3 and Question 1. Benchmark implementation

The analysis of the reviewer is correct in every point. The benchmark logs of the submission show that medium-mpa-0 was loaded, whereas the MACE reference hull had been computed with medium-0b3. For 20 hull entries, the median deviation from the stored hull energy is 0.0345 eV/atom with medium-mpa-0 and 4e-7 eV/atom with medium-0b3.

We evaluated all saved structure sets again with the fixed benchmark (commit fbf1ba48), without regenerating any structure. The reviewer asks whether a chemistry-dependent MACE offset changes the ranking, and it does not. mSUN rises slightly for every model in the table of Section 1, from 13.4% to 14.3% for the pretrained model, from 36.6% to 37.4% for Chemeleon2, from 43.7% to 44.3% for A′ and from 45.5% to 46.3% for OMatGRPO. The SUN counts fall, for example from 7 to 1 for OMatGRPO.

The description of the aggregation was inaccurate, as the reviewer states. All tables were produced with the criterion of the benchmark, which applies the threshold to the mean EhullE_\mathrm{hull} over the potentials that returned a value and requires at least two of them. The other files are mace-torch 0.3.13, orb_v3_conservative_inf_omat and uma-s-1, the reference hull LeMaterial/LeMat-Bulk-MLIP-Hull at revision 70d505bb, and the novelty reference LeMat-Bulk compatible_pbe at revision 0dc17eea. The revision replaces all benchmark values, corrects Sec. 4.1 and the convention table, and lists versions and checksums in App. B.

Minor comments

  • DFT agreement. A recount confirms the values of the reviewer for the DFT workflow of the submission (63 of 67, 30 of 32 and 12 of 13). Reviewer V7ba pointed out that the Materials Project uses a double relaxation, and the revision now uses that workflow. With it, 64 of 67 threshold classifications agree, and the two subsets remain at 30 of 32 and 12 of 13 (details in our response to Reviewer V7ba). The revision uses "threshold classifications agree" throughout, since Co10Ho lies above 0.1 eV/atom under both methods.
  • DFT hull (Question 4). The DFT hull is the dft split of LeMaterial/LeMat-Bulk-MLIP-Hull at revision 70d505bb (144,127 entries). Its energy column equals the uncorrected VASP energy of LeMat-Bulk, which we verified for 2,000 entries, so both sides of the comparison are uncorrected. App. E now names the file, the revision and the column, and describes the MP2020 shift as a sensitivity check.
  • Guards. We agree and present the single-element guard as a restriction of the task to compounds. The revision no longer describes elemental structures as meaningless, and it no longer describes a value below the hull as a failure of the potential.
  • Validity. The decrease comes from the charge-neutrality check of the benchmark. OMatGRPO generates anion-rich oxides, halides and chalcogenides that fail the charge-neutrality test of validity check while the two geometry checks stay above 99%. A penalty for such structures in the reward restores the validity to 96.6% and raises mSUN to 53.2% (Section 5).
Reviewer_gT4nOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper applies reinforcement learning to the discrete composition channel of a crystal generation model. The authors use GRPO to directly optimize atom-type transitions, while keeping the generation of atomic positions and lattice parameters in a continuous framework.

They also design a reward function to improve stability, novelty, and diversity, and study how the model can exploit weaknesses in the reward function. Experiments show that the mSUN score improves from 13.4% to 45.5%. DFT calculations are also used to check the stability of some generated structures.

Critical strengths
  1. The paper directly optimizes discrete element choices during crystal generation, which is different from previous methods that mainly focus on continuous structure generation. This is an interesting and useful direction.

  2. The authors pay attention to reward hacking and provide experiments to show how it happens. The proposed safeguards make the results more convincing.

Critical weaknesses
  1. The reward function and evaluation metric share similar objectives, especially novelty and uniqueness. It is therefore unclear how much of the improvement reflects genuine progress in material discovery rather than optimization toward the benchmark.

  2. The structural validity rate drops from 94.1% to 69.8% after RL. This is a noticeable decrease, and the paper does not fully explain this trade-off or its impact on practical use.

Questions for authors
  1. Could the authors provide an evaluation that is more independent of the training reward, to better demonstrate the actual improvement in material discovery?

  2. What are the main reasons for the decrease in structural validity after RL? Can the authors clarify whether this is mainly caused by changes in composition or by issues in geometry generation?

  3. How were the structures selected for DFT validation? Are these samples representative of the overall generated structures?

  4. Could the authors explain whether jointly optimizing composition and geometry might help improve structural validity in future work?

2 comments

We thank the reviewer for the assessment and for the four questions, which helped us to state more precisely what our results show. "Section N" refers to our general response and "Sec. N" to the paper. All values are from the corrected benchmark version (Section 1).

Weakness 1 and Question 1. Shared objectives of reward and metric

We agree that the reward and the metric share a novelty test, a uniqueness test and the UMA potential. For each shared component we have an evaluation that does not contain it.

  • Novelty and uniqueness term. A training run without this term in the reward reaches 40.7% mSUN (1,017 structures), against 14.3% (357) for the pretrained model. The value lies inside the range of the five training runs with the complete reward (34.1% to 46.3%, Section 3).
  • UMA potential. With the Orb and MACE potentials only, OMatGRPO reaches 46.8% (1,171) and the pretrained model 14.4% (360) (response to reviewer SeBm).
  • Novelty reference. The reward tests novelty against the 27,136 structures of the MP-20 training set, whereas the benchmark tests it against the 5,335,299 entries of LeMat-Bulk.
  • Machine-learned potentials. DFT, which neither the reward nor the benchmark uses, gives the same classification at 0.1 eV/atom as UMA for 21 of 22 structures of the final model on hulls with at least 12 reference entries.

The gain is therefore not a direct product of overlap between reward and metric. Whether the gain is progress in materials discovery is a separate question, and our answer after the discussion period is that it is so only in part. We address this question in our 2nd comment below. The revision lists the shared components and these evaluations in Sec. 4.3.

Weakness 2 and Question 2. Validity

The submission did not explain the decrease in validity, and we thank the reviewer for raising it. The decrease is caused by the composition and not by the geometry. The benchmark calls a structure valid if it passes a charge-neutrality check and two geometry checks. More than 99% of the OMatGRPO structures pass the two geometry checks, while the fraction that passes the charge-neutrality check is 70.3%, because OMatGRPO generates anion-rich oxides, halides and chalcogenides. A matched control in which RL acts only on positions and lattice keeps a validity of 93.9% (Section 2).

A penalty in the reward for structures that fail the charge-neutrality check restores the validity to between 96.6% and 98.6% and raises mSUN to between 53.2% and 66.8% in three training runs (Section 5). The revision explains the decrease in Sec. 4.2 and adds these runs to Table 1.

Question 3. Selection of structures for DFT

The 90 structures were drawn in three batches from the structures that UMA places at or below 0.1 eV/atom above the hull. The strata (number of elements, number of reference entries of the hull, compositions with fluorine or oxygen) were fixed before any calculation. Within each stratum the structures were drawn at random with a fixed seed, except for three structures that were selected manually. The sample is therefore representative of the structures that we report as metastable and not of all generated structures.

To cover the other side of the threshold, we computed a stratified random sample of 20 OMatGRPO structures with UMA EhullE_\mathrm{hull} between 0.1 and 0.2 eV/atom. DFT places 19 of the 20 above 0.1 eV/atom, with a mean absolute difference of 8 meV/atom to UMA. DFT tests the energy classification and gives no information on novelty. The revision gives the strata and both samples in App. E.

Question 4. Joint optimization of composition and geometry

OMatGRPO already optimizes composition, positions and lattice jointly with one reward. The validity is lost through the composition and is recovered by the composition-based penalty described above, so the joint optimization is not what limits the validity. The geometry is the weaker part of the method in a different respect. Every structure is relaxed before it is scored, so the reward does not act on the generated geometry, and the generated structures remain about 1 Å from their relaxed geometry before and after RL. A reward term for the displacement under relaxation is the natural extension, and the revision names it as future work in the conclusion.

Broader findings of the study

The first weakness of the review asks how much of the improvement is progress in materials discovery and how much is optimization toward the benchmark. The experiments of the discussion period allow a direct answer, which we summarize in this comment.

The gain on the benchmark is robust. RL on the discrete atom-type channel raises mSUN from 14.3% for the pretrained model to between 34.1% and 46.3% in five training runs, and to between 53.2% and 66.8% with the charge-neutrality penalty. The gain remains without the novelty and uniqueness term in the reward, without UMA in the benchmark, and under DFT. A matched control without RL on the atom-type channel stays at 14.3%. The direct optimization of the element choices, which the reviewer names as a strength, is therefore what produces the gain.

The gain in mSUN is nevertheless not a gain of the same size in new materials. OMatGRPO moved toward alloys of chemically similar metals. 96.3% of its mSUN structures contain metallic elements only, against 44.0% for the pretrained model, and their median formation energy is +0.05 eV/atom. The hulls of such systems are shallow, so an ordered arrangement on a known lattice lies within 0.1 eV/atom of the hull and passes the novelty test because its formula is new. When only the mSUN structures without a known parent lattice are counted, OMatGRPO retains 391 structures against 218 for the pretrained model (Section 6). A real gain remains, and it is smaller than the gain in mSUN.

This limitation lies in how mSUN counts novelty and is not specific to our method. For 39% to 81% of the mSUN structures of every model that we evaluated, including MatterGen, DiffCSP and Chemeleon2, a known phase of the same chemical system has the same lattice. A model that optimizes mSUN directly reaches these structures first.

We believe that both results together are the reason the study should be published. The paper provides a method that optimizes the element choices of a crystal generator directly, with a large gain on the community benchmark that is reproduced over several training runs. The paper also shows, with the reward-hacking analysis that the reviewer names as a strength and that the revision extends from single-element structures to alloys, that such a gain can pass both novelty tests and DFT validation without a matching gain in new compounds. Methods that optimize mSUN directly are appearing quickly, and we regard this second result as being as useful to the community as the first. The revision proposes the checks of this study as a standard before a gain in mSUN is read as discovery. These checks are the composition classes and formation energies of the mSUN structures, a comparison with known parent lattices, and a DFT sample on both sides of the stability threshold. We thank the reviewer for the questions that led to this analysis.

Reviewer_yHFwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper applies GRPO to the discrete atom-type decisions of OMatG alongside its continuous position and lattice channels. Under the masked discrete sampler, schedule factors cancel in realized commit-event likelihood ratios, enabling updates through categorical element probabilities. A reward combines near-hull stability, novelty, uniqueness and composition diversity with guards against sparse-reference exploitation and single-element concentration. The complete recipe improves nominal LeMat-GenBench mSUN from 13.4% to 45.5%, and the proposed reward also improves Chemeleon2. The paper documents reward failures and provides DFT spot-checks. The contribution is therefore a discrete-channel implementation and an empirical study of reward design; the main questions are how much benefit is specific to that channel and what level of materials novelty the evaluation establishes.

Critical strengths
  1. A substantial benchmark gain survives important restrictions. Table 1 reports mSUN over all nominal 2,500 generations, so the final model's lower structural validity is included in its yield. The increase from 336 to 1,138 mSUN structures also survives excluding single-element structures. Combining the reported counts with the sparse-reference shares gives approximately 10.5% better-referenced mSUN yield for the prior and 34.8% for OMatGRPO, still about a 3.3-fold increase. These results support the complete recipe's effectiveness under the stated evaluation (Sec. 4.2; Tables 1 and 8).

  2. The reward audit and transfer provide substantive empirical knowledge. The exact categorical event ratio and KL are appropriate for the stated masked sampler (Sec. 3.2; Appendix A). The study documents optimization-induced failure modes and targeted interventions, directly addressing the need to scrutinize automated discovery criteria emphasized by [1]. Without explicit novelty/uniqueness reward, the guarded recipe still reaches 39.3% mSUN; the proposed reward also raises Chemeleon2 from 36.6% to 43.7%. Reporting adverse relaxation interactions, DFT disagreements and sparse-hull limitations strengthens the study (Secs. 4.3–4.5; Appendices B–E).

Critical weaknesses
  1. The main controls do not isolate the benefit of discrete composition updates. The introduction and Sec. 4.2 infer that large improvements require composition-channel RL and attribute most of the gain to it. However, the frozen-composition control also removes creativity, occurrence and MMD terms, shares one sampled composition across a group, relaxes positions only during reward evaluation, and changes continuous-channel weights from 0.1 to 1 (Appendix F, p. 26, lines 1378–1386). The 9.8% versus 45.5% comparison therefore mixes several factors. Best-of-N similarly uses stability-only ranking and position-only scoring relaxation, while RL receives novelty/diversity information and cell relaxation (lines 1388–1395). The results establish useful recipe comparisons, but leave the contribution of the paper's principal methodological component insufficiently isolated. A matched channel intervention, or a clearly narrower attribution, is needed.

  2. The evaluation supports benchmark-defined candidates more clearly than the broader discovery and superiority claims. Leeman et al. [1] explain why database absence alone does not resolve novelty relative to known disordered phases and solid-solution families. Here, the reward and internal evaluation match structures within the same reduced formula (Sec. 3.3; Appendix B), while the formal benchmark uses a broader reference and its own matching procedure. The paper does not establish how ordered configurations, composition variants and known disordered parent phases are distinguished. DFT label agreement on selected structures addresses energetics under the specified reference protocol, rather than this broader novelty question; the pooled 63/67 trusted-tier agreement includes 22 trusted-tier checks from the final model, with 20 agreements (Table 10). I regard the extent of any novelty overcounting as unresolved; clarification of matching scope and a small candidate audit could help bound the discovery claims. Comparative interpretation is also sensitive to the authors' own reference tier: Tables 1, 7 and 8 imply approximately 34.8% better-referenced mSUN for OMatGRPO, versus 42.8% for Chemeleon2 with the proposed reward, reversing their raw 45.5% versus 43.7% ordering. Together with the single-seed swaps, this limits the “best overall” claim. The important distinction is between a verified gain in the defined benchmark and an established advantage in meaningful materials discovery.

Reference:

[1] Josh Leeman et al. “Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis.” PRX Energy 3, 011002 (2024). https://doi.org/10.1103/PRXEnergy.3.011002. Relevant discussion: Secs. II–III on novelty and alternative explanations, and Sec. VI on compositional disorder. The experimental identification errors analyzed there concern A-lab; no corresponding experimental error is alleged for the present computational paper.

Questions for authors
  1. Can you provide a comparison that freezes the categorical sampling policy while preserving the reward, cell-relaxation convention, group sampling and continuous optimization settings, explaining how shared parameters are handled? If that intervention is infeasible, which narrower conclusion follows from the existing control? A reward-aware selector on the existing candidate pool would be useful as an optional diagnostic; the present selection result should otherwise be identified as stability-only best-of-N.

  2. Following [1], how does the formal benchmark distinguish new ordered phases from ordered configurations or new compositions within known disordered solid-solution families? Please clarify its matching procedure separately from the same-formula reward/internal check. A small audit of representative mSUN candidates against their closest known parent phases would help determine the scope of the discovery claims. An ordered phase can itself be new; the issue is establishing that distinction. Please also separate final-model DFT label agreement from pooled agreement and from novelty validation.

  3. Please report exact better-referenced mSUN counts across the reward/pipeline comparisons and explain how the interpretation changes. The estimates above use rounded sparse shares; the reference-count threshold is itself a heuristic. How were Figure 2's 95% intervals constructed, and are additional final-policy or reward-swap seeds available?

5 comments

Response to Reviewer yHFw (1/4)

We thank the reviewer for the assessment. Addressing the second weakness in particular has made the paper scientifically more sound, and the revision now states that an improvement in a benchmark does not always translate into the discovery of new materials. "Section N" refers to our general response. All values are from the corrected benchmark version (Section 1).

Weakness 1 and Question 1. Controls

We agree that the comparison of 45.5% with 9.8% in the submission combines several factors. After re-scoring, these values are 46.3% and 10.4%. We therefore trained the control that the reviewer describes. RL on the atom-type channel is switched off, and the reward, the cell relaxation, the group sampling and the settings of the continuous channels are those of OMatGRPO. The only difference is that the weight of the atom-type term in the policy loss is zero.

On shared parameters, the three output heads read from one network. In the control, the atom-type head receives a gradient only from its KL term to the pretrained model, while the shared network continues to change through the position and lattice terms. A strict freeze requires compositions drawn from the pretrained model, which changes the group sampling. We trained that control as well, with the full reward.

ModelmSUN
Pretrained model14.3% (357)
Strict freeze, compositions drawn from the pretrained model12.5% (312)
Matched control, atom-type term of the policy loss set to zero14.3% (357)
OMatGRPO46.3% (1,158)

The matched control has the same count as the pretrained model by coincidence, since it generates different structures. Its composition distribution stays that of the pretrained model, with a Jensen-Shannon divergence of the element frequencies of 0.003, against 0.179 for OMatGRPO. In the tested setting the gain is therefore obtained only when RL acts on the atom-type channel. Each control is a single training run. The gain that disappears in the control is a change in the generated compositions, which we examine under Weakness 2.

We also evaluated best-of-N with the full reward criteria on the same 47,795 samples of the pretrained model. It reaches 57.6% mSUN and exceeds OMatGRPO, but it keeps 3.0% of the generated structures against 46.3% for OMatGRPO (Section 2).

The revision removes "require" from the introduction. Table 1 reports the matched control and best-of-N with the full reward, and the appendix reports the strict freeze and the stability-only best-of-N of the submission.

Comment (2/4)

Weakness 2 and Question 2. Novelty and discovery

We agree, and we thank the reviewer for this important point. Neither novelty test makes the distinction that the reviewer describes. We performed the suggested audit and two further tests.

The benchmark compares each structure with the 5,335,299 computed entries of LeMat-Bulk, using the StructureMatcher of pymatgen with ltol = 0.1. Our reward and internal evaluation compare it with the 27,136 structures of the MP-20 training set, using the StructureMatcher with default settings. Both tests compare a structure only with entries of the same reduced formula.

LeMat-Bulk contains ordered, computed structures only. We tested the consequence with the benchmark's own metric. The LeMat-Bulk entry HoEr$_4$ is returned as not novel, while the same structure with one Ho site replaced by Er (HoEr$_9$) is returned as novel. A new composition on a known lattice, or an ordered variant of a known solid solution, is therefore counted as novel by both tests.

Compound mSUN structuresPretrained modelMatched controlOMatGRPO
Total3493471,153
Without a nonmetal1511511,110
Rare-earth elements only69340
With a nonmetal19819643

The gain of OMatGRPO is almost entirely in alloy compositions without a nonmetal. The shift is present in both repeated training runs and absent in the matched control. The charge-neutrality penalty (Section 5 of the general response) does not remove it.

Tests. All tests use the structures that the benchmark scored, and each rule was fixed before the comparison was run.

  1. Known parent lattice. For each mSUN structure we searched all LeMat-Bulk entries of its chemical system for an entry with the same lattice once one pair of elements, or all elements, are treated as one species.
  2. Disordered-structure matcher of MatterGen (Zeni et al.), with default settings, against LeMat-Bulk and the MatterGen reference set.
  3. Audit of 60 sampled candidates against experimental entries with mixed site occupancy in the Crystallography Open Database (next comment).

Seed 1 and seed 2 are two further training runs of OMatGRPO with different random seeds and otherwise identical settings (Section 3 of the general response).

Pretrained modelOMatGRPO, submitted runOMatGRPO, seed 1OMatGRPO, seed 2
mSUN3571,1581,051852
with a known parent lattice139 (38.9%)767 (66.2%)643 (61.2%)240 (28.2%)
matched by the MatterGen matcher57 (16.0%)245 (21.2%)161 (15.3%)59 (6.9%)
without a known parent lattice218391408612

The values for the other models are given in Section 6 of the general response.

Comment (3/4)

Weakness 2 and Question 2, continued

For 66.2% of the mSUN structures of OMatGRPO, a known entry of the same chemical system has the same lattice, including 339 of the 340 structures with only rare-earth elements. The effect is not specific to our method, since the share is 38.9% for the pretrained model, 48.1% for MatterGen, 49.2% for DiffCSP, 70.5% for Chemeleon2 and 81.0% for our reward in the Chemeleon2 pipeline. The MatterGen matcher finds a known entry for only 21.2%, because it requires the same reduced formula and therefore does not detect a new composition on a known lattice. Counted without a known parent lattice, OMatGRPO remains above the pretrained model in all three audited training runs, by a factor of 1.8 to 2.8 against 2.4 to 3.2 on mSUN. When the dilute all-metal structures are also removed, the submitted run keeps 88 structures against 185 for the pretrained model (our reply to Reviewer P1HN).

We also drew 60 of the 1,153 compound mSUN structures of OMatGRPO at random, stratified by reference count and chemistry, with the sample and the rule fixed before any result was read. For each candidate we searched the Crystallography Open Database for experimental entries with mixed site occupancy in its chemical system. For 12 of the 60 such an entry exists on the lattice of the candidate or of its closest parent, and for 9 of these it matches the candidate itself, for example Cu4Ni, Mg11Ag, Tb3Pr2Ho and PrGd6Er. For 32 a known parent lattice exists but no experimental entry was found, and 16 have no known parent lattice.

Limits. As the reviewer notes, an ordered phase can itself be new. Cu3Au, CuAu and Ni3Al are ordered variants of fcc with their own properties, and an ordered structure is a distinct phase when its energy gain relative to the random alloy of the same composition exceeds the mixing-entropy term. For a dilute alloy with 10% solute, TSmixT S_\mathrm{mix} is about 28 meV/atom at 1000 K, while the ordering energy of mixed rare earths is a few meV/atom, so such ordered cells are snapshots of a solid solution. A matching parent lattice therefore does not show that a candidate is known, and the absence of a match does not show that it is new. The comparison with merged species is permissive for alloys on close-packed lattices, and the Crystallography Open Database is incomplete for intermetallic solid solutions.

DFT tests the energy classification and gives no information on novelty. For the final model, DFT and UMA agree for 21 of 22 structures on hulls with at least 12 reference entries, and the pooled value of 64 of 67 includes structures from other runs. Of the 30 DFT structures of the final model, 18 have a known parent lattice.

Change in the manuscript. The main text states both matching procedures and adds a subsection on the composition of the mSUN structures, with these counts for all models. It describes the result as a gain in the benchmark metric that does not always correspond to new materials, and it proposes the comparison with the random alloy as a check. Leeman et al. and Zeni et al. are cited, and the appendix contains the full audit.

Comments on the paper

8 comments

We thank the reviewers for their careful assessment. The reviews identified errors in the submission and asked for controls that were missing. We have corrected the errors and run the requested experiments, which are listed below.

One point changes the message of the paper. A gain in the benchmark metric mSUN does not by itself mean that new materials have been discovered. The revision states this explicitly (Section 6).

Points raised by more than one reviewer are answered in this general response. The individual responses refer to its section numbers.

Experiments run during the discussion period

ExperimentRequested by
Re-scoring of all 43 benchmark evaluations with the corrected benchmark versionP1HN
Matched control with RL on the atom-type channel switched offP1HN, yHFw
Best-of-N selection with the full reward criteriaP1HN, yHFw
Two repeated OMatGRPO training runs with different seedsP1HN, yHFw
Training with a charge-neutrality penaltySeBm, gT4n
Baselines MatterGen and DiffCSPSeBm
Benchmark without the UMA potentialSeBm, gT4n
Training from the Alex-MP-20 pretrained modelSeBm
Training without relaxation before the rewardSeBm
Benchmark at intermediate training checkpointsSeBm
Classification of all single-element structures, with DFTV7ba
Removal of reference entries from well-populated hullsV7ba
Addition of generated structures to sparse hullsV7ba
Training with a reference-count threshold that depends on the number of elementsV7ba, yHFw
DFT with the Materials Project double-relaxation workflow, 88 structuresV7ba
DFT for 20 structures above the stability thresholdV7ba, gT4n
Audit of mSUN structures against known parent lattices and experimental disordered phasesyHFw
Novelty under the disordered-structure matcher of MatterGenyHFw

Section 1. Benchmark re-scoring (P1HN). The benchmark version behind the submission loaded the MACE checkpoint medium-mpa-0, while the MACE reference hull of the benchmark was computed with medium-0b3. The maintainers of LeMat-GenBench corrected this in commit fbf1ba48 (Sep 15, 2026). We re-scored all saved structure sets with that commit. No structure was regenerated.

ModelmSUN, submittedmSUN, re-scored
Pretrained model13.4%14.3% (357 structures)
OMatGRPO45.5%46.3% (1,158)
Chemeleon236.6%37.4% (936)
Our reward in the Chemeleon2 pipeline (A′)43.7%44.3% (1,107)

mSUN is a fraction of the 2,500 generated structures. The ordering of the methods is unchanged, and all values in our responses are re-scored. Sec. 4.1 also described the stability criterion of the benchmark inaccurately. The benchmark applies the threshold to the mean EhullE_\mathrm{hull} over the potentials and does not use a vote. The revision corrects Sec. 4.1, App. B and all tables and figures.

Section 2. Contribution of RL on the atom-type channel (P1HN, yHFw, SeBm). The control of the submission differed from OMatGRPO in several settings. We trained a matched control in which the weight of the atom-type term in the policy loss is zero and every other setting is that of OMatGRPO.

ModelValidNovelMetastablemSUN
Pretrained model2,3521,930707357 (14.3%)
Matched control, RL on the atom-type channel switched off2,3482,010626357 (14.3%)
OMatGRPO1,7451,5411,3551,158 (46.3%)

All values are counts out of 2,500 generated structures.

The matched control and the pretrained model are different structure sets, as the other columns show, and they share only 4 mSUN structures. That both contain exactly 357 mSUN structures is a coincidence. The matched control shows no gain in mSUN, and its composition distribution is that of the pretrained model. In our setting the gain is therefore obtained only when RL acts on the atom-type channel. This is one training run. The word "require" in the introduction is replaced by this narrower statement.

We also ran best-of-N selection with the full reward criteria. From 47,795 samples of the pretrained model we selected the 2,500 highest-ranked structures that are also novel and unique. This selection reaches 57.6% mSUN (1,441 structures) and exceeds OMatGRPO. Our statement that oversampling cannot reach the gain of RL was wrong and is removed. The two approaches differ in yield. The selection keeps 1,441 mSUN structures out of 47,795 generated (3.0%), and OMatGRPO keeps 1,158 out of 2,500 (46.3%).

Both results are in the revised introduction, Sec. 4.2, Table 1 and App. F. The responses to Reviewers P1HN and yHFw give the details.

Reviewer_V7baOpenReview ↗

Missing general comment Sections 3-6

Dear authors, might you by any chance have forgotten to post Sections 3-6 of the general comment? Some of the experiments listed in the "Experiments run during the discussion period" table are missing, and some of the individual reviewer responses mention "Section 5/6 of general comment".

Section 3. Rankings and training runs with different seeds (P1HN, yHFw). OMatGRPO, Chemeleon2 and A′ (Chemeleon2 trained with our reward) cannot be ranked on mSUN fraction of the 2,500 generated structures. We trained OMatGRPO four more times with the same settings and a different random seed, and the five runs give between 34.1% and 46.3% mSUN, with a mean of 41.1%. These four additional runs were trained during the discussion period.

ModelmSUN, all hullsmSUN, hulls with at least 12 reference entries
Pretrained model14.3% (357)11.2% (281)
Chemeleon237.4% (936)34.6% (865)
A′44.3% (1,107)43.3% (1,082)
OMatGRPO, submitted run46.3% (1,158)35.5% (888)
OMatGRPO, seed 142.0% (1,051)37.1% (927)
OMatGRPO, seed 234.1% (852)26.7% (667)
OMatGRPO, seed 346.0% (1,150)36.2% (905)
OMatGRPO, seed 437.0% (926)29.0% (725)

The second column counts only structures whose convex hull has at least 12 reference entries, the "trusted" tier of the submission.

Every OMatGRPO run is far above the pretrained model. On all hulls, Chemeleon2 and A′ lie inside the range of the OMatGRPO runs. In the second column A′ is above every OMatGRPO run. A′ generates structures with more elements (higher arity) than OMatGRPO, and such structures reach 12 reference entries more easily, because the count includes the entries of all subsystems (Section 4). This is the point raised by Reviewer V7ba, and it means the second column favors the pipeline that generates more elements per structure. The runs differ by up to 12 percentage points, while the intervals of Fig. 2 extend about 2 points to each side, because they describe only the sampling of 2,500 structures from one trained model. The revision removes the claim of "best overall performance", does not rank the three methods, and reports all five runs in a new appendix table.

Section 4. Reference-count threshold (V7ba, yHFw). The threshold of 12 reference entries is a coarse heuristic, as both reviewers note. The reference count of a structure includes the entries of all subsystems. As a result no elemental system reaches 12, while 92% of the ternary and all quaternary systems reach it through subsystem entries alone. The revision states this and reports the number of mSUN structures under two further definitions.

ModelAt least 12 reference entries (submission)At least one reference entry that contains all elements of the structureReference count at or above the median of all chemical systems with the same number of elements
OMatGRPO888755667
Chemeleon2865559564
Our reward in the Chemeleon2 pipeline (A′)1,082317543

The ordering of OMatGRPO and A′ depends on the definition. The agreement between DFT and UMA is 93 to 94% under all three.

Following the point of Reviewer V7ba that one threshold for every number of elements is too coarse, we trained OMatGRPO once with the medians of the last column as thresholds. These are 2, 9, 23, 48 and 108 reference entries for structures with one, two, three, four, and five or more elements. mSUN is 32.1% (803 structures), against 42.0% (1,051) for the run with the threshold of 12 and the same random seed (seed 1 in Section 3). In this one training run the element-dependent threshold did not improve the result. We keep the threshold of 12 for the model of the paper and report this run in the appendix. The revision changes Sec. 4.5 and adds two appendix tables. More discussion can be found in the response to Reviewer V7ba comment 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.