acceptodds
Under review as a conference paper at ICLR 2027

GLASS: Global Latent Aggregation with Slot-based Set Decoding for Scalable All-Atom Crystal Generation

Abstract

Generative models for crystals enable the discovery of novel structures, but scaling all-atom generation to larger systems such as metal–organic frameworks remains challenging. We connect this difficulty to the correspondence problem of particle-space generation. Even on a single fixed target set, index-free permutation-equivariant particle flows require substantially more training for reliable generation as set size and density increase, under both independent and optimal-transport couplings. To resolve this challenge, we introduce GLASS—Global Latent Aggregation with Slot-based Set Decoding, which encodes structures in a permutation-invariant global latent space and learns their distribution via flow matching. A learned-slot decoder constructs all atoms in parallel, removing atom-wise correspondence from generative transport. On MP20, GLASS is competitive with particle-space models, and flow training can reach the validity of the training data at every structure size. On a QMOF subset, GLASS generates MOFs with up to 150 atoms per unit cell without conditioning on building blocks, topology, or composition, and approaches the structural validity of the training data. On both datasets, flow training exposes a validity–novelty tradeoff, and MOF novelty remains limited by autoencoder generalization on the available data. These results show that separating correspondence assignment from generative transport provides a simple route toward high-validity generation of larger atomistic systems.

Primary area:
generative models
Keywords:
generative models, flow matching, set generation, metal-organic frameworks, latent flow matching, crystal generation

Reviews

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

This paper addresses the scaling problem of 3D crystal material generation. Existing generative models, during generation implicitly model how a noise particle should become which atom, and since there are no atom indices, this matching gets harder as the number of atoms grows.

Authors proposed GLASS as a fix to this problem, where they first encode the crystal structure into a single global latent vector that doesn’t depend on atom ordering. Next, a flow matching model learns the latent space distribution and generates new latent vector. Following that, a slot-based decoder then turns the latent into all atoms at once.

Critical strengths
  1. Overall, the problem is interesting and prominent. Existing generative models suffer as the crystal structure grows larger.
  2. The paper is very well written with clearly motivating the problem and enugh background about crystal generation
  3. The idea of a slot-based decoder is really novel and interesting.
  4. The authors perform detailed experiments with MP-20 and MOFs datasets to empirically evaluate their model.
  5. Codebase provided.
Critical weaknesses

See the Questions.

Questions for authors
  1. Crystal Materials are translationally invariant as well; how does the AE encoder maintain that? Is translation augmentation or canonicalization used?

  2. The autoencoder is deterministic with no KL Loss; how smooth is the latent space? Is there any way to analyse that?

  3. What's the significance of Pair-distance loss? What happens if we remove it? It would be great if the authors do an ablation study on that and checked the stability, novelty, and uniqueness of the model.

  4. What is d_z for a single vector encoding up to 150 atoms, and what is the held-out reconstruction accuracy?

  5. In Fig-3, authors stated GLASS at 50k and 1M flow-training steps. What are the steps for other models? Are they also similar or very less? If less, then the comparison is unfair.

  6. Tables 1 and 11 use different models in the pre-relaxed and unrelaxed groups. The same baselines should be evaluated under both.

  7. Stability, uniqueness, and SUN look good, but novelty of the model is only 45–53%. Any justification from authors would be good.

  8. Finally, the paper misses some important papers on material generation like DiffCSP++, SymmCD, CrysLLMGen, CrysBFN, TGDMat, FlowMM, FlowLLM, OXtal,Wyckoff Transformer

  9. Finally, what is the sampling time and computation needed for GLASS?

5 comments

Dear Reviewer vpVs,

Thank you for the careful evaluation of our manuscript and for the positive assessment of the problem and the slot-based decoder. We address each question below.

Q1: Crystal Materials are translationally invariant as well; how does the AE encoder maintain that? Is translation augmentation or canonicalization used?

Thank you for raising this important point. The encoder is not translation invariant, and we use neither translation augmentation nor canonicalization (Appendix D.1). We use the fractional coordinates exactly as stored in the datasets, without any shift. Structures are typically stored with the cell origin at a standard origin of their space group. In MP20 and LeMat-Bulk, many atoms occupy special Wyckoff positions, so fractional coordinates recur across structures, and the decoder slots specialize to these positions (Figure 12). In QMOF150, atoms mostly occupy general positions, and the shared structure is the placement of symmetry elements. Translation augmentation removed this anchor in our tests: shifting inputs and targets together made training collapse, and shifting only the inputs lowered held-out reconstruction.

Q2: The autoencoder is deterministic with no KL Loss; how smooth is the latent space? Is there any way to analyse that?

Thank you for this interesting question. While defining smoothness in this latent space is not a trivial task, we have investigated the behavior of decoded structures around encoded data latents. We find that several behaviors emerge in the latent space: Some latent directions primarily activate/deactivate slots or change the chemistry. Other directions mostly change the cell. For larger structures, latent variables can also shift the atomic structure by moving several atoms. However, targeted displacements of individual atoms are challenging. When editing a specific position in a target atom of a training sample and subsequently encoding and decoding it, the decoded structure exhibits additional artifacts, such as an altered chemical composition or several displaced atoms. Therefore, changes in the latent vectors allow for structured changes in decoded crystals, but precise edits can be challenging.

In addition, we also compare a KL-regularized version in Table 8. For a KL weight of 1e-5, we find a slight gain in the valid and novel yield, while a higher 1e-4 weight increases validity at the expense of novelty.

Q3: What's the significance of Pair-distance loss? What happens if we remove it? It would be great if the authors do an ablation study on that and checked the stability, novelty, and uniqueness of the model.

Thank you for this suggestion. We ran the requested ablation with three training runs per setting. On MP20, the pair-distance loss has no significant effect, and mean SUN and MSUN are slightly higher without it (Appendix E.4):

Pair lossInputValidUniqueNovelMetastableSUNMSUN
onraw89.29 ± 0.5988.32 ± 0.6845.65 ± 1.2042.19 ± 1.630.07 ± 0.0611.68 ± 0.48
offraw88.95 ± 0.4288.11 ± 0.6246.68 ± 1.5742.88 ± 0.730.20 ± 0.1112.53 ± 0.34
onrelaxed96.56 ± 0.3995.48 ± 0.6652.71 ± 1.4063.63 ± 1.151.13 ± 0.1622.31 ± 0.86
offrelaxed96.44 ± 0.1495.60 ± 0.2854.23 ± 1.8963.91 ± 0.521.40 ± 0.1424.08 ± 1.45

On QMOF150, the pair-distance loss has a mild effect on validity, but no noticeable effect on novelty, uniqueness, or VNU:

Pair lossValidNovelUniqueVNU
on74.79 ± 0.6615.27 ± 0.4959.28 ± 0.148.76 ± 0.15
off73.98 ± 0.8515.46 ± 0.5959.29 ± 0.218.82 ± 0.12

Overall, this loss term has a marginal effect and is not essential to the performance of GLASS. We added the ablation to Appendix E.4 (Tables 10 and 11) and keep the original recipe for all other results.

Q4: What is d_z for a single vector encoding up to 150 atoms, and what is the held-out reconstruction accuracy?

For QMOF150, we use a latent dimensionality of 64, reaching a held-out reconstruction RMSD of 1.25 Å and 58.68% type reconstruction accuracy, compared to 0.019 Å RMSD and 99.99% accuracy in training. A comparison of different latent dimensionalities is shown in Table 9, and shows a negligible impact.

Q5: In Fig-3, authors stated GLASS at 50k and 1M flow-training steps. What are the steps for other models? Are they also similar or very less? If less, then the comparison is unfair.

We compared the reported training lengths and found a large range. Zatom-1 uses about 43k steps, MatterGen 47k, and DiffCSP 106k. ADiT trains for about 488k steps, Crystalite for 1.33M, and OMatG for 1.7M. The GLASS 50k checkpoint is at the short end of this range, and the 1M checkpoint is below the two longest, so GLASS is not favored by training length. However, we also note that step counts are not directly comparable across batch sizes and model sizes.

Q6: Tables 1 and 11 use different models in the pre-relaxed and unrelaxed groups. The same baselines should be evaluated under both.

The reference values in Tables 1 and 14 (Table 11 in the original submission) are taken from the LeMat-GenBench leaderboard, where each model appears once, as submitted by its authors. The number of released models and the cost of evaluating the benchmark with relaxation prevented us from re-running every model under both settings. We have amended the relevant section to mark this more clearly. Table 16 complements this with a selection of models evaluated under the unified relaxation and evaluation pipeline of the Crystalite protocol.

Q7: Stability, uniqueness, and SUN look good, but novelty of the model is only 45–53%. Any justification from authors would be good.

Thank you for raising this point. Across the flow training, the model moves along a validity-novelty frontier: Early in training, validity is lower, but many samples are novel. As training progresses and the model closely approximates the training distribution, samples become more valid but less novel, shifting the point on the frontier. This transition is shown in Figure 6, showing a peak and eventual decline in novel and valid structures. Therefore, we treat the novelty of a structure set as an optimizable parameter via the training steps.

Furthermore, the selected checkpoint exhibits a novelty comparable to other models such as Crystalite or OMatG, showcasing its strong positioning in the high novelty and high validity regime. We also note that novelty is computed against the LeMat-Bulk reference set, not the MP20 training set.

Q8: Finally, the paper misses some important papers on material generation like DiffCSP++, SymmCD, CrysLLMGen, CrysBFN, TGDMat, FlowMM, FlowLLM, OXtal, Wyckoff Transformer.

Thank you for raising this point. DiffCSP++, FlowMM, and Wyckoff Transformer were already cited. We have incorporated the relevant missing literature in the appropriate section (Appendix A).

Q9: Finally, what is the sampling time and computation needed for GLASS?

On a single H100 GPU, sampling 1,000 MP20-size crystals takes 0.21 s, and sampling 1,000 QMOF150 MOFs takes 0.32 s (Table 21). Training the autoencoder takes about 20 minutes for MP20 and 3.5 hours for QMOF150. Flow training takes about 50 minutes for 1M steps on either dataset (Appendix I). For comparison, Table 16 lists 5 to 22 s per 1,000 structures for Crystalite and 85 s for ADiT.

Thank you for the detailed responses. Some concerns remain:

Q1 (Translation invariance). I consider this a significant limitation: the same crystal under a different origin is physically identical, yet the encoder treats it as a different input. I would ask the authors to quantify this directly by translating held-out structures with random shifts and reporting (i) the latent distance between original and shifted copies, relative to typical distances between different crystals, and (ii) the reconstruction accuracy of the shifted inputs. Without this analysis, it is unclear whether the latent space represents crystal structure or a dataset-specific coordinate convention. In addition, the authors state that shifting inputs and targets together made training collapse, and shifting only the inputs lowered held-out reconstruction. Could the authors explain why this happens? It suggests that the model relies on the absolute coordinate frame of the stored structures rather than on the crystal structure itself.

Q2 (Latent smoothness). Thank you for the qualitative analysis and the KL-regularized comparison. The observation that a small edit to a single atom changes the composition or displaces several atoms after re-encoding suggests the latent space may not be locally smooth. Could the authors provide a PHATE visualization of the encoded latents, colored by properties such as composition, space group or formation energy, to show whether similar crystals lie close together?

Q4 (Reconstruction). The gap between training and held-out reconstruction (0.019 Å vs. 1.25 Å RMSD; 99.99% vs. 58.68% type accuracy) indicates strong overfitting of the autoencoder, which may also explain the high fraction of generated MOFs that replicate training structures. Could the authors report validation curves and test whether early stopping on a held-out validation set, or regularization (e.g., weight decay, dropout, or the KL term of Table 8), improves held-out reconstruction, and how this affects the novelty of generated structures? Could they also report the corresponding held-out reconstruction numbers for MP-20?

Q7 (Novelty). The decline in novelty as training proceeds suggests that the model increasingly reproduces training-like structures, consistent with the reconstruction gap noted in Q4. I would encourage the authors to discuss this connection in the revised paper.

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

The paper introduces GLASS, a latent flow-matching model for inorganic crystal and MOF generation. Motivated by the correspondence problem in particle-space generation, the authors propose using a global latent representation. The method consists of two parts: (1) the GLASS autoencoder, which encodes crystal structures into permutation-invariant global latent representations. Its decoder learns to generate a variable number of atoms by predicting an additional empty class and uses Hungarian matching to align predicted and target atoms; and (2) a latent flow-matching model trained on these global representations. Experimental results demonstrate competitive performance on MP20 and high structural validity on QMOF150.

Critical strengths

Thorough ablations and analysis: The authors clearly report the reconstruction gap of the autoencoder between training and validation splits on different datasets, as well as the sensitivity of performance to the latent dimension (dzd_z).

Critical weaknesses

Overall, while the paper is well motivated, its main empirical contribution focuses on achieving high structural validity. However, structural validity alone is insufficient to establish the usefulness of a generative model for materials discovery, where novelty and diversity are also essential. GLASS achieves competitive performance on MP20 but shows limited novelty on QMOF150, raising concerns about its effectiveness for discovering new MOF structures.

  • Low novelty on QMOF150: While the authors report high structural validity on QMOF150, the VNU rate is only 8.76%, compared with 42.4% reported for Mofasa, although the models use different training splits. More importantly, the authors report that 98.3% of valid GLASS generations match training structures, suggesting that the high validity largely reflects the reproduction of familiar frameworks rather than generalization to unseen structures. This concern is further supported by the large autoencoder reconstruction gap between training and validation structures. Thus, despite its high structural validity, the practical benefit of GLASS for discovering novel MOFs remains unclear.
  • Missing MOF generation baselines: The paper reports results for Mofasa despite differences in training datasets but does not provide comparable generation results for other relevant methods, such as MOFDiff, MOFFlow-2, and MOF-BFN. These methods report substantial VNU yields in their respective evaluation settings (30.0%, 31.35%, and 40.7%, respectively). Although these results are not directly comparable due to differences in datasets and evaluation protocols, all three methods support unconditional MOF generation, regardless of the representations they use. At minimum, the authors should provide reasonable justifications for excluding these baselines and preferably compare them fairly under shared training and evaluation settings.

Minor weakness:

  • Incorrect and inconsistent references: Many references contain bibliographic errors or inconsistent formatting, including incorrect publication years and page ranges, missing conference or proceedings information, and inconsistent citation formats. For example, DiffCSP++ is cited as an ICLR 2024 paper but incorrectly lists 2023 as the publication year. The authors should carefully check and standardize the bibliography.
Questions for authors
  1. How does the method's VNU compare with other baselines under comparable settings?
  2. Have the authors explored directions to improve novelty?
2 comments

Dear Reviewer qtPE,

Thank you for the detailed evaluation of our manuscript and for raising important points. We address your concerns and questions below.

W1: Validity alone is insufficient; low novelty on QMOF150.

We agree that materials discovery requires novel structures, and the current GLASS model on QMOF150 does not find a large number of structures beyond its training data. This is acknowledged in the abstract, Section 5, and the limitations paragraph in Section 6, with a detailed analysis in Appendix G.2–G.3.

The focus of this work is the decline in validity as structure size increases. In Section 2, we identify the correspondence problem as a contributor to this decline, and we show that removing correspondence from generative transport restores validity: 74.8% on QMOF150 against 80.3% for the training data, 52.9% for Mofasa, and 15% to 16% for ADiT, Zatom-1 and SinAE. Most prior all-atom models do not approach training-level validity at this size, so we consider this a necessary step toward novel generation, though not a sufficient one.

The low novelty has an identifiable cause. The QMOF150 autoencoder reconstructs training structures at 0.019 Å, but held-out structures at 1.25 Å, so decoding fails away from training latents. Where the autoencoder generalizes, the same flow generates novel structures (MP20, and the new LeMat-Bulk results in Appendix E.5, Table 13).

Q1: How does the method's VNU compare with other baselines under comparable settings?

We agree that MOFDiff, MOFFlow-2, and MOF-BFN are relevant baselines for unconditional MOF generation. Our comparison focused on methods that generate complete atomic structures without explicitly decomposing the MOF into molecular building blocks. This scope follows the central question of our work: whether separating atom-wise correspondence from generative transport can support scalable all-atom generation.

The distinction concerns the representation and modeling assumptions. MOFDiff and MOF-BFN generate building-block identities and arrangements, while MOFFlow-2 generates building-block SMILES and subsequently models their placement and torsional flexibility. These approaches exploit an explicit decomposition of chemical components and framework geometry. GLASS instead decodes the complete atom set from a global latent representation, without a building-block generation or assembly stage. This motivated prioritizing Mofasa as a closely related all-atom baseline, while explicitly acknowledging differences in its training data.

We agree, however, that these representational differences do not diminish the value of comparing overall MOF generation performance. We have clarified the focus on all-atom methods in Section 5 and listed the reported VNU reference values for MOFDiff, MOF-BFN, and MOFFlow-2 for context. To enable a more direct comparison on shared data, we are also training GLASS on BW-DB, restricted to structures with at most 300 atoms (about 92% of the dataset). We will report the result in this thread if it completes within the discussion period.

Q2: Have the authors explored directions to improve novelty?

Yes. We believe the poor held-out autoencoder performance limits novelty, because the decoder is unable to decode latents outside the training domain. We have explored two directions:

  • Increase the latent width. Increasing the latent width to allow for more information to pass through the bottleneck leaves performance largely unchanged (Tables 9 and 19).
  • Large-scale data. Building on the observed closing of the validation gap in the LeMat-Bulk experiments (0.064 vs 0.058 Å at 64 dimensions, Table 12), we are exploring larger MOF datasets and systematic pretraining approaches for a generalizable autoencoder.

Minor: bibliography.

We corrected the bibliography, including the DiffCSP++ year.

Thank you for the response. While the authors acknowledge the low novelty and identify a potential cause, the issue has not been resolved, nor has the proposed cause been verified experimentally. For example, the findings from LeMat-Bulk, which covers structures of at most 20 atoms, may not carry over to the QMOF150 setting.

I also agree with the authors that representational and methodological differences do not diminish the value of comparing overall MOF generation performance with existing baselines. However, such a comparison is not yet available.

Given that my main concerns remain unaddressed at this point, I will keep my score.

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

The paper argues that a major obstacle to scaling all-atom generation to larger systems such as metal–organic frameworks is the correspondence problem. In particle-space flow matching, the model has to implicitly learn which noise particle moves to which target atom, and atoms have no natural order. To resolve this challenge, the authors propose GLASS—Global Latent Aggregation with Slot-based Set Decoding, which encodes structures in a permutation-invariant global latent space and learns their distribution via flow matching. A learned-slot decoder constructs all atoms in parallel, removing atomwise correspondence from generative transport. The authors evaluate GLASS on MP20 and QMOF datasets. On MP20, GLASS is competitive with particle-space models. On a QMOF subset, GLASS generates MOFs with up to 150 atoms per unit cell without conditioning on building blocks, topology, or composition, and approaches the structural validity of the training data.

Critical strengths
  1. The paper proposed a hypothesis to the difficulty of scaling up all-atoms generation to larger systems - Index-free permutation-equivariant particle flows need much more training for reliable generation as set size and density grow, under both independent and optimal-transport pairing. The authors also designed experiments to test two root causes.

  2. On the basis of the hypothesis, the authods proposed GLASS by running generation entirely in an order-independent global latent space. The GLASS method achives competitive performance on the MP20 dataset. One the QMOF dataset, it reaches near taining validity.

Critical weaknesses
  1. The method trades novelty for validity. The validity mostly come from recalling from the training set.
Questions for authors

The authors identified the autoencoder as a bottleneck and trained one on LeMatBulk. However the authors did not report the generation quality of this autoencoder. How does the GLASS autoencoder trained on LeMatBulk work in a flow?

1 comment

Dear Reviewer Cwmj,

Thank you for the careful review of our manuscript and for outlining positive aspects as well as weaknesses of the present work. We address your concern and your question below.

W1: The method trades novelty for validity and achieves validity by recalling the training set.

We agree that limited novelty is an important limitation of our current MOF results. This limitation is explicitly acknowledged in the abstract, Section 5, and the limitations paragraph in Section 6, with detailed novelty analyses in Appendix G.2–G.3. In particular, we report that most valid MOF generations reproduce frameworks present in the training data.

The results on small inorganic crystals (MP20) demonstrate a more favorable balance between validity and novelty. At the 50k checkpoint, GLASS achieves 96.56% validity and 52.71% novelty after relaxation, together with competitive stable and metastable unique-and-novel yields (Table 1).

We view our current MOF results as representing the high-validity, limited-novelty end of an empirical spectrum. They demonstrate the potential of the approach for generating larger all-atom structures with high structural validity, while highlighting novel framework generation as a challenge that is still unresolved. Exploring and improving this tradeoff, particularly through better autoencoder generalization to increase novelty while retaining high validity, is an important direction for future work.

Q1: How does the GLASS autoencoder trained on LeMatBulk work in a flow?

We trained a flow model on the 64-dimensional LeMat-Bulk autoencoder, using the same experimental setting as for MP20 (50k steps) and an additional later checkpoint (100k steps). We report generation results on LeMat-GenBench, which we have also incorporated into the revised manuscript (Appendix E.5, Table 13).

Flow stepStructuresValid (%)Unique (%)Novel (%)Metastable (%)SUN (%)MSUN (%)
50kraw86.45 ± 0.4286.44 ± 0.4265.24 ± 0.336.55 ± 0.580.00 ± 0.003.33 ± 0.65
50kNequIP92.33 ± 0.4092.31 ± 0.4073.65 ± 0.9224.03 ± 0.720.20 ± 0.0815.65 ± 0.64
100kraw87.60 ± 1.5687.59 ± 1.5360.99 ± 1.069.32 ± 0.680.00 ± 0.004.73 ± 0.79
100kNequIP92.07 ± 0.6192.03 ± 0.5769.88 ± 0.7426.80 ± 0.660.25 ± 0.1016.61 ± 0.84

LeMat-GenBench measures novelty against LeMat-Bulk, which contains this model's full training set. At 50k steps, 65% of the raw samples are novel (74% after relaxation), so the model's validity cannot be explained by recalling training data.

Compared with the MP20 model at 50k steps (relaxed: 96.56% valid, 63.63% metastable, 22.31% MSUN), validity, metastability, and MSUN are lower. We attribute this to an earlier position on the validity-novelty frontier. At batch size 1024, 50k steps correspond to about 1,900 passes over MP20 but only about 10 passes over the 5.1M LeMat-Bulk structures. The 100k checkpoint is consistent with this: novelty decreases (73.65% to 69.88% relaxed), while metastability (24.03% to 26.80%) and MSUN (15.65% to 16.61%) increase.

Overall, these results show that with an autoencoder that generalizes, GLASS can generate mostly novel structures, including metastable ones, beyond its training data.

Comments on the paper

1 comment

We thank all reviewers for their careful and constructive reviews. The reviewers found the problem interesting and well motivated (qtPE, vpVs) and the slot-based decoder novel (vpVs). They appreciated the hypothesis for why all-atom generation is hard to scale, together with the experiments that test it (Cwmj), and the ablations and transparent reporting of the autoencoder reconstruction gap (qtPE).

The main concerns were the limited novelty on QMOF150 (Cwmj, qtPE) and the missing comparison with building-block MOF methods (qtPE). The reviewers also asked how the LeMat-Bulk autoencoder performs in a flow (Cwmj), what the pair-distance loss contributes (vpVs), and for additional related work (vpVs).

New experiments

  • Flow on the LeMat-Bulk autoencoder (Cwmj Q1, qtPE W1). At 50k steps, the model reaches 86.5% raw validity and 65.2% novelty. Novelty is measured against LeMat-Bulk, which contains the model's training set, so these samples are not recalled. This supports our view that the low MOF novelty stems from the autoencoder's poor generalization on the small QMOF150 dataset, and is not inherent to the approach.
  • Pair-distance loss ablation (vpVs Q3). Over three seeds on MP20 and QMOF150, the loss has a marginal effect and is not essential.

Changes in the revision

  • Appendix E.5 (Table 13) reports the LeMat-Bulk generation results.
  • Appendix E.4 (Tables 10 and 11) reports the pair-distance loss ablation.
  • Section 5 lists the published VNU of MOFDiff, MOFFlow-2, and MOF-BFN and states why they are not directly comparable (qtPE).
  • Appendix A includes the requested related work, and the bibliography is corrected (vpVs, qtPE).
  • Appendix F.3 contains additional sampling details for baseline methods.

Correction of the external LeMat-GenBench benchmark

Independently of our work, a bug was identified in the LeMat-GenBench evaluation framework. Due to this bug, an incorrect evaluation model was selected, which affected the energy-based metrics of all models on the leaderboard. The benchmark maintainers have corrected the leaderboard, and we have re-evaluated GLASS with the corrected benchmark and updated baseline values from the corrected leaderboard (Tables 1 and 14). Validity, uniqueness, and novelty are unchanged.

Across models, the stable fraction dropped, SUN decreased, and MSUN increased slightly. For example, with relaxation, GLASS changed from 1.92% to 1.13% SUN and from 21.16% to 22.31% MSUN, and Crystalite from 1.50% to 1.20% SUN and from 22.60% to 23.60% MSUN. Our conclusion that GLASS is competitive on MP20 is unchanged.

We believe the new results strengthen the central claim of the paper: separating atom-wise correspondence from generative transport enables high-validity all-atom generation at larger structure sizes. We are happy to answer further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.