acceptodds
Under review as a conference paper at ICLR 2027

MiAD: Mirage Atom Diffusion for De Novo Crystal Generation

Abstract

In recent years, diffusion-based models have demonstrated exceptional performance in searching for simultaneously stable, unique, and novel (S.U.N.) crystalline materials. However, most of these models don't have the ability to change the number of atoms in the crystal during the generation process, which limits the variability of model sampling trajectories. In this paper, we demonstrate the severity of this restriction and introduce a simple yet powerful technique, mirage infusion, which enables diffusion models to change the state of the atoms that make up the crystal from existent to non-existent (mirage) and vice versa. We show that this technique improves model quality by up to compared to the same model without this modification. Mirage infusion also improves MatterGen and FlowMM, suggesting that the benefits are not limited to a single generative architecture. The resulting model, Mirage Atom Diffusion (MiAD), is an equivariant joint diffusion model for de novo crystal generation that is capable of altering the number of atoms during the generation process. MiAD achieves an DFT S.U.N. rate on MP-20 under our evaluation protocol, the highest among the models in Table-1.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion models, de novo crystal generation, generative models, material science, deep learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_eTetOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, the authors address a key limitation of existing equivariant diffusion and flow-based models for crystal generation: they do not allow the number of atoms to change during sampling. Typically, these models sample the number of atoms N_{atoms} from the empirical training distribution and then perform diffusion on atom types and coordinates.

To address this limitation, the authors propose MiAD (Mirage Atom Diffusion), a novel generative framework that extends DiffCSP by allowing the number of atoms to change dynamically during generation.

The central idea is simple. They introduce mirage atoms (atom type 0), which can be treated as virtual or masked atoms. These atoms can appear or disappear during the generation process, allowing the model to adjust the number of atoms dynamically.

The authors claim and empirically demonstrate that this flexibility improves the SUN metric compared to existing bas

Critical strengths
  1. The paper is very well written, very easy to go through and the core idea is well-motivated. Also, the authors did a thorough literature survey, citing most of the existing papers.
  2. The idea of Mirage Atom Infusion and reduction is very simple, but empirically shown to be very effective, particularly in improving the SUN score.
  3. The experimental section of the paper is very strong. They have done denovo generation on both the MP-20 and MPTS-52 datasets and compared against both stability and validity metrics (in the Appendix). Also, for stability, they have evaluated using both DFT and MLIP, which makes the evaluation setup complete.
Critical weaknesses

I guess the major weakness of the paper is its very limited novelty. They only add flexibility to change atoms in an equivariant joint diffusion model for de novo crystal generation. Also, this idea of adding a MASK node or virtual nodes is being explored in other domains like Molecules or Social Networks. So overall, though the idea seems to be effective, but the methodological contribution is very limited.

Questions for authors
  1. Table 1,2: # of Unique and Novel Materials by MiAD is lower than in other baseline models. Ideally, MiAD gives more flexibility in choosing atom types. The authors also claim: "this constraint reduces the model’s flexibility and hinders its ability to explore a broader range of plausible crystal structures, impacting both the diversity and quality of the generated outputs." However, it seems to reduce diversity. Why? There is no justification for that.

  2. Table 3: Transferability of Mirage infusion across joint (L, F, A) generators - This experiment setup is good, but I would love tp check how more sophisticated models like Only LLM or ADiT or FlowLLM or CrysLLMGen perform with Mirage infusion. They have shown better SUN than DiffCSP, MatterGen, or FlowMM.

  3. From Table 4: Ablation study on the number of Mirage atoms in MiAD- there is no consistency. With N_m = 30, performance degrades, but N_m = 35 improves. This part is not clear. Authors claim: “Beyond a certain number, the neural network loses the capacity to manage them effectively”, however, it seems to be not correct.

  4. What is the distribution of # of atoms generated in 10000 materials, like how many are less than 5, less than 10, 15, and 20 for MP-20 and MPTS-52

  5. Authors have chosen Nm = maximum number of atoms+5. What if we choose N_atom ~ P(N_atom) + 5? Will it improve the SUN for DiffCSP and MatterGen?

  6. They assume L remains fixed; however, adding or removing new atoms may need to adjust L a bit as well.

  7. Codebase not provided. I would love to reproduce the results.

4 comments

We thank the reviewer for the detailed questions and for praising the writing, motivation, literature coverage, and the strength of the experimental section. We clarify the novelty point below and then answer each question.

W1. Limited novelty / MASK nodes in other domains

We agree that MASK/virtual nodes appear in other domains. Inventing placeholders in the abstract is not our claim.

We respectfully disagree that this makes the methodological contribution limited. Equivariant joint (L,F,A)(L,F,A) crystal generators typically keep the atom count fixed along the sampling trajectory. MiAD removes that restriction with uniform mirage infusion and masked coordinate loss. Appendix B shows that the usual placeholder recipes completely break for de novo crystals: the drug-design-style Center + NonMask setting makes the model worse than DiffCSP, and only Unif + Mask delivers the gains. That working, crystal-specific formulation is the contribution.

Q1. Why are Unique & Novel counts lower in Tables 1-2?

This is the intended stability-diversity trade-off under S.U.N., already discussed in Sec. 4. Isolated Unique&Novel rates (or isolated stability rates) can be gamed. S.U.N. rewards the joint objective. MiAD often trades some raw U&N among all samples for substantially higher stability among generated crystals, and still achieves the best overall S.U.N. For example, relative to WyFormer, MiAD has lower Unique&Novel among stable crystals but higher stability, and wins on S.U.N. Relative to methods with very high stability but low novelty among stable crystals (e.g. some ADiT settings), MiAD keeps a stronger Unique&Novel rate among stable structures. The claim about flexibility refers to exploring compositions/structures along a trajectory by adding/removing atoms, not to maximizing the unconditional U&N rate of every sample.

Q2. Apply mirage infusion to LLM / ADiT / FlowLLM / CrysLLMGen, etc.

We cannot add those backbones in this rebuttal. The paper already shows transfer across three independently reproduced (L,F,A)(L,F,A) generators (DiffCSP, MatterGen, FlowMM). Extending mirage infusion to Only-LLM / ADiT / FlowLLM / CrysLLMGen is interesting future work, but it is a much larger experimental campaign (different representations, conditioning, and training stacks).

Q3. Inconsistent dynamics in Table 4 after Nm=25N_m=25

We disagree that Table 4 shows a contradictory "degrades then improves" law that refutes the capacity claim. Beyond the optimum Nm=25N_m=25, performance does not keep increasing:

NmN_meq-V2 S.U.N.CHGNet S.U.N.
205.312.0
255.512.9
304.711.3
354.611.8

CHGNet rises slightly from 30 to 35 (11.3 to 11.8), but eq-V2 continues downward (4.7 to 4.6). There is no consistent recovery that would motivate further increases of NmN_m. Both metrics remain below the Nm=25N_m=25 optimum. Past that point, the trend plateaus / fails to improve. It is not a strictly monotone collapse at every step, and we do not claim that.

Q4. Distribution of #atoms in 10k materials for MP-20 and MPTS-52

For MP-20, this is already in the paper: Appendix J / Figure na_dist reports the atom-count distribution for S.U.N. crystals and discusses how often the model changes atom count during generation.

For MPTS-52, the fraction of structures in each atom-count bin is:

number of atomsMPTS-52MiAD trained on MPTS-52
1-50.1620.131
6-100.2040.228
11-150.1400.142
16-200.1360.140
21-250.0860.088
26-300.0800.077
31-350.0390.063
36-400.0820.065
41-450.0280.034
46-520.0420.029

MiAD's generated atom-count distribution closely tracks the MPTS-52 training distribution across bins.

Q5. What if Natom∼P(Natom)+5N_{\mathrm{atom}} \sim P(N_{\mathrm{atom}})+5 instead of max⁡+5\max+5?

If the suggestion is to run mirage infusion with a small additive mirage budget on top of P(N)P(N) rather than padding to max⁡+5\max+5, that is exactly MiAD-WRNA (Appendix H). Per crystal, we use Nm=Natoms+Unif[0,10]N_m = N_{\mathrm{atoms}} + \mathrm{Unif}[0,10] (+5 mirages on average), and at generation we sample from the training atom-count distribution plus that small mirage budget.

On MPTS-52 with the DiffCSP backbone, this already improves CHGNet S.U.N. over DiffCSP (8.2% to 9.2%) while cutting the training-time multiplier from about ×2.26 (fixed Nm=57N_m=57) to about ×1.14, though it is slightly below full MiAD (9.9%). So for DiffCSP-style joint generators the answer is yes: the P(N)+5P(N)+5-style schedule helps, with a compute-quality trade-off versus max-padding.

We did not rerun that exact WRNA schedule on MatterGen. Given that standard mirage infusion already transfers to MatterGen under otherwise identical configs (Table 3), we expect the same budget trade-off to apply, but we do not have that MatterGen-WRNA number.

Q6. Does LL need adjustments when atoms are added/removed?

No additional special adjustment is required. Lattice LL is already generated jointly by the diffusion/flow process. The model learns to produce lattices that are consistent with the final (post-mirage-resolution) composition. Empirically we did not find that mirage infusion requires a separate lattice-correction stage. The lattice head is trained end-to-end to match lattices of complete crystals, including those whose atom count changed relative to intermediate states.

Q7. Codebase

An anonymized zip of the codebase is linked in Appendix G ("Code").

Thank you for the response. I remain unconvinced about the methodological novelty. While MiAD improves S.U.N. over DiffCSP empirically, the core idea of placeholder atoms is not new, and the contribution lies in a specific formulation.

Moreover, its effectiveness is shown only for equivariant diffusion and flow models (DiffCSP, MatterGen, FlowMM):

  • In Table 1 (DFT), transformer-based and Wyckoff-based models such as ADiT and WyFormer achieve higher stable, metastable and unique/novel rates.
  • Autoregressive and LLM-based models handle variable atom counts natively, so the problem MiAD addresses may not arise for them.

Could the authors clarify whether mirage infusion is a technique specific to equivariant diffusion/flow models, or whether it can benefit crystal generation models more generally, for example by applying it to ADiT?

Thank you for the follow-up.

Novelty. We still disagree with this assessment of our contribution. Could you point us to the exact reference that you think undermines its novelty, so that we can position our work fairly? The closest works we are aware of (Schneuing et al., CrystalDiT, UniMat, WyckoffDiff) are already discussed and compared against in Section 3 and Appendices B.3 and C. In any case, we believe that the simplicity of the proposed approach, its significant improvement over the base model, and its transferability across training methods and architectures are valuable for the field.

Table 1. We strongly disagree with this reading. Comparing the columns one by one is not meaningful, since each can be trivially maximized: stability by copying the training set, novelty by generating random structures. No single model is better than MiAD on all of these rates: ADiT is more stable but much less novel, and WyFormer is more novel but far less stable. We state this in Section 4, and this is exactly why S.U.N. and M.S.U.N. are the standard metrics for comparing de novo crystal generation models, and why we use them. On these metrics, MiAD is ahead of both ADiT and WyFormer.

Autoregressive and LLM-based models. We would like to clarify a misunderstanding about the problem MiAD addresses. These models produce crystals of different sizes, but so does DiffCSP. The limitation we address is different: the atom count cannot be revised during generation. This is also true for autoregressive models, which add atoms one by one and never remove or change an atom once it is generated. Mirage atoms are not padding. They let the model add and remove atoms at any step, looking at the whole structure, and the model uses this to correct its own mistakes (Appendix L and our response to Reviewer ZyR8).

Moreover, the LLM-based and autoregressive models in Table 1 (FlowLLM and WyFormer) are well below MiAD (S.U.N. 4.7% and 4.8% vs. 8.2%).

ADiT. Yes, we have now tested it. ADiT is a latent diffusion model, so we kept its VAE frozen and added mirages only to the latents denoised by the Transformer. Each crystal is extended to a fixed number of slots. The extra slots are mirages: they hold latents of random atoms, so they look like ordinary atoms, and every slot carries an atom/mirage indicator that is noised and denoised together with its latent. Sampling starts from noise in all slots, with no pre-sampled atom count; slots can switch between atom and mirage along the way, and mirage slots are removed before decoding. The main ingredients of MiAD thus carry over directly: we tested all four settings of Table 6 and, as in MiAD, Center placeholders do not work, while Unif works both with and without the mask (we report Unif + NonMask). A fine-tune of the released checkpoint of jointly trained ADiT improves both metrics under the protocol of Table 3:

M.S.U.N. (%)S.U.N. (%)
ADiT46.910.1
ADiT + Mirage Infusion54.311.9

The key switching behaviour of MiAD is preserved, and slots can change state in both directions, as in Appendix L.

We will add this result to Table 3 and update the paper.

Reviewer_gS3nOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MiAD. MiAD introduces mirage infusion, a technique that augments crystal structures with placeholder atoms that can transition between existent and non-existent states, allowing diffusion models to dynamically add or remove atoms during generation. The work identifies effective design choices—including uniformly initialized mirage coordinates and masked coordinate losses—and validates the method through extensive ablations, DFT-based evaluation, and experiments on MP-20, MPTS-52, and Alex-MP20. It also shows that mirage infusion consistently improves different generative backbones, including DiffCSP, MatterGen, and FlowMM.

Critical strengths
  1. Simple and effective core idea. Mirage atoms allow the model to add or remove atoms during diffusion, directly addressing the fixed-atom-count limitation of many crystal generators.
  2. Strong physical validation. The main comparison uses DFT evaluation on 10,000 generated structures, providing stronger evidence of stability than MLIP-only experiments.
  3. Good transferability and ablations. Mirage infusion improves DiffCSP, MatterGen, and FlowMM, while the paper carefully studies atom count, initialization, loss masking, and loss weighting.
Critical weaknesses
  1. High computational overhead. The main MP-20 configuration substantially increases per-step time and memory and requires many more training epochs than the DiffCSP baseline.
  2. Heterogeneous baseline comparisons. Some DFT baseline results come from prior publications rather than a completely unified reproduction and evaluation pipeline.
  3. Insufficient uncertainty reporting. The main results do not include multiple training seeds, confidence intervals, or statistical significance tests.
  4. The novelty and mechanism claims are somewhat overstated. Related dummy-atom ideas already exist, and the “error-correction” analysis is based on a selectively chosen subset of successful trajectories.
Questions for authors

Same as weakness

3 comments

We thank the reviewer for the clear summary and for highlighting simplicity, DFT validation, transferability, and ablations as strengths. We agree with several of the weaknesses and clarify below.

W1. High computational overhead

We agree that the overhead is real: a MiAD training step is about 4.3× slower than a DiffCSP step, and MiAD needs more epochs to reach its best quality (Appendix G.5). Three points put this cost into perspective.

More training does not let the baseline catch up. In Figure 2, both models are trained for up to 3000030000 epochs. DiffCSP reaches its best S.U.N. early, at around 20002000 epochs, and degrades with longer training, while MiAD stays above DiffCSP at every epoch. In particular, DiffCSP trained for 3000030000 epochs uses roughly the compute of MiAD trained for 70007000 epochs and remains well below it. The gain therefore comes from mirage infusion rather than from the larger training budget.

The overhead is small compared with validation. In a discovery pipeline, candidates are typically filtered for uniqueness and novelty first, and only then validated with DFT, which is far more expensive than sampling. Among unique and novel structures, the DFT stability rate is 11.2%11.2\% for MiAD against 3.7%3.7\% for DiffCSP, so MiAD needs about 3×3\times fewer DFT runs per stable unique-and-novel material.

The overhead can be reduced. For datasets with a wide range of atom counts, MiAD-WRNA (Appendix H) adds a random number of mirage atoms to each crystal, five on average, instead of padding every crystal to a fixed size. On MPTS-52, this lowers the training-time multiplier relative to DiffCSP from 2.26× to 1.14× and keeps more than half of the gain: eq-V2 S.U.N. is 2.5% for MiAD-WRNA, against 1.6% for DiffCSP and 3.2% for MiAD.

W2. Heterogeneous baseline DFT comparisons

We agree that some DFT baseline numbers come from prior publications. Our policy is: we include DFT results only from papers that document a configuration matching ours (Materials Project-style settings), so those numbers are as comparable as the literature allows. Fully re-running unified DFT for all baselines on 10k structures each is prohibitively expensive for us. Re-running DFT for only a subset of models would still leave a heterogeneous table and could introduce new inconsistencies. Given that constraint, citing matched published DFT configs while running our own DFT for MiAD (and matched MLIP protocols for controlled ablations/transfer) is the inevitable practical compromise rather than a selective reporting choice.

W3. Insufficient uncertainty reporting

We agree. There are two sources of variance, sampling and training, and we address them separately.

Sampling. To measure this variance directly, we sampled 1010 independent sets of 1000010000 crystals from MiAD and from baseline models and evaluated each set with CHGNet. The run-to-run standard deviation of S.U.N. is 0.20.2–0.40.4 percentage points, in line with the binomial standard error at n=10000n = 10000, which corresponds to a 95%95\% interval of about ±0.5–0.7\pm0.5–0.7 points for a single table entry. The improvements on which our main claims rest are considerably larger: mirage infusion adds 3.23.2–4.34.3 points of S.U.N. to each backbone in Table 3, and MiAD's DFT S.U.N. exceeds DiffCSP by 4.94.9 points and the next best model by 1.71.7 points (Table 1).

Training. The final model was trained with a single seed. As a proxy for training variance, the five further MiAD models in Tables 4 and 5, trained independently with different NmN_m and loss weights, reach CHGNet S.U.N. between 11.3% and 12.9% (standard deviation 0.5 points), all well above DiffCSP (8.9%), and MiAD stays above DiffCSP at every checkpoint in Figures 2 and 3.

W4. Novelty / mechanism claims somewhat overstated

Part 1 (related dummy-atom ideas). We agree that placeholder / MASK / dummy-atom ideas exist in other domains and that CrystalDiT also pads with dummy atoms. We do not claim to invent "virtual atoms" in the abstract. Our contribution is a working expansion of the dominant equivariant joint (L,F,A)(L,F,A) crystal generators, with crystal-specific choices (uniform infusion + masked coordinate loss) that we show are necessary (Appendix B), plus transfer across DiffCSP / MatterGen / FlowMM and DFT-backed S.U.N. gains. Related work already discusses these distinctions.

Part 2 (error-correction on successful trajectories). We address this point in detail in our response to Reviewer ZyR8 (W1). In short, we repeated the analysis of Appendix L on 600600 random late-pruning events from 2000020000 new trajectories, regardless of the final outcome. Removing the atom improves the intermediate structure as often in non-S.U.N. crystals as in S.U.N. ones (median EhullE_{\mathrm{hull}} 0.197→0.0490.197 \to 0.049 and 0.094→−0.0190.094 \to −0.019 eV/atom, improved in 84%84\% and 82%82\% of cases), so the effect is not an artefact of selecting successful trajectories. We also agree that this analysis illustrates a mechanism rather than proving the source of the gains, which rests on the controlled comparisons in Tables 1–3, and we will rephrase Appendix L accordingly.

Thank you I keep the positie reviews.

Reviewer_ZyR8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes MiAD, which aims to improve the ability of crystal generative models to discover novel, stable structures by allowing atoms to be added or removed during generation. The authors encode the presence/absence of an atom as a special atom type, use uniformly placed virtual ("mirage") atoms, and mask the coordinate-reconstruction loss for them. Experiments show the benefit of this design on DiffCSP, as well as its transfer to MatterGen and FlowMM. Overall, the paper offers a simple, reusable modification of the generative space and supports its effectiveness with fairly thorough experiments.

Critical strengths
  1. The method is simple and reusable. By tying atom existence into type generation, the model can add or remove atoms while adjusting a structure; this modification yields gains across three backbones.
  2. The experiments fairly thoroughly identify the effective design choices. The initialization and loss ablations show that naively adding virtual atoms does not guarantee improvement — only the combination of uniform placement and masking the coordinate-reconstruction loss achieves the best results.
Critical weaknesses
  1. The error-correction explanation relies mainly on hand-picked successful samples. The authors show that these samples become more stable after atoms are removed, but do not check whether the same phenomenon also appears in failure cases. On the basis of these successful cases alone, it is still hard to tell whether adding/removing atoms is the main reason for the improved generation quality.
  2. The key ablations are placed mostly in the appendix and are not presented sufficiently in the main text. Uniformly placing virtual atoms and masking the coordinate loss are important designs of this paper. The corresponding comparison results are mostly in the appendix, while the main text only briefly summarizes them, so the reader has to flip back and forth to judge the effect of these two choices.
Questions for authors

Clarification of results: the paper notes that some structures from CrystalDiT lose novelty after relaxation. Are there DFT S.U.N. results that re-assess novelty based on the post-relaxation structures? If so, please report that number, so that the gap between the two methods can be judged under the same criterion.

2 comments

We thank the reviewer for the careful reading and for recognizing the simplicity, reusability, and thorough ablations of mirage infusion. We address the weaknesses and the clarification question below.

W1. Error-correction analysis / whether add/remove is the main reason for gains

Part 1 (hand-picked successful samples).

The 5757 crystals in Appendix L are not hand-picked: they are all S.U.N. crystals among 10001000 samples whose last real → mirage transition occurred at t≤150t \le 150. The reviewer is right, however, that selecting on a successful outcome cannot tell whether the effect is specific to successes.

To check this, we generated 2000020000 new trajectories from the same checkpoint. In 59.8%59.8\% of them, the last real → mirage transition occurs at t≤150t \le 150, so late pruning is the typical behaviour of the model rather than a rare event. Then, we sampled 600600 random crystals with the late pruning event, regardless of the final outcome, and repeated the comparison of Appendix L (out of 600600, 9191 have overlapping atoms and cannot be relaxed, almost always the counterfactual MtM_t).

Final crystalnMedian EhullE_{\mathrm{hull}} (eV/atom), MtM_t → M^t\hat M_tImprovedNon-trivial symmetry
S.U.N.620.094 → −0.01982.3%8.1% → 17.7%
Not S.U.N.4470.197 → 0.04984.1%6.1% → 13.7%

The Table shows that removing the atom improves the intermediate structure just as often in non-S.U.N crystals as in S.U.N crystals, so the effect is a systematic property of the sampler rather than of the selected examples. The removed atom is not arbitrary either: in 39.6%39.6\% of events it is the atom with the largest CHGNet force in its crystal, whereas chance would give about 9%9\%.

Part 2 (main concern: "hard to tell whether adding/removing atoms is the main reason").

We agree that the success of MiAD cannot be attributed solely to the error-correction mechanism, and we do not claim that pruning is the only source of the gains. Appendix L illustrates one channel through which the model improves generation. We will rephrase Appendix L accordingly.

The evidence we rely on is the controlled quantitative comparison:

  1. Same backbone, only mirage infusion changed. Holding DiffCSP fixed and enabling dynamic atom count raises CHGNet S.U.N. from 8.9% to 12.9% and DFT S.U.N. to 8.2% (Tables 1-2).
  2. Design ablations isolate how add/remove must be implemented. Uniform infusion + masked coordinate loss is necessary. Naive placeholders (e.g. center-of-mass / unmasked) do not deliver the gains and can hurt the baseline (Appendix B).
  3. Transfer across generators. The same change improves MatterGen (6.1% to 10.4%) and FlowMM (7.0% to 10.2%) under otherwise identical configs (Table 3).
  4. Capacity / compute controls. Gains are not explained by merely spending more compute or by generating >20-atom crystals (Appendix J).

If the reviewer has a specific protocol in mind, we are happy to discuss it.

W2. Key ablations mostly in the appendix

We agree. Uniform placement and coordinate-loss masking are the central design choices, and their comparison was previously only summarized in the main text. In the revision, we move the main design-choice ablation (Table 6: Unif/Center × Mask/NonMask) into the Experiments section, with a short discussion of why only Unif + Mask works. Further ablations (NmN_m, loss reweighting) remain in Appendix B.

Q. CrystalDiT DFT S.U.N. with novelty re-assessed after relaxation

Yes. This concern is exactly why we flag CrystalDiT's protocol in Sec. 4. CrystalDiT filters uniqueness/novelty before prerelaxation and then estimates stability on that subset. In our evaluation, about 11.6% of CrystalDiT crystals that are novel before prerelaxation are no longer novel afterward, which inflates reported novelty/S.U.N. Under a matched eq-V2 protocol, MiAD remains ahead (5.5% vs. 3.1%). We do not have a full re-run of CrystalDiT's published DFT results with post-relaxation novelty recomputed for every structure, as it would require regenerating or re-relaxing their DFT set under a unified pipeline.

I thank the authors for their detailed response. My concerns have been addressed, and I maintain my rating of 3 (weak acceptance).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.