- 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
- 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.
- The idea of Mirage Atom Infusion and reduction is very simple, but empirically shown to be very effective, particularly in improving the SUN score.
- 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
-
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.
-
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.
-
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.
-
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
-
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?
-
They assume L remains fixed; however, adding or removing new atoms may need to adjust L a bit as well.
-
Codebase not provided. I would love to reproduce the results.
4 comments
Q6. Does need adjustments when atoms are added/removed?
No additional special adjustment is required. Lattice 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. (%) | |
|---|---|---|
| ADiT | 46.9 | 10.1 |
| ADiT + Mirage Infusion | 54.3 | 11.9 |
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.
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) 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) 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=25
We disagree that Table 4 shows a contradictory "degrades then improves" law that refutes the capacity claim. Beyond the optimum Nm=25, performance does not keep increasing:
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 Nm. Both metrics remain below the Nm=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_distreports 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:
MiAD's generated atom-count distribution closely tracks the MPTS-52 training distribution across bins.
Q5. What if Natom∼P(Natom)+5 instead of max+5?
If the suggestion is to run mirage infusion with a small additive mirage budget on top of P(N) rather than padding to max+5, that is exactly MiAD-WRNA (Appendix H). Per crystal, we use Nm=Natoms+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=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)+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.