The literature review and empirical comparisons are seemingly incomplete. Firstly, there is a noticeable lack of discussion on Chemeleon2, published in Nature Machine Intelligence, which also uses a VAE-based latent diffusion framework and demonstrates strong metastable-material generation after reinforcement learning. Chemeleon also applies the same latent diffusion model that the proposed work does, showing a lack of novelty in the proposed work. A critical ablation of their joint latent space could be performed by comparing embeddings from Chemeleon and CrysUniNet to further show the importance of the ComFormer + Transformer representation. Furthermore, more credit must be assigned to existing work on transformer-based generators such as Crystalite which serve as the SOTA for SUN performance as far as the reviewer knows.
The authors should acknowledge recent crystal generators (CSP and DNG alike) such as Crystalite, OMatG, and Kinetic Langevin Diffusion, alongside symmetry-aware approaches including SymmCD, SGEquiDiff, Space Group Conditional Flow Matching and MaskGXT. These works provide context for assessing the contribution of their geometric representation and generation framework. The discussion of property-guided generation should also include OMatGRPO which improves MSUN yield and demonstrates how single-element repackings and insufficient reference-hull coverage can inflate apparent discovery performance. These findings motivate additional diagnostics of the generated chemical and structural distributions.
The authors should evaluate their pretrained generator on LeMat-GenBench under a matched sampling budget of k=2500 structures as opposed to their 10,000. The benchmark forces a standard thermodynamic stability definition and proper novelty assessment allowing the authors to more effectively place themselves relative to other generators. We also recommend the authors sample at least 10 independent k=2500 batches to permit reporting mean performance and error bars, while retaining comparisons at the same batch size rather than padding predictions up to 10000. This requires additional inference and benchmark evaluation, but no retraining. Matching the benchmark’s relaxation, novelty, and stability protocols would make comparisons substantially more informative than juxtaposing percentages from incompatible pipelines. If time permits, comparisons against MP-20-only models do not isolate the contribution of the proposed representation. A matched-data ablation is needed, this means training the VAE and latent flow model on only MP20 is needed for a proper comparison.
If the claimed improvements survive matched data, evaluation protocols, and uncertainity estimates I will consider raising my score.