- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes SGBNet, which aims to improve geometric-to-scalar interactions while achieving edge-efficient computation. On the SPICE-MACE-OFF and MLIPAudit benchmarks, SGBNet demonstrates state-of-the-art or competitive performance. In terms of computational efficiency, SGBNet outperforms eSEN and establishes a new Pareto frontier in the accuracy–efficiency trade-off.
Critical strengths
- Designing edge-efficient MLIPs with scalar–geometric interaction capabilities is an important and meaningful research.
- The paper is well-organized and clearly written, providing clear descriptions of the different modules of SGBNet.
Critical weaknesses
- The analysis of "low-rank readout coupling" appears to identify a potential structural limitation rather than a sufficiently validated practical bottleneck. Eq. (1) only establishes an upper bound on the rank determined by the dimension (K) of the intermediate statistics, but does not demonstrate whether this bound actually limits the representational capacity of existing models in practice. The paper would be more convincing if it could establish that the accuracy of existing MLIPs is indeed constrained by (K). Unfortunately, no theoretical analysis or experimental evidence is currently provided to support this claim.
- The performance of SGBNet is not particularly impressive, especially in terms of accuracy. The paper should include comparisons with more recent MLIPs, such as EquiformerV3[1], DPA4[2], and AllScAIP[3]. Furthermore, there is already a substantial body of work exploring scalar–higher-order geometric interactions (e.g., GeoMFormer[4] and HIENet[5]). Similarly, the proposed geometric-to-scalar interaction mechanism shares similarities with HEGNN[6] and GotenNet[7]. I believe the paper should provide a more thorough methodological comparison with these existing approaches and include them in the experimental evaluation.
[1] Liao et al. EquiformerV3: Scaling Efficient, Expressive, and General SE(3)-Equivariant Graph Attention Transformers. arXiv preprint.
[2] Li et al. DPA4: Pushing the Accuracy–Cost Frontier of Interatomic Potentials with EMFA SO(2) Convolution. arXiv preprint.
[3] Qu et al. A recipe for scalable attention-based MLIPs: unlocking long-range accuracy with all-to-all node attention. ICML2026.
[4] Chen et al. GeoMFormer: A General Architecture for Geometric Molecular Representation Learning. ICML2024.
[5] Yan et al. Learning Materials Interatomic Potentials via Hybrid Invariant-Equivariant Architectures. TMLR.
[6] Cen et al. Are High-Degree Representations Really Unnecessary in Equivariant Graph Neural Networks? NeurIPS2024.
[7] Aykent at al. GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks. ICLR2025.
Questions for authors
- See Weakness 1. Could the authors provide theoretical justification or experimental evidence demonstrating that the representational capacity of existing equivariant MLIPs is indeed limited by low-rank readout coupling?
- Compared with recent MLIPs (e.g., EquiformerV3, DPA4, AllScAIP, GeoMFormer, and GotenNet), does SGBNet offer advantages in terms of accuracy and computational efficiency?
- I believe a model should only be considered "edge-efficient" if its edge-level computation and memory consumption are lower than those at the node level. Under an experimental setting where the number of atoms is fixed while the average number of neighbors varies, how do the runtime and memory consumption of SGBNet scale?
- I suggest reporting the training efficiency of different models under the same hardware and computational backend to enable a more comprehensive comparison.
- For Matbench Discovery, could the authors provide comparisons using the officially recommended training datasets (MPTrj and OMat+sAlex+MPTrj)?
2 comments
3. I believe a model should only be considered "edge-efficient" if its edge-level computation and memory consumption are lower than those at the node level. Under an experimental setting where the number of atoms is fixed while the average number of neighbors varies, how do the runtime and memory consumption of SGBNet scale?
We thank the reviewer for pointing out the ambiguity in our description and apologize for the lack of clarity. We do not claim that SGBNet reduces the overall complexity to , or that edge-level costs are lower than node-level costs. Rather, "edge-efficient" refers to reducing edge-dependent computation and memory.
With 1,818 atoms fixed, we varied the cutoff radius and fitted each metric against the number of edges. The resulting slopes are:
| Metric | eSEN-3.2M | SGBNet-S | Reduction |
|---|---|---|---|
| Runtime (μs/edge) | 1.172 | 1.145 | 2.3% |
| Memory (KiB/edge) | 127.80 | 85.43 | 33.2% |
| FLOPs (MFLOPs/edge) | 15.95 | 9.75 | 38.9% |
These results show that SGBNet reduces edge-dependent memory and computation while retaining complexity. Detailed fitting results are provided in the appendix (Figure 4 and Table 9). We have also clarified this definition in the Introduction to avoid this ambiguity.
4. I suggest reporting the training efficiency of different models under the same hardware and computational backend to enable a more comprehensive comparison.
We appreciate this suggestion. We will include an additional comparison of training efficiency in the appendix, using the same hardware and computational backend to ensure a more consistent evaluation.
5. For Matbench Discovery, could the authors provide comparisons using the officially recommended training datasets (MPTrj and OMat+sAlex+MPTrj)?
We appreciate this suggestion. Due to the limited rebuttal time, we are unable to conduct the additional experiments with the two requested training configurations. The results reported in our paper are obtained using sAlex+MPTrj, which shares the sAlex and MPTrj components with the officially recommended OMat+sAlex+MPTrj configuration and is also adopted in prior works such as MACE and OrbV2. We acknowledge that evaluating the model on MPTrj alone and with the full OMat+sAlex+MPTrj configuration would provide a more comprehensive comparison. We will consider these additional experiments in future work.
We appreciate the reviewer raising this concern. We establish the rank limitation analytically and demonstrate its practical impact empirically: the effective Jacobian rank increases with the readout rank bound, accompanied by consistent improvements in force accuracy. Thus, the low-rank readout coupling is not merely a theoretical limitation but a practical representational bottleneck in eSEN. The results are as follows:
Please see Common Response 1 for details.
We thank the reviewer for the suggestions. We have conducted additional experiments and expanded the comparisons with recent MLIPs. The results show that SGBNet-S achieves the fastest inference, while maintaining better accuracy than GotenNet, EquiformerV3, and DPA4-Neo. We acknowledge that EquiformerV3 achieves strong results on materials benchmarks, although its performance is less competitive on the molecular benchmark considered here. DPA4-Air achieves higher accuracy, but at a substantially higher computational cost. We have expanded the Related Work and Discussion sections to discuss the methods mentioned, and included the models we could evaluate in the quantitative comparison (GotenNet is listed in Appendix).
All our runtime measurements were conducted on an A100 GPU. Our cross-framework eSEN comparison indicates that implementation differences can account for approximately 24% of runtime, which is insufficient to explain the observed speed gaps.