acceptodds
Under review as a conference paper at ICLR 2027

SGBNet: Scalar-Geometric Bidirectional Network for Edge-Efficient Interatomic Potentials

Abstract

Machine learning interatomic potentials often suffer from restricted geometric-to-scalar information flow and edge-dominated computation. We propose SGBNet, an equivariant architecture that improves the accuracy-efficiency trade-off through scalar-geometric bidirectional (SGB) interactions and edge-efficient convolution. SGB interactions improve geometric-to-scalar information flow by introducing explicit pathways from higher-order geometric features to invariant scalar features. In the SO(2) branch, conjugate multiplication cancels shared rotation-induced phases to construct scalar features. Edge-efficient SGB convolution leverages the structure of scalar features to move channel mixing from edges to nodes, reducing edge-domain computation. Experiments demonstrate clear accuracy-efficiency gains on molecular benchmarks, slightly improved performance on materials benchmarks, and scalability to large-scale systems. On SPICE-MACE-OFF, all SGBNet variants lie on the accuracy-efficiency Pareto frontier. The compact SGBNet-N maintains comparable MLIPAudit performance to eSEN-3.2M while reducing runtime and memory consumption by 47% and 60%, respectively. Moreover, SGBNet-N is both faster and substantially more accurate than the Euclidean-based ViSNet, narrowing the divide between efficient Euclidean models and more expressive higher-order equivariant architectures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine learning potentials, machine learning force fields, graph neural network, equivariance

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_4PENOpenReview ↗
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
  1. Designing edge-efficient MLIPs with scalar–geometric interaction capabilities is an important and meaningful research.
  2. The paper is well-organized and clearly written, providing clear descriptions of the different modules of SGBNet.
Critical weaknesses
  1. 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.
  2. 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
  1. 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?
  2. Compared with recent MLIPs (e.g., EquiformerV3, DPA4, AllScAIP, GeoMFormer, and GotenNet), does SGBNet offer advantages in terms of accuracy and computational efficiency?
  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?
  4. I suggest reporting the training efficiency of different models under the same hardware and computational backend to enable a more comprehensive comparison.
  5. For Matbench Discovery, could the authors provide comparisons using the officially recommended training datasets (MPTrj and OMat+sAlex+MPTrj)?
2 comments

1. The analysis of "low-rank readout coupling" appears to identify a potential structural limitation rather than a sufficiently validated practical bottleneck. ... 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?

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:

ConfigurationSeparated normOriginSGB head, dSGB = 8d_\mathrm{SGB} = 8SGB head, dSGB = 32d_\mathrm{SGB} = 32SGB head, dSGB = 128d_\mathrm{SGB} = 128
Rank bound011765128
Effective rank0.0 ± 0.01.0 ± 0.013.0 ± 1.144.8 ± 3.494.1 ± 4.2
Force MAE4.674.704.464.254.11
#Params3,175,5533,175,5533,181,6973,200,1293,278,857
Time38.0538.3737.8037.5537.68

Please see Common Response 1 for details.

2. The performance of SGBNet is not particularly impressive, especially in terms of accuracy. ... I believe the paper should provide a more thorough methodological comparison with these existing approaches and include them in the experimental evaluation. Compared with recent MLIPs (e.g., EquiformerV3, DPA4, AllScAIP, GeoMFormer, and GotenNet), does SGBNet offer advantages in terms of accuracy and computational efficiency?

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).

SGBNet-SGotenNetEquiformerV3DPA4-Neo†DPA4-Air†
Force MAE3.918.165.404.752.85
Energy MAE0.200.841.020.220.12
Time32.4175.3572.4882.12139.10
Memory3.22813.66512.9224.1946.841
† MAE reported in the respective DPA4 paper.
  1. GotenNet: SGBNet-S achieves substantially better accuracy and is approximately 2.3× faster. We use the official small model with one fewer interaction layer to keep its model scale closer to SGBNet-S; it remains considerably slower.
  2. EquiformerV3: SGBNet-S achieves better accuracy and is approximately 2.2× faster. Since no parameter-matched official configuration is available, we use the official channel-width ratios with two interaction layers and lmax⁡=2l_{\max}=2.
  3. DPA4: The reported DPA4-Neo results show that SGBNet-S achieves better accuracy and is approximately 2.5× faster. DPA4-Air achieves better accuracy, but its reported runtime is approximately 4.3× higher than SGBNet-S.
  4. AllScAIP: Its global all-to-all attention targets long-range interactions. The official small model runs out of memory beyond approximately 6,000 atoms and is over 4× slower than SGBNet-S on the 1,818-atom system, precluding a fair large-scale comparison.
  5. GeoMFormer: Its official repository does not provide executable model code, making reproducible quantitative evaluation infeasible. We therefore include a methodological comparison in the revised manuscript.

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.

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 O(N)\mathcal{O}(\mathcal{N}), 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:

MetriceSEN-3.2MSGBNet-SReduction
Runtime (μs/edge)1.1721.1452.3%
Memory (KiB/edge)127.8085.4333.2%
FLOPs (MFLOPs/edge)15.959.7538.9%

These results show that SGBNet reduces edge-dependent memory and computation while retaining O(E)\mathcal{O}(\mathcal{E}) 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.

Reviewer_ZYjfOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes SGBNet, aiming to enhance the conversion and utilization of geometric feature into scalar feature to improve accuracy. Furthermore, it makes several interesting attempts regarding efficiency, such as concentrating scalar operations on node features to reduce computational complexity, and leveraging the high utilization of geometric-to-scalar conversion to substantially reduce costly geometric feature computations, thereby significantly improving efficiency while maintaining high accuracy.

Critical strengths
  1. The paper shows a good understanding of existing problems, and its design is well motivated. The authors analyze network information flow to find current model weaknesses before designing the architecture, so the logic is clear.

  2. The method addresses efficiency at both the mathematical and implementation levels. The results suggest a promising accuracy–efficiency trade-off compared with eSEN.

Critical weaknesses
  1. As a reader less familiar with equivariant geometric networks, I found the motivation difficult to follow. The abstract and introduction would benefit from more background and references, particularly around line 41.

  2. The current ablation studies do not clearly show how much each component contributes. For example, SO(2)-feedback-only compared to the original eSEN.

  3. The theoretical proof of the "low-rank bottleneck" is not solid enough, and there is no experimental evidence showing that the gains come from solving this problem. Still, I find this idea and starting point very interesting.

  4. Details to reproduce the code are missing: the appendix only lists main channel sizes, not the full architecture. Also, please clarify the exact rule for loading the previous checkpoint and skipping the current batch after a loss spike.

  5. Memory and FLOPs are measured differently across frameworks. The main efficiency claims should therefore rely primarily on same-framework comparisons.

Questions for authors

See weaknesse.

3 comments

1. As a reader less familiar with equivariant geometric networks, I found the motivation difficult to follow. The abstract and introduction would benefit from more background and references, particularly around line 41.

We thank the reviewer for this suggestion. We have added brief background on the symmetry principles underlying equivariant GNNs, together with additional references, to improve the accessibility of the Introduction. The corresponding revisions are highlighted in blue.

2. The existing ablations do not clearly separate each component’s contribution. Please compare the original eSEN, head-only, SO(2)-feedback-only, SO(3)-FFN-feedback-only, and full-model variants.

We appreciate this suggestion. We have added separate ablations of the SO(2)-feedback, SO(3)-FFN-feedback, and SO(3)-head modules on eSEN. All three modules independently improve accuracy. The results are as follows:

eSEN+SO(2)+FFN+Head+SGB full
Force MAE4.703.963.984.113.83
Time38.3741.5438.2837.6836.48
Memory4.7816.2944.7834.7825.444
#Params3,175,5534,076,6733,372,1613,278,8573,102,121

Please see Common Response 2 for detailed analysis.

3. How severe is the low-rank bottleneck in practice? Please provide more experimental evidence to support the proposed explanation.

We appreciate the reviewer raising this concern. Our new rank-controlled experiments confirm that the low-rank readout coupling is a substantial practical bottleneck, as evidenced by the consistent improvements in effective Jacobian rank and force accuracy with increased coupling capacity. The results are as follows:

ConfigurationSeparated normOriginSGB head, dSGB = 8d_\mathrm{SGB} = 8SGB head, dSGB = 32d_\mathrm{SGB} = 32SGB head, dSGB = 128d_\mathrm{SGB} = 128
Rank bound011765128
Effective rank0.0 ± 0.01.0 ± 0.013.0 ± 1.144.8 ± 3.494.1 ± 4.2
Force MAE4.674.704.464.254.11
#Params3,175,5533,175,5533,181,6973,200,1293,278,857
Time38.0538.3737.8037.5537.68

Please see Common Response 1 for detailed analysis.

4. Important details needed for reproducibility are missing or insufficiently specified. Although the appendix lists the main architectural widths, I could not find a complete specification. In addition, the checkpoint rollback and batch-skipping procedure used after training loss spikes requires clarificatio.

We thank the reviewer for pointing out the insufficient implementation details. We have expanded the appendix to provide explicit definitions of the architectural hyperparameters. We have clarified the loss-spike detection criteria and the complete checkpoint rollback and batch-skipping procedure. We have also added the radial basis and envelope configurations, cutoff radii, learning rate, gradient clipping threshold, and EMA settings. We will make the implementation code publicly available after the review process. All changes in the revised manuscript are highlighted in blue.

5. Efficiency comparisons: Memory and FLOPs are measured differently across frameworks. The main efficiency claims should therefore rely primarily on same-framework comparisons.

Our primary Pareto-frontier comparison focuses on SGBNet, eSEN, and ViSNet, which are already evaluated under the same-framework. These comparisons form the basis of our main accuracy-efficiency analysis and are trustworthy (Critical Strengths 1 by reviewer 8HFj). We report cross-architecture comparisons with additional methods that are not on the established Pareto frontier. These results are not used to construct the primary Pareto frontier.

Thanks to the authors for their response, which has addressed my concerns. I find the motivation and design of this paper solid, though the initial version was somewhat unpolished. With the added revisions, I have raised my score accordingly.

We sincerely thank the reviewer for the positive feedback and for increasing the score. We are glad that the revisions have addressed the concerns.

Reviewer_8HFjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes SGBNet, an SO(2)-based equivariant interatomic potential in the eSCN/eSEN family with separate scalar and geometric pathways. Its first component, scalar-geometric bidirectional (SGB) interactions, builds invariant scalars from higher-order features. The SGB SO(2) linear takes the conjugate product of two SO(2)-linear maps within the same frequency mm in the edge-aligned frame, and the SGB SO(3) linear forms node-wise l⊗l→0l\otimes l\to 0 contractions. Its second component, the edge-efficient convolution, moves scalar channel mixing from edges to nodes and reallocates channels from geometric to scalar features. Experiments in a shared JAX codebase cover efficiency profiling, SPICE-MACE-OFF, MLIPAudit, Matbench Discovery/sAlex and OMat24. SGBNet-S is 16% faster and uses 33% less memory than eSEN-3.2M, with lower force MAE on all seven SPICE subsets. SGBNet-N is 47% faster than eSEN-3.2M with comparable MLIPAudit scores but higher SPICE errors. Results on materials are roughly on par with eSEN.

Critical strengths
  1. Credible gains in a controlled setting. SGBNet-S vs eSEN-3.2M (−16%-16\% runtime, −33%-33\% memory, lower force MAE on 7/7 subsets) and SGBNet-M vs eSEN-6.5M (−16%-16\% runtime, −33%-33\% memory, better or tied force MAE on 7/7 subsets) are both measured in the same JAX codebase, so the comparison is trustworthy.

  2. A simple and provably invariant module. The conjugate-product construction is cheap, comes with a complete invariance proof, and can be dropped into other SO(2)-based models.

  3. Evaluation beyond MAE. MLIPAudit baselines are retrained on the same data and protocol. The appendix covers conformers, non-covalent interactions, tautomers, RDFs and protein sampling, and the efficiency profiling spans system sizes and frameworks.

Critical weaknesses
  1. The novelty is limited. The SGB SO(3) linear is a learnable l⊗l→0l\otimes l\to 0 invariant, close to PaiNN's ⟨Uv,Vv⟩\langle U\mathbf{v},V\mathbf{v}\rangle, MACE's product basis and ViSNet's inner products. Moving scalar linear layers to nodes is the standard gather/scatter commutation (as in SchNet), and it saves only about 5% runtime (Table 6). How does SGB differ from these constructions, and how does it compare with a PaiNN-style norm or inner-product baseline?

  2. The motivation is not validated. The m=0m=0 sector of SO(2) convolutions already passes l≥1l\ge1 information to scalars, and SGB adds about 34% runtime (32.07 vs 23.89) without a compute-matched control. Does SGBNet keep the m=0m=0 coupling, and does SGB still help at equal compute?

  3. The efficiency gain is not attributed to the architecture. Most of it comes from channel reallocation (−30%-30\% runtime, while force MAE rises from 3.40 to 3.75), and no parameter counts are given. What are the parameter counts, and how does an eSEN scaled to the same depth, width and runtime perform?

Questions for authors

Please see the questions in the Weaknesses section.

1 comment

1. The novelty is limited. The SGB SO(3) linear is a learnable l⊗ l → 0l\otimes l \to 0 invariant, close to PaiNN's ⟨ Uv, Vv ⟩\langle U\mathbf{v}, V\mathbf{v} \rangle, MACE's product basis and ViSNet's inner products. How does SGB differ from these constructions, and how does it compare with a PaiNN-style norm or inner-product baseline?

We clarify that SGBNet makes both methodological and architectural contributions. Its methodological novelty lies primarily in the SO(2) conjugate-product construction, which the reviewer also recognized as a simple, provably invariant, and readily reusable module (Critical Strengths 2). For the SO(3) branch, we acknowledge that contraction is an established mathematical operation; its contribution is architectural, introducing an explicit feedback pathway to incorporate higher-order geometric information into scalar representations. SGBNet differs from PaiNN and ViSNet, which lack features with l>1l>1, and MACE, which does not employ distinct left- and right-hand features.

Moving scalar linear layers to nodes is the standard gather/scatter commutation (as in SchNet).

While mathematically related to the gather/scatter commutation in SchNet, our design addresses a fundamentally different setting: SchNet operates solely on invariant features, whereas our Edge-efficient convolution handles coupled scalar and higher-order geometric features with edge-dependent modulation, which does not commute with scalar transformations. We develop a reformulation that preserves the original model's effectiveness by separately applying modulation to target and source features before summation.

2. The motivation is not validated. The m=0m=0 sector of SO(2) convolutions already passes l≥ 1l\geq 1 information to scalars. Does SGBNet keep the m=0m=0 coupling?

Our motivation specifically concerns the limited rank of the readout coupling, which is distinct from the existing m=0m=0 coupling. The additional experiments in Common Response 1 directly validate this motivation: increasing the readout coupling rank consistently increases the measured effective Jacobian rank and improves force accuracy. The results are as follows:

ConfigurationSeparated normOriginSGB head, dSGB = 8d_\mathrm{SGB} = 8SGB head, dSGB = 32d_\mathrm{SGB} = 32SGB head, dSGB = 128d_\mathrm{SGB} = 128
Rank bound011765128
Effective rank0.0 ± 0.01.0 ± 0.013.0 ± 1.144.8 ± 3.494.1 ± 4.2
Force MAE4.674.704.464.254.11
#Params3,175,5533,175,5533,181,6973,200,1293,278,857
Time38.0538.3737.8037.5537.68

Regarding the m=0m=0 pathway, SGBNet indeed retains the original coupling in SO(2) convolutions; the SGB SO(2) module introduces an additional geometric-to-scalar interaction rather than replacing the existing one.

SGB adds about 34% runtime (32.07 vs 23.89) without a compute-matched control. Does SGB still help at equal compute?

We thank the reviewer for raising this concern. Our new compute-controlled experiment shows that SGB does indeed improve Force MAE by 18.5% while running approximately 4.9% faster than the original eSEN.

eSEN+SGB full
Force MAE4.703.83
Time38.3736.48
Memory4.7815.444
#Params3,175,5533,102,121

Please see Common Response 2 for the experimental details.

3. The efficiency gain is not attributed to the architecture. Most of it comes from channel reallocation (runtime, while force MAE rises from 3.40 to 3.75). Moving scalar linear layers to nodes ... saves only about 5% runtime (Table 6). How does an eSEN scaled to the same depth, width and runtime perform?

The effectiveness of an architecture should not be evaluated in terms of efficiency alone, as runtime can always be reduced at the expense of accuracy. As demonstrated in Common Response 2, SGB modules could improve both accuracy and runtime, establishing its practical effectiveness. We acknowledge that channel reallocation accounts for a larger share of the runtime reduction, while reordering also contributes to the speedup. Both are integral components of our edge-efficient design. eSEN does not support independently configuring scalar and geometric channels. This flexibility is enabled by our dual-path architecture, so we cannot construct a directly corresponding scaled eSEN baseline.

No parameter counts are given. What are the parameter counts?

We have added the parameter counts of all model variants to the Hyper-parameters section of the appendix, with the changes highlighted in blue. However, our design primarily targets practical runtime rather than parameter reduction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response

Low-rank readout coupling

We thank all the reviewers for raising related concerns about the motivation and practical significance of the low-rank readout coupling (Q2 of reviewer 8HFj, Q3 of reviewer ZYjf, and Q1 of reviewer 4PEN). Our theoretical analysis and new experiments show that the low-rank readout coupling in eSEN is a substantial practical limitation, rather than merely a theoretical concern.

We conducted controlled experiments using eSEN with different readout coupling capacities. Specifically, we compared (i) separately handling the scalar and geometric components in the readout norm (rank 0), (ii) the original eSEN-3.2M readout (rank 1), and (iii) SGB energy heads with SGB hidden dimensions of 8, 32, and 128, respectively. The corresponding rank bounds are given by min⁡(2dSGB+1,dhead)\min(2d_\mathrm{SGB}+1,d_\mathrm{head}). For each variant, we compute the Jacobian ∂ h/∂ χ\partial h/\partial \chi, where hh denotes the 128-dimensional output (dhead = 128d_\mathrm{head} = 128) of the first head MLP layer and χ\chi denotes the geometric features before the norm. We quantify the effective coupling rank using the entropy-based effective rank of the Jacobian's singular values [1].

The results are as follows:

ConfigurationSeparated normOriginSGB head, dSGB = 8d_\mathrm{SGB} = 8SGB head, dSGB = 32d_\mathrm{SGB} = 32SGB head, dSGB = 128d_\mathrm{SGB} = 128
Rank bound011765128
Effective rank0.0 ± 0.01.0 ± 0.013.0 ± 1.144.8 ± 3.494.1 ± 4.2
Force MAE4.674.704.464.254.11
#Params3,175,5533,175,5533,181,6973,200,1293,278,857
Time38.0538.3737.8037.5537.68

Increasing the rank from 1 to 17, 65 and 128 reduces Force MAE from 4.70 to 4.46, 4.25 and 4.11, respectively, with only 0.2%, 0.8% and 3.3% more parameters, respectively, with essentially unchanged runtime. The substantial gains cannot be simply explained by a small increase in the parameter numbers and demonstrates that the low-rank coupling is a practical representational bottleneck, not merely a theoretical limitation. We have added the new experiments to Introduction and highlighted in blue.

[1] Roy & Vetterli.The Effective Rank: A Measure of Effective Dimensionality. EUSIPCO 2007.

Additive ablation of SGB modules

We thank the reviewers for raising concerns regarding the performance of the SGB modules (Q2 of Reviewer 8HFj and Q2 of Reviewer ZYjf). Our compute-controlled experiment demonstrates that SGB modules, even without the Edge-efficient design, achieve clear accuracy improvements while running slightly faster.

We conducted experiments by augmenting eSEN with SGB modules. Specifically, we first added each of the three modules individually (SO(2)-feedback-only, SO(3)-FFN-feedback-only, and SO(3)-head-only) to assess their respective contributions. We then incorporated all three modules while reducing the hidden channels from 128 to 104 (approximately 19%) to control computational cost.

eSEN+SO(2)+FFN+Head+SGB full
Force MAE4.703.963.984.113.83
Time38.3741.5438.2837.6836.48
Memory4.7816.2944.7834.7825.444
#Params3,175,5534,076,6733,372,1613,278,8573,102,121

The individual ablations show that all three SGB modules independently improve accuracy, reducing Force MAE from 4.70 to 3.96, 3.98, and 4.11, respectively. More importantly, eSEN augmented with all three SGB modules achieves an 18.5% reduction in Force MAE while running approximately 4.9% faster and using fewer parameters, despite having 19% fewer hidden channels. We have added these experiments and the corresponding analysis to the revised manuscript (highlighted in blue).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.