acceptodds
Under review as a conference paper at ICLR 2027

When Does Equivariance Help? Canonical Alignment in Neural Fluid Surrogates

Abstract

Neural surrogates can accelerate computational fluid dynamics (CFD) simulations by orders of magnitude, but practical deployment in engineering and healthcare applications requires architectures that scale to high-resolution meshes and learn effectively from limited data. Explicit equivariance offers a principled inductive bias, yet its accuracy benefits may depend on the prediction task and the distribution of anatomical orientations. We investigate this dependence across three hemodynamic benchmarks with different degrees of natural canonical alignment. To support this study, we introduce the Anchored-Branched Geometric Algebra Transformer (AB-GATr), an -equivariant surrogate that efficiently predicts coupled surface and volume quantities. Across these benchmarks, AB-GATr consistently outperforms the evaluated non-equivariant models, including variants trained with rotational augmentation, while achieving accuracy competitive with -equivariant LaB-GATr at substantially lower training cost. In comparison, rotational augmentation provides inconsistent benefits across architectures and can reduce accuracy. A controlled experiment on ShapeNet-Car shows that strong canonical alignment can favor non-equivariant models, but their accuracy generally deteriorates as training orientations broaden and can decline sharply under broader test rotations. We further investigate these patterns using extended symmetry-breaking diagnostics and probes of the predictive information associated with canonical alignment across all benchmarks. Together, these results support explicit equivariance for the evaluated hemodynamic tasks with natural orientation variation, while showing that its accuracy benefits depend on the task and orientation distribution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
neural surrogates, geometric deep learning, equivariance, fluid dynamics

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_qvhCOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how canonical alignment affects the performance of equivariant models in fluid prediction tasks. The authors propose AB-GATr, which combines geometric algebra representations with anchor attention and a multi-branch architecture. Experiments on three hemodynamic datasets and ShapeNet-Car suggest that explicit equivariance is generally beneficial for the evaluated hemodynamic tasks, but may not always help when the data are strongly canonically aligned.

Critical strengths
  1. The research question is interesting. Rather than assuming equivariance is always beneficial, the authors investigate how orientation distributions affect model performance. The controlled experiments on ShapeNet-Car are particularly informative.

  2. The experimental evaluation is extensive, covering multiple hemodynamic datasets, rotational augmentation, and different architectures. AB-GATr also achieves a promising balance between prediction accuracy and computational efficiency.

Critical weaknesses
  1. The contribution of equivariance is not fully isolated. Although AB-GATr performs well, it also differs from the non-equivariant models in parameter count, feature representation, and normalization. Differences across datasets further make it difficult to determine how much of the improvement comes from equivariance itself and the role of canonical alignment.

  2. Generalization to unseen base anatomies is not sufficiently clear. aneumo contains many synthetic geometries derived from a limited number of original anatomies. The paper mentions stratification by base anatomy, but it is unclear whether the splits are strictly disjoint at the original anatomy level. This may affect how the generalization results should be interpreted.

Questions for authors
  1. Does aneumo use strictly disjoint splits at the original anatomy level? Are all synthetic geometries derived from the same base anatomy assigned to the same split? Could the authors provide results on entirely unseen base anatomies? If the method can be shown to generalize well to completely unseen original anatomical structures, I would be willing to raise the score.

  2. AB-GATr and AB-UPT differ in model configuration and normalization. How can the performance gains be attributed specifically to equivariance? Could the authors conduct a more controlled experiment by varying canonical alignment within the same hemodynamic dataset?

  3. For ShapeNet-Car, are the geometry, inflow direction, and CFD labels rotated together during evaluation? Given that the model uses only coordinates as input, how does this relate to the discussion that boundary conditions must transform jointly with geometry?

Minor Comments: what is the statistical unit used in the Wilcoxon tests in Table 1, individual test cases or independent runs? Also, the reference information for Suk et al. (2024a) appears to mix the 2024 preprint year with the 2026 journal publication details. In the attention formula at line 230, both inputs are queries. Should the second one be a key? Also, Theorem A.1 seems to require a non-degeneracy assumption, since the zero function provides a counterexample. Could the authors clarify these points?

5 comments

We thank the reviewer for recognising the extent of the evaluation and AB-GATr’s balance between computational cost and performance.

Question 1

Yes, the ANEUMO splits are strictly disjoint at the original anatomy level: all synthetic geometries derived from the same base anatomy are assigned to a single split, with no overlap between training, validation and test sets. Table 1 therefore already assesses generalisation to entirely unseen original anatomies.

Question 2

A) Matching model configuration and training budget

We agree that the models should be made as comparable as possible to ensure a fair comparison. As described in Appendix A.3, AB-UPT and AB-GATr share the number of anchors, dropout rate, pooling radius, and block layout. Differences remain in settings that are model-specific, i.e. their hidden representations and attention configurations due to differences in standard and geometric algebra, respectively. AB-GATr has substantially fewer parameters than AB-UPT (1.6M vs 7M), but incurs ~4 times higher computational costs per epoch and approximately 3–4 times the VRAM usage.

The normalization choices reflect the requirement to preserve equivariance: axis-wise normalisation of vectors generally breaks rotational equivariance, whereas scaling vector targets by a common scalar, such as magnitude, preserves it. We also evaluated magnitude normalization for AB-UPT, but it performed worse than per-axis normalization. We retained the latter to provide the stronger baseline.

As we noted in Appendix A.5, the models required different numbers of epochs to achieve their best observed performance. Following reviewer's suggestions, we have extended this section to include a numerical summary. In the table below, we provide the AB-UPT (rot) performance with doubled training duration from 500 to 1000 epochs on a single AAA-100 fold, retaining the same architecture. We report relative $ L_2 $ errors (%) for velocity ($ u )andWSS() and WSS ( \tau $). Doubling the training did not lead to improvement in performance.

ModelEpochsParamsTimeVRAM$ u $$ \tau $$ u $ (rot)$ \tau $ (rot)
AB-UPT (rot)5007M1 h1.9 GB41.739.141.639.8
AB-UPT (rot)10007M2 h1.9 GB42.740.742.440.5
AB-GATr10001.6M11 h7 GB33.834.233.834.2

For architectural and training choices that could not be matched directly, we selected the best-performing options among those evaluated for each model.

B) Controlled orientation experiment for AAA-100

Following the reviewer’s suggestion, we extended the controlled orientation study in Fig. 5 to AAA-100. On a single fold, we trained AB-UPT with five rotation-augmentation ranges (0°, 30°, 60°, 90° and 180°) and evaluated each model across the same test ranges. The 0° setting retains native orientations.

As on ShapeNet-Car, AB-UPT generally performs worse when test rotations exceed its training range. However, broader augmentation also improves performance on the native AAA-100 orientations. This differs from ShapeNet-Car and is consistent with AAA-100’s weaker canonical alignment. AB-GATr maintains constant and superior performance across the evaluated rotations.

WSS ($ \tau )/Velocity() / Velocity ( u $): relative $ L_2 $ error (%)

AB-UPT train / Test0°30°60°90°180°
AB-GATr34.2 / 33.834.2 / 33.834.2 / 33.834.2 / 33.834.2 / 33.8
180°39.3 / 41.939.2 / 42.239.4 / 41.839.1 / 41.739.2 / 41.6
90°41.3 / 41.441.2 / 41.441.0 / 41.141.0 / 41.240.6 / 42.3
60°41.1 / 42.741.1 / 42.842.2 / 43.142.9 / 44.547.9 / 50.3
30°43.9 / 43.944.2 / 44.345.9 / 46.649.8 / 53.261.4 / 73.3
0°71.7 / 54.578.3 / 65.587.9 / 76.292.7 / 90.9100.9 / 106.5

Question 3

In the ShapeNet-Car experiment (Fig. 5), the inflow direction is rotated jointly with the geometry. Spatial coordinates, velocity and pressure targets are transformed consistently. The models receive coordinates without a separately encoded inflow direction. Consequently, geometry orientation also provides implicit information about the boundary conditions, and the experiment does not independently isolate these two sources of information. Due to this implicit boundary-conditions encoding the complete problem (geometry + boundary conditions) transforms consistently with respect to E(3).

Minor

  1. The Wilcoxon tests use paired errors from individual test cases, rather than independent training runs.
  2. We will update this bibliography entry to reflect the 2026 journal publication.
  3. Yes, the second argument should refer to keys; thanks for pointing this out. We will correct this typo.
  4. The non-degeneracy assumption is indeed required. We will add the assumption that $ f(x) \neq 0 $ and revise the proof accordingly.

Thank you for the clarifications and additional experiments; I will maintain my original score, as the contribution of equivariance remains difficult to separate from other architectural differences. I understand the focus on vector-valued fields, but pressure is also important in hemodynamic modeling, and its omission from the hemodynamic evaluation should be reflected in the scope of the conclusions. Please also clarify the exact definition of the velocity relative l2 error and whether it is evaluated on all volumetric mesh nodes or a sampled subset.

Thank you for considering our clarifications and additional experiments.

Here, we provide clarification on the relative $ L_2 $ error definition. For each test case $ i $, the relative $ L_2 $ is:

L2(yi,y^i)=∑t=1Ti∑n∈Vi∥y^i,n,t−yi,n,t∥22∑t=1Ti∑n∈Vi∥yi,n,t∥22L_2(\mathbf{y}_i,\hat{\mathbf{y}}_i)=\frac{\sqrt{\sum_{t=1}^{T_i}\sum_{n\in\mathcal{V}_i}\Vert\hat{\mathbf{y}}_{i,n,t}-\mathbf{y}_{i,n,t}\Vert_2^2}}{\sqrt{\sum_{t=1}^{T_i}\sum_{n\in\mathcal{V}_i}\Vert\mathbf{y}_{i,n,t}\Vert_2^2}}

where ($ \mathcal V_i )denotestheevaluatednodes,() denotes the evaluated nodes, ( T_i $) the number of evaluated time points, and $ \mathbf{y}, \mathbf{\hat{y}} $ the target and estimated fields, respectively. Evaluation uses all volumetric mesh nodes for volumetric quantities and all surface mesh nodes for surface quantities.

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

This paper asks when building rotation symmetry (equivariance) into neural networks that predict fluid flow actually helps, and argues the answer depends on how consistently oriented the dataset's shapes are. The authors introduce AB-GATr, an equivariant model that combines an efficient transformer (AB-UPT) with geometric algebra layers (GATr) to predict blood flow in volumes and on vessel walls. On three blood-flow datasets, it outperforms non-equivariant models, even ones trained with rotation augmentation, and matches an earlier equivariant model (LaB-GATr) at much lower training cost. On a car-aerodynamics dataset where all cars share one orientation, the ordinary model does better in-distribution but fails badly on unseen rotations, which the authors use to support their orientation-consistency argument.

Critical strengths
  1. The question matters and the framing is useful. Whether equivariance pays off in practice is contested. The paper separates symmetry of the PDE from symmetry of the full problem (inputs plus boundary conditions plus data distribution).

  2. The study tests both a non-equivariant baseline (AB-UPT) and an independent one (Transolver++), each with and without SO(3) augmentation. The finding that augmentation helps inconsistently, and even hurts Transolver++ on SynthBifCor, is informative.

  3. The ShapeNet-Car experiment is clean and honest. The rotation-range grid in Fig. 5 shows both sides of the story. Equivariance loses in-distribution under strong alignment but gives flat, reliable error under any rotation. The paper reports this against its own model.

  4. AB-GATr is practically useful. It is 4–6× cheaper to train than LaB-GATr (Table 3), uses less memory, and keeps anchor attention’s resolution flexibility. It also has fewer parameters (1.6M vs 7M for AB-UPT) and is exactly equivariant.

  5. The setup is transparent. The paper reports hyperparameters, compute, a statistical test, and a limitations section that admits there is no principled criterion yet.

Critical weaknesses
  1. Equivariance wins on all three hemodynamic datasets regardless of alignment (which the paper acknowledges), making the central claim of the paper weakly supported.. The thesis rests on one contrasting dataset from a different domain. The relative gain over AB-UPT^rot is not monotone in the integrated symmetry response (velocity: about 28%, 15% and 16% for SynthBifCor, AAA-100 and ANEUMO), though with only three datasets this is suggestive at best.

  2. The diagnostic “correlation” (Table 5) rests on three datasets. It omits ANEUMO and uses a different target for ShapeNet-Car (p) than for the others (τ).

  3. In the car dataset, every car faces the same way and the wind always blows from the same direction. So when the ordinary model does better there, it's possible that we can't tell whether that's because of the consistent car orientation or the fixed wind direction. A cleaner test would vary orientation consistency within a single blood-flow dataset.

  4. Model selection and budgets are unclear. AAA-100 has no validation split (Table 4), and the paper doesn’t say how hyperparameters and epochs were chosen, so test-fold selection can’t be ruled out. Furthermore, epochs, learning rates and surface/volume handling differ across models. The compute-matched runs (A.5) are described only qualitatively.

  5. The contribution is modest, and the title overpromises. AB-GATr is largely AB-UPT plus GATr layers, and the diagnostic extension is incremental over Lawrence et al. The paper concedes it offers no criterion for when equivariance helps.

Questions for authors
  1. Can you vary orientation consistency within a single blood-flow dataset, for example by limiting rotations to a growing range, to show that the benefit of equivariance changes with alignment?

  2. The AAA-100 dataset has no validation set (Table 4). How were learning rates, epochs, and the "best-performing" configurations chosen, and was any of that done using the test folds?

  3. On ShapeNet-Car, orientation and inflow direction are fixed together. How do you know alignment, and not the fixed boundary conditions, explains the ordinary model’s in-distribution advantage? In the Fig. 5 rotation experiments, is the inflow rotated with the car?

  4. Table 5 covers only three datasets and uses different targets (p vs τ). Can you use a common target, such as velocity, and show how the scores relate to the actual gains in Table 1 and Fig. 5?

3 comments

We thank the reviewer for recognizing the importance of the research question and the computational benefits of the proposed AB-GATr. We address all the questions below:

Question 1

Following the reviewer’s suggestion, we extended the controlled orientation study in Fig. 5 to AAA-100. On a single fold, we trained AB-UPT with five rotation-augmentation ranges (0°, 30°, 60°, 90° and 180°) and evaluated each model across the same test ranges. The 0° setting retains native orientations.

As on ShapeNet-Car, AB-UPT generally performs worse when test rotations exceed its training range. However, broader augmentation also improves performance on the native AAA-100 orientations. This differs from ShapeNet-Car and is consistent with AAA-100’s weaker canonical alignment. AB-GATr maintains constant and superior performance across the evaluated rotations.

WSS ($ \tau )/Velocity() / Velocity ( u $): relative $ L_2 $ error (%)

AB-UPT train / Test0°30°60°90°180°
AB-GATr34.2 / 33.834.2 / 33.834.2 / 33.834.2 / 33.834.2 / 33.8
180°39.3 / 41.939.2 / 42.239.4 / 41.839.1 / 41.739.2 / 41.6
90°41.3 / 41.441.2 / 41.441.0 / 41.141.0 / 41.240.6 / 42.3
60°41.1 / 42.741.1 / 42.842.2 / 43.142.9 / 44.547.9 / 50.3
30°43.9 / 43.944.2 / 44.345.9 / 46.649.8 / 53.261.4 / 73.3
0°71.7 / 54.578.3 / 65.587.9 / 76.292.7 / 90.9100.9 / 106.5

Question 2

For AAA-100, the ten held-out patients from one designated fold were used for hyperparameter selection. The selected configurations were then fixed across the remaining folds. Below, we provide an updated AAA-100 comparison (using relative $ L_2 $ error (%)) excluding the development fold (10/100 cases), using the remaining nine evaluation folds (90/100 cases):

Model$ u $$ \tau $$ u $ (rot)$ \tau $ (rot)
Transolver++†71.782.1123110
AB-UPT57.971.4109109
Transolver++†, rot92.766.788.965.5
AB-UPT rot46.645.245.745.1
LaB-GATr†42.348.142.348.1
AB-GATr39.538.839.538.8

Additionally, we provide a table including performance on the development fold only:

Model$ u $$ \tau $$ u $ (rot)$ \tau $ (rot)
Transolver++†68.179.3122108
AB-UPT54.371.6110101
Transolver++†, rot87.266.087.364.0
AB-UPT rot41.739.141.639.8
LaB-GATr†38.839.838.839.8
AB-GATr33.834.233.834.2

Generally, the performance on the development fold is better with models following similar trends. We will expand Appendix A.3 to include these details.

Question 3

In the ShapeNet-Car experiment (Fig. 5), the inflow direction is rotated jointly with the geometry. Spatial coordinates, velocity and pressure targets are transformed consistently. The models receive coordinates without a separately encoded inflow direction. Consequently, geometry orientation also provides implicit information about the boundary conditions, and the experiment does not independently isolate these two sources of information.

We recognize that varying the car’s orientation within a fixed wind-tunnel configuration would provide additional insight into the respective roles or orientation and boundary conditions. However, such setup defines a different physical problem and thus requires new CFD solutions that are not available in ShapeNet-Car.

SynthBifCor provides partial complementary evidence concerning alignment. As illustrated in Figs. 3 and 7, SynthBifCor geometries share a common inlet orientation and prescribed inflow velocity, while their branch orientations, vary substantially (particularly in the XY plane). Equivariant models outperform the non-equivariant alternatives in this setting, showing that fixed inflow conditions do not uniformly favour non-equivariant models. This observation is consistent with a role of canonical alignment, although differences between the datasets prevent generalising the outcomes to alignment alone. As acknowledged in our limitations, the findings provide empirical evidence rather than a general criterion for when equivariance is beneficial.

Question 4

We have extended the experiment to the common target velocity and appended the results to Table 5. We find the same trend as with the surface quantities.

MetricShapeNet-CarSynthBifCorAAA-100
$ \int m(p_X \circ r_\Theta)-\frac{\pi}{2} $0.9390.8630.666
$ \tilde{t}(p_{X,Y}) $$ p:\ 29.23 $, $ u:\ 43.32 $$ \tau:\ 8.52 $, $ u:\ 8.58 $$ \tau:\ 2.35 $, $ u:\ 1.84 $

The reviewer raises a good point about comparing the monotonicity of observed performance with the descriptive metrics. Relative accuracy of AB-UPT^rot over AB-GATr (on "rotated") is 1.07, 1.38 and 1.16 for ShapeNet-Car (Figure 5), SynthBifCor and AAA-100 (Table 1), respectively.

Integrated symmetry response and task-useful distributional symmetry breaking $ t(p_{X,Y}) $ aim to isolate two mechanisms: 1) alignment in the data distribution and 2) correlation of alignment and labels. We argue that observed relative accuracy should correlate with $ t(p_{X,Y}) $ when the degree of alignment is accounted for. Integrated symmetry response is thus a "necessary but not sufficient" component.

In practice, observations might be obfuscated by additional factors such as training dynamics and inductive bias (e.g., geometric algebra). Empirically finding a formula that accommodates both the components above while being robust to practical factors is an interesting idea. This would require to test more datasets and might favour a more idealised setting. We have added the following to future work:

We expect [a criterion determining when equivariance is beneficial] could be established as a function of integrated symmetry response ∫m(pX∘rΘ)\int m(p_X \circ r_\Theta) and task-useful distributional symmetry breaking t(pX,Y)t(p_{X,Y}) (Table 5) under consideration of a diverse collection of datasets.

We have adapted our language from "correlates" to "partly explains" when addressing diagnostic performance of distributional symmetry breaking and have added the following comment to the appendix:

However, no monotonicity can yet be established between observed relative accuracy of AB-GATr over AB-UPT with respect to t(pX,Y)t(p_{X, Y}) and ∫m(pX∘rΘ)\int m(p_X \circ r_\Theta) as the metrics isolate 1) alignment in the data distribution and 2) correlation of alignment and labels with possibly non-trivial interplay and obfuscation by practical factors such as training dynamics and inductive bias.

Regarding the raised 'Weakness 2', we omit $ t(p_{X,Y}) $ for ANEUMO because the dataset is not aligned ($ m(p_X) = 0.5 $), following the flow chart proposed by Lawrence et al.

In response to Weakness 4 regarding compute burdens and hyperparameters across the models.

We provide the compute-budget comparison on the same AAA-100 fold, including a larger AB-UPT variant with 1920 latent channels instead of 192 and a standard variant with doubled training. The table reports relative $ L_2 $ errors (%), alongside total training time, parameter count and peak VRAM usage:

ModelEpochsChannelsParamsTimeVRAM$ u $$ \tau $$ u $ (rot)$ \tau $ (rot)
AB-UPT (rot)5001927M1 h1.9 GB41.739.141.639.8
AB-UPT (rot)10001927M2 h1.9 GB42.740.742.440.5
AB-UPT (rot)5001920693M8 h21 GB54.660.247.760.8
AB-GATr100016 mv + 64 s1.6M11 h7 GB33.834.233.834.2

For AB-GATr, mv and s denote multivector and scalar channels, respectively.

Increasing AB-UPT’s latent channel size brings its training time closer to AB-GATr’s (8 versus 11 hours), while requiring three times the VRAM and over 300x more parameters. However, errors increase for both targets relative to the original AB-UPT configuration. Thus, neither longer training nor greater capacity closes the gap to AB-GATr in the tested configurations.

These exploratory results on AAA-100 indicate that increasing baseline compute alone does not necessarily close the equivariance performance gap to AB-GATr. However, this analysis is not exhaustive; alternative compute budget allocations across different datasets may behave differently. We will expand the compute-matched section in the Appendix to include these details.

For architectural and training choices that could not be matched directly, we selected the best-performing options among those evaluated for each model.

Comments on the paper

3 comments

We thank the reviewer for their positive assessment of the experimental settings and the practical contribution of AB-GATr. We address all the questions below:

Question 1

In the ShapeNet-Car experiment (Fig. 5), the inflow direction is rotated jointly with the geometry. Spatial coordinates, velocity and pressure targets are transformed consistently. The models receive coordinates without a separately encoded inflow direction. Consequently, geometry orientation also provides implicit information about the boundary conditions, and the experiment does not independently isolate these two sources of information. Due to this implicit boundary-conditions encoding, the complete problem (geometry + boundary conditions) transforms consistently with respect to E(3).

We recognize that varying the car’s orientation within a fixed wind-tunnel configuration would provide additional insight into the respective roles of orientation and boundary conditions. However, such a setup defines a different physical problem and thus requires new CFD solutions that are not available in ShapeNet-Car.

However, our SynthBifCor experiment provides partial complementary evidence concerning alignment. As illustrated in Figs. 3 and 7, SynthBifCor geometries share a common inlet orientation and prescribed inflow velocity, while their branch orientations vary substantially (particularly in the XY plane). Equivariant models outperform the non-equivariant alternatives in this setting, showing that fixed inflow conditions do not uniformly favor non-equivariant models. This observation is consistent with the role of canonical alignment, although differences between the datasets prevent generalising the outcomes to alignment alone. As acknowledged in our limitations, the findings provide empirical evidence rather than a general criterion for when equivariance is beneficial.

Question 2

We agree that the models should be made as comparable as possible to ensure a fair comparison. As described in Appendix A.3, AB-UPT and AB-GATr share the number of anchors, dropout rate, pooling radius, and block layout. Differences remain in settings that are model-specific, i.e. their hidden representations and attention configurations due to differences in standard and geometric algebra, respectively. AB-GATr has substantially fewer parameters than AB-UPT (1.6M vs 7M), but incurs ~4 times higher computational costs per epoch and approximately 3–4 times the VRAM usage.

The normalization choices reflect the requirement to preserve equivariance: axis-wise normalisation of vectors generally breaks rotational equivariance, whereas scaling vector targets by a common scalar, such as magnitude, preserves it. We also evaluated magnitude normalization for AB-UPT, but it performed worse than per-axis normalization. We retained the latter to provide the stronger baseline.

As we noted in Appendix A.5, the models required different numbers of epochs to achieve their best observed performance. Following reviewer's suggestions, we have extended this section to include a numerical summary. In the table below, we provide the AB-UPT (rot) performance with doubled training duration from 500 to 1000 epochs on a single AAA-100 fold, retaining the same architecture. We report relative $ L_2 $ errors (%) for velocity ($ u )andWSS() and WSS ( \tau $). Doubling the training did not lead to improvement in performance.

ModelEpochsParamsTimeVRAM$ u $$ \tau $$ u $ (rot)$ \tau $ (rot)
AB-UPT (rot)5007M1 h1.9 GB41.739.141.639.8
AB-UPT (rot)10007M2 h1.9 GB42.740.742.440.5
AB-GATr10001.6M11 h7 GB33.834.233.834.2

For architectural and training choices that could not be matched directly, we selected the best-performing options among those evaluated for each model.

I thank the authors for their clarifications. The response explains that the inflow direction rotates jointly with the geometry and acknowledges that the ShapeNet-Car experiment does not independently isolate alignment from implicitly encoded boundary conditions. This appropriately qualifies its interpretation.

The additional training results show that doubling AB-UPT’s training duration does not improve performance on the evaluated fold. This is useful evidence, although it does not constitute a strict compute-matched comparison or fully isolate equivariance from other architectural differences.

These remaining limitations do not outweigh the useful empirical findings and practical contribution of AB-GATr. I encourage the authors to incorporate these qualifications into the final manuscript. I maintain my rating of 3 (Weak acceptance).

We extended the compute-budget comparison on the same AAA-100 fold to include a larger AB-UPT variant with 1920 latent channels instead of 192. The table reports relative $ L_2 $ errors (%), alongside total training time, parameter count and peak VRAM usage:

ModelEpochsChannelsParamsTimeVRAM$ u $$ \tau $$ u $ (rot)$ \tau $ (rot)
AB-UPT (rot)5001927M1 h1.9 GB41.739.141.639.8
AB-UPT (rot)10001927M2 h1.9 GB42.740.742.440.5
AB-UPT (rot)5001920693M8 h21 GB54.660.247.760.8
AB-GATr100016 mv + 64 s1.6M11 h7 GB33.834.233.834.2

For AB-GATr, mv and s denote multivector and scalar channels, respectively.

Increasing AB-UPT’s latent channel size brings its training time closer to AB-GATr’s (8 versus 11 hours), while requiring three times the VRAM and over 300x more parameters. However, errors increase for both targets relative to the original AB-UPT configuration. Thus, neither longer training nor greater capacity closes the gap to AB-GATr in the tested configurations.

These exploratory results on AAA-100 indicate that increasing baseline compute alone does not necessarily close the equivariance performance gap to AB-GATr. However, this analysis is not exhaustive; alternative compute budget allocations across different datasets may behave differently. We will expand the compute-matched section in the Appendix to include these details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.