acceptodds
Withdrawn from ICLR 2027

AURA: Graph Learning via Superposed Neural Fields Beyond Message Passing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Message-passing graph neural networks update each node from a fixed-radius neighbourhood, so a depth- model cannot use information more than hops away. Many relational systems, molecules among them, instead interact through fields that superpose contributions from distributed sources. We introduce AURA, a graph layer built on source-medium-response computation: nodes emit source coefficients, a learned sparse medium operator propagates them through a truncated resolvent solve, and each node updates by probing the value, gradient, and curvature of the resulting field. Local linear message passing is a degenerate case of AURA, an exact solve aggregates walks of every length, and for every depth we construct tasks on which all depth- local message-passing networks are at chance while a constant-depth AURA model is exact. We test the mechanism with controls. On three interference tasks, local baselines stay at chance, full-attention graph transformers solve every task (a positive control), and AURA reaches 79–90% mean accuracy; replacing only AURA's learned medium with a uniform one, at identical depth, receptive field, and parameter count, returns all three tasks to chance. On a shared QM9 protocol and a QM9-derived field-reconstruction diagnostic, which bound the scope of our claims, AURA has the lowest mean error among the evaluated models, at 1.6 the per-run wall-clock cost of a GPS-style transformer on QM9. Code and all per-run result files are provided as supplementary material.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
geometric deep learning, beyond message passing, long-range dependencies, GNN expressivity, resolvent operators, molecular property prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_34wPOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AURA organises graph computation as source emission, learned sparse-medium propagation, and differential response. The exact resolvent aggregates all walk lengths, while the implementation uses truncated propagation and therefore has finite reach. The interference results seem to show a benefit from accessing remote information relative to the tested local baselines. Reduced-budget QM9 and an analytic field diagnostic suggest favourable mean errors against compact evaluated models, alongside higher computational cost and unresolved architectural confounds.

Critical strengths
  • [Important] A concrete and inspectable source-medium-response architecture. The paper specifies scalar, dipole, and quadrupole-like sources, state/geometry-dependent sparse edge weights, truncated scalar resolvent propagation, and local value/gradient/curvature probes. The implementation aggregates beyond an individual local update within a layer, while sparse operator application avoids constructing a dense node-to-node attention matrix. Equations and appendix details distinguish this finite implementation from the exact-resolvent idealisation. This appears to provide a well-specified architectural hypothesis about structured field aggregation, with receptive-field and convergence assumptions that can be examined experimentally. It does not yet establish an operational advantage over other propagation mechanisms.
  • [Important] The remote-information controls seem to make the synthetic tasks interpretable. The tasks place label-relevant sources outside the local baselines' accessible neighbourhoods while preserving local probe information. On Remote Charge, AURA appears stronger than the tested local models, although it remains less reliable than the global-attention controls. The attention results seem to show that the labels are learnable when remote information is available, supporting a locality-based interpretation of the local models' failures. The exact-solve separation supplies a formal counterpart within its solver and comparator restrictions. These controls help frame the question, even though they do not isolate the value of the learned medium from every reach-matched alternative.
  • Geometric prediction and field reconstruction seem promising within the tested protocol. The reduced-budget QM9 study appears to favour AURA's mean error against the listed compact baselines, with clearer relative improvements on dipole and orbital-gap prediction than on polarizability and spatial extent. The analytic field diagnostic also seems to favour AURA over Kernel-Sum in aggregate error and correlation, while some regional comparisons are less decisive. These trends suggest that the architecture may transfer beyond synthetic classification and merits further testing. Differences in parameter count, coordinate access, training budget, and specialised readout nevertheless prevent a confident claim of superiority over matched geometric models.
Critical weaknesses
  • [Important] Exact-solve analysis may not support the practical reach claim. The implemented truncated propagation and stacked local probes have a finite receptive field, so sufficiently distant sources cannot affect the probe. The exact Neumann-series argument additionally requires a contractive operator, whose trained norm was neither enforced nor measured. Although the paper discloses these boundaries, the constant-depth narrative appears to need a clearer distinction between an ideal exact solve and the finite solver budget actually used. Reach-matched experiments with longer bridges and measurements of the learned operator would help establish which theoretical regime is realised.
  • [Important] The uniform-medium control may confound state dependence with propagation scale. Assigning identical small weights to every channel changes signal magnitude as well as removing the learned medium's state dependence. Repeated propagation can attenuate an individual path contribution, before accounting for walk multiplicity and subsequent gain learning. Dormant medium parameters also do not match the active trainable functionality of the full model. The reported collapse towards chance therefore seems to show that this particular uniform setting fails; it does not exclude tuned fixed diffusion or reach-matched learned propagation. Scale-tuned, capacity-matched alternatives would provide a stronger mechanism test.
Questions for authors
  • Does the uniform medium's apparent weakness persist after validation-tuning β and fixed/normalised weights, across independent seeds and comparable active capacity?
Reviewer_CLB1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The manuscript introduces a source-medium response mechanism in the scope of graph learning. The first step is the emission where nodes generated source coefficients. Those signals are propagate and superpose through a medium, defining the second step. Lastly, each node measures the field and then updates its representation.

Critical strengths

S1: Learned medium space

What I do like about the approach is the idea of learning a medium space, through which the nodes propagate their information.

S2: Propagation mechanism

The usage of the truncated neumann expansion for the computation of the scalar channels leads to a reduction of the computational overhead as it prevents the computation of the inverse of equation (3). The only question regarding this formulation would be if the authors tested to isolate the contribution of q in each step. Maybe they can add a discussion for that.

Critical weaknesses

W1: no clear + wrong formalism

After equation (2), the authors define S_i(x), where it is not denoted what the delta function refers to. Moreover, the second term shows us that the gradient is multiplied with the transpose d^T. Applied to the three coordinates of the delta input, the gradient has dimensions 3x1 to be multiplied with the transpose Bx3. So far so good, the last term computes the second-order derivative which should have the dimensions 3x3. However, M has previously defined as 5xB in equation (2). Therefore there is a dimensionality mismatch.

W2: Equation 4

Equation 4 is followed by the “directional and second-order channels enter as one-medium-weighted hop” where Y_2 is introduced as the degree-2 real spherical harmonics. To fit the dot product in the third term, it should be of dimension 5, however, the five spherical-harmonic basis functions are not defined.

W3: No HPO for competitors

table 1 seems unfair against competitors as there was not hpo conducted for the other baselines. the authors state in the setup fixed learning rate, weight decay, and gradient clipping. Therefore, we have fix configurations without any further hpo. As previous studies (Dwivedi et al.) have shown that different GNN architectures exhibit substantially different optimization characteristics, using just a single configuration across all baselines might not adquately reflect the best performances of the models.

W4: Long-range relationships

A well known problem of GNN is the over-smoothing problem. Therefore, defining a medium space in order to overcome an alignment of representations across several hops, seems like an appealing approach. However, I’m missing a deeper inspection of the long range dependencies in the evaluation part.

W5: Restricted benchmark selection

In order to assess the model’s benefits across a broader range of benchmark datasets and to see its generalization capabilities, I suggest to include a broader set of benchmark datasets.

The authors restricted their analysis just to some tasks of the QM9 dataset (https://pytorch-geometric.readthedocs.io/en/2.6.1/generated/torch_geometric.datasets.QM9.html)

W6: Field reconstruction

In Q3, the authors show the results on the field reconstruction where the pearson correlations of aura and the kernel-sum heuristic are quite similar. Inspecting further the results of figure 3, the example shows reconstruction capabilities where both approaches pretty much align with the ground truth. Did the authors also inspect an example where we see a more distinctive behavior? Aura also has nearly 15 times more parameters, where the benefit over the kernel-sum according to the pearson correlation seems relatively modest. Now, the question is, does this benefit really come from the underlying source-medium-response architecture or simply from the increased representational capacity. Here, I’d like to discuss the results under a fair budget for each competitor. We also do not have a robustness analysis for the field reconstruction as I could imagine that the proposed architecture might produce different predictions for rotated configurations.

W7: Over-smoothing

How does their approach deal with the over-smoothing effect? A discussion on that point would be interesting to see if the claim is to have better long-range capabilities. Even though it might not be theoretically justified as stated in section 3. Is there any empirical evidence?

Questions for authors
  • Q1: Could the authors clarify the notational flaws and missing defintions? (W1+W2)
  • Q2: Do the authors have conducted experimental setups with a greater search space for configurations for the competitors? (W3)
  • Q3: Did the authors tested their approach also on other benchmark datasets to check on the model’s capabilities across a broader set of benchmark datasets? What was the reason for restricting only on QM9? (W5)
  • Q4: Can the authors elaborate on the field reconstruction mentioned above? (W6)
  • Q5: For the long-range aspect, I would love to see a broader inspection on these effects and a potential discussion on the over-smoothing effect, and how the proposed approach might overcome this issue that often arises in conventional gnn models. (W4+W7)
Reviewer_3MaXOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
3
Read review
Paper summary

AURA is a research paper in the domain of message-passing graph neural networks. The authors present an alternative approach to propagating information in such networks diverging from the common local paradigm of spreading information one hop at a time. They presented a physics inspired method formulated as a global field solve from all node's sources over a learned medium. Every node then measures that field at its position.

They show that tasks can be constructed where their constant depth method is exact (under the assumption of an exact field solve) while local mpnn of fixed depth L perform at chance level. They further present control experiments replacing the learned component with a uniform output showing this intervention results in chance level performance. Finally the present promising performance on a QM9 subset molecular task.

AURA blocks in more detail: In the standard setting message-passing gnns work iteratively, a node collects information from its direct neighbors and updates its own information. Thus, information can only propagate one neighbor per round and after n rounds each node has received information from nodes at most n hops away.

The authors' contribution is a new architectural block called AURA layer. Aura layers follow the following 5 steps:

  1. A learned neural component called emitter predicts a scalar (q_i), a vector (d_i) in R^3 and a vector (M_i) in R^5 based on the nodes state for each of the fields channels. These predictions define the node's source.
  2. for every directed graph edge an mlp predicts an edge weight based on pairwise distance and the nodes' pairwise states. (again per channel)
  3. Beginning from the sources q_i of step 1, the field is updated K times based on the original scalar q_i and a damped, edge weighted field of the neighbors. After the K loops direction dependent scalar contributions from neighbors d_i and M_i are added.
  4. Each node then probes its own field and calculates gradient-like and curvature-like features from differences with its neighbors field values
  5. The nodes update their state based on another MLP that receives probe and old state, the output is added to the old state.

Three of these layers are stacked (with K = 6), which gives a receptive field of up to 21 hops. The model is trained end to end from a single loss.

Critical strengths

The core idea is interesting and to my knowledge new in this form. It shows promising results in the limited experimental setting the authors have conducted. Most notably, the experiments on QM9 molecular properties have promising performance under small training budgets. Here, AURA has the lowest MAE on all four targets.

Critical weaknesses

While I like the approach, the script lacks sufficient arguments for the merit of the new method. As the authors admit themselves, the QM9 experiments are only conducted on a small subset. I miss experiments isolating the benefit of the proposed architecture from propagation reach and model capacity. I do believe the manuscript would benefit significantly from some of the additions the authors list in their limitations such as long range benchmarks or local baselines with 21 hop reach. The geometric baseline could be done at matched capacity. Without these additional controls, the reported QM9 gains cannot be attributed to the field mechanism rather than to AURA's larger reach (21 vs. 3 hops) or its larger parameter count. While the uniform medium control keeps both fixed setting every weight to 1/32 may also strongly influence the signal at every hop unless beta compensates. With a 21-hop receptive field the trained model is itself a fixed-depth local model, so the construction of Theorem 3.5 (with L>= 21) applies to it as well. Theory currently differs from what is being trained. The separation result assumes an exact field solve, whereas the implemented model uses a fixed-degree polynomial with bounded receptive field. Ideally the authors could extend theoretical results so that they are applicable to the trained implementation. I do however want to positively point out that the authors state the mismatch themselves. It would be nice to have the exact solve condition also mentioned in the title or abstract.

Questions for authors

Eq. (4) and Algorithm 1 suggest that the implemented scalar solve is a fixed degree-6 polynomial in a learned operator. The dipole and quadrupole terms add one weighted hop. Do I understand this correctly? If yes please clarify how this differs from polynomial propagation (e.g. APPNP or GPR-GNN) with learned edge weights? All of the six steps is a sum over graph neighbours, correct? Then how is this beyond message passing? (I have seen your related work section ll. 400 to 407 this doesn't clarify it enough.)

Suggestion: A baseline using the same edge MLP and the same K but with standard propagation would help separate the contribution of your design from that of the propagation depth.

Comments on the paper

1 comment

Q1: Thanks for correcting the notation flaws and checking for any inconsistencies.

Q2: The authors acknowledged the fairness to other competitors. With the new setup, we see a substantial decrease in the overall performance gap across various settings. One follow-up to the provided answer is “lowest error on all four targets”, where in the first version of the manuscript told us in the abstract “all three tasks”. Or are you referring to the four properties of QM9? I’m a bit confused, which forth task you are referring here to. However, a proper tuning of the competitors led to a massive downgrade of the initial claims that from my perspective needs further investigation for a strong methodological contribution.

Q3: Thanks for providing the answer. The manuscript stays restricted to the QM9 benchmark.

Q4: Thanks for an budget-aware analysis. It seems like the increase of the capacity did not help for the MAE but the matched readout largely closes the gap leaving the question of the results of AURA are really statistical significant. And thank you also for not refraining from reporting results where the proposed method is beaten by other competitors like the kernel-sum in figure 4. For the molecule setting and providing a method that is not rotation invariant leaves also room for further investigation on how one can improve on that.

Q5: Thanks for that one. I do have a follow-up question here for the Table 29 showing the aura learned vs the aura fixed setting. The aura fixed setting seems to work up to a distance of ten whereas the learned collapses after D=5. However, the argumentation is that the max receptive field works up to 21 hops. But effectively it seems like to collapse much earlier. And at the end of the day, the initial concern then holds that the long-range dependencies are actually not resolved.


I’d like to thank the authors for their ambitious effort in answering all questions that apparently led to almost a complete revision of their manuscript.

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

Related papers

Discussion (0)

Sign in to comment.