acceptodds
Under review as a conference paper at ICLR 2027

DRGFM: Dynamic Riemannian Graph Foundation Models via Learnable Curvature Spectrum

Abstract

Graph foundation models (GFMs) require structural knowledge that transfers across domains and remains useful as tasks evolve. Yet implicit structural encodings make it difficult to specify what should transfer and what should be preserved. Curvature and spectral analysis offer complementary descriptions of local geometry and multi-scale propagation, but predefined geometric operators cannot themselves adapt to learning objectives across domains. We propose DRGFM, a Dynamic Riemannian Graph Foundation Model that models structural knowledge through a learnable curvature spectrum. A learned curvature-weighted Laplacian provides graph-dependent spectral coordinates and filters, coupling mixed-curvature representations with spectral attention to accommodate geometric heterogeneity. This formulation further supports structural knowledge evolution: during continual learning, importance-weighted regularization constrains spectral-coordinate drift on bounded anchor subgraphs in frozen reference bases, while the current curvature estimator and backbone remain adaptable. A conditional local analysis relates this displacement to historical-loss drift. Experiments on few-shot classification, zero-shot link prediction, and temporal continual learning evaluate transfer and retention, while matched ablations support the contribution of learnable curvature beyond a fixed prior. Together, these results support a unified approach to learning transferable structural representations and preserving them during subsequent adaptation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Foundation Models, Riemannian Geometry, Curvature Spectrum, Graph Neural Networks, Spectral Graph Learning, Continual Learning, Cross-Domain Transfer, Graph Transformer

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_wxNKOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors propose a dynamic Riemannian graph foundation model that models structural knowledge using a learnable curvature spectrum. The method uses a designed graph transformer to learn mixed-curvature representations. Experiments on various tasks demonstrate the effectiveness of the proposed approach.

Critical strengths

S1. This paper is well-motivated.

S2. Experiments show that DRGFM achieves the best model quality in most cases.

S3. The method is clearly described.

Critical weaknesses

W1. Some recent geometry-based GFM baselines are missing.

W2. The experimental results lack sufficient explanation.

Questions for authors

Q1. In the experimental section, the authors use some recent baselines. However, recent geometry-based baselines (e.g., R-GoG [1] and GAUGE [2]) are missing, which also improve the model quality based on Riemannian geometry. Could the authors add a comparison with these baselines to better demonstrate the effectiveness of the proposed method.

Q2. In some cases, the proposed method falls behind the baselines. For example, in Table 1 (training on Facebook), DRGFM is not better than GCOPE and GraphAny. Could the authors provide the deeper explanation for these results.

[1] Learning Graph Foundation Models on Riemannian Graph-of-Graphs. ICML 26

[2] Are Common Substructures Transferable? Riemannian Graph Foundation Model with Neural Vector Bundles. ICML 26

3 comments
AuthorsOpenReview ↗

Added R-GFM and GAUGE comparisons and explanation of domain-dependent results

Thank you for the positive assessment and for suggesting recent geometry-based GFMs. We added R-GFM and GAUGE comparisons and clarified the domain-dependent results.

1. Comparisons with the suggested methods. We evaluated R-GFM, from Learning Graph Foundation Models on Riemannian Graph-of-Graphs, and GAUGE, from Are Common Substructures Transferable? Riemannian Graph Foundation Model with Neural Vector Bundles. Both use the same five source graphs, training-step budget, and evaluation protocol as DRGFM; targets are unseen during pretraining.

The tables report mean±SD. NC uses 100 episodes with N=min(5,C); LP uses 10 negative-sampling runs. Target test edges are excluded from representation computation, and zero-shot LP uses no target parameter updates.

Target / NC settingR-GFMGAUGEDRGFM
Facebook / 1-shot.3560±.0555.2922±.0797.4671±.0976
Facebook / 5-shot.4873±.0434.3096±.0787.5936±.0558
GitHub / 1-shot.5485±.1269.5739±.2722.7007±.1950
GitHub / 5-shot.5930±.0834.5780±.2741.7649±.1195
LastFM / 1-shot.3117±.1446.2544±.2615.3290±.0847
LastFM / 5-shot.5471±.0988.4072±.2280.6307±.0916
PubMed / 1-shot.4364±.0738.4030±.0786.5782±.0731
PubMed / 5-shot.5495±.0509.6111±.0958.6709±.0623
Target / LP metricR-GFMGAUGEDRGFM
Facebook / AUC.9228±.0007.9234±.0004.9820±.0062
Facebook / AP.9229±.0012.9434±.0006.9848±.0051
GitHub / AUC.7054±.0008.7275±.0012.9546±.0082
GitHub / AP.6771±.0007.7190±.0015.9565±.0076
LastFM / AUC.8810±.0015.8071±.0025.9751±.0065
LastFM / AP.8958±.0023.8552±.0030.9844±.0063
PubMed / AUC.7610±.0021.6242±.0020.9217±.0039
PubMed / AP.8078±.0021.7144±.0031.9226±.0043

DRGFM has the highest mean in all eight NC and eight LP entries against these two additions. We will incorporate these results and the protocol details into the experimental discussion.

2. Why does DRGFM not lead every Facebook NC setting? Facebook is an unseen target, not a pretraining source; only episode support labels are used for adaptation. In 1-shot NC, DRGFM (.4671) is below GCOPE (.4852), but above GraphAny (.4365). In 5-shot, DRGFM (.5936) is below GraphAny (.6122) and SAMGPT (.6111), and close to GCOPE (.5922).

Our curvature-spectral interface learns structural representations, while NC also depends on alignment with target classes and sampled support labels. Better structural transfer need not yield the best class boundary in every setting. GCOPE's cross-domain graph composition and GraphAny's classification-oriented design provide different inductive biases. This is a plausible interpretation, not a measured causal explanation for Facebook.

The controls give within-architecture evidence: on Facebook, fixed Forman reduces 1-/5-shot accuracy from .4671/.5936 to .4050/.5379; generic edge weighting gives .3630/.4375, and removing spectral bias gives .3247/.3898. Thus, the proposed components improve Facebook classification within a controlled architecture even though DRGFM does not lead every alternative pipeline. DRGFM reaches .9820 AUC and .9848 AP in zero-shot LP on Facebook, illustrating that relative performance depends on task as well as domain. We will revise the discussion to state both the gains and remaining variation without claiming uniform superiority.

AuthorsOpenReview ↗

Revision available: R-GFM/GAUGE comparisons and Facebook discussion

Thank you again for your constructive review. The uploaded revision incorporates the results and discussion reported in our rebuttal. For easy reference:

  • Q1: recent geometry-based baselines (Section 6.2, Tables 1-2). We added R-GFM, from Learning Graph Foundation Models on Riemannian Graph-of-Graphs, and GAUGE under the same source graphs, training-step budget, and evaluation protocol. DRGFM has higher mean performance than both additions in all eight NC and eight LP entries across the four targets.
  • Q2: Facebook results (Section 6.2; Appendix J, "Task-dependent results"). We now explicitly discuss the settings where DRGFM trails GCOPE or GraphAny. Classification depends on alignment with target classes and support labels as well as structural transfer; we present this as a plausible interpretation, not an established causal explanation. Tables 5-6 also show that the proposed components improve Facebook NC within the controlled architecture, while the discussion avoids claiming uniform superiority across methods or tasks.

Could you briefly confirm whether these additions address your two concerns, or indicate any point that still needs clarification? Even a short response would help us understand whether further clarification is needed during the remaining discussion period.

Thank you for your response. My concerns are fully addressed. I have improved the overall score.

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

This paper proposes DRGFM (Dynamic Riemannian Graph Foundation Model), which represents structural knowledge through a learnable "curvature-weighted Laplacian operator" and its corresponding spectral coordinates. The model also combines Forman curvature priors with feature-dependent residuals, uses the obtained spectral information to adjust the mixed curvature representation and Chebyshev spectral attention mechanism, and introduces an anchor-based spectral coordinate regularizer to support continual learning. The authors evaluated the model in scenarios involving few-shot node classification, zero-shot link prediction, and continual learning, and conducted curvature ablation experiments, spectral analysis, and a conditional local analysis of historical-loss drift.

Critical strengths
  1. The curvature-spectral interface provides a unified structural representation that links local geometric structure, spectral coordinates, and representation learning. This is more explicit than common approaches in which curvature is used as a fixed descriptor or auxiliary signal.
  2. The empirical evaluation covers several transfer settings, including few-shot node classification, zero-shot link prediction with held-out edges, temporal continual learning, curvature ablations, and a curvature-spectrum analysis. The authors also explicitly state their information-visibility assumptions.
Critical weaknesses
  1. Tables 4 and 5 compare the full model with the w/o learnable curvature and w/o curvature variants, but these comparisons do not isolate learnable curvature. The first replaces the residual curvature estimator with fixed Forman curvature while retaining the curvature-weighted Laplacian and spectral channels, whereas the second removes both the curvature mechanism and the spectral bias. The paper does not report parameter matching across these variants, and the near-chance AUC of 0.4991 on GitHub makes the w/o curvature comparison especially ambiguous. Since learnable curvature is the paper’s central claim, the current results do not establish that this mechanism, rather than added capacity or other architectural changes, drives the reported gains. Parameter-matched Euclidean, random-curvature, and shuffled-curvature controls are therefore necessary to evaluate the claim itself.
  2. The continual-learning evidence is too limited to support the paper’s retention claim. Table 3 evaluates only one five-task Elliptic stream and reports point estimates without standard deviations, confidence intervals, or variation across task orders. The paper acknowledges in Appendix E.2 that the comparison with ER-GNN and FROMP does not establish an equal-memory advantage unless storage budgets are matched, but no matched-budget experiment is reported. Because retention is a primary contribution, a single five-task stream with point estimates cannot establish the claimed advantage. Multiple runs, multiple task sequences, and a memory-matched comparison are therefore required to support the central retention claim.
  3. CUSP (ICLR 25) already combines a curvature-weighted Laplacian, mixed-curvature representations, and learnable spectral filtering. Although cited, it is not included in the comparisons. A comparison under a matched training and evaluation protocol would help establish what DRGFM adds beyond this closely related approach.
  4. Section 4.2 performs the main transformations in tangent space, but no Euclidean variant retaining the same curvature-dependent modulation and routing is evaluated. A parameter-matched model with generic learnable edge weights is also missing. Furthermore, the “w/o curvature” variant disables spectral attention, confounding the effects of these two components. The current ablations therefore do not establish whether the gains require the proposed geometric formulation or arise from adaptive weighting and spectral processing.
  5. Table 3 lacks comparisons against variants without regularization, with uniform spectral weights, or with representation distillation on the same anchors. These controls are needed to determine whether curvature-based spectral coordinates and energy-based weighting improve retention beyond simply preserving historical representations.
  6. Continual learning is evaluated only on a five-task Elliptic stream with a fixed binary label space. The reported AA improvement over FROMP is 0.010, but no results across independent training runs or uncertainty estimates are provided(e.g. standard deviations). This leaves both the statistical reliability of the gain and its robustness across different graph domains and longer task sequences unclear.
Questions for authors
  1. Can you provide a parameter-matched comparison between the full model, a standard Euclidean graph Transformer, a fixed-curvature model, a random-curvature model, and a shuffled-curvature model? This would clarify whether the gains come from learnable geometry or from additional parameters and spectral attention.
  2. Which part of the mixed-curvature representation is responsible for the reported gains? Please provide an ablation that fixes the representation to Euclidean space while keeping the rest of the architecture and training budget unchanged.
  3. Why is the w/o curvature variant almost random on GitHub link prediction (AUC 0.4991), while the full model reaches 0.9546? Is this caused by training instability, the removal of spectral bias, or the curvature mechanism itself? A matched control is needed to rule out an under-trained baseline.
  4. How many independent runs, random seeds, and task orders were used for Table 3? Please report mean ± standard deviation and compare DRGFM with EWC, SI, MAS, ER-GNN, and FROMP under equal anchor-memory and tuning budgets.
  5. Were the transfer baselines pre-trained on the same source graphs and with comparable computational budgets? If not, the few-shot and zero-shot comparisons may reflect differences in pre-training data rather than the proposed curvature-spectral mechanism.
  6. How are the anchor subgraphs selected, and how sensitive are the results to anchor size, the number of retained Lanczos modes, and the choice of spectral regularization weight? These details are important for both reproducibility and the claimed “bounded memory” property.
5 comments
AuthorsOpenReview ↗

Matched geometry controls and repeated, memory-budgeted continual learning

Thank you. We added controlled comparisons and repeated CL evaluations.

1. Capacity and geometry (Q1-3). All controls retain FiLM, routing, FFN gating, weighted Laplacian, and spectral attention. Generic weighting removes the Forman prior and curvature supervision, retaining the same MLP and weighting pipeline. Full and generic each have 791,786 trainable backbone parameters. The Euclidean control changes manifold maps/distances; two registered manifold-curvature scalars become inactive. Fixed/random/shuffled controls freeze the 33,890-parameter estimator (757,896 trainable parameters), so they are structural, not trainable-capacity-matched controls.

LP ROC-AUCFacebookGitHubLastFMPubMed
Euclidean.7561.6416.6835.6230
Generic weighting.8497.6698.8173.7816
Fixed Forman.9315.8407.9152.8876
Random curvature.6930.6619.6400.5876
Shuffled curvature.6433.5659.6219.6359
DRGFM.9820.9546.9751.9217

DRGFM also leads all eight NC entries. For brevity, 5-shot NC accuracy on Facebook/GitHub is .5936/.7649 (full), .3744/.5184 (Euclidean), .4375/.7033 (generic), and .5379/.7043 (fixed Forman). Random/shuffled controls use fixed-seed moment-matched noise/permuted Forman values; all variants are independently pretrained. The original GitHub .4991 ablation removed multiple components and cannot isolate curvature. New GitHub AUC controls give .6362 without spectral bias, .6416 with Euclidean representations, .8407 with fixed Forman, and .7618 with equal routing.

2. Repeated CL and task granularity (Q4). Table 3 was one run. New results use five complete sequential runs (seeds 24-28), paired across methods, with the same pretrained backbone, reinitialized task heads, and fixed splits. Entries are accuracy-AA, mean±SD.

Method5 tasks10 tasks5 tasks, 12 MiB
EWC.9017±.0095.9345±.0114.8997±.0112
SI.9031±.0073.9217±.0112.8942±.0028
MAS.9064±.0089.9259±.0060.9035±.0089
ER-GNN.9176±.0057.9173±.0125.9336±.0241
FROMP.9219±.0055.9429±.0092.9334±.0173
DRGFM.9307±.0109.9607±.0057.9375±.0195

The 10-task stream divides the same 40 Elliptic steps into 10x4 rather than 5x8, preserving chronology and a maximum of 1,250 optimization steps. This tests temporal granularity. Its AP-AA is .7719±.0122 for DRGFM vs .6720±.0807 for FROMP; accuracy forgetting is .0159±.0056 vs .0334±.0103.

Five-task BWT is -.0425±.0168; forgetting is .0508±.0135 vs FROMP's .0566±.0052. We will replace Table 3 with repeated estimates and revise the discussion to emphasize reduced forgetting, rather than stable positive BWT. Under 12 MiB, DRGFM has the highest mean accuracy-AA; FROMP has lower accuracy forgetting (.0314 vs .0399) and higher AP-AA (.7387 vs .7336). We will describe this as competitive retention under matched storage, with evidence scoped to Elliptic.

The 12 MiB cap applies to retained CL tensor state after task-boundary eviction, including anchor structures, bases, coordinates, importance weights, and replay data. All methods meet it; DRGFM's largest state is 11,576,580 bytes. It is not a peak GPU-memory cap.

3. Spectral preservation. On identical anchors, energy, uniform, and representation penalties use coefficient .5, without separate tuning. Accuracy-AA is .9307±.0109, .9199±.0061, and .9161±.0088; no regularization gives .8956±.0081. Forgetting is .0508, .0600, .0629, and .0849. These comparisons support energy-weighted spectral preservation under the tested settings.

4. Protocol and CUSP (Q5). Transfer GFMs, including R-GFM and GAUGE, use the same five source graphs, training-step budget, and evaluation protocol; no class-text loss is used. CUSP is designed for supervised graph learning, not source pretraining and few-shot transfer. Applying DRGFM's pretraining objective would therefore require adapting the original method. We train CUSP with the same target support-label budget: its 5-shot accuracy is .3260/.5552/.2769/.6683 on Facebook/GitHub/LastFM/PubMed, vs DRGFM's .5936/.7649/.6307/.6709. This compares complete pipelines; the matched controls above assess mechanism contributions. Fully supervised CUSP is not ranked against few-shot methods.

5. Anchors and sensitivity (Q6). Anchors use label-free random starts, 1-hop BFS expansion, and induced edges, capped at 1,024 nodes. We retain 32 lowest-frequency Lanczos modes, including zero modes. Accuracy-AA is .9280/.9126/.9307 for 256/512/1,024 nodes and .9044/.9119/.9307/.9129 for 8/16/32/64 modes. For weights .5/1/2 it is .9307/.9289/.9299, vs .8956 without regularization. The results clarify sensitivity to modes and regularization strength.

I appreciate the authors' detailed response and additional experiments, which have addressed several of my concerns. While the experimental evaluation is reasonably comprehensive, my main concern remains the limited methodological novelty. The proposed learnable curvature-spectral formulation appears to be an incremental extension of existing geometric graph learning methods, and its motivation is not sufficiently compelling. Overall, I remain unconvinced that the contribution meets the novelty expectations of ICLR. Therefore, I will maintain my original score.

Q1. Parameter Matching and Geometric Structure

The authors now show that the full model and the generic edge-weighting control have the same number of trainable backbone parameters. This addresses part of my previous concern about whether the performance gain could simply result from differences in model capacity. The remaining controls examine the role of geometric structure, among which the fixed, random and shuffled controls freeze the estimator, and their trainable parameters are fewer than those of the full model. The authors also acknowledge that these are not capacity-matched controls. The performance gap therefore cannot be attributed to geometric structure alone, since the differences in trainable capacity may also contribute to the results. Regarding the Euclidean control, the authors do not report its parameter count, so I cannot confirm whether its parameter count is consistent with that of the full model. At present, only the generic edge-weighting control can be confirmed as parameter-matched. The authors should state this explicitly in the revised manuscript and report the parameter count of the Euclidean control.

Q2. Mixed-Curvature Representation and Component Ablations

The authors add a Euclidean representation variant that retains curvature modulation, routing, gating and spectral bias. They also report separate ablation results for these components and remove the Ollivier warm-up. The added experiments provide the component-level comparisons I requested in my previous review. The point estimates alone are not sufficient to assess the stability of these ablations. The authors need to report the number of pre-training runs and the variance across runs in the revised manuscript.

Q3. Interpretation of the Near-Chance Result without Curvature

The newly added control experiments show that the earlier near-chance result came from removing several components at once. The fixed-Forman, Euclidean, equal-routing and no-spectral-bias controls provide a more appropriate basis for evaluating the curvature claim. These controls also show why the original ablation was insufficient. I would be cautious about retaining the original near-chance result as evidence for the curvature claim. The authors should either remove this comparison or explicitly acknowledge that the result is confounded by the simultaneous removal of multiple components.

Q4. Continual-Learning Evaluation under a Matched Memory Budget

In this part, the authors add multiple complete sequential runs and a variant with finer-grained tasks, compare each continual learning method under a unified storage cap, and report the standard deviations. This addresses my earlier concern that the single-run evaluation and different memory budgets made the comparisons difficult to interpret. I still cannot tell from the current response how sensitive the results are to the order in which tasks are presented. Please list the task-order variations used in the experiments and provide the detailed memory accounting in the appendix.

Q5. Baseline Comparability and the CUSP Comparison

The authors state that R-GFM, GAUGE and DRGFM use the same source graphs, training-step budget and evaluation protocol, and they add a comparison with CUSP under the same target support-label budget. However, this is not a like-for-like comparison, because CUSP is designed for supervised learning, whereas DRGFM uses source pre-training followed by few-shot transfer. Their training objectives also differ. I do not think that matching the target support-label budget is sufficient to make these results directly comparable. The authors should state this limitation in the main text and treat the CUSP result as a supplementary comparison, not as evidence that DRGFM outperforms CUSP in the setting it was originally designed for.

Q6. Anchor Selection and Sensitivity Analysis

The authors have now clarified the construction of the anchors and also reported the sensitivity of the results to anchor size, the number of retained Lanczos modes and the spectral regularization weight. Please place the anchor-selection procedure and the full sensitivity results in a separate appendix subsection so that readers can locate these implementation details and experimental results together.

Complexity and Reproducibility

In the revised version, the authors provide additional information on computational complexity, pre-training time, hardware configurations, model architecture and optimization settings. These additions help address the previous lack of information on computational overhead. These additions should improve the reproducibility of the research. In addition, I suggest that the authors further clarify whether the reported time includes data preprocessing, metric computation, and task-boundary anchor construction, as the computational overhead of these components varies across different continual learning methods.

Reviewer_P8WfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DRGFM to address the challenges posed by the implicit encoding of structural knowledge in existing graph foundation models and the limited adaptability of fixed geometric operators to cross-domain structural differences. The method explicitly models structural knowledge through a learnable curvature spectrum, combining local geometric information with multi-scale spectral information to learn transferable representations that adapt to different graph structures. In addition, DRGFM employs importance-weighted regularization to constrain spectral-coordinate drift and preserve historical structural knowledge during continual learning. Experiments on few-shot node classification, zero-shot link prediction, and temporal continual learning demonstrate the effectiveness of the proposed method.

Critical strengths
  1. Modeling transferable structural knowledge through a learnable curvature spectrum, rather than relying solely on fixed geometric descriptors, is an interesting perspective and is potentially useful for cross-domain graph learning tasks.
  2. The curvature-spectral interface combines local geometric information with multi-scale spectral representations and supports the preservation of historical structural knowledge during continual learning, providing a unified modeling framework.
  3. The experiments cover both cross-domain transfer and continual learning, demonstrating the method's applicability across different learning settings.
Critical weaknesses
  1. The novelty of the proposed learnable curvature spectrum is not sufficiently established. Although the ablation studies demonstrate the benefits of learnable curvature over fixed Forman curvature, the paper lacks comparisons with general learnable edge-weighting mechanisms. It therefore remains unclear whether the improvements arise from the proposed geometric formulation or simply from the additional flexibility of adaptive edge weights.
  2. The experimental evaluation leaves several key claims insufficiently supported. The paper provides neither a computational complexity analysis nor comparisons of training and inference times with the baselines. The ablation studies mainly focus on learnable curvature, without sufficiently isolating the contributions of mixed-curvature representations, curvature-conditioned modulation and gating, and curvature-spectral attention. The spectral-coordinate regularizer is also not independently evaluated in the continual learning setting. Furthermore, several key hyperparameters and implementation details are not adequately documented, making the results difficult to reproduce.
Questions for authors
  1. How can the authors demonstrate that the performance improvements arise from the proposed curvature–spectral mechanism rather than the additional flexibility of learnable edge weights?
  2. Can the authors provide ablation studies on individual model components and loss functions to validate the necessity of each design?
  3. How does the proposed curvature–spectral mechanism maintain its effectiveness under substantial structural discrepancies between source and target graphs? Can the authors provide an analysis of its performance under cross-domain topological shifts?
4 comments
AuthorsOpenReview ↗

New mechanism controls, component ablations, and cross-domain analysis

Thank you for recognizing the potential of the curvature-spectral interface and identifying evidence needed to assess it. We added controlled comparisons, component ablations, and repeated continual-learning runs.

1. Geometry versus generic edge weights. We independently pretrained a generic edge-weight variant with the same features, estimator MLP, backbone, modulation, routing, gating, and spectral attention. It retains residual scaling, temperature, exponential mapping, and clipping, but removes the Forman prior and curvature supervision. Both models have 791,786 trainable backbone parameters.

LP ROC-AUCFacebookGitHubLastFMPubMed
Generic weights.8497.6698.8173.7816
Fixed Forman.9315.8407.9152.8876
DRGFM.9820.9546.9751.9217

DRGFM also improves all eight NC entries over generic weighting. This supports the geometrically guided formulation beyond the flexibility of an edge MLP, without attributing the entire gain to a single component.

2. Component and objective controls. Each variant was independently pretrained. The Euclidean control replaces manifold maps/distances while retaining curvature modulation, routing, gating, and spectral bias; only two registered manifold-curvature scalars become inactive. Other rows disable only the named component.

VariantFacebook LP AUCGitHub LP AUCFacebook NC 5-shot ACCGitHub NC 5-shot ACC
Euclidean representations.7561.6416.3744.5184
Equal routing weights.8561.7618.4699.6604
Without FiLM.8961.8210.4598.6507
Without FFN gating.8818.8185.4902.6496
Without spectral bias.7603.6362.3898.5916
Without node-invariance loss.8936.8354.4817.6456
Without Ollivier warm-up.9552.9175.5580.6836
DRGFM.9820.9546.5936.7649

For brevity, NC columns show 5-shot results; all seven controls also reduce 1-shot accuracy on both graphs. These ablations support the tested components across LP and NC.

On identical anchors over five training seeds, energy-weighted spectral regularization gives accuracy-AA .9307±.0109, vs .8956±.0081 without regularization, .9161±.0088 with representation distillation, and .9199±.0061 with uniform spectral weights. Forgetting is .0508, .0849, .0629, and .0600, respectively. Each regularized condition uses coefficient .5, without separate tuning. These comparisons support energy-weighted spectral preservation under the tested settings. We will incorporate the repeated estimates and scope the paper's retention conclusions to this evidence.

3. Source-target structural differences. The protocol transfers from product, image, encyclopedia, and coauthorship graphs to unseen social, music, and citation graphs. We transfer estimator/filter parameters, not a source eigenbasis; the target's visible topology and features define its curvature-weighted Laplacian. Frozen parameters thus produce graph-dependent operators without target-label optimization in zero-shot LP. Local curvature conditioning and polynomial filters respond to graph structure without aligning source and target eigenvectors.

The normalized Laplacian spectrum lies in [0,2], and the shifted Chebyshev filter obeys ∥g(L~)∥2≤∑k=0K∣αk∣\|g(\widetilde L)\|_2\leq\sum_{k=0}^K|\alpha_k|. This bounds filter amplification independently of graph size. The empirical results support effectiveness on the tested transfers; the bound does not guarantee predictive robustness under arbitrary topology shifts. We will make this distinction explicit in the paper.

4. Computation and reproducibility. For n nodes, e edges, width d, H heads, R probes, and degree K, dense attention costs O(n^2 d) and stores O(H n^2) scores; sparse probe filtering costs O(K e H R), alongside O(n d^2) projections. These are the principal per-layer costs. Ordinary attention filtering requires no full eigendecomposition; anchor modes are computed at task boundaries, with subsequent projection costing O(n_a m d).

On the same single 48-GB RTX 4090 and five source graphs, pretraining took 15,588 s for DRGFM, 17,413 s for R-GFM, and 40,808 s for GAUGE, measured from the first training step and excluding preprocessing. Baselines use the same source graphs, training-step budget, and evaluation protocol.

We will document the 3-layer, 128-hidden, 4-head architecture; manifold dimensions 80/24/24; K=3 and four probes; AdamW at 5e-4; and reconstruction/Forman/link/spectral/node loss weights .1/.1/1/1/1. Class-text supervision is disabled; Ollivier warm-up decays from 1 to 0 over 3,000 steps. We will add these hyperparameter settings to the appendix.

AuthorsOpenReview ↗

Revision available: matched edge-weight control and component ablations

Thank you again for your review. The uploaded revision now incorporates the evidence reported in our rebuttal. The main locations are:

  • Q1: geometry versus generic weighting (Section 6.4, Table 5; Appendix G). Full and generic models each have 791,786 trainable backbone parameters and share the weighting pipeline and downstream modules. DRGFM has higher mean NC and LP performance on all four targets; for example, GitHub LP AUC is .9546 versus .6698 for generic weighting.
  • Q2: components and preservation (Tables 5-6; Appendix I.4). Geometry and component controls cover both NC and LP. Separate continual-learning comparisons test energy-weighted spectral preservation against uniform weights, representation distillation, and no regularization over five training seeds.
  • Q3 and reproducibility (Appendices J, B.2, and F). We clarify how the operator adapts to each target graph and distinguish bounded filter amplification from a predictive robustness guarantee. Architecture and training settings, computational costs, and measured pretraining times are now documented. Could you briefly indicate whether the matched generic-weighting control addresses your main concern about additional edge-weight flexibility, and whether any remaining issue is decisive for your assessment? A short indication would help us focus further clarification.
AuthorsOpenReview ↗

For the AC: revised evidence relevant to Reviewer P8Wf's concerns

To facilitate the meta-review, we summarize the evidence incorporated into the revision in response to Reviewer P8Wf, together with the remaining limitations.

1. Generic edge-weight flexibility (Section 6.4, Table 5; Appendix G.1, Tables 10-11). Full and generic models each have 791,786 trainable backbone parameters and share the estimator architecture, edge-weight mapping, and downstream modules. Generic removes the Forman prior and curvature supervision. DRGFM has higher mean performance in all eight NC and eight LP entries across four targets; GitHub LP AUC is .9546 versus .6698, and 5-shot NC accuracy is .7649 versus .7033. This supports the geometrically guided formulation beyond the flexibility of the tested generic edge MLP, without attributing the gain to the prior alone.

2. Component and preservation controls (Tables 5-6; Appendices G and I.4). The Euclidean control retains curvature conditioning and spectral attention. Independently pretrained ablations separately test routing, FiLM, FFN gating, spectral bias, node-invariance loss, and Ollivier warm-up on NC and LP. These isolate the listed components and two pretraining objectives.

On identical anchors over five complete sequential runs, energy-weighted preservation gives mean accuracy-AA .9307, versus .8956 without regularization, .9161 with representation distillation, and .9199 with uniform spectral weights. Accuracy forgetting is .0508, .0849, .0629, and .0600, respectively. Tables 22-23 report SDs and AP-based metrics. Energy, uniform, and representation use coefficient .5 without separate tuning; the conclusion concerns these tested settings.

3. Cross-domain interpretation (Appendices J and B.2). Estimator/filter parameters transfer, while each target's visible topology and features define its own curvature Laplacian, without source-target eigenvector alignment. Appendix J clarifies why the operator remains graph-dependent with frozen parameters; Appendix B.2 bounds filter amplification independently of graph size. These explain the transfer mechanism but do not establish predictive robustness as topology-shift severity increases. No controlled severity experiment is added.

4. Reproducibility and cost (Appendix F, Tables 7-9). The revision documents architecture, loss weights, optimization, anchors, statistical protocols, and principal computational costs. Pretraining times on the same RTX 4090, source graphs, and step budget are 15,588 s (DRGFM), 17,413 s (R-GFM), and 40,808 s (GAUGE), excluding preprocessing and download; inference-time comparisons are not reported. Tables 3-4 and Appendix I replace single-run CL results with mean and SD over five complete runs, conditional on a common pretrained checkpoint and fixed split.

The revision strengthens the mechanism and preservation evidence within the tested settings. We would appreciate consideration of these additions and their stated limits in the meta-review, and remain available to clarify unresolved points.

AuthorsOpenReview ↗

Follow-up on the edge-weight comparison

Thank you again for suggesting the generic edge-weight comparison in Q1. In the experiment reported in our rebuttal, both models use the same edge MLP and have the same trainable backbone parameter count. DRGFM achieved higher mean 5-shot node classification accuracy and link prediction AUC on each of the four target graphs.

As the discussion draws to a close, could you let us know whether this comparison addresses your concern about edge-weight flexibility, or whether there is a specific point we could still clarify?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.