acceptodds
Withdrawn from ICLR 2027

Towards Downstream Tuning-Free Graph Foundation Models with Manifold-based Structural Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph foundation models (GFMs) demonstrate strong potential in cross-domain graph learning by pre-training on multiple source domains and adapting to unseen target domains. Many existing GFMs are generally prompt-based and address graph domain gaps through downstream tuning of additional prompt embeddings, and therefore focus primarily on representation-level adaptation under a Euclidean geometric assumption. However, real-world graphs often share transferable structural patterns across domains, such as tree-like and cyclic substructures, that are naturally non-Euclidean and thus cannot be sufficiently exploited by existing Euclidean-based prompt-centric GFMs. To address this issue, this paper presents GFManifold, a graph structure-centric framework that adopts hyperbolic and spherical Riemannian manifold branches with learnable curvatures to learn manifold-induced structural relations across source domains for efficient parameter tuning-free GFM adaptation. During pre-training, GFManifold learns cross-domain graph structural patterns through domain alignment and geodesic link prediction across diverse source-domain graphs. During downstream adaptation, GFManifold exploits manifold-induced geodesic relations to adapt the target graph structure for downstream tasks, enabling efficient target-domain parameter tuning-free adaptation and avoiding downstream gradient-based parameter updates in most existing GFMs. Extensive experiments on 14 benchmarks showcase its effectiveness and efficiency on cross-domain graph adaptation tasks. Code is available upon acceptance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Data-centric Methods, Parameter Tuning-Free Adaptation, Cross-Domain Graph Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_QjkWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes GFManifold, a graph foundation model for downstream parameter tuning-free adaptation. It employs hyperbolic and spherical constant-curvature manifolds with globally shared learnable curvature parameters. During pretraining, the model jointly optimizes a GNN encoder and manifold branches through MMD-based domain alignment and geodesic link prediction. During downstream inference, manifold-induced distances are used to reweight the target graph without updating pretrained parameters.

While the proposed adaptation strategy is practically relevant, I have substantial concerns about the methodological novelty, the justification of the geometric construction, and the validity and significance of the theoretical claims.

Critical strengths

S1. Practical downstream adaptation without parameter updates.

The graph-reweighting strategy avoids gradient-based optimization on unseen target domains while retaining a pretrained GNN encoder. This is a practically relevant design that may improve adaptation efficiency. However, its practical utility alone does not establish the claimed geometric or theoretical contribution.

Critical weaknesses

W1. Limited geometric novelty and an unsubstantiated connection to domain-invariant structural learning.

The central motivation assumes that transferable graph motifs, such as trees, cycles, and cliques, are insufficiently captured by Euclidean GFMs and can be better exploited through non-Euclidean geometry. However, the existence of recurring motifs does not imply that they are domain-invariant or label-relevant. Moreover, Euclidean-coordinate representations do not necessarily imply intrinsically flat data geometry, just as explicitly parameterizing manifold curvature does not establish that the intrinsic geometry of graph data has been learned.

GFManifold employs only two globally shared scalar curvature parameters, corresponding to predefined hyperbolic and spherical constant-curvature spaces. These parameters determine the curvature of the model spaces, not the intrinsic curvature of individual graphs. Although different embedding configurations can represent heterogeneous relations, the model provides no explicit mechanism for adapting curvature or geometric mixing to individual nodes, regions, or domains, nor a mathematical argument that its globally shared construction faithfully preserves heterogeneous transferable structures.

Furthermore, Eq. (6) combines two geodesic distances into a composite dissimilarity, rather than defining a new Riemannian manifold or unified geodesic geometry. The construction therefore amounts to combining two homogeneous geometric views through globally learned weights. Global parameter sharing does not, by itself, establish structural invariance.

This is particularly concerning given prior work such as RiemannGFM [1], GraphGlue [2], and GAUGE [3], which have already investigated non-Euclidean geometry and transferable graph structures. The main differences of GFManifold lie in global curvature parameterization, pretraining objectives, and downstream edge reweighting. The paper does not demonstrate a fundamentally new geometric capability or explain why this construction should identify domain-invariant, label-relevant structures beyond existing approaches.

W2. The theoretical analysis assumes the desired invariance and does not establish a substantive guarantee for the proposed method.

The actual learning objective combines MMD-based distribution alignment and geodesic link prediction. Neither objective explicitly separates domain-invariant from domain-specific structural information. MMD aligns marginal representation distributions without guaranteeing label-conditional invariance, while link prediction can preserve domain-specific connectivity. Since the GNN encoder processes the original graph before manifold mapping, the proposed objectives do not establish that the resulting distances isolate transferable structural components.

The theoretical analysis does not resolve this issue. Appendix E assumes, for an idealized conditional analysis, that the manifold-induced transformation corresponds to the shared structural component and that this component is invariant and sufficient for prediction. The authors acknowledge that the practical objectives are not guaranteed to recover this decomposition.

More importantly, Eq. (18) is essentially a marginalization identity for a predictor operating on the assumed shared representation, while Eqs. (19)–(20) follow directly from the definition of a minimax-optimal predictor. These results do not establish that the actual GFManifold optimization approaches the minimax solution or recovers invariant structure on unseen domains.

An additional mathematical concern arises in Eqs. (7)–(8), where a Gaussian-type kernel is constructed from squared composite geodesic distances and used for MMD. Positive semidefiniteness is required for the standard RKHS interpretation of MMD, but it is not automatically guaranteed for Gaussian kernels based on geodesic distances in curved spaces. The paper does not establish this property for its proposed composite distance.

Consequently, the theoretical analysis provides no substantive connection between the learned global curvatures, the actual optimization objectives, and the claimed cross-domain structural invariance. This is a fundamental gap in the paper's central argument rather than a minor technical omission.

Questions for authors

The following questions concern issues that could materially affect my assessment.

  1. Geometric novelty and expressiveness. What fundamental geometric modeling capability does GFManifold introduce beyond RiemannGFM [1], GraphGlue [2], and GAUGE [3]? In particular, what justifies the sufficiency of two globally shared constant-curvature parameters for learning transferable structures across heterogeneous graphs, and how can their geometric contribution be distinguished from the expressivity of the jointly trained GNN encoder?

  2. Structural invariance. What mathematical property or explicit training mechanism ensures that the proposed MMD and link-prediction objectives recover domain-invariant, label-relevant structures rather than generic or domain-specific graph similarities? Can the authors provide a nontrivial justification connecting the actual learned representations to the assumed invariant decomposition?

  3. Mathematical validity of the distance-based objective. Are the aggregation weights in Eq. (6) constrained to be nonnegative? More importantly, can the authors establish that the Gaussian-type kernel in Eq. (7) is positive semidefinite under the proposed composite geodesic distance, thereby justifying its use in the MMD objective?

  4. Theoretical significance. Beyond marginalization in Eq. (18) and the minimax definition in Eqs. (19)–(20), what nontrivial generalization property does Theorem 4.1 establish for GFManifold? How does the actual source-domain optimization connect to the stated minimax formulation involving unseen target domains?

References

[1] RiemannGFM: Learning a Graph Foundation Model from Riemannian Geometry. WWW 2025.

[2] Multi-Domain Riemannian Graph Gluing for Building Graph Foundation Models. ICLR 2026.

[3] Are Common Substructures Transferable? Riemannian Graph Foundation Model with Neural Vector Bundles. ICML 2026.

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

This paper proposes GFManifold, a manifold-based tuning-free adaptation method for graph foundation model. This method uses hyperbolic and spherical manifold branches with learnable curvatures to learn graph structural relation across different source domain. During downstream adaptation period, it uses the geodesic relation to adjust the edge weights of target graph without updating model parameters. Experiments are conducted on 14 graph datasets. The results show GFManifold achieves improvement in node and graph classification.

Critical strengths

S1. This paper studies a meaningful problem: how to adapt graph foundation models to a unseen domain without fine-tuning model parameters.

S2. The method is clear and easy to understand. It is an interesting idea that using manifold-based structural relations to adjust the target graph can improve adaptation.

S3. Experiments are extensive. The paper evaluates both node and graph classification across several datasets.

Critical weaknesses

W1. The complementary between hyperbolic and spherical hasn't be fully demonstrated. Although Fig. 3 provides several ablation studies, it does not include variants that only use hyperbolic or spherical manifold. Because the two branches are derived from the same GNN representation, it is not clear whether we really need two types of manifolds.

W2. The performance improvement might partly come from the simple graph structure refinement. According to Alg. 1, every edge with below 0.1 weights will be removed during downstream adaptation. I believe that compare GFManifold with simple edge-pruning methods with the same sparsity level will make the conclusion more convincing.

W3. The theoretical analysis does not fully explain. Theorem 4.1 gives a cross-domain risk objective across both source domain and target domain. However, the model is trained using MMD and Geodesic Link Prediction loss, while target domain is unseen during pre-training. I believe a proof of how these training objective achieve the theoretical optimum would make the analysis more complete.

Questions for authors

See W1~3

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

The paper proposes GFManifold, a GNN-based graph foundation model that aims to achieve downstream parameter tuning-free adaptation. The core idea is to pre-train two complementary constant-curvature Riemannian branches with learnable scalar curvatures jointly with a GNN encoder across multiple source domains, using MMD-based domain alignment and geodesic link prediction. At downstream time, the pre-trained model is frozen; instead of tuning prompts/classifiers, the target graph is transformed into a geodesic-weighted graph by reweighting edges according to manifold-induced structural dissimilarity, and then encoded by the fixed GNN for few-shot inference.

Critical strengths
  1. A novel structure-centric GFM framework is proposed to learnmanifold-induced structural relations across source domains. It exploits manifold-induced geodesic relations for target graph structural adaptation.
  2. It enables parameter tuning-free GFM adaptation by adapting the graph structure from a data centric aspect. Reframing GFM adaptation from the parameter space to the graph-structure space is a clearly positioned and practically motivated alternative, and Table 1 makes this distinction legible.
Critical weaknesses
  1. The paper's novelty over RiemannGFM is (a) learnable vs. fixed curvature and (b) structure reweighting vs. representation-level use. However, the ablation in §5.2 is presented only as a numeric-free radar plot in Figure 3, with three variants (i.e., w/o GeoGraph, w/o Manifold, w/o κ). The problem is that the w/o GeoGraph variant is visually close to the full model on several datasets in Figure 3 (e.g., Wisconsin, Computers, Pubmed), while the §5.2 text concludes that "the non-Euclidean manifold component has the most significant impact." Because Figure 3 has no numeric table, no standard deviations, and no y-axis units, a reader cannot verify this claim. From the presented evidence, the paper cannot determine whether the gain comes from learnable curvature, from the two-branch composite metric, or simply from the act of reweighting and pruning target-graph edges.
  2. Theorem 4.1 is a restatement of assumptions, not theoretical support for the method, yet it is listed as a primary contribution. Theorem 4.1 states that optimal parameters minimize the sup of expected risk over source and target domains, under Assumptions E.1 (independence of shared and domain-specific components) and E.2 (sufficiency of the shared component). Theorem 4.1 is essentially definitional and does not connect to the actual training objective.
  3. The ablation does not separate "learnable curvature" from "two branches" from "edge reweighting + pruning." The paper's central novelty claim is learnable curvature (Table 1), yet the ablation only compares full model vs. w/o κ (fixed κ = ±1). It does not include (a) a single-branch variant (hyperbolic-only or spherical-only), (b) a variant with learnable curvature but a single shared scalar for both branches, (c) a variant with fixed curvature but the full two-branch composite metric, and (d) a variant with geodesic reweighting but no pruning threshold (0.1). Without (a)–(c), the claim that "complementary branches" and "learnable curvature" each matter is unsupported. Without (d), the gain may be attributable to pruning alone, which is a well-known trick for heterophilic graphs.
  4. Efficiency claims need a fairer accounting. §5.5 and Figure 5 compare wall-clock and GPU memory "during downstream by each method's pre-trained model." However, GFManifold's downstream step still requires (i) a forward pass of the GNN on the target graph, (ii) exponential maps to two manifolds, (iii) pairwise geodesic computations for edge reweighting, and (iv) a second GNN forward pass on the reweighted graph (Algorithm 1, lines 14–18). The paper should report the end-to-end downstream time including both forward passes and the geodesic computation, and separate pre-training cost from adaptation cost. Also, Figure 5 uses point size for GPU peak memory but reports only two datasets (Cornell, Citeseer); please report the full table for all 14 benchmarks.
Questions for authors
  1. The paper's novelty over RiemannGFM is (a) learnable vs. fixed curvature and (b) structure reweighting vs. representation-level use. However, the ablation in §5.2 is presented only as a numeric-free radar plot in Figure 3, with three variants (i.e., w/o GeoGraph, w/o Manifold, w/o κ). The problem is that the w/o GeoGraph variant is visually close to the full model on several datasets in Figure 3 (e.g., Wisconsin, Computers, Pubmed), while the §5.2 text concludes that "the non-Euclidean manifold component has the most significant impact." Because Figure 3 has no numeric table, no standard deviations, and no y-axis units, a reader cannot verify this claim. From the presented evidence, the paper cannot determine whether the gain comes from learnable curvature, from the two-branch composite metric, or simply from the act of reweighting and pruning target-graph edges.
  2. Theorem 4.1 is a restatement of assumptions, not theoretical support for the method, yet it is listed as a primary contribution. Theorem 4.1 states that optimal parameters minimize the sup of expected risk over source and target domains, under Assumptions E.1 (independence of shared and domain-specific components) and E.2 (sufficiency of the shared component). Theorem 4.1 is essentially definitional and does not connect to the actual training objective.
  3. The ablation does not separate "learnable curvature" from "two branches" from "edge reweighting + pruning." The paper's central novelty claim is learnable curvature (Table 1), yet the ablation only compares full model vs. w/o κ (fixed κ = ±1). It does not include (a) a single-branch variant (hyperbolic-only or spherical-only), (b) a variant with learnable curvature but a single shared scalar for both branches, (c) a variant with fixed curvature but the full two-branch composite metric, and (d) a variant with geodesic reweighting but no pruning threshold (0.1). Without (a)–(c), the claim that "complementary branches" and "learnable curvature" each matter is unsupported. Without (d), the gain may be attributable to pruning alone, which is a well-known trick for heterophilic graphs.
  4. Efficiency claims need a fairer accounting. §5.5 and Figure 5 compare wall-clock and GPU memory "during downstream by each method's pre-trained model." However, GFManifold's downstream step still requires (i) a forward pass of the GNN on the target graph, (ii) exponential maps to two manifolds, (iii) pairwise geodesic computations for edge reweighting, and (iv) a second GNN forward pass on the reweighted graph (Algorithm 1, lines 14–18). The paper should report the end-to-end downstream time including both forward passes and the geodesic computation, and separate pre-training cost from adaptation cost. Also, Figure 5 uses point size for GPU peak memory but reports only two datasets (Cornell, Citeseer); please report the full table for all benchmarks.
Reviewer_YVheOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GFManifold, a structure-centric graph foundation model. It learns cross-domain shared non-Euclidean structural patterns during pretraining through hyperbolic and spherical manifold branches with learnable curvature. In the downstream stage, it directly leverages geodesic relations to reweight the edges of the target graph, thereby achieving efficient cross-domain adaptation without parameter fine-tuning. It significantly outperforms existing prompt-tuning-based GFMs on 14 benchmarks while incurring lower computational overhead.

Critical strengths

It proposes a structure-centric adaptation mechanism that requires no target-domain parameter fine-tuning, enabling effective cross-domain transfer without updating model parameters. This design simultaneously yields significant performance gains and efficiency advantages.

Critical weaknesses

The cross-domain generalization analysis (Theorem 4.1) relies on relatively strong independence and sufficiency assumptions. Its main proof is relegated to the appendix, and its supporting relationship with the experiments is not sufficiently tight.

Questions for authors

1.Why is it necessary to use both hyperbolic and spherical manifolds? Can a single manifold achieve similar results?

2.How can the authors demonstrate that the model indeed captures cross-domain shared structural patterns rather than merely learning node similarities?

3.What motivates selecting the domain with the smallest average geodesic edge distance as the pivot for MMD alignment? How sensitive is the model to different pivot selection strategies?

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

Related papers

Discussion (0)

Sign in to comment.