{"id":"9827bd7b-aca9-4f81-8585-7fb4a85361ba","arxiv_id":"2502.03251","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RiemannGFM learns transferable graph structure by pretraining on tree and cycle substructures embedded in hyperbolic and spherical spaces.","lead":"RiemannGFM pretrains a graph neural network from structure only, treating trees and cycles as a shared vocabulary embedded in curved spaces. It reports strong cross-domain transfer on graphs without text features, but the evidence has gaps in error bars, baseline details, and a geometry proof.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transfer claim depends on unaligned Laplacian eigenvector inputs; sign and basis ambiguity across graphs is never handled or tested.","rationale":"The reader's weakest_assumption correctly identifies the Laplacian eigenvector input as the load-bearing point: it is the only input to the model, so if it is not comparable across graphs, the pretrained structural vocabulary cannot transfer. The paper fixes K but does not address sign ambiguity or eigenspace basis ambiguity, and no experiment tests sensitivity to these choices. This directly threatens the central claim of generating informative node encodings for an arbitrary graph. The additional theorem issue (midpoint proof using chordal rather than geodesic distance) is a real correctness concern, but it does not by itself invalidate the empirical transfer results; the eigenvector standardization problem is more decisive. The proposed concrete test is straightforward and would settle whether the concern lands. Since the reader already reached CONDITIONAL on essentially this basis, the verdict remains unchanged: the paper is promising but requires this robustness check and a corrected or qualified theoretical claim before the universality statement can be accepted.","tokens_in":24384,"tokens_out":4975,"duration_ms":52393,"concrete_test":"For each target graph (Citeseer, Pubmed, GitHub, Airports), recompute the top-K eigenvectors of the normalized graph Laplacian, then flip the sign of each eigenvector with probability 1/2 before running the pretrained RiemannGFM and the downstream link-prediction evaluation. Repeat with 10 independent random sign masks. If the mean AUC/AP in Table 1 changes by more than about 1 percentage point, or if any mask changes the ranking against baselines, the model is not invariant to an unavoidable degree of freedom in its input and the reported transfer results are not robust. For a stronger check, also apply random orthogonal rotations within numerically degenerate eigenspaces and verify stability. If the results are stable across all masks, the eigenvector-transfer concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The model's only input is the top-K eigenvectors of the normalized graph Laplacian (Sec 4.1.3 and App. E.3.2). These eigenvectors are defined only up to sign, and for degenerate eigenspaces up to arbitrary orthonormal basis rotations. The paper 'normalizes' different datasets only by fixing the dimension K; it does not align signs or bases across graphs, and it describes no sign-robustness mechanism such as using absolute values, sign-invariant encodings, or sign-flip data augmentation. The proposed attention and bundle convolution are not sign-invariant functions of their inputs, so a global sign flip of the input eigenvectors will generally change the node encodings and downstream predictions. Since pretraining on ogbn-arxiv, Physics, and Amazon-Computers exposes the model to one arbitrary eigenvector sign convention per dataset, the target graphs (Citeseer, Pubmed, GitHub, Airports) may carry different conventions, and the claimed cross-domain structural transfer could be an artifact of accidentally consistent sign choices. No experiment in the paper checks this. Separately, the proof of Theorem 2 in App. B.2 identifies the squared distance with the chordal expression 2/κ − 2⟨x, y⟩_κ, which is not the squared geodesic distance on the hyperboloid or sphere; the 'geometric midpoint' is therefore not the Fréchet mean claimed. That mathematical issue is serious but secondary to the core universality claim, which stands or falls on whether the eigenvector input is a reliable shared structural representation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RiemannGFM, a self-supervised graph foundation model that uses only structural information, without textual attributes. The model is built around a 'structural vocabulary' of trees and cycles, which are respectively embedded in hyperbolic and spherical spaces via a product bundle that pairs manifold coordinates with tangent-space node encodings. The architecture consists of a cross-geometry attention module, a bundle convolution based on parallel transport, and a geometric contrastive loss between hyperbolic and spherical views. The model is pretrained on ogbn-arxiv, Physics, and Amazon-Computers and evaluated on Citeseer, Pubmed, GitHub, and Airports for node classification and link prediction, with additional few-shot, ablation, and pretraining-dataset studies. The central claim is that a single structure-only model can transfer across arbitrary graph domains by learning shared geometric structure.","tokens_in":24690,"tokens_out":6258,"duration_ms":59932,"significance":"If the claims hold, RiemannGFM would be a genuinely novel direction for graph foundation models: it moves beyond LLM-based, text-attributed approaches and provides a structure-only pretraining paradigm grounded in Riemannian geometry. The paper deserves credit for proposing a concrete structural-vocabulary hypothesis (trees to hyperbolic, cycles to spherical), releasing code, and including ablations over geometry choices and pretraining datasets. However, the significance is conditional on two load-bearing points: whether unaligned Laplacian eigenvectors can serve as a transferable input representation, and whether the geometric midpoint operation is mathematically correct as stated. The current manuscript does not establish either point, and the main experimental table lacks error bars, so the empirical support is weaker than the text suggests.","major_comments":[{"comment":"The model's only input is the top-K eigenvectors of the normalized graph Laplacian, but the paper neither aligns eigenvector signs/bases across graphs nor uses any sign-invariant mechanism. Since eigenvectors are defined only up to sign (and, for degenerate eigenspaces, up to arbitrary orthogonal rotations), a global sign flip of the input will generally change the outputs of the attention in Eq. (6) and the bundle convolution in Eq. (7), because these operations are not sign-invariant. Pretraining on ogbn-arxiv, Physics, and Amazon-Computers therefore fixes one arbitrary sign convention per dataset, while target graphs may carry different conventions, so the reported cross-domain transfer could be an artifact of accidentally consistent sign choices. The authors should test this directly: for example, apply random sign flips (and random basis rotations within degenerate eigenspaces) to the input eigenvectors at evaluation time and report the variance in downstream metrics, or train with sign-flip augmentation and verify robustness. Alternatively, they should replace eigenvector inputs with sign-invariant spectral features. This issue is load-bearing because the universality claim rests entirely on the transferability of these input features.","section":"Sec. 4.1.3 and App. E.3.2"},{"comment":"Theorem 2 is not proved as stated. The proof identifies the squared distance with the chordal expression 2/κ − 2⟨x, y⟩_κ, but this is not the squared geodesic distance on the hyperboloid or sphere. The weighted arithmetic mean in Eq. (4) is the normalized projection of a weighted average in the ambient space, and it minimizes the sum of squared chordal distances, not the sum of squared geodesic distances. Therefore the claim that Eq. (4) is the geometric midpoint (Fréchet mean) with respect to geodesic distance is incorrect, and Eq. (5) does not implement the stated geometric midpoint. This matters because the vocabulary learning module and graph-level alignment both rely on the midpoint operation. The authors should either prove the result for the true geodesic distance, replace the claim with a chordal-distance centroid and adjust the narrative, or implement a Fréchet mean and re-derive the attention update accordingly.","section":"Sec. 3.1.2 and App. B.2, Eq. (4) and Eq. (5)"},{"comment":"The manifold-preserving linear operation in Eq. (3) is undefined whenever W x_s = 0, since the re-scaling factor α divides by ||W x_s||. This includes the north pole x = [1/sqrt(|κ|), 0, ..., 0] for any W, so Theorem 1's claim 'for any W' is not literally true. The authors should state the domain where the operation is defined and explain how the model handles nodes whose spatial coordinate is zero or whose transformed spatial coordinate vanishes, especially during initialization from zero input encodings.","section":"Sec. 3.1.2, Eq. (3) and Theorem 1"},{"comment":"The main transfer-learning table reports no standard deviations or significance tests, and the downstream protocol for link prediction is under-specified. In particular, the text claims 'consistently achieves the best results' for link prediction, but some node-classification gaps are small (e.g., GitHub ACC 85.96 vs 85.68 for GCN; Airport ACC 55.29 vs 52.34 for GraphMAE2), and without error bars or significance tests these differences are not evaluable. Furthermore, the paper does not describe how link prediction heads are trained or evaluated on new graphs (negative sampling, train/val/test split, frozen vs fine-tuned encodings), nor how featureless baselines (GCN, SAGE, DGI, GraphMAE2) are applied to GitHub and Airports, which have feature dimension 0 in Table 6. The authors should add error bars/significance tests to Table 1 and fully specify the evaluation protocols in Appendix E.","section":"Sec. 4.2.1, Table 1, and App. E"}],"minor_comments":[{"comment":"There are several typos and citation inconsistencies: 'RemannGFM' in the conclusion, 'prformance' in the Table 4 caption, 'eignvectors' in App. E.3.2, and 'Importation Notations' in Table 5. In Table 3, OpenGraph is cited as [59] but the correct reference appears to be [50].","section":"Throughout"},{"comment":"The metric of the product bundle is written with I_{d_H+1} for both the hyperbolic and spherical tangent factors; the second identity should likely be I_{d_S+1}.","section":"Eq. (2)"},{"comment":"The definition of the curvature-aware cosine says 'cos_κ(·) = cosh(·) with κ > 0', but for the hyperbolic case one would expect κ < 0; this appears to be a typo.","section":"App. B.3"},{"comment":"The proof of the midpoint switches between indices i and j (e.g., 'Σ_i ν_i x_j' and 'Σ_{j∈N_i} ν_{ij} h_j'), which makes the derivation hard to follow; the notation should be cleaned up.","section":"App. B.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the venue and addresses a timely topic. In my view, the revision should focus on two things before reconsideration: an empirical demonstration that the Laplacian-eigenvector input is stable under sign/rotation ambiguity (or a change to sign-invariant inputs), and a correction of the geometric midpoint theorem. The experimental tables outside Table 1 include standard deviations, so the missing error bars in Table 1 are likely fixable rather than fatal. I would not recommend acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new thing here is the combination: a structure-only graph pretraining model whose shared vocabulary is trees and cycles, embedded in a product of hyperbolic and hyperspherical bundles with tangent-space encodings. I don't know another GFM that does this without text. The paper ships code, gives ablations for the geometry choices, and shows impressive link prediction transfer on non-text graphs (Airports, GitHub). Credit where due: the architecture is inventive and the empirical wins are at least plausible.\n\nThe soft spots are real, though. The \"any graph\" claim rests on top-K eigenvectors of the normalized graph Laplacian as the only input. Those eigenvectors are defined only up to sign, and up to basis rotation in degenerate eigenspaces. The paper normalizes only by fixing K; there is no sign alignment, no sign-invariant encoding, no sign augmentation. Since the attention and bundle convolution are not sign-invariant functions of their inputs, a global sign flip changes the encodings. Pretraining on three source graphs uses one arbitrary sign convention per graph; target graphs can carry different ones. The paper never checks this. That is a load-bearing unknown, not a nit. Also, using the largest eigenvectors of the normalized Laplacian is unusual—I'd want to know whether they meant the smallest.\n\nSecond, Theorem 2's proof identifies squared distance with 2/κ − 2⟨x,y⟩_κ, which is the chordal distance, not the squared geodesic distance on the sphere or hyperboloid. So the \"geometric midpoint\" is not the Fréchet mean it is claimed to be. The aggregation may still work in practice, but the theorem as stated does not follow from that proof.\n\nThird, Table 1 has no error bars or significance tests, and the downstream heads are under-specified. Those are fixable, but they matter for the headline transfer numbers. The universality claim also rests on only four target datasets, which is thin for \"any graph.\"\n\nThe central idea is good enough to deserve referee time. I'd send it to review, but with a clear request: handle or justify the eigenvector sign/basis issue, correct or downgrade the midpoint theorem, and add variance and error bars to the main table. The paper is for graph-ML readers working on pretraining and geometric GNNs; it does not reorganize a field, but it is a step beyond text-coupled GFMs.","headline":"A genuinely novel structure-only pretraining architecture whose 'any graph' claim currently rests on unaligned spectral inputs and a shaky geometric proof; worth reviewing but not ready as stated.","tokens_in":25204,"tokens_out":3489,"would_cite":true,"duration_ms":35367,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RiemannGFM claims that a single model pretrained only on the structural vocabulary of trees and cycles, embedded in appropriately curved Riemannian spaces, can transfer across arbitrary graphs—including graphs with no textual attributes.","keywords":["graph foundation model","Riemannian geometry","structural vocabulary","constant curvature space","hyperbolic space","hyperspherical space","product bundle","cross-domain transfer learning"],"falsifier":"A concrete test: take the frozen pretrained RiemannGFM and, for each target graph, randomly flip the signs or permute the rows of its top-$K$ Laplacian eigenvectors before generating encodings. If link-prediction AUC on Citeseer or Airport stays near the reported values, the eigenvectors are not the carrier of transfer; if it collapses, the method depends on a per-graph spectral alignment that the paper does not provide.","tokens_in":24179,"feed_emoji":"🕸️","tokens_out":8527,"duration_ms":74579,"temperature":0.7,"pith_summary":"This paper tries to establish that a graph foundation model does not need text attributes or language models to transfer across domains: the shared 'vocabulary' of trees and cycles is enough. The proposed model, RiemannGFM, learns that vocabulary in curved geometric spaces, placing trees in hyperbolic space and cycles in hyperspherical space, and is pretrained with a geometric contrastive objective. If the claim is right, one pretrained structure-only encoder can serve arbitrary graphs, which would matter for the many real graphs that carry no useful text. The paper supports the claim with cross-domain link prediction and few-shot node classification experiments, including a Citeseer link-prediction AUC of 99.40 against 96.90 for a strong self-supervised baseline.","feed_headline":"Trees and cycles give one graph model cross-domain transfer","feed_subtitle":"Pretrained on graph geometry alone, RiemannGFM beats text-based rivals on link prediction without any node text.","key_machinery":"The load-bearing object is the product bundle: a manifold paired with its tangent spaces, formed as the Cartesian product of a hyperbolic factor (for trees) and a hyperspherical factor (for cycles), written $\\mathcal{P}^{d_P} = (\\mathbb{H}^{d_H}_{\\kappa_H}\\otimes T\\mathbb{H}^{d_H}_{\\kappa_H})\\otimes(\\mathbb{S}^{d_S}_{\\kappa_S}\\otimes T\\mathbb{S}^{d_S}_{\\kappa_S})$. Node coordinates on the manifolds carry local substructure geometry, while node encodings in the tangent spaces carry global structure; bundle convolution resolves the incompatibility of different tangent spaces by parallel transport, and the geometric midpoint aggregates coordinates across sampled substructures. The vocabulary learning module uses a manifold-preserving linear operation and cross-geometry attention (query from one geometry, key and value from the other), and the geometric contrastive loss contrasts the two views in the shared tangent space of the north pole.","core_discovery":"On the paper's own terms, the discovery is that the graph domain has a structural vocabulary—trees and cycles—and that this vocabulary maps onto constant-curvature Riemannian geometry: trees align with hyperbolic space, cycles with hyperspherical space. The model pairs each node with a coordinate on a manifold (its position inside a sampled substructure) and an encoding in the tangent space at that coordinate (its global-structure information), combined into a product bundle $\\mathcal{P}^{d_P} = (\\mathbb{H}^{d_H}_{\\kappa_H}\\otimes T\\mathbb{H}^{d_H}_{\\kappa_H})\\otimes(\\mathbb{S}^{d_S}_{\\kappa_S}\\otimes T\\mathbb{S}^{d_S}_{\\kappa_S})$. A universal Riemannian layer alternates cross-geometry attention, which updates coordinates within substructures, and bundle convolution, which transports encodings between tangent spaces via parallel transport. Pretraining is self-supervised through a contrastive loss between the hyperbolic and hyperspherical views, with only the top-$K$ eigenvectors of the normalized graph Laplacian as input; the same frozen parameters then generate node encodings for an unseen graph.","pith_inferences":["(Editorial inference) If eigenvectors are truly the only input, the paper implies that graph structure alone carries a universal signal; a natural next test is whether random or sign-flipped eigenvectors destroy transfer, which would reveal how much of the result depends on exact spectral alignment.","(Editorial inference) The tree-and-cycle vocabulary is minimal, and the same product-bundle design can accommodate additional substructures or curvatures, so the framework may extend to motif-rich graphs beyond the two-geometry case.","(Editorial inference) The geometric contrastive loss can be read as a parameter-free form of multi-view contrastive learning, suggesting a bridge between geometry-aware self-supervision and standard graph contrastive methods that rely on augmentation.","(Editorial inference) A practical consequence the authors do not spell out is that the method applies to graphs whose node features are missing or privacy-sensitive, since node identity is represented purely through spectral structure."],"forward_implications":["A graph foundation model can be built without text: structure alone transfers across citation, social, collaboration, and airport networks, so text-free graphs are no longer excluded from pretraining.","The same frozen encoder competes with specialized models on link prediction; on Citeseer it reaches 99.40 AUC, above DGI's 96.90, and it leads the reported baselines on Pubmed, GitHub, and Airport.","Few-shot node classification improves most on non-attributed graphs, for example 77.83 accuracy in the 1-shot setting and 84.19 accuracy in the 5-shot setting on GitHub, where text-based models show negative transfer.","The geometry pairing is doing real work: trees in hyperbolic space and cycles in hyperspherical space beat the Euclidean, all-hyperbolic, and all-spherical alternatives in the paper's ablations.","Because the only input is a fixed-size set of Laplacian eigenvectors, the model applies to any graph regardless of node feature dimension, and it supports minibatch training."],"supporting_citations":[{"why":"Supplies the geometric fact that trees embed in the hyperbolic plane with bounded distortion, grounding the tree-hyperbolic pairing.","marker":"[31]"},{"why":"Supplies the Riemannian-geometry machinery (constant curvature spaces, tangent bundles, parallel transport) that the product bundle and bundle convolution are built on.","marker":"[28]"},{"why":"Supplies the volume-growth evidence that hyperbolic space aligns with hierarchical, tree-like structures.","marker":"[26]"},{"why":"Supplies the product-of-model-spaces mixed-curvature formulation that RiemannGFM extends by adding tangent spaces.","marker":"[12]"},{"why":"Supplies the deep-graph-infomax contrastive baseline whose Citeseer link-prediction score (96.90 AUC) is compared against RiemannGFM's 99.40.","marker":"[46]"},{"why":"Supplies the cross-domain graph pretraining baseline that motivates and benchmarks the claimed transferability.","marker":"[58]"},{"why":"Supplies the LLM-based open graph foundation model baseline, which fails on non-attributed graphs and frames the text-free gap.","marker":"[50]"}],"fun_headline_variants":["Graph geometry alone powers cross-domain model","Trees and cycles: a universal graph vocabulary","No text needed: Graph model learns from geometry","Riemannian geometry gives graphs a shared vocabulary","Universal graph pretraining via trees and cycles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pretraining signal rests on the assumption that the top few eigenvectors of a graph's normalized Laplacian—computed independently for each graph, with arbitrary signs and orderings—encode a reusable structural vocabulary that transfers across very different graphs.","fun_headline_variants_meta":{"raw":{"variants":["Graph geometry alone powers cross-domain model","Trees and cycles: a universal graph vocabulary","No text needed: Graph model learns from geometry","Riemannian geometry gives graphs a shared vocabulary","Universal graph pretraining via trees and cycles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1599,"prompt_tokens":1063,"completion_tokens":536,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":469}},"tokens_in":679,"tokens_out":536,"duration_ms":4756,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:22:19.299057+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take the frozen pretrained RiemannGFM and, for each target graph, randomly flip the signs or permute the rows of its top-$K$ Laplacian eigenvectors before generating encodings. If link-prediction AUC on Citeseer or Airport stays near the reported values, the eigenvectors are not the carrier of transfer; if it collapses, the method depends on a per-graph spectral alignment that the paper does not provide.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the geometric fact that trees embed in the hyperbolic plane with bounded distortion, grounding the tree-hyperbolic pairing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Riemannian-geometry machinery (constant curvature spaces, tangent bundles, parallel transport) that the product bundle and bundle convolution are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the volume-growth evidence that hyperbolic space aligns with hierarchical, tree-like structures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the product-of-model-spaces mixed-curvature formulation that RiemannGFM extends by adding tangent spaces."},{"cited_title":"Hamilton, Pietro Liò, Yoshua Bengio, and R","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-graph-infomax contrastive baseline whose Citeseer link-prediction score (96.90 AUC) is compared against RiemannGFM's 99.40."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cross-domain graph pretraining baseline that motivates and benchmarks the claimed transferability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LLM-based open graph foundation model baseline, which fails on non-attributed graphs and frames the text-free gap."}],"review_version":1}