{"id":"22e2b435-448e-4917-a26d-32cbbd15a3b3","arxiv_id":"2412.11085","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GraphMoRE learns per-node mixtures of Riemannian expert embedding spaces, reducing distortion on graphs that mix tree-like, cycle-like, and flat substructures.","lead":"GraphMoRE is a graph neural network that lets each node pick its own mix of curved embedding spaces (hyperbolic, spherical, Euclidean) using a routing network, instead of forcing the whole graph into one space. It reports higher accuracy on link prediction and node classification benchmarks and lower embedding distortion than existing Riemannian graph models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (10) is a pair-dependent weighted average of squared expert distances, not a geodesic distance on any constructed manifold; the central 'heterogeneous manifold' claim is asserted, not demonstrated.","rationale":"Good-faith reading: the paper's empirical contribution is substantial—consistent link-prediction gains over nine baselines, ablations, released code, and lower distortion—and I see no reason to doubt the reported numbers. The concern is specifically that the paper's headline theoretical framing, the construction of a heterogeneous manifold with pointwise varying curvature, rests entirely on Eqs. (8)-(10), and those equations do not define a Riemannian geometry. Eq. (8) concatenates scalar-rescaled embeddings from different manifolds; Eq. (10) reweights per-expert squared distances with pair-dependent softmax weights. A distance of this form is not shown to satisfy metric axioms, let alone to arise from a Riemannian metric; the paper switches between d and d^2 without defining which quantity is used in the Fermi-Dirac decoder and distortion loss. The reader's weakest-assumption analysis identified exactly this point; I agree. The fix is not to reject the paper: the empirical results are separable from the geometric rhetoric. But the condition for acceptance should be either (i) provide a formal construction of the heterogeneous manifold and show Eq. (10) is its geodesic distance, or (ii) remove or soften the manifold claims and present the method as a learned weighted combination of Riemannian expert distances. Since the reader already set CONDITIONAL on essentially this issue, my read leaves the verdict unchanged.","tokens_in":15552,"tokens_out":6524,"duration_ms":65558,"concrete_test":"Use the released implementation to train GraphMoRE on Cora. Freeze the learned expert embeddings and gating weights, compute d(u,v) = sqrt(Eq. (10)) for all pairs in a random sample of 10^4 node triples, and test the triangle inequality d(u,v) <= d(u,w) + d(v,w). If a material fraction (say more than 1%) of triples violate it, Eq. (10) cannot be a Riemannian geodesic distance on any single manifold, and the heterogeneous-manifold claim should be revised to 'learned pair-dependent weighted distance without a constructed manifold.' Additionally, fit fixed per-expert scales a_eps by least squares to the squared distances and report the residual; if the residual is large, the pair-dependent alignment is doing essential work, so Eqs. (9)-(10) cannot be interpreted as a fixed product metric.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central advertised contribution is that fusing expert outputs 'construct[s] personalized mixed curvature spaces' and embeds the graph 'into a heterogeneous manifold with varying curvatures at different points' (Sec. 4.3, Abstract). The only distance supplied for this space is Eq. (10), d^2(u,v) = Sum_eps W^eps_(u,v) d_eps^2(Z_u_eps, Z_v_eps), with W_(u,v) = Softmax(W_u . W_v) in Eq. (9). This is a pair-dependent convex combination of squared geodesic distances from the individual expert manifolds. No Riemannian metric, tangent-space identification, or geodesic is defined for a space containing the concatenated representations of Eq. (8), and none follows automatically: for fixed weights, sqrt(Sum w_eps d_eps^2) is a product-space metric, but here the alignment weights change with (u,v), so there is no single metric whose distance function equals Eq. (10) for all node pairs. Eq. (9) is also ambiguous: if the dot is the scalar product, softmax of a scalar cannot yield the per-expert vector required by Eq. (10); if it is elementwise multiplication, the resulting 'alignment' is a heuristic reweighting, not a map between tangent spaces. Thus the claim that the graph is embedded in a heterogeneous Riemannian manifold is asserted rather than derived. The strong link-prediction numbers may survive as evidence for a weighted-distance interpolation scheme, but the geometric grounding advertised in the title and abstract is not currently supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"GraphMoRE proposes a mixture-of-experts framework for graph representation learning in which each expert is a Riemannian GNN operating in a space of a different curvature, and a gating network assigns per-node weights based on local topological encodings. The per-node weighted fusion of expert embeddings is claimed to embed the graph into a heterogeneous manifold with point-wise varying curvature. To compute distances between nodes in this mixed space, the paper introduces an alignment strategy that combines per-expert squared distances with pair-dependent Softmax weights. The model is trained by combining a downstream-task loss with an embedding-distortion loss, and is evaluated on link prediction and node classification on five real-world datasets plus synthetic graphs, reporting gains over Euclidean and Riemannian baselines, along with ablations and a distortion comparison.","tokens_in":15898,"tokens_out":4728,"duration_ms":41521,"significance":"If the geometric construction were rigorous, the paper would address a real limitation of single-curvature and product-manifold models: topological heterogeneity within a graph. The empirical results are consistently strong: GraphMoRE improves link prediction over all nine baselines on all five real datasets, with ten-run averages and released code, and the ablations identify each component's contribution. However, the advertised novelty—the 'heterogeneous manifold with varying curvatures at different points'—is asserted rather than mathematically established. The actual distance used in training and evaluation, Eq. (10), is a pair-dependent weighted average of squared expert distances, which is not shown to be the geodesic distance of any Riemannian manifold that the construction defines. This weakens the paper's central claim, though the empirical contribution could survive as a weighted-distance mixture model if the narrative is revised accordingly.","major_comments":[{"comment":"Equation (9) as written is ill-defined: if '·' denotes the standard inner product, then W_u · W_v is a scalar, and Softmax of a scalar is identically 1, so W_(u,v) would be a scalar, not the per-expert vector required by Eq. (10). If '·' denotes elementwise multiplication, that must be stated explicitly and the alignment rationale must be justified; currently no definition is given for how the weight vector W_(u,v) is produced. This ambiguity makes the distance in Eq. (10) ill-posed and directly affects the loss in Eq. (7), the link-prediction decoder, and Table 5.","section":"Sec. 4.3, Eq. (9)"},{"comment":"The central claim that the construction embeds the graph 'into a heterogeneous manifold with varying curvatures at different points' is not substantiated. No Riemannian metric, tangent-space identification, geodesic, or curvature tensor is defined for the space of concatenated embeddings in Eq. (8). Eq. (10) defines a pair-dependent convex combination of squared distances from different expert manifolds; for fixed weights it is a product-space metric, but the weights change with the node pair, so no single Riemannian distance function is shown to equal Eq. (10). The paper needs either to provide a concrete manifold construction (e.g., a warped product or a manifold with a position-dependent metric) and prove Eq. (10) is its distance, or to revise the abstract and Section 4.3 to describe a weighted-distance mixture model rather than a heterogeneous manifold.","section":"Sec. 4.3 and Abstract"},{"comment":"The complexity analysis in Sec. 4.5, stated as O(|R||D| + |Ds| + K|D|), omits the pairwise loop explicitly shown in Algorithm 1 ('for each node pair (u, v) in G') and the all-pairs sum in Eq. (7). This pairwise computation is O(N^2) and dominates on large graphs, contradicting the stated linear complexity. The authors should either include the quadratic term in the analysis or describe how the distortion loss and alignment are computed (e.g., sampling pairs), since the current presentation does not support the scalability implied by the linear bound.","section":"Sec. 4.5 and Algorithm 1"},{"comment":"The embedding-distortion comparison in Table 5 is partly circular: Eq. (7) defines LD, which is minimized as part of the training objective in Eq. (11), and Table 5 reports the same measure as the evaluation metric. Since the baselines (Q-GCN, MofitRGC) do not optimize LD, the fact that GraphMoRE achieves lower average distortion is expected by construction and does not independently demonstrate superior structure preservation. To support the claim, the authors should report distortion for a GraphMoRE variant trained without the distortion loss (e.g., the 'w/o distortion' ablation), or use an independent measure such as stress or a shortest-path-distance prediction task.","section":"Sec. 5.2, Table 5 and Eq. (7)"}],"minor_comments":[{"comment":"There is a typo: 'etopological heterogeneity' should be 'topological heterogeneity'.","section":"Sec. 2.2"},{"comment":"The section heading 'Comlexity Analysis' is misspelled; it should be 'Complexity Analysis'.","section":"Sec. 4.5 title"},{"comment":"The caption contains 'Rimannian', which should be 'Riemannian'.","section":"Fig. 1 caption"},{"comment":"The text says 'Additionly' instead of 'Additionally'.","section":"Sec. 5.2"},{"comment":"The multi-resolution sampling scale set R is never specified; the appendix mentions 'induced subgraphs of ego networks with different radius' but does not report the actual radii used. This hyperparameter should be listed for reproducibility.","section":"Sec. 5.1 / Appendix B"},{"comment":"The description states GraphMoRE achieves the best results on the vast majority of datasets, but on PubMed the Q-GCN baseline has higher Weighted-F1 and Micro-F1 than all three GraphMoRE variants; the text should be more precise about this exception.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's strongest asset is its empirical evaluation: consistent gains across multiple datasets, ablations, and released code. The main risk is that the title, abstract, and Section 4.3 promise a 'heterogeneous manifold' that is never actually constructed; the distance in Eq. (10) is a heuristic weighted average. If the authors cannot supply a rigorous geometric derivation, the paper would need to be reframed as an empirical mixture-of-experts approach with weighted distances, which would reduce the novelty but still be a plausible contribution. The distortion evaluation in Table 5 is also methodologically circular as currently presented. The complexity analysis omission is easy to fix but should not be overlooked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"GraphMoRE is a genuinely new combination—MoE gating over Riemannian GNN experts with different curvatures—and the empirical work is solid: five real datasets, ten-run averages, ablations that isolate gating, diversity, alignment, and distortion guidance, plus synthetic graphs and released code. The link-prediction gains over MotifRGC are large and consistent, and the node-classification improvements, while smaller and backbone-dependent, are mostly there. This is a practical step toward handling topological heterogeneity, and the multi-resolution local-topology encoding as a routing signal is a nice idea.\n\nSoft spots are mostly about packaging. The central claim that the method embeds the graph into a 'heterogeneous manifold with varying curvatures' is asserted, not demonstrated. Eq. (10) defines a pair-dependent weighted average of squared expert distances; the weights change with (u,v), so there is no single Riemannian metric whose distance function equals this for all pairs. That does not kill the method—the downstream numbers stand as evidence for a weighted-distance interpolation scheme—but the geometric framing should be either constructed or substantially softened. Eq. (9) is also ambiguous: if Wu·Wv is a scalar product, Softmax of a scalar does not give the per-expert vector used in Eq. (10); I assume elementwise multiplication is intended, but it needs to be written out. The complexity analysis in Sec. 4.5 omits the O(N^2) pairwise term from Eq. (7) and the alignment loop; for a paper about foundation models, that is a real omission. Finally, Table 5 is partly self-fulfilling—the distortion objective is itself a training signal, so lower distortion is expected; the fact that downstream tasks improve independently is the more convincing evidence, and the paper should lean on that.\n\nNone of this is fatal. The empirical claims appear reproducible and the novelty is real. A serious referee should engage; the revision needs to clarify Eq. (9), either construct the mixed-space metric or reframe the contribution as an alignment/interpolation method, and give an honest complexity count.","headline":"Real novelty and consistent empirical wins, but the 'heterogeneous manifold' is a label, not a constructed geometry; fixable in revision.","tokens_in":16438,"tokens_out":2291,"would_cite":true,"duration_ms":19578,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By routing every node through a learned mixture of hyperbolic, Euclidean, and spherical spaces, GraphMoRE claims to embed graphs into a point-dependent curvature manifold and to beat all compared baselines on link prediction.","keywords":["graph representation learning","topological heterogeneity","Riemannian manifold","mixture of experts","κ-stereographic model","link prediction","node classification","graph neural networks"],"falsifier":"Compute the aligned distance from Eq. (10) for every triple in a small graph and check the metric axioms: if any triple violates the triangle inequality or has $d(u,u)\\neq 0$ for embedded nodes, the mixed space is not a Riemannian metric space and the heterogeneous-manifold claim fails. A second decisive test is to ablate Eq. (9) by fixing $W_{(u,v)}$ to the uniform weight vector while retraining; if AUC on Cora or Citeseer does not drop, the alignment strategy is not load-bearing for the reported gains.","tokens_in":15335,"feed_emoji":"📈","tokens_out":16616,"duration_ms":133130,"temperature":0.7,"pith_summary":"GraphMoRE is an attempt to solve the graph-embedding problem caused by topological heterogeneity by giving each node its own mixture of constant-curvature spaces instead of embedding the whole graph in one space. The authors build a Mixture-of-Experts model in which every expert is a graph neural network that operates in the κ-stereographic model with a fixed curvature, spanning hyperbolic, Euclidean, and spherical geometries. A topology-aware gating network estimates the local geometry around each node from multi-resolution sampled subgraphs and produces expert weights for that node; the node's embedding is the gated fusion of all expert outputs, and an alignment rule defines pairwise distances between nodes living in different mixtures. The model is trained jointly on the downstream task and a distortion-minimization term, and the paper reports that it beats nine baselines on link prediction on all five real datasets, wins most node-classification settings, and achieves the lowest average embedding distortion on every dataset. The broader interest is that a single heterogeneous geometry of this kind could let one graph foundation model handle many graph types uniformly.","feed_headline":"Per-node curvature routing beats all baselines on link prediction","feed_subtitle":"Each node chooses its own mixture of hyperbolic, Euclidean, and spherical experts, lowering distortion below every baseline.","key_machinery":"The central object is the personalized mixed-curvature embedding space, produced by a gated mixture of Riemannian experts and measured by an aligned weighted distance. Each expert is a graph neural network built on the κ-stereographic model, which is the named unifying object: one smooth manifold family $M^d_\\kappa$ that becomes the Poincaré ball for $\\kappa<0$, Euclidean space for $\\kappa=0$, and the stereographic sphere for $\\kappa>0$, so exponent and logarithm maps are available across all experts. The topology-aware gating mechanism, composed of multi-resolution local-topology encoding plus a distortion-guided gating network, carries the routing; the distortion loss $L_D = \\frac{1}{|V|^2} \\sum_{i,j} \\left| \\left(\\frac{d(i,j)}{g(i,j)}\\right)^2 - 1 \\right|$ steers each node toward the expert mixtures that preserve its local geometry. The alignment strategy in Eq. (9)-(10) is the mechanism that makes distances between nodes in different mixtures comparable.","core_discovery":"The central claim is that local topological heterogeneity is better served by personalized mixed-curvature spaces than by any globally uniform space, including a product manifold. GraphMoRE realizes this claim by treating each curvature sign and curving degree as an expert: the experts are Riemannian GNNs with fixed curvatures $\\{-3,-1,0,1,3\\}$, the gating network reads sampled subgraphs at multiple radii and outputs per-node weights, and the node embedding is the fusion $Z_v = \\mathrm{concat}_{\\varepsilon\\in\\mathcal{E}}(W^\\varepsilon_v \\otimes_\\varepsilon Z^\\varepsilon_v)$. To compare nodes in different mixtures, the paper proposes $W_{(u,v)} = \\mathrm{Softmax}(W_u \\cdot W_v)$ and $d^2(u,v) = \\sum_{\\varepsilon\\in\\mathcal{E}} W^\\varepsilon_{(u,v)} d^2_\\varepsilon(Z^\\varepsilon_u, Z^\\varepsilon_v)$. The authors argue that this construction embeds the graph into a heterogeneous manifold whose curvature varies from point to point, and they report that it outperforms all nine baselines on link prediction on Cora, Citeseer, Airport, PubMed, and Photo, with the lowest average embedding distortion on each.","pith_inferences":["Editorial inference: the method's geometric core is testable in a formal follow-up: one could define the tangent space and geodesic of the mixed space from the gated expert maps and verify that the claim of point-varying curvature is a genuine Riemannian structure rather than a weighted-distance heuristic.","Editorial inference: feeding explicit discrete curvature estimators, such as graph sectional curvature, into the gating network would make the routing interpretable and reveal whether the learned weights track geometric measures; this is not tested in the paper.","Editorial inference: zero-shot routing to unseen graphs would determine whether GraphMoRE's per-node curvature mixture generalizes across datasets, a property graph foundation models would need.","Editorial inference: the alignment rule Eq. (9) can be read as learned metric interpolation on the simplex; replacing it with a parameterized bilinear form over expert spaces might yield a smoother notion of node-to-node distance and extend the method to edge-weighted graphs."],"forward_implications":["If the claims hold, graph embeddings no longer need a globally fixed curvature: every node can receive its own mixture, and the model reports higher link prediction AUC than all nine baselines on Cora, Citeseer, Airport, PubMed, and Photo.","On Cora, Citeseer, and PubMed the reported AUC gains over the runner-up are 2.01, 1.94, and 1.92 percentage points, so the advantage is not marginal on citation networks.","The ablation study ties every major component to the gains: removing the distortion loss, gating, expert diversity, or alignment each lowers AUC and F1, with loss of expert diversity producing the largest drop (Cora AUC from 97.91 to 94.34).","Since GraphMoRE keeps its accuracy on larger synthetic graphs while baselines degrade, per-node routing appears to decouple performance from graph scale.","The same architecture wins most node-classification settings with GCN, GAT, and SAGE backbones, indicating the method is not tied to one classifier."],"supporting_citations":[{"why":"Supplies the κ-stereographic model and the exponential/logarithmic maps used by every Riemannian expert.","marker":"Bachmann, Bécigneul, and Ganea 2020"},{"why":"Provides the hyperbolic GCN architecture and the link-prediction protocol and Airport dataset used in the experiments.","marker":"Chami et al. 2019"},{"why":"Introduces mixed-curvature product-space embeddings and the sectional-curvature parallelogram law used to analyze topological heterogeneity.","marker":"Gu et al. 2019"},{"why":"Establishes the mixture-of-experts routing principle that the topology-aware gating mechanism adapts.","marker":"Jacobs et al. 1991"},{"why":"Supplies the sparsely-gated MoE formulation that the architecture follows for expert weighting.","marker":"Shazeer et al. 2017"},{"why":"Provides the Poincaré embedding setting and the Fermi-Dirac decoder used to evaluate link prediction.","marker":"Nickel and Kiela 2017"},{"why":"Defines the MotifRGC baseline whose results GraphMoRE must beat and whose embedding distortion is compared in Table 5.","marker":"Sun et al. 2024"},{"why":"Defines the Q-GCN pseudo-Riemannian baseline that also targets topological heterogeneity and is used for comparison.","marker":"Xiong et al. 2022"},{"why":"Supports the claim that Riemannian GNNs adjust curvature only in a limited range, motivating diverse fixed initial curvatures for the experts.","marker":"Fu et al. 2021"}],"fun_headline_variants":["GraphMoRE: per-node curvature mixture handles topological heterogeneity","Mixture of Riemannian experts reduces graph embedding distortion","Each node chooses its own curvature space for lower distortion","GraphMoRE: topology-aware gating picks optimal curvature per node","Personalized mixed-curvature spaces outperform uniform product manifolds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on treating the weighted combination of expert embeddings plus the alignment rule as a real geometric space; no metric, distance-triangle law, or coordinate geometry is ever proved for that space, so if those operations do not form a genuine geometry, the method's deeper claim reduces to a useful weighted-distance trick.","fun_headline_variants_meta":{"raw":{"variants":["GraphMoRE: per-node curvature mixture handles topological heterogeneity","Mixture of Riemannian experts reduces graph embedding distortion","Each node chooses its own curvature space for lower distortion","GraphMoRE: topology-aware gating picks optimal curvature per node","Personalized mixed-curvature spaces outperform uniform product manifolds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00041,"raw_usage":{"total_tokens":2174,"prompt_tokens":1040,"completion_tokens":1134,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1052}},"tokens_in":656,"tokens_out":1134,"duration_ms":9556,"temperature":1.0,"reasoning_tokens":1052,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:18:30.902415+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the aligned distance from Eq. (10) for every triple in a small graph and check the metric axioms: if any triple violates the triangle inequality or has $d(u,u)\\neq 0$ for embedded nodes, the mixed space is not a Riemannian metric space and the heterogeneous-manifold claim fails. A second decisive test is to ablate Eq. (9) by fixing $W_{(u,v)}$ to the uniform weight vector while retraining; if AUC on Cora or Citeseer does not drop, the alignment strategy is not load-bearing for the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the κ-stereographic model and the exponential/logarithmic maps used by every Riemannian expert."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the hyperbolic GCN architecture and the link-prediction protocol and Airport dataset used in the experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sparsely-gated MoE formulation that the architecture follows for expert weighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Poincaré embedding setting and the Fermi-Dirac decoder used to evaluate link prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the MotifRGC baseline whose results GraphMoRE must beat and whose embedding distortion is compared in Table 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Q-GCN pseudo-Riemannian baseline that also targets topological heterogeneity and is used for comparison."}],"review_version":1}