{"id":"2cadbc9c-b668-4133-acf0-406ee3c93a3e","arxiv_id":"2507.07335","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"R-SGFormer pairs GraphMoRE embeddings with SGFormer, but its own tables show the full model fails to consistently beat SGFormer or GraphMoRE baselines.","lead":"This paper combines GraphMoRE's Riemannian manifold embeddings with the SGFormer graph transformer and tests Stiefel/Grassmann projections. The headline claim of a 3 percent accuracy lift is contradicted by the paper's own tables on several datasets, and the full model often trails its own backbone.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that R-SGFormer consistently surpasses both SGFormer and GraphMoRE is unverifiable and internally contradicted, because Table 1 omits SGFormer and shows R-SGFormer below GraphMoRE adapters on CITESEER.","rationale":"The reader's weakest assumption correctly identifies the reliance on literature baselines as the key vulnerability. My stress test goes further: the paper's own Table 1 contradicts the central claim on CITESEER, where R-SGFormer's weighted F1 (66.26) is below several GraphMoRE adapters (e.g., GraphMoRE-GCN at 69.73). Additionally, SGFormer is absent from Table 1 and the combined R-SGFormer is absent from Table 2, so the claimed superiority over SGFormer cannot be checked on any consistent metric. The Section 6 disclosure that baselines are 'best numbers available' from prior literature means cross-table comparisons conflate different splits and metrics. Together, these issues place the strongest claim—consistent, reproducible gains over both components on all four benchmarks—beyond what the evidence supports. The paper's Stiefel and Grassmann variants show plausible small gains in Table 2, but those are not the headline claim. I agree with the reader's REJECT verdict: the central empirical claim is internally contradicted and not verifiable as presented. A unified rerun could potentially rehabilitate the idea, but that would be a conditional contribution, not the current submission. No ad hominem is intended; the critique is on the comparability and consistency of the reported numbers.","tokens_in":9915,"tokens_out":3165,"duration_ms":32621,"concrete_test":"Re-run a unified node-classification protocol with identical data splits, evaluation metrics, and hyperparameter ranges for SGFormer, GraphMoRE-GCN/GAT/SAGE, and R-SGFormer on CORA, CITESEER, AIRPORT, and PUBMED, and include SGFormer in the Table 1 F1 table. If R-SGFormer's CITESEER weighted F1 does not exceed 69.73 (GraphMoRE-GCN) and does not beat SGFormer under the same metric, the 'consistently surpasses both' claim is false; this single check settles whether the central claim survives.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim, 'The combined model consistently surpasses both SGFormer and the original GraphMoRE adapters' (Section 6), is not supported by the paper's own tables. Table 1 reports weighted/macro F1 for GraphMoRE-based baselines and R-SGFORMER but omits SGFormer entirely; Table 2 reports accuracy and includes SGFormer but omits the combined R-SGFormer. On CITESEER, Table 1 shows R-SGFORMER at 66.26 weighted F1, below GraphMoRE-GCN (69.73), GraphMoRE-GAT (68.59), GraphMoRE-SAGE (69.40), Q-GCN (68.89), and MOTIFRGC (68.31). Thus the claimed consistent superiority over GraphMoRE adapters is contradicted by the authors' own numbers on one of the four headline benchmarks, and the claimed superiority over SGFormer cannot be checked because the two methods appear in different tables with different metrics. Section 6 also states that baselines are 'best numbers available' from prior literature, so cross-table comparisons mix different data splits and evaluation protocols. This is a correctness risk at the core of the contribution, not a stylistic issue. Missing hyperparameters (e.g., K in Algorithm 2) and the absence of code further prevent verification, but the internal contradiction alone is decisive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes R-SGFormer, a graph transformer that augments the SGFormer backbone with Riemannian manifold structure. Two variants, R-SGFormer(S) and R-SGFormer(G), project query/key matrices onto Stiefel or Grassmann manifolds and add an orthogonality regularizer; a third variant, called R-SGFormer, prepends GraphMoRE's mixture-of-Riemannian-experts embeddings to SGFormer via a cross-attention block. The authors claim consistent accuracy improvements of up to 3% over both SGFormer and GraphMoRE adapters on four node-classification benchmarks, and they report additional results on seven datasets for the projection variants.","tokens_in":10158,"tokens_out":4096,"duration_ms":47310,"significance":"If the empirical claims were reproducible, combining node-level mixed-curvature embeddings with a scalable linear-attention graph transformer would be a useful engineering contribution, and the Stiefel/Grassmann ablation is a sensible study of geometric inductive biases. However, the paper provides no code, no machine-checked proofs, and no parameter-free derivations; its value rests entirely on the experimental section. The central comparative claim is currently not supported by the paper's own tables, so the significance cannot be assessed as stated.","major_comments":[{"comment":"The sentence \"The combined model consistently surpasses both SGFormer and the original GraphMoRE adapters\" is contradicted by Table 1 on CITESEER: R-SGFORMER(Ours) obtains 66.26 weighted F1, which is below GraphMoRE-GCN (69.73), GraphMoRE-GAT (68.59), and GraphMoRE-SAGE (69.40). The text acknowledges the lower value but calls it \"competitive\"; this does not support the claim of consistent superiority.","section":"Section 6 / Table 1"},{"comment":"The headline accuracy lift cannot be verified from the tables because the relevant methods are split across tables with different metrics. Table 1 reports weighted/macro F1 for R-SGFormer and GraphMoRE baselines but omits SGFormer entirely, while Table 2 reports accuracy for SGFormer and the projection variants but omits the combined R-SGFormer. Consequently, the abstract's claim of \"up to 3%\" improvement over SGFormer is not directly checkable from any single table.","section":"Section 6 / Tables 1 and 2"},{"comment":"The evaluation relies on \"best numbers available\" from the SGFormer and GraphMoRE literature rather than re-running baselines under identical data splits, metrics, and training protocols. The tables mix weighted F1 and accuracy, and the dataset sets overlap only partially (Table 1 has CORA, CITESEER, AIRPORT, PUBMED; Table 2 adds ACTOR, SQUIRREL, CHAMELEON, DEEZER). Without a unified protocol or released code, the relative gains over the baselines are not established.","section":"Section 6 / first paragraph"},{"comment":"The paper states that \"node–feature matrices in standard benchmarks naturally cluster near orthogonal subspaces\" and that replacing X by the orthogonal factor from QR leaves accuracy \"virtually unchanged,\" but no supporting experiment or table is provided. Additionally, Algorithm 2 depends on the number of experts K and curvature set C, yet the experiments do not report K, the gating network architecture, or the expert regularization details, making the experiments irreproducible.","section":"Section 5.1 / Algorithm 2 / Section 6 hyperparameters"}],"minor_comments":[{"comment":"The text contains an unresolved cross-reference \"Sec. ??\" when discussing Riemannian Adam; this should be fixed.","section":"Section 4.4"},{"comment":"There is a typo: \"emprical\" should be \"empirical.\"","section":"Section 5.2"},{"comment":"The table header contains spacing artifacts such as \"SGF ORMER\" and \"R-SGF ORMER\"; these should be corrected to \"SGFormer\" and \"R-SGFormer.\"","section":"Table 2"},{"comment":"Algorithm 2 refers to a \"cross-attention block,\" while Section 5.3 describes the same component as the \"linear-attention layer of SGFormer\"; the terminology should be reconciled.","section":"Section 5.3 / Algorithm 2"},{"comment":"Some citations are malformed, e.g., \"(Kipf & Welling, 2017(@)\", and the reference list contains entries such as Hassani and Khasahmadi that do not appear to be cited in the text; the bibliography should be cleaned and cross-checked.","section":"References"}],"recommendation":"reject","confidential_remarks":"In my view, the manuscript is a straightforward combination of two published models, and the evaluational setup is not internally consistent. The CITESEER result in Table 1 directly contradicts the paper's central claim, and the absence of a unified comparison makes the claimed gain unverifiable. I would not invite a revision at this venue; the comparison would need to be redone from scratch with a single protocol and all relevant baselines rerun."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2507.07335. The paper combines GraphMoRE's Riemannian mixture-of-experts embeddings with the SGFormer backbone, and also tries two cheaper variants that project attention queries and keys onto Stiefel or Grassmann manifolds with an orthogonality regularizer. The specific combination is new, and the Stiefel/Grassmann comparison in Table 2 is the most useful piece: those projections give a consistent, modest lift over plain SGFormer across seven datasets (roughly 0.5–2.2 points). That part is worth a footnote.\n\nThe problem is the central claim. The abstract says the combined model \"lifts accuracy by up to 3%\" and Section 6 says it \"consistently surpasses both SGFormer and the original GraphMoRE adapters.\" Table 1 directly contradicts that on CITESEER, where R-SGFORMER scores 66.26 weighted F1 against 69.73 for GraphMoRE-GCN. The text even admits the CITESEER number is lower, then claims consistency anyway. That is an internal contradiction, not a stylistic issue.\n\nThe evaluation design makes it worse. Table 1 reports F1 for GraphMoRE baselines and R-SGFORMER but omits SGFormer entirely; Table 2 reports accuracy for SGFormer and the Stiefel/Grassmann variants but not the combined R-SGFormer. So the \"up to 3%\" lift over SGFormer cannot be checked in any single table. Baselines are borrowed from the literature rather than re-run, so cross-table comparisons mix different splits and metrics. There is no code or data, and key hyperparameters such as the number of experts K are missing. The manuscript also has unfinished references and typos, which adds noise.\n\nSo: the projection idea is plausible, but the headline result is unsupported. If the authors redo the experiments under a uniform protocol, include SGFormer and the combined model in the same table, and scale back the claims, this could become a modest empirical contribution. As it stands, I wouldn't cite it, and I wouldn't send it to a serious referee. It is a good example of why you should sanity-check your own tables against your prose, but that is about the only lesson here.","headline":"The Stiefel/Grassmann projection study is a reasonable lightweight idea, but the headline R-SGFormer claim is contradicted by the paper's own Table 1 and unverifiable against SGFormer.","tokens_in":10723,"tokens_out":4558,"would_cite":false,"duration_ms":47015,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Riemannian mixture-of-experts front end, routing each node to the curved space that best fits its local topology, lifts graph-transformer accuracy by up to 3 percent on four node-classification benchmarks.","keywords":["graph transformers","Riemannian manifolds","mixture of experts","node classification","curvature-adaptive embeddings","Stiefel manifold","Grassmann manifold","linear attention"],"falsifier":"Re-run SGFormer, GraphMoRE-adapted GCN/GAT/SAGE, and R-SGFormer under identical splits, metrics, and hyperparameter sweeps on CORA, CITESEER, AIRPORT, and PUBMED; if the up-to-3% gap over both components shrinks to within one standard deviation, the central lift claim is not supported.","tokens_in":9643,"feed_emoji":"🪐","tokens_out":6707,"duration_ms":63465,"temperature":0.7,"pith_summary":"This paper tries to establish that graph transformers lose information by embedding every node in a single flat Euclidean space, and that a lightweight manifold-routing front end can fix it. The proposed R-SGFormer prepends a mixture-of-experts layer from GraphMoRE that inspects each node's local topology and assigns it to the Riemannian expert—spherical, flat, or hyperbolic—whose curvature fits best. The curved embeddings are fused with raw features through a cross-attention block before the SGFormer backbone runs its linear-attention and GNN branches. The paper reports up to 3 percentage points of accuracy gain over strong baselines on four node-classification benchmarks, and argues that the explicit geometric projections also make the latent space more interpretable.","feed_headline":"Geometry-aware router lifts graph transformers by up to 3%","feed_subtitle":"Sending each node to its best-fitting curved space sharpens node classification on four benchmarks.","key_machinery":"The load-bearing object is a Riemannian mixture-of-experts layer (from GraphMoRE) used as a front end: local topological subgraphs are sampled, a gating network assigns each node to one of several constant-curvature experts with curvatures in {-3, -1, 0, 1, 3}, and an alignment module normalises pairwise distances between experts so the resulting node embedding is geometrically consistent. A cross-attention block fuses these curved embeddings with raw node features before SGFormer's linear-attention module and GNN branch combine local and global signals. The Stiefel and Grassmann variants instead project query-key matrices via QR or SVD and add an orthogonality penalty to the final representation, testing whether orthogonal-complement feature sets carry the same benefit.","core_discovery":"The central claim is that curvature-adaptive node embeddings and a scalable transformer backbone are complementary, and that combining them yields a model—R-SGFormer—that beats both components on their own. In the paper's experiments, R-SGFormer records 82.44/80.66 weighted/macro F1 on CORA, 93.53/93.18 on AIRPORT, and 81.01/80.21 on PUBMED, surpassing both SGFormer and GraphMoRE-adapted GCN, GAT, and SAGE models, while remaining competitive on the sparse CITESEER benchmark. The paper also claims that the simpler Stiefel and Grassmann projection variants with an orthogonality regularizer improve on the plain SGFormer backbone on every dataset tested, which supports the general thesis that geometric bias helps graph transformers even without full curvature routing.","pith_inferences":["A natural next test is whether the same Riemannian front end transfers to link prediction or graph-level tasks, since the paper only evaluates node classification.","The paper's tables mix F1 and accuracy metrics, so a reader cannot directly compare models across Table 1 and Table 2; running every method under one protocol would clarify whether the gains are metric-specific.","If the routing idea is what matters, it could be dropped into any attention-based graph model, not just SGFormer, wherever local topology is heterogeneous.","The conclusion's suggestion of a fixed library of Riemannian node embeddings, analogous to pretrained word vectors, points toward treating geometry as a reusable input representation rather than a per-model choice."],"forward_implications":["Substituting or concatenating GraphMoRE's curvature-adaptive embeddings with raw features produces a consistent uplift over the vanilla SGFormer backbone.","The combined R-SGFormer surpasses both SGFormer and GraphMoRE adapters for GCN, GAT, and SAGE on CORA, AIRPORT, and PUBMED, while remaining competitive on CITESEER.","Stiefel and Grassmann projections plus an orthogonality regularizer improve on plain SGFormer on every dataset in Table 2, so geometric bias helps even without full curvature routing.","Because the transformer branch keeps linear attention, the gains come without giving up SGFormer's near-linear scalability."],"supporting_citations":[{"why":"Supplies GraphMoRE, the mixture-of-Riemannian-experts front end whose curvature-adaptive embeddings form the core of R-SGFormer.","marker":"Guo et al., 2025"},{"why":"Supplies SGFormer, the scalable linear-attention-plus-GNN backbone that R-SGFormer extends and compares against.","marker":"Wu et al., 2023b"},{"why":"Provides the GCN baseline reported in both comparison tables.","marker":"Kipf & Welling, 2017"},{"why":"Provides the GAT baseline used in the comparisons.","marker":"Veliˇckovi´c et al., 2018"},{"why":"Provides the HGCN hyperbolic baseline listed in Table 1.","marker":"Chami et al., 2019"},{"why":"Supplies the Riemannian Adam optimizer used for GraphMoRE's curved experts.","marker":"B´ecigneul & Ganea, 2019"},{"why":"Supplies the Adam optimizer used to train the overall model.","marker":"Kingma & Ba, 2015"},{"why":"Provides the NodeFormer scalable graph transformer baseline in Table 2.","marker":"Wu et al., 2022"}],"fun_headline_variants":["Curved-space router boosts graph transformers by 3%","Riemannian routing sharpens graph transformer accuracy","Adaptive curvature embeddings lift graph transformer F1","Mixture-of-manifolds layer boosts graph transformer accuracy","Each node picks its best space for sharper graph learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed lifts rest on 'best numbers available' for strong baselines from the SGFormer and GraphMoRE literature rather than re-running every method under identical data splits and metrics, so the gains are meaningful only if those published baselines were measured under the same protocol.","fun_headline_variants_meta":{"raw":{"variants":["Curved-space router boosts graph transformers by 3%","Riemannian routing sharpens graph transformer accuracy","Adaptive curvature embeddings lift graph transformer F1","Mixture-of-manifolds layer boosts graph transformer accuracy","Each node picks its best space for sharper graph learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000979,"raw_usage":{"total_tokens":4092,"prompt_tokens":813,"completion_tokens":3279,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":3203}},"tokens_in":429,"tokens_out":3279,"duration_ms":26238,"temperature":1.0,"reasoning_tokens":3203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:43:46.635487+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run SGFormer, GraphMoRE-adapted GCN/GAT/SAGE, and R-SGFormer under identical splits, metrics, and hyperparameter sweeps on CORA, CITESEER, AIRPORT, and PUBMED; if the up-to-3% gap over both components shrinks to within one standard deviation, the central lift claim is not supported.","supporting_citations":[{"cited_title":"Graphmore: Mitigating topological heterogeneity via mixture of riemannian experts","cited_arxiv_id":null,"evidence_quote":"Supplies GraphMoRE, the mixture-of-Riemannian-experts front end whose curvature-adaptive embeddings form the core of R-SGFormer."},{"cited_title":"Hyperbolic graph convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the HGCN hyperbolic baseline listed in Table 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Adam optimizer used to train the overall model."}],"review_version":1}