acceptodds
Under review as a conference paper at ICLR 2027

R-MESD: Riemannian Multi Expert Learning for Structural-Semantic Decoupling in Text-Attributed Graphs

Abstract

Representation learning on text-attributed graphs (TAGs) is crucial for modeling rich node semantics and complex graph structure. Nevertheless, this task faces structural-semantic mismatch arising from divergent modality distributions, as well as heterogeneous local topologies that a single geometric space cannot adequately capture. Existing approaches typically propagate semantic features along the given topology or rely on a fixed representation space, potentially amplifying modality interference and overlooking local geometric diversity. To address these issues, we propose R-MESD, a Riemannian Multi Expert Structural-Semantic Decoupling model in TAGs. Specifically, R-MESD initializes structural positions from Laplacian spectral signals and uses local structural descriptions to route nodes to Hyperbolic and Spherical experts. The expert representations from different manifolds are transformed into a shared latent space within each geometric branch and subsequently fused according to node-level routing weights, followed by cross-geometry attention.Meanwhile, node semantics are represented as tangent vectors anchored at the learned structural positions and further refined through cross-geometry semantic alignment and local tangent-space attention. Extensive experiments on five public benchmarks demonstrate that R-MESD consistently outperforms state-of-the-art baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Text-Attributed Graphs, Graph Representation Learning, Graph Neural Network

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ctioOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on the fact that existing text-attributed graph (TAG) methods suffer from two issues: (1) structural-semantic mismatch arising from divergent modality distributions, and (2) heterogeneous local topologies that a single geometric space cannot adequately capture. To address these issues, this paper proposes Riemannian Multi-Expert Structural-Semantic Decoupling (R-MESD) to separate structural and semantic representations. The structural branch employs adaptive Hyperbolic and Spherical experts, with an LLM-based structural description for calculating node-level routing weights. The semantic branch encodes textual information as tangent-space vectors anchored at learned structural positions, and further applies cross-geometry semantic alignment and tangent-space semantic attention. Experiments on five datasets show improvements over existing GNNs and TAG baselines.

Critical strengths
  1. The paper tackles a practical problem where structural-semantic mismatch arises from divergent modality distributions in text-attributed graph learning.
  2. The proposed framework is technically sound and includes meaningful experimental analysis, such as the performance comparison and the module ablation study.
Critical weaknesses

None

Questions for authors
  1. Figure 1 and Table 4 discuss the model performance w.r.t. heterophily intervals, this makes me doubt whether advanced heterogeneous graph models can also show good performance. So the authors are supposed to provide comparisons with advanced heterogeneous graph models.
  2. The training objective in Eq. (32) doesn’t include a classification loss or a loss using the labels. Is R-MESE an unsupervised model? If so, it should be compared with more unsupervised methods, not only supervised methods.
  3. The ablation should provide the standard deviations.
  4. Why do the authors only use two geometric branches in Table 3, but not three branches with H, S, and E?
  5. The function of structural positions needs more explanation. Why is it useful to be concatenated with the semantic representations in Eq. (28)?
  6. In Section 4.4, the authors state that R-MESD yields clearer class separation, indicating more discriminative node representations. However, I just find little difference between (b) and (c).
  7. I wonder whether the promising model performance is attributed to the powerful LLM. Do the authors use the same LLM as the baseline models or a more powerful one?
3 comments

Q1: Following your suggestion, we compare our method with two advanced heterophilic graph models, H2GCN and GPR-GNN. As shown in Table 1, our method outperforms both models on all five datasets, confirming its advantage over methods designed for heterophilic graphs.

Table 1. Comparison with Heterophily-Oriented Methods on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
H2GCN89.85 ± 0.4275.26 ± 0.7388.85 ± 0.2861.72 ± 1.1667.83 ± 0.59
GPR-GNN89.21 ± 0.3675.94 ± 0.9187.92 ± 0.4760.78 ± 0.8468.32 ± 0.63
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

 

Q2: R-MESD uses unsupervised encoder pretraining followed by supervised downstream classification. The pretraining objective in Eq. (32) uses no node labels. After pretraining, we freeze the encoder and train a downstream GNN classifier using the training labels.

Following your suggestion, we include DGI, GraphMAE2, and BGRL as additional unsupervised methods. As shown in Table 2, our method achieves higher accuracy on all five datasets than these unsupervised methods.

Table 2. Comparison with Unsupervised Methods on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
DGI86.07 ± 1.2072.26 ± 0.3883.90 ± 0.2364.50 ± 0.8765.66 ± 0.21
GraphMAE284.32 ± 0.5564.66 ± 1.2082.87 ± 0.1164.21 ± 0.0664.56 ± 0.08
BGRL87.55 ± 0.0973.91 ± 0.2384.97 ± 0.4965.94 ± 0.1669.13 ± 0.13
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

 

Q3: Following your suggestion, we report the mean accuracy and standard deviation for the ablation experiments.

Table 3. Ablation Results on Accuracy (%).

VariantsCoraPubMedArXiv-2023
w/o Lstr\mathcal L_{\mathrm{str}}87.16 ± 0.8990.76 ± 0.9074.65 ± 0.82
w/o Lsem\mathcal L_{\mathrm{sem}}86.53 ± 1.4086.43 ± 1.5173.74 ± 0.89
w/o Cross-Geometry Structural Attention85.73 ± 0.8290.91 ± 1.0275.87 ± 0.93
w/o Cross-Geometry Semantic Alignment88.59 ± 1.3887.74 ± 0.7374.32 ± 1.01
Uniform Routing87.13 ± 0.8289.97 ± 0.8074.81 ± 0.84
ALT Replacement86.94 ± 0.7889.14 ± 0.8874.86 ± 0.70
R-MESD92.37 ± 0.3192.59 ± 0.6277.98 ± 1.62

 

Q4: Our R-MESD omits a separate Euclidean branch because hyperbolic and spherical experts with curvature near zero can approximate Euclidean geometry. Experts with different curvatures thus capture both curved and nearly Euclidean structures. Table 3 in our paper further supports this choice: H/S outperforms configurations with a Euclidean branch.

 

Q5: Concatenating structural positions in Eq. (28) gives the classifier direct access to topology, helping distinguish nodes with similar texts but different structures. It also preserves structural features as separate inputs, allowing the classifier to learn their importance for the task.

 

Q6: There seems to be a misunderstanding about the comparison. The clearer class separation refers to the improvement of both (b) and (c) over the initial representations in (a), rather than a large difference between (b) and (c).

Q7: Our gains do not rely solely on the LLM. Qwen3-8B has fewer parameters than the baselines’ LLMs. To examine its role, we compare two routing variants without LLM descriptions. Structure-only routing uses structural statistics, while Direct-feature routing also uses text embeddings. As shown in Table 4, both perform well, showing that our model remains effective without description generation.

Table 4. Comparison of Routing Inputs on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
Structure-only routing87.76 ± 0.7480.81 ± 1.3688.06 ± 0.5773.09 ± 1.9272.22 ± 1.08
Direct-feature routing89.33 ± 0.6379.81 ± 2.1991.06 ± 0.4074.59 ± 1.8374.22 ± 1.35
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

Thanks for the authors, all my issues have been addressed. I will maintain my score.

Reviewer_3dAwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes R-MESD, a Riemannian multi-expert framework for text-attributed graphs. It employs Hyperbolic and Spherical experts with adaptive routing to capture heterogeneous graph structures, while representing node semantics as tangent vectors. Cross-geometry attention and semantic alignment are introduced to integrate structural and semantic information. Experiments on five benchmarks report improvements over existing baselines.

Critical strengths
  1. R-MESD achieves competitive results across five benchmarks, and the ablation studies provide evidence for the effectiveness of its components.

  2. Addressing structural-semantic mismatch and heterogeneous local topology is relevant to text-attributed graph representation learning.

Critical weaknesses

W1. Limited methodological novelty and theoretical contribution.

My primary concern is the limited novelty of the proposed framework. Mixed-curvature representation learning, geometric mixture-of-experts, adaptive expert routing, and cross-geometry interactions have been extensively explored in prior work [1–5].

In particular, CRGFM [1] already incorporates geometric mixture-of-experts with node-level routing, Augmented Lorentz Transformation (ALT), manifold-based structural representations, tangent-space attribute representations, cross-geometry attention, and parallel transport. RiemannGFM [2] also employs manifold–tangent-space representations and cross-geometry interactions. Furthermore, GeoMoE [3] explicitly studies curvature-guided adaptive routing over geometric experts.

Compared with these approaches, R-MESD appears to primarily combine established techniques within a text-attributed graph setting, without introducing a sufficiently distinct methodological contribution.

The theoretical contributions are also limited. The two theorems mainly establish Lipschitz continuity of the initialization and target-manifold preservation under ALT. These are relatively standard mathematical properties and do not provide substantial new insights into geometric expert routing, structural-semantic decoupling, or the expressive advantages of the proposed architecture.

W2. Insufficient evidence supporting the claimed contributions and experimental reliability.

Although the experimental results are promising, they do not convincingly establish that the improvements originate from the proposed structural-semantic decoupling or adaptive geometric modeling.

First, the routing mechanism incorporates node and neighborhood textual information, meaning that structural representations remain influenced by semantics. The paper does not clearly distinguish architectural separation from genuine structural-semantic disentanglement. Moreover, the downstream GNN classifier introduces additional neighborhood aggregation, further complicating this claim.

Second, the method relies on Qwen3-8B to generate local structural descriptions, but the contribution of this additional semantic information is not sufficiently isolated. Direct comparisons with closely related Riemannian frameworks and simpler Euclidean alternatives under matched settings are also missing.

Third, the experimental protocol raises reproducibility and fairness concerns. Some baseline results exhibit unusually large performance differences, and OGBN-ArXiv uses random splits instead of its official evaluation protocol.

Finally, certain mathematical and implementation details remain unclear. The cross-geometry attention is formulated over all graph nodes, whereas the implementation adopts neighborhood sampling. Additionally, the well-definedness conditions of ALT, particularly for Spherical target manifolds, are not clearly guaranteed during optimization.

These issues make it difficult to attribute the reported improvements to the claimed methodological contributions.

References

[1] Sun, L., & Yu, P. S. (2026). A Riemannian Perspective on Graph Foundation Models: Curvature as a Guiding Principle. Frontiers of Computer Science. (CRGFM)

[2] Sun, L., et al. (2025). RiemannGFM: Learning a Graph Foundation Model from Riemannian Geometry. Proceedings of The Web Conference 2025.

[3] Cao, H., et al. (2026). Geometric Mixture-of-Experts with Curvature-Guided Adaptive Routing for Graph Representation Learning. arXiv preprint. (GeoMoE)

[4] Sun, L., et al. (2022). A Self-Supervised Mixed-Curvature Graph Neural Network. AAAI. (SelfMGNN)

[5] Bachmann, G., et al. (2020). Constant Curvature Graph Convolutional Networks. ICML.

Questions for authors
  1. Methodological novelty: What are the fundamental differences between R-MESD and existing frameworks such as CRGFM [1], RiemannGFM [2], and GeoMoE [3]? In particular, which components constitute genuinely new methodological contributions beyond adapting existing techniques to text-attributed graphs?

  2. Theoretical contribution: Can the authors provide stronger theoretical justification for adaptive curvature routing and structural-semantic decoupling? How do the current theorems establish properties beyond those already known for manifold initialization and geometric transformations?

  3. Decoupling and experimental attribution: How is structural-semantic decoupling ensured when textual information influences expert routing and a downstream GNN performs neighborhood aggregation? Can the authors provide controlled comparisons with simpler dual-branch Euclidean models, direct structural-statistics-based routing without LLM descriptions, and closely related mixed-curvature approaches?

  4. Experimental fairness: Were all baselines evaluated using equivalent textual features, pretrained encoders, and computational resources? How do the authors explain the unusually low results of several baselines, and why was the official OGBN-ArXiv split not used?

  5. Implementation and mathematical validity: Is cross-geometry attention implemented globally or over sampled neighborhoods? How is its computational complexity managed? Additionally, how are the domain constraints required by ALT guaranteed during optimization, especially for Spherical manifolds? Can the authors provide runtime and memory comparisons?

References

[1] Sun, L., & Yu, P. S. (2026). A Riemannian Perspective on Graph Foundation Models: Curvature as a Guiding Principle. Frontiers of Computer Science. (CRGFM)

[2] Sun, L., et al. (2025). RiemannGFM: Learning a Graph Foundation Model from Riemannian Geometry. Proceedings of The Web Conference 2025.

[3] Cao, H., et al. (2026). Geometric Mixture-of-Experts with Curvature-Guided Adaptive Routing for Graph Representation Learning. arXiv preprint. (GeoMoE)

[4] Sun, L., et al. (2022). A Self-Supervised Mixed-Curvature Graph Neural Network. AAAI. (SelfMGNN)

[5] Bachmann, G., et al. (2020). Constant Curvature Graph Convolutional Networks. ICML.

3 comments

W1&Q1: Thanks! Our R-MESD and the other methods (i.e., CRGFM, RiemannGFM, and GeoMoE) follow common geometric modeling principles, but there are two main differences.

First, our motivations are totally different. CRGFM and RiemannGFM use Riemannian geometry to build graph foundation models, while GeoMoE uses geometric experts to capture diverse graph structures. Our R-MESD addresses structural-semantic mismatch: graph connections do not always reflect semantic relations between node texts. Our goal is to combine structural and semantic information to learn more discriminative node representations.

Second, the methodologies are different. CRGFM standardizes expert representations into a unified product space. RiemannGFM learns from tree and cycle substructures and aligns their representations, while GeoMoE combines expert outputs in a shared tangent space. Our R-MESD couples adaptive geometric modeling with separate structural and semantic processing. Local descriptions guide geometric expert selection, while each branch retains structural positions and semantic tangent vectors. Cross-geometry structural attention integrates complementary topology from hyperbolic and spherical branches without merging their representations into a single space. On the semantic side, cross-geometry fusion is followed by neighborhood aggregation under the guidance of the target node's own semantics. This allows R-MESD to adapt to local topology and weigh neighboring content by semantic relevance, helping address structural-semantic mismatch.

 

W1&Q2: Following your suggestion, we provide two additional theoretical analyses.

First, we provide theoretical support for adaptive curvature routing by analyzing how changes in local evidence affect expert weights. The routing weights are defined as

wi=softmax⁡(MLP⁡route(ci)).w_i=\operatorname{softmax}(\operatorname{MLP}_{\mathrm{route}}(c_i)).

Here, cᵢ and cᵢ′ encode two local descriptions of the same node, with corresponding weights wᵢ and wᵢ′. If the routing MLP has Lipschitz constant Lᵣ from the Euclidean norm on its input to the maximum norm on its output, then

∥wi′−wi∥1≤2Lr∥ci′−ci∥2.\lVert w_i'-w_i\rVert_1\le 2L_r\lVert c_i'-c_i\rVert_2.

This result shows that small changes in local descriptions lead to controlled changes in expert weights. Our router adapts to local evidence without abrupt changes in expert weights. In addition, the full model outperforms Uniform Routing in Table 2 of our paper, supporting the effectiveness of adaptive routing.

Second, we theoretically support structural-semantic decoupling by analyzing how semantic changes affect aggregation with fixed structural positions. The structural branch derives node positions from graph topology, while the semantic branch encodes node texts as semantic vectors. During aggregation, our R-MESD uses structural positions to guide the parallel transport of semantic vectors, which determine message content and relevance. Text may influence structural expert selection but does not enter their message passing as a node attribute. Thus, decoupling separates structural and semantic encoding while allowing their interaction.

Suppose neighbor vjv_j's semantic vector changes from ujHu_j^H to ujH+δjHu_j^H+\delta_j^H, where δjH\delta_j^H represents the change. We recompute the attention weights and aggregated output as βiH′\beta_i^{H\prime} and ziH′z_i^{H\prime}, respectively. Assume the value map UVHU_V^H has operator norm at most LVL_V, and ∥ujH∥pjH≤M\lVert u_j^H\rVert_{p_j^H}\le M. Since parallel transport preserves Riemannian norms, we obtain

∥ziH′−ziH∥piH≤LV∑j∈N(vi)βijH′∥δjH∥pjH+LVM∥βiH′−βiH∥1.\lVert z_i^{H\prime}-z_i^H\rVert_{p_i^H}\le L_V\sum_{j\in\mathcal{N}(v_i)}\beta_{ij}^{H\prime}\lVert\delta_j^H\rVert_{p_j^H}+L_VM\lVert\beta_i^{H\prime}-\beta_i^H\rVert_1.

The two norm terms measure changes in neighbor semantics and attention weights, respectively. Low attention limits a neighbor’s direct contribution, and the same bound applies to the spherical branch. By separating structure and semantics, we can bound the effect of neighboring semantic changes with fixed structural positions, supporting the stability of our design.

 

W2&Q3: Thanks! First, textual information influences expert routing, but does not enter structural message passing as node attributes. Structural experts process graph spectral signals, while the semantic branch encodes node texts as tangent vectors. For downstream GNN neighborhood aggregation, we freeze the encoder and train the GNN on its outputs. The GNN combines these representations for prediction without updating either encoder. Thus, our structural-semantic decoupling means separate encoding, with interaction during routing and prediction.

Second, we compare with a simpler dual-branch Euclidean model, represented by the E/E configuration in our ablation study. As shown in Table 1, H/S achieves higher accuracy on all three datasets. These results show that complementary curvatures also contribute to the improvement.

Table 1. Comparison of Geometric Configurations on Accuracy (%).

Geometric configurationCiteSeerPubMedArXiv-2023
E/E73.20 ± 0.3587.04 ± 2.5674.82 ± 0.89
R-MESD (H/S)81.65 ± 1.1792.59 ± 0.6277.98 ± 1.62

Third, we evaluate direct structural-statistics-based routing without LLM descriptions through two variants: Structure-only routing uses structural statistics directly, while Direct-feature routing combines them with text embeddings. As shown in Table 2, both perform well, but the full model achieves higher accuracy on all five datasets.

Table 2. Comparison of Routing Inputs on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
Structure-only routing87.76 ± 0.7480.81 ± 1.3688.06 ± 0.5773.09 ± 1.9272.22 ± 1.08
Direct-feature routing89.33 ± 0.6379.81 ± 2.1991.06 ± 0.4074.59 ± 1.8374.22 ± 1.35
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

Finally, we compare with RiemannGFM as a closely related mixed-curvature approach. As shown in Table 3, our method achieves higher accuracy on all five datasets.

Table 3. Comparison with RiemannGFM on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
RiemannGFM88.12 ± 0.3274.95 ± 0.7788.09 ± 0.3559.35 ± 1.4662.43 ± 2.12
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

 

W2&Q4: Thanks! First, we evaluate the baselines and R-MESD using equivalent textual features, pretrained text encoders, and computational resources.

Second, dataset characteristics may affect baseline performance: PubMed has an average semantic similarity of 0.356 between neighboring texts, while the Arxiv datasets contain 40 classes. These characteristics highlight the challenges of class distinction and neighborhood aggregation under structural-semantic mismatch.

Finally, we also report results on the official OGBN-ArXiv split. As shown in Table 4, R-MESD achieves higher accuracy than the strongest baseline in this comparison.

Table 4. Comparison on the Official OGBN-ArXiv Split on Accuracy (%).

MethodsOGBN-ArXiv
GCN71.31 ± 0.11
GAT70.68 ± 0.05
LLM4NG72.26 ± 0.10
AD-GCN73.98 ± 0.09
R-MESD75.39 ± 1.07

 

W2&Q5: Thanks! First, cross-geometry structural attention is computed within sampled two-hop subgraphs, not over the full graph (Eqs. 18–21). Thus, the node set VBV_B in Eqs. (20)–(21) refers to the sampled subgraph.

Second, attention has a computational complexity of O(nB2d)O(n_B^2d) and requires O(nB2)O(n_B^2) space for attention weights, where nBn_B is the number of sampled nodes and dd is the representation dimension. Each batch contains 32 target nodes, with at most 20 and 10 neighbors sampled per node at the first and second hops, respectively (Appendix F.3).

Next, the domain constraints required by ALT follow from Eq. (13) and Appendix C.2. The identity 1−κ∥xs∥22=∣κ∣xt21-\kappa\lVert x_s\rVert_2^2=\lvert\kappa\rvert x_t^2 requires xt≠0x_t\neq 0 for Eq. (13) to be defined. For spherical targets, ∥Wκxs∥22≤1κˉ\lVert W_\kappa x_s\rVert_2^2\le\frac{1}{\bar{\kappa}} keeps the square root real, while for hyperbolic targets, the numerator is always positive. The projection in Eq. (14) also requires 1+∣κˉ∣ yt≠01+\sqrt{\lvert\bar{\kappa}\rvert}\,y_t\neq 0. Under these conditions, Theorem 2 ensures that the temporal correction keeps the transformed point on the target manifold.

Finally, Table 5 compares training time and memory usage with GCN and GAT on OGBN-ArXiv. The one-time offline generation of Qwen3-8B descriptions (~2.4 days) is excluded from training time, as the descriptions are reused across experiments. Our method achieves higher accuracy with similar GPU memory usage during training.

Table 5. Comparison of Accuracy, Training Time, and Memory Usage on OGBN-ArXiv.

MethodAcc (%)Training time (s)GPU allocated (GiB)GPU reserved (GiB)CPU RSS (GiB)
GCN70.76 ± 0.165821.583.201.84
GAT68.63 ± 0.116091.772.781.86
R-MESD76.11 ± 1.4521021.603.222.33
Reviewer_6BJJOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes R-MESD for learning on text-attributed graphs with structural-semantic mismatch. It represents graph structure using hyperbolic and spherical experts, routes nodes using LLM-generated local descriptions, and encodes textual semantics in tangent spaces anchored at structural positions. Cross-geometry interactions and parallel transport combine these representations. Experiments report improved node classification accuracy on five citation networks.

Critical strengths
  1. R-MESD achieves the highest mean accuracy on all five evaluated datasets. Results across five random seeds provide useful evidence for the consistency of these gains.
  2. Comparisons against Euclidean and single-geometry variants support the benefit of combining hyperbolic and spherical representations within the proposed architecture.
  3. The motivation analysis and heterophily-stratified evaluation connect the proposed approach to observed model behavior. Strong improvements on highly heterophilic nodes provide evidence relevant to the structural-semantic mismatch the paper aims to address.
Critical weaknesses
  1. CRGFM (Sun & Yu, 2026) already combines geometric experts, ALT, manifold-based structural representations, tangent-space attribute representations, and attention through parallel transport. Although cited, it is not evaluated, and the methodological differences remain insufficiently explained.
  2. The paper does not provide sufficient empirical evidence for the necessity of Qwen3-8B, such as comparisons with routing directly from the same structural statistics and text embeddings. It therefore remains unclear whether using an LLM to generate descriptions is necessary.
  3. All evaluated datasets are citation networks, which raises concerns about the method’s generality. These datasets may share similar text characteristics and edge semantics, whereas the relationship between connectivity and textual meaning could differ substantially in domains such as product and social networks.
  4. The pipeline requires per-node Qwen3-8B description generation, six geometric experts, and encoder pretraining, yet no end-to-end runtime or memory measurements are reported. It remains unclear whether the accuracy gains justify the additional computational cost over simpler alternatives.
Questions for authors
  1. What are the substantive methodological differences from CRGFM, and how does R-MESD compare under matched input and training conditions?
  2. How does routing from Qwen3-8B descriptions compare with routing directly from the same structural statistics and text embeddings?
  3. Do you have evidence that the gains extend beyond citation networks, where the relationship between graph connectivity and textual semantics may differ?
  4. What are the preprocessing time, training time, and peak memory usage on OGBN-ArXiv, including description generation, and how do these compare with simpler baselines?
1 comment

W1&Q1: Thanks! Our R-MESD and CRGFM follow common geometric modeling principles, but there are two main differences. First, our motivations are different. CRGFM focuses on structural diversity and complexity to build a graph foundation model for different graphs and tasks. R-MESD addresses structural-semantic mismatch: graph connections do not always reflect semantic relations between node texts. Our goal is to combine structural and semantic information to learn more discriminative node representations.

Second, the methodologies are different. CRGFM standardizes expert representations into a unified product space and uses manifold prompt learning for downstream adaptation. Our R-MESD couples adaptive geometric modeling with separate structural and semantic processing. Local descriptions guide geometric expert selection, while each branch retains structural positions and semantic tangent vectors. Cross-geometry structural attention integrates complementary topology from hyperbolic and spherical branches without merging their representations into a single space. On the semantic side, cross-geometry fusion is followed by neighborhood aggregation under the guidance of the target node's own semantics. This allows R-MESD to adapt to local topology and weigh neighboring content by semantic relevance, helping address structural-semantic mismatch.

Since CRGFM's code is not publicly available, we compare with the closely related RiemannGFM. As shown in Table 1, R-MESD achieves higher accuracy on all five datasets.

Table 1. Comparison with RiemannGFM on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
RiemannGFM88.12 ± 0.3274.95 ± 0.7788.09 ± 0.3559.35 ± 1.4662.43 ± 2.12
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

 

W2&Q2: Following your suggestion, we compare routing from Qwen3-8B descriptions with Direct-feature routing using the same structural statistics and text embeddings. As shown in Table 2, Direct-feature routing performs well without LLM descriptions, while the full model achieves higher accuracy on all five datasets.

Table 2. Comparison of Routing Inputs on Accuracy (%).

MethodsCoraCiteSeerPubMedArXiv-2023OGBN-ArXiv
Direct-feature routing89.33 ± 0.6379.81 ± 2.1991.06 ± 0.4074.59 ± 1.8374.22 ± 1.35
R-MESD92.37 ± 0.3181.65 ± 1.1792.59 ± 0.6277.98 ± 1.6276.11 ± 1.45

 

W3&Q3: Thanks! We evaluate our method beyond citation networks on Computers and Products, where node texts describe products and edges represent co-purchase relationships. As shown in Table 3, our method outperforms the strongest baselines on both datasets, supporting its effectiveness in this different domain.

Table 3. Comparison on Product Networks on Accuracy (%).

MethodsComputersProducts
AD-GCN80.48 ± 2.3785.38 ± 2.81
GAT74.05 ± 1.6480.91 ± 0.46
AdaSpec71.49 ± 0.6281.18 ± 0.29
SLGCN80.15 ± 0.2585.98 ± 0.07
R-MESD81.67 ± 1.2989.70 ± 1.83

 

W4&Q4: Following your suggestion, we report training time and memory usage on OGBN-ArXiv, with GCN and GAT as lightweight baselines. The one-time offline generation of Qwen3-8B descriptions (~2.4 days) is excluded from training time. We reuse the generated descriptions in subsequent experiments. As shown in Table 4, our method achieves higher accuracy while keeping training-stage GPU memory usage close to that of the lightweight baselines.

Table 4. Comparison of Accuracy, Training Time, and Memory Usage on OGBN-ArXiv.

MethodAcc (%)Training time (s)GPU allocated (GiB)GPU reserved (GiB)CPU RSS (GiB)
GCN70.76 ± 0.165821.583.201.84
GAT68.63 ± 0.116091.772.781.86
R-MESD76.11 ± 1.4521021.603.222.33

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.