One for All: Universal Topological Primitive Transfer for Graph Structure Learning

Thanks for your positive comments on theoretically guarantee, and our first large-scale benchmark. For your concerns, below we make the responses.

Q1: Topological primitives are limited to two-hop neighborhoods, potentially missing higher-order structural patterns.

A1: Thanks for your comments. Although the topological primitives focus on two-hop neighborhoods, the GSN framework ensures effective learning of high-order graph information—both theoretically and empirically—without relying on excessive high-order motifs. The reason is three-folds:

• Based on first- and second-order motifs, we performed multi-layer SSM learning, which implemented weighted union driven by graph structure sequences. This process actually corresponds to the neighborhood-selective aggregation mechanism of GAT, thus achieving high-order topological pattern learning through low-order motifs.
• Excessive high-order motifs can lead to high computational complexity, and we need to balance the relationship between performance and complexity.
• The excellent performance on the LRGB benchmark also evaluate that the proposed method can effectively learn high-order graph topology (please refer to Table 1).

Q2: LLM-generated annotations require manual validation, introducing human oversight in dictionary construction.

A2: Thanks for your comments. We acknowledge the LLM-generated annotations require manual validation. Our design alleviates human oversight by efforts of these two aspects: 

• Manual Checking: The Topograph mapping dictionary generated by pretrained models were verified manually, specifically, per mapping (from motifs to text) was checked by at least 3 annotators. The mutual check between multiple annotators ensures consistency, thereby reducing the possibility of human oversight. 
• Efficient Workflow: Each annotator does not check the annotation information of all nodes, but only checks the annotations of the mapping dictionary, which includes 14 primitives * 10 descriptive variants. Thus, less verification content also reduces the possibility of human oversight.

Q3: How does the framework perform on graphs with dynamic topological changes over time?

A3: Thanks for your question. According to your concerned dynamic graph tasks, we would will to make classifications as follows:

• Firstly, our work mainly focuses on universal knowledge transfer tasks for public knowledge graphs (mostly static). while the temporal graph learning you mentioned are promising research fields, dynamic graph tasks are not the research focus.

• Especially, following your suggestion, it may be feasible to handle dynamic graph topology inputs by modifying the SSM backbone network of G²SN. Specifically, SSM requires merging a new temporal branch and an existing topological sequence branch, so that G²SN can capture the dynamic patterns of the topological structure. In this way, for the temporal motif sequence $\mathbf{m}\_t$and the temporal text sequence $\mathbf {u}\_t$, the G²SN learns dynamic node embeddings through SSM and selects important temporal topological relationships through a structure selection mechanism. Thereby obtain the similar conclusion through theoretical deduction in Appendix C.2 (please refer to the pages 14 lines 496-510).

Thanks for your question again. We will defer this design to the future work.

Q4: Can the transfer mechanism be extended to heterogeneous graphs with diverse node/edge types?

A4: Thank you for your question. We have successfully applied it to HGs. Specifically, knowledge transfer experiment on heterogeneous graph dataset ogbn-mag achieves +5.01% accuracy gain (please refer to Table 2, pages 8).

Q5: The computational overhead of constructing topological-text annotations increases with graph size, affecting scalability to extremely large datasets. The reliance on LLM annotations may introduce bias if textual descriptions do not fully capture structural nuances.

A5: Thanks for your comment. We will address your concerns from the following two aspects:

i) Scalability: The complexity of topological mapping is $\mathcal{O}(N\cdot d\^2)$ ($d$: average degree), which can extend to large-scale graph. And the dataset we are currently using has reached a scale of billion-edges (please refer to Table 4, Pages 12).

ii) Deviation mitigation: As the universal SOTA model, GPT-4 have been extensively validated for robustness and applied on various benchmark tasks. Further, we have taken specific mechanisms to reduce bias:

• Prompt engineering: Prompt templates could enforce structured outputs (e.g., "Node degree: [value]; Triangles: [count]") , enhancing noise resilience.
  
• Manual Checking: The Topograph mapping dictionary generated by pretrained models were further verified manually, specifically, per mapping (from motifs to text) was checked by at least 3 annotators. This human verification process can further reduce bias or noise.  By incorporating prompt design and manual verification, we can reduce bias from large foundation models. Following your suggestions, we make sure to add the above discussion in the next version.

Thanks for your positive comments on interesting problem domain and model novelty. For your concerns, below we make the responses.

Q1: The paper introduces State Space Models (SSM) without sufficiently explaining why SSM is necessary for graph structure learning. Traditional GNN models or Graph Transformers could potentially handle the graph structure transfer task without the complexity introduced by SSM. The lack of motivation for choosing SSM over more standard models makes the innovation feel less compelling.

A1: Thanks for your insightful comments. We choose SSMs due to considerations of performance and efficiency.

• Compared with Traditional GNN models or Graph Transformers, firstly, SSM has lower time complexity while possessing equivalent long-range learning capabilities, which is beneficial for large-scale tasks during pre training. Specifically, for input sequences of length $L$, G²SN processes sequences in $O(Ld\^2)$ time, while transformers and GNNs in $O(L^2d)$ time ($d$ represents the hidden layer dimension).

• Furthermore, our G²SN has a more flexible adaptive topology selection mechanism. We performed multi-layer SSM learning, which implemented weighted union driven by graph structure sequences. This process actually corresponds to the neighborhood-selective aggregation mechanism of GNN, thus achieving high-order topological pattern learning through low-order motifs. This promotes the mining and transfer of universa structure knowledge.

• Following your suggestion, we conducted a further experiment to validate the advantages of our G²SN model. We used transformer as the representation learning model and applied it to the frozen transfer scenarios on the ogbn-arxiv and Pubmed datasets. The topological primitive sequence and text sequence are concatenated together as transformer inputs. The experimental results are as follows: |Methods|ogbn-arxiv|Pubmed| |---|---|---| |G²SN Frozen|3.95↑|1.91↑| |Transformer Frozen|2.86↑|0.72↑|

The experimental results show that the transformer underperforms G²SN.

Q2: Despite the paper's claim of addressing domain generalization, it lacks any formal theory or analysis regarding generalization bounds or invariant representation in the context of graph learning. There is no mention of foundational work such as IRM (Invariant Risk Minimization), domain adaptation theory, or generalization bounds. This gap makes it difficult to assess the actual effectiveness of the proposed approach in handling domain shifts.

A2: Thanks for your comments. We acknowledge that our initial submission emphasized empirical validation, and we agree that incorporating foundations like IRM and generalization bounds would strengthen the work. Following your suggestions, we have derived the theoretical connection to domain adaptation/generalization:

Our AdaCross-Transfer mechanism implicitly aligns with IRM’s core principle of learning invariant representations across domains. Specifically:

• Invariant Representation Learning: The cross-attention module (Eqs. 10–15) dynamically aligns structural-textual features using a data-adaptive weight vector $\hat{\mathbf{w}}$, which reweights features based on topological primitive distributions. This forces the model to prioritize domain-agnostic structural semantics (motif distributions) over spurious domain-specific correlations, analogous to IRM’s invariance objective.

• Generalization Bounds: Theorem 5.1 (and Appendix C.3) proves that G²SN’s output $\mathbf{Y}(t)$ exhibits Lipschitz-continuous gradient flow under cross-domain shifts. This provides a theoretical foundation for stability during transfer, as bounded gradient sensitivity implies robustness to input perturbations (e.g., structural distribution shifts). Formally, the Lipschitz constant $L$ (Eq. 44) upper-bounds the feature distortion $\|\mathbf{Y}\_{\mathcal{D}\_S}(t) - \mathbf{Y}\_{\mathcal{D}\_T}(t)\|$ by the MMD distance between source ($\mathcal{D}\_S$) and target ($\mathcal{D}\_T$) domains (Eq. 45). This aligns with classical domain adaptation theory [Ben-David et al., 2010], where generalization error depends on domain discrepancy and feature stability.

Q3: The paper does not cite or compare the proposed method to any established domain generalization approaches in the context of graph learning, such as the Multi-Domain Graph Foundation Model. Without this context, it is unclear how this method advances or improves upon the current state-of-the-art in domain generalization for graph data [1][2].

A3: Thanks for your insightful comments. Following your suggestions, we have carefully investigated these works of your concern. We have found that although the two methods you mentioned have significant effects, they are inherently different from our transfer settings and application scenarios. Specifically:

• Different Transfer Settings: These two methods emphasize cross-domain transfer tasks under few-shot settings, which is different from our supervised transfer tasks. Therefore, directly comparing performance is unfair.

• Fusion of Node Features: From the methodological perspective, these two methods emphasize the deep fusion of structural information and node features, which is fundamentally different from our pure structure driven knowledge transfer. Our method can be applied to graph data without node features. Therefore, it should not be compared with these two methods together.

In summary, due to the inherently different transfer settings and application scenarios, these methods are not suitable as baselines for comparison.

Thanks for your positive comments on our framework's innovative concept, the first graph-text benchmark, theoretical convergence guarantees, and consistent performance improvements. For your concerns, below we make the responses.

Q1: Automatically generated textual descriptors may embed bias or noise, yet no human evaluation or further verification confirms their fidelity or usefulness for transfer.

A1: We sincerely thank you for this suggestion. As the universal SOTA model, GPT-4 have been extensively validated for robustness and applied on various benchmark tasks. Further, we have taken specific mechanisms to reduce bias or noise:

i) Prompt engineering: Prompt templates could enforce structured outputs (e.g., "Node degree: [value]; Triangles: [count]") , enhancing noise resilience.

ii) Manual Checking: The Topograph mapping dictionary generated by pretrained models were further verified manually, specifically, per mapping (from motifs to text) was checked by at least 3 annotators. This human verification process can further reduce bias or noise.

By incorporating prompt design and manual verification, we can reduce bias/noise from large foundation models. Following your suggestions, we make sure to add the above discussion in the next version.

Q2: It is unclear how the theoretical time/space complexity and the empirical runtime of G²SN-Transfer compare with standard GNN baselines such as GCN and GraphSAGE, especially when all components (motif extraction, dual-stream encoder, AdaCross) are included.

A2: Thanks for your comments. We hope to address your concerns through the following three points:

i) Empirical Runtime of Motif Extraction: The complexity of motif extraction is $O(|V| \cdot d\^k)$ ($d$=avg. degree, $k$=2-hops). For Cora (2,708 nodes) and UniHG-STA (77.3M nodes), this takes 7.07s and 214.22h (please refer to Table 5, Pages 12).

ii) AdaCross Overhead: The parameter count of AdaCross is 9% lower than the average downstream baseline models used (please refer to pages 2, lines 55-63).

iii) G²SN Inference: For an input sequence of length $L$, G²SN processes sequences in $O(Ld\^2)$ time, while GraphSAGE and GCN in $O(L\^2d$) time ($d$ represents the hidden layer dimension). Empirical runtime of pretraining is 5.06 hours on single 4090 (please refer to Appendix G.1).

In next version, we will emphasize above theoretical time/space complexity and the empirical runtime.

Q3: It is unclear how the number of primitives is selected, what exact algorithm is used to compute each graph’s motif-frequency vector, and how the computation scales (e.g., to k-node motifs) on large graphs.

A3: Thanks for your comments.

• Firstly, this choice is the result of balancing performance and complexity. Considering the balance between expressive ability and computational complexity, we determine the motif as within a 2-hop neighborhood. The spatiotemporal cost of motif statistics in the three hop neighborhood is too high. 

• Additionally, capturing long-range information does not require excessive higher-order motif design. The reason is three-folds: a) Based on first- and second-order motifs, we performed multi-layer SSM learning, which implemented weighted union driven by graph structure sequences. This process actually corresponds to the neighborhood-selective aggregation mechanism of GAT, thus achieving high-order topological pattern learning through low-order motifs. b) Excessive high-order motifs can lead to high computational complexity, and we need to balance the relationship between performance and complexity. c) The excellent performance on the LRGB benchmark also evaluate that the proposed method can effectively learn high-order graph topology (please refer to Table 1).

• Further, for the motif-frequency vector calculation problem that you concerned, primitives were selected via statistical recurrence analysis (Blanche et al., 2020) for each node's 1-hop and 2-hop subgraphs. To ensure numerical accuracy, we traverse all nodes on the large-scale graph. Complexity is $O(|V| \cdot \Delta\^2)$ for 2-hop primitives ($\Delta$=max degree), thus it can be directly used for large-scale graphs.

Q4: In Figure 3, which is the textual sequence $u$ and LLM-Topomotif sequence $m$?

A4: Thanks for your questions. The upper input is textual sequence $u$, and the lower input is LLM-Topomotif sequence $m$. We will clarify captions in the next version.

Q5: How might the learned topological-textual latent space enable inverse problems—for example, generating synthetic graphs that satisfy a natural-language specification of desired structural properties?

A5: Thanks for your insightful comments. This is a very interesting question but not the focus of this research. Based on the experience of text-graph generation, it may be feasible to solve related problems by establishing a generation algorithm and alignment benchmarks between graph motifs and text pairs. We will defer this design to the future work.

Q6: What theoretical or empirical considerations led to a state-space model over other more common graph-aware transformers or message-passing models? Have you evaluated G²SN against a parameter-matched Transformer encoder that is fed the same motif distribution and text embeddings?

A6: Thanks for your insightful comments. We choose SSMs due to considerations of performance and efficiency.

• Compared with graph-aware transformer or message-passing models, firstly, SSM has lower time complexity while possessing equivalent long-range learning capabilities, which is beneficial for large-scale tasks during pre training. Specifically, for input sequences of length $L$, G²SN processes sequences in $O(Ld\^2)$ time, while transformers and GNNs in $O(L^2d)$ time ($d$ represents the hidden layer dimension). 
• Furthermore, our G²SN has a more flexible adaptive topology selection mechanism. We performed multi-layer SSM learning, which implemented weighted union driven by graph structure sequences. This process actually corresponds to the neighborhood-selective aggregation mechanism of GNN, thus achieving high-order topological pattern learning through low-order motifs. This promotes the mining and transfer of universa structure knowledge. 
• Following your suggestion, we conducted a further experiment to validate the advantages of our G²SN model. We used transformer as the representation learning model and applied it to the frozen transfer scenarios on the ogbn-arxiv and Pubmed datasets. The topological primitive sequence and text sequence are concatenated together as transformer inputs. The experimental results are as follows: |Methods|ogbn-arxiv|Pubmed| |---|---|---| |G²SN Frozen|3.95↑|1.91↑| |Transformer Frozen|2.86↑|0.72↑|

The experimental results show that the transformer underperforms G²SN.

Q7: Motif distributions are treated as fixed-length vectors of size $M$. Did you test the sensitivity of the whole pipeline to $M$ (e.g., 16 vs. 64 primitives)?

A7: Thanks for your insightful comments. We have added a sensitivity experiment, considering the high time overhead of statistically analyzing more motifs on the pre training dataset UniKG (currently requiring 214.22 hours), we have used motif vectors of different lengths (10, 12, 14, 16), and we also attempted to add some new motif (17=16+1(a 3-star/4-star/4-clique motif)). The results of the sensitivity experiment are as follows:

The experimental results show that the model is robust to the length of motif vectors, and introducing new motif (e.g. 3-star) can improve performance.

Thanks for your positive comments on technical soundness, empirical rigor, consistent performance boosts, releases code and data. For your concerns, below we make the responses.

Q1 [Weaknesses1 & Questions1]: The fixed dictionary of seven primitives may underspecify higher-order motifs, limiting generality. How sensitive are the results to the chosen set of seven primitives? Would including non-cyclic motifs such as stars or cliques improve transfer?

A1: Thanks for your comments. Regarding your concerns about i) fixed length, ii) sensitivity, and iii) more topological primitives including non-cyclic motifs, we provide the following response:

• Fixed Length: Although the topological primitives vector has a fixed length after setting, the GSN framework ensures effective learning of high-order graph information—both theoretically and empirically—without relying on excessive high-order motifs. The reason is three-folds: a) Based on first- and second-order motifs, we performed multi-layer SSM learning, which implemented weighted union driven by graph structure sequences. This process actually corresponds to the neighborhood-selective aggregation mechanism of GAT, thus achieving high-order topological pattern learning through low-order motifs. b) Excessive high-order motifs can lead to high computational complexity, and we need to balance the relationship between performance and complexity. c) The excellent performance on the LRGB benchmark also evaluate that the proposed method can effectively learn high-order graph topology (please refer to Table 1).

• Sensitivity of Topological Primitives Vector Length：We select topological primitive vectors of different lengths to test sensitivity. Specifically, We compared the original performance against performances selecting 8, 10, and 12 length. The experimental results are as follows:

Experimental results show that increasing motif vector length improves model performance but introduces additional computational overhead. To balance performance and complexity, we selected 14 in our manuscript.

• More Topological Primitives: Following your suggestions, we conducted additional experiments incorporating 3-star, 4-star, and 4-clique structures as new topological primitives. The experimental results are as follows: |Methods|ogbn-arxiv|Pubmed| |---|---|---| |Original Frozen|3.95↑|1.91↑| |+3-star Frozen|4.11↑|2.30↑| |+4-star Frozen|3.92↑|2.16↑| |+4-clique Frozen|3.41↑|1.78↑|

The experimental results demonstrate that incorporating additional motifs does not consistently enhance performance. The 4-star and 4-clique structures may introduce redundant information.

Q2: Theoretical analysis assumes linear time-invariant dynamics; applicability to dynamic or weighted graphs remains unverified.

A2: Thanks for this insightful comment. While our current experiments focus on static graphs (the primary domain of STA-18), our framework does support dynamic or weighted graphs through the following mechanisms:

• Weighted Graphs: Our topological primitives explicitly encode edge-type diversity (Section 3), which can naturally treate edge weight as extra motif. Specifically, edge weights are used as new motifs to expand the text sequence $\mathbf{u}$ to $\mathbf{u}\_{ew}$ and the topological primitive sequence $\mathbf{m}$ to $\mathbf{m}\_{ew}$. Thereby obtain the similar conclusion through theoretical deduction (please refer to the pages 14 lines 496-510).

• Dynamic Graphs: Dynamic graph data can be naturelly handled through G²SN's SSM backbone. Specifically, Input the temporal motif sequence $\mathbf{m}\_t$and the temporal text sequence $\mathbf {u}\_t$. The G²SN learns dynamic node embeddings through SSM and selects important temporal topological relationships through a structure selection mechanism. Thereby obtain the similar conclusion through theoretical deduction (please refer to the pages 14 lines 496-510).

The above specific derivation will be included in the appendix of the next version.

Q3: The writing requires non-trivial improvement. There are too many rare words/complex sentences used in the paper, which does not align with the goal of academic writing.

A3: Thanks for your comments. In next version, we will simplify complex sentences and replace rare words to improve readability.

Q4: Representative GFM + LLM studies (related works) not discussed sufficiently.

A4: Thanks for your comments. Following your suggestions, we here discuss more GFM + LLM studies, including LangGFM, GFM-RAG and PromptGFM:

• LangGFM: Integrates formal grammars (e.g., Context-Free Grammars) with LLMs to enforce syntactic validity in generated text. Unlike standard LLMs that rely solely on statistical patterns, LangGFM uses grammar parsers as decoders to constrain outputs to linguistically valid structures. 
• GFM-RAG: Augments Retrieval-Augmented Generation (RAG) with grammatical knowledge graphs. By structuring retrieved content using grammar frameworks before feeding it to the LLM, GFM-RAG improves factual coherence and reduces hallucination. 
• PromptGFM: Embeds grammatical rules directly into LLM prompts via constrained decoding templates. This lightweight approach steers generation toward grammatically valid outputs without modifying the model architecture. In multilingual applications, PromptGFM dynamically switches grammar templates based on input language, significantly reducing structural errors in low-resource languages where LLMs typically struggle.

We will add the above discussion to related work in the next version.

Q5: LLM-generated descriptors can vary stylistically; what mechanisms ensure annotation consistency?

A5: We sincerely thank you for this suggestion. As the universal SOTA model, GPT-4 have been extensively validated for robustness and applied on various benchmark tasks. Further, we have taken specific mechanisms to enhance annotation consistency:

i) Prompt engineering: Prompt templates could enforce structured outputs (e.g., "Node degree: [value]; Triangles: [count]") , enhancing noise resilience.

ii) Manual Checking: The Topograph mapping dictionary generated by pretrained models were further verified manually, specifically, per mapping (from motifs to text) was checked by at least 3 annotators. This human verification process can further ensure annotation consistency.

By incorporating prompt design and manual verification, we can reduce the risk of vary stylistically from foundation models. Following your suggestions, we make sure to add the above discussion in the next version.

Q6: Have you observed negative transfer on heterogeneous datasets ?

A6: Thanks for your question. We have discussed the negative transfer of heterogeneous graph data (please refer to the Homogeneity Superiority section in Section 6.4, pages 9, lines 302-319). According to the experimental results, the heterogeneous graph ogbn-mag underperforms homogeneous graphs with similar scale by 2.3–4.8%, indicating that the heterogeneous structure introduces semantic fragmentation and hinders cross graph alignment. To mitigate negative transfer, careful parameter tuning and extended iteration steps are necessary when working with heterogeneous graphs.

Thanks for your positive comments on our theoretical rigor, detailed technical descriptions, broad impact potential, and innovative architecture. For your concerns, below we make the responses.

Q1: Limited Baseline Comparisons: The evaluation includes only GCN and GIN as baselines (Table 1, Page 8), omitting comparisons with contemporary graph neural network (GNN) architectures (e.g., GraphSAGE, GAT) or transfer learning methods (e.g., GraphCL, DANN). This restricts the ability to position GSN-Transfer’s performance relative to the state-of-the-art, potentially undermining claims of superiority (Page 3).

A1: Thanks for your comment. The compared baselines include seven graph learning methods, including three popular ones named GCN, GIN, GatedGCN, and four recent SOTA models GPS+Transformer, GPS+Performer, GPS+BigBird, and Exphormer, rather than only GCN and GIN. Our G²SN outperforms all these seven baselines, please refer to Table 1 (Page 8). In addition, GraphSAGE was reported in Table 2. Following your suggestion, we also used GAT, GraphCL, and DANN to handle structure migration tasks, but encountered out-of-memory issues.

Q2: Insufficient Ablation Analysis: Ablation studies (Table 3, Page 9) are conducted on only six datasets and focus solely on transfer methods, neglecting to dissect contributions of key components (e.g., TopoGraph Mapping, dual-stream architecture, motif selection). This limits the evidence for the necessity and optimality of each module.

 A2: Thanks for your comment. These six datasets used in ablation studies include five large-scale datasets, and cover three fields, including citation classification (node classification), molecular property prediction (link prediction), and protein property prediction (graph classification). This should be sufficient to evaluate the contributions of our designed modules in ablation study.

Following your suggestions, we have added ablation studies on motif selection and dual-stream architecture. Due to the rebuttal period limitation, these ablation studies are conducted on two datasets. The experiment results are as below:

i) Motif (Topological Primitive) Selection: We conduct the motif selection ablation experiments on the ogbn-arxiv and Pubmed datasets. Specifically, We compared the original performance against performances selecting 8, 10, and 12 motifs. The experimental results are as follows:

Experimental results show that increasing motif types improves model performance but introduces additional computational overhead. To balance performance and complexity, we selected 14 motifs in our manuscript.

ii) Dual-Stream Architecture: We independently ablated both topological primitive sequence branch and text sequence branch in the dual-stream architecture. The ablation experiment results are as follows:

The experimental results show that each branch contributes to model performance.

Q3: Incomplete Complexity Analysis: While computational complexity is briefly addressed (Page 18), the manuscript lacks quantitative metrics (e.g., time complexity for pre-training or inference). Given the framework’s reliance on large-scale pre-training and cross-attention, this omission raises concerns about scalability and practical deployment.

A3: Thanks for your comments. Following your suggestion, we provide the time complexity analysis of the pre-training. The time complexity of G²SN is mainly determined by the SSM backbone, which has $O(Ld^2)$ complexity ($L$=seq len, $d$=hidden dim). And it is lower than transformer’s $O(L^2)$. On UniKG-STA (77M nodes), pre-training takes 5.06 hours (using single RTX 4090, please refer to Appendix G).

Q4: Over-Reliance on Appendices: Critical details, such as G²SN block implementation (Page 17) and hyperparameter configurations (Appendix H, Page 23), are deferred to appendices. This disrupts the main text’s self-containment, potentially challenging readers seeking a cohesive narrative.

A4: Thanks for your comments. We will migrate critical details to the main text: G²SN implementation (Appendix C.1) and Hyperparameters (Appendix H) in the next version.

Q5: Narrow Task Focus: The evaluation primarily focuses on structural tasks (e.g., motif prediction) and standard graph tasks (node classification, link prediction) (Page 7). The manuscript does not explore more complex applications, such as graph generation or temporal graph analysis, limiting the perceived breadth of impact.

A5: Thanks for your comments. It should be emphasized that the proposed knowledge transfer framework has already modeled 13 downstream tasks, which exceeds other similar works [GraphControl WWW’24, G-Adapter AAAI’24].

While the temporal graph learning and graph generation you mentioned are promising research fields, they are not suitable in this work. Specifically, our data lacks temporal information, leading to a fundamental mismatch to temporal graph learning. Moreover, achieving graph generation requires redesigning the model architecture.

Q6: Incremental Relation to Prior Work: While the topological primitive concept is novel, the manuscript builds on existing ideas like state space models (SSMs) (Page 3) and graph motif analysis (Page 10, references to Choromanski et al., 2020). The distinction from prior motif-based methods (e.g., Graph2Vec, Chen & Koga, 2019) is not fully clarified, raising questions about incremental novelty.

A6: Thanks for acknowledging the novelty of our topological primitive. Regarding relation to prior works, we would like to clarify your concerns as follows:

• Firstly, a fundamental contribution of our work is construction and release of the 18 graph datasets (please refer to lines 72-75) to promote the graph structure transfer task, which differs from the previous tasks you mentioned. 
• Secondly, we summarize the differences between our method and the two ones you mentioned as follows:

i)Our G²SN vs. SSMs: We proposed the structural control mechanism to inject the graph structure knowledge through a dual-stream framework. In this framework, topological primitives can act as structural control flow, guiding attention to topology-critical regions (please refer to Theorem 4.2). In contrast, SSMs cannot explicitly and adequately model graph structure.

ii)Our G²SN vs. Graph2Vec [Chen & Koga, 2019]: Our G²SN algorithm incorporates sequence learning derived from ordinary differential equations (ODEs), providing theoretical guarantees for capturing contextual graph structures (see Appendix C.2, lines 496-510). In contrast, Graph2Vec's approach relies on dynamic neighborhood sampling during training rather than employing predefined motifs, which differs from our method in terms of mechanism.

Q7: Terms like “topological primitives” and “structural textures” (Pages 1–2) are introduced without intuitive explanations. Could the authors provide clearer definitions or examples in the introduction or Section 2, targeting readers unfamiliar with graph learning? Please revise the text to include these clarifications.

A7: Thank you for your suggestion. We will add intuitive definitions of these terms in the revised version. Specifically:

• Topological Primitives: representative local connectivity patterns (e.g., triangles, stars) and their distribution.

• Structural Textures: Transferable semantic units (analogous to CV textures) formed by primitive distributions.

We will provide a clear explanation of these terms in the next version.

Thanks for your positive feedback and confirmation that all comments have been addressed. Following your suggestion, we will further refine the paper’s overall clarity in the revised version. We appreciate your time and valuable comments.