acceptodds
Under review as a conference paper at ICLR 2027

Bridging the Gap to Unseen Graphs: Target-Side Adaptation for Frozen Graph Foundation Models

Abstract

Graph foundation models (GFMs) are emerging as a key paradigm for cross-domain graph learning. However, most GFMs typically face two limitations during the adaptation phase: (i) Firstly, the adaptation method of GFMs is closely coupled with the internal architecture of the model, making it difficult for users to adapt published models to new target graphs in black-box scenarios. Furthermore, (ii) most existing methods rely on a limited number of labels to optimize parameters during the adaptation phase. When the target graph differs from the pre-training domain, limited supervision may not effectively guide adaptation, potentially leading to negative transfer. To address these limitations, this paper proposes a target-side adaptation framework for frozen GFMs to improve the practicality and effectiveness of GFMs. Specifically, we design a feature adapter and a structural adapter at the model’s input to enable the frozen GFM to adapt to the features and structure of the target data, allowing users to perform adaptation without needing to understand the model’s internal design. Subsequently, we employ a maximal coding rate reduction loss to exploit unlabeled target nodes and promote compact but diverse adapted features, while using a spectrum alignment loss to adjust the target graph such that its information spectrum matches the encoding preference of the frozen GFM. Additionally, we theoretically analyzed that our adaptation method achieves a lower generalization error bound. Comprehensive experiments demonstrate the effectiveness and generalization capability of this method on various downstream tasks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Learning, Graph Foundation Models, Test-time Adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Ret3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an input-side adaptation framework for frozen graph foundation models named TAGFM. This framework employs a feature adapter that is trained with the goal of minimizing the maximum encoding rate, as well as a structural adapter guided by the spectral alignment between the target graph and the estimated response of the frozen model. Experiments were conducted under the setting of few samples, covering five GFM backbones and seven target datasets. The average performance improvement is encouraging, but currently, it is impossible to verify its claimed black-box applicability, the explanatory power of spectral responses as an effective encoding preference, or the claimed generalization advantage.

Critical strengths
  1. Adapting an open-source, frozen GFM to an unseen graph with few labels is practically important.
  2. This paper evaluates broad backbones. Results on five GFMs and 50 random splits provide a useful starting point for assessing portability. In Table 1, TAGFM improves average one-shot accuracy over each backbone’s original adaptation method by 7.21–11.00 percentage points.
  3. The feature and structure adapters can be evaluated separately, and the paper includes initial component ablations.
Critical weaknesses
  1. The paper describes TAGFM as a black-box adaptation method, but optimizing the feature adapter with the supervised objective in Eq. (14) appears to require gradients through the frozen GFM with respect to its input. Freezing the backbone parameters prevents updates to those parameters; it does not eliminate the need for this gradient. If “black-box” means access to model predictions only, Algorithm 1 is incomplete without a specified zeroth-order or surrogate optimization procedure. If input gradients are available, the method is feasible, but the claim should be framed more precisely as architecture-agnostic adaptation of a frozen, differentiable backbone. The authors should state which operations the GFM interface exposes and report query and runtime costs under the claimed access setting.

  2. The spectral objective lacks a validated link to predictive utility. Equations (7)–(8) treat the magnitude of the model’s response to a frequency-localized perturbation as its preference for that frequency. A large response need not imply that the frequency helps classification. Moreover, structure optimization measures spectral energy using the original features X, whereas final prediction uses adapted features \tilde X. The paper does not show that alignment persists after feature adaptation.

  3. Theorem 2.1 provides a conditional upper bound on query risk. Its central link between spectral alignment and prediction risk is assumed in Assumption 2, while Assumption 3 assumes that the black-box spectral estimate is close to an unknown ideal preference. These assumptions are not justified or empirically assessed. Moreover, minimizing the spectral-alignment loss reduces only one term in the bound; the support risk and other terms may also change. The paper provides no comparable bound for competing adaptation methods, so it cannot conclude that TAGFM has a tighter upper bound. The MCR loss does not appear as an improving term in the theorem, leaving its claimed theoretical contribution to generalization unsupported. The authors should qualify the claim, examine the assumptions, and provide a matched comparison if they wish to claim a tighter bound.

  4. Figures 4–5 remove entire adapters but do not test MCR against an otherwise identical feature adapter, or model-guided spectral alignment against random or simpler structural objectives. Comparisons with input-side graph prompting and transformation methods also need sharper treatment. GPF and GTrans are cited but not directly evaluated; IGAP’s treatment of graph signal and structure gaps from a spectral perspective warrants explicit positioning.

  5. The paper does not discuss DAPrompt (AAAI 2026), which jointly adapts graph structure and semantics for few-shot learning, or Graph Sparse Prompting (TPAMI 2026), which provides input-side feature prompting for pretrained GNNs. These works are relevant to TAGFM’s claims about joint feature–structure adaptation and portability across backbones. Although their training settings are not identical, the paper should explain the differences explicitly and compare against them under matched backbones and label budgets where feasible. Without this positioning, it is difficult to determine which gains arise from TAGFM’s proposed objectives rather than from input-side adaptation more generally.

[1] DAPrompt: Dual Alignment Prompt of Structure and Semantics for Few-shot Graph Learning. AAAI 2026.
[2] Reliable and Compact Graph Fine-Tuning via Graph Sparse Prompting. TPAMI 2026.

Questions for authors

See weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Ret3 (Part1)

We thank the reviewer for the careful reading of our paper and constructive comments in detail.

W1. The paper describes TAGFM as a black-box adaptation method, but optimizing the feature adapter with the supervised objective in Eq. (14) appears to require gradients through the frozen GFM with respect to its input. Freezing the backbone parameters prevents updates to those parameters; it does not eliminate the need for this gradient. If “black-box” means access to model predictions only, Algorithm 1 is incomplete without a specified zeroth-order or surrogate optimization procedure. If input gradients are available, the method is feasible, but the claim should be framed more precisely as architecture-agnostic adaptation of a frozen, differentiable backbone. The authors should state which operations the GFM interface exposes and report query and runtime costs under the claimed access setting.

We thank the reviewer for the constructive comments. We acknowledge that the term “black-box” was imprecise and have replaced it throughout the revised manuscript with “architecture-agnostic adaptation of frozen GFMs.” Specifically, TAGFM requires access to the model’s forward output embeddings and, during feature adaptation, input gradients computed through the frozen backbone to update the feature adapter. The backbone parameters remain unchanged throughout. We have explicitly clarified this access requirement (Appendix A).

Regarding computational overhead, we provide a time complexity analysis in Appendix D.4 and compare TAGFM with the native adaptation strategies in terms of runtime and peak GPU memory in Appendix D.5. On Pubmed (19,717 nodes), the largest graph by node count in our evaluation, TAGFM incurs costs comparable to those of the native strategies, with lower costs in some settings. For BRIDGE, TAGFM and native adaptation take 10.25 s and 10.22 s, respectively, with peak GPU memory of 301 MB and 274 MB. For SAMGPT, TAGFM reduces runtime from 11.87 s to 9.75 s and peak GPU memory from 446 MB to 311 MB. For MDGFM, runtime and peak GPU memory decrease by approximately 75% and 71%, respectively. These results demonstrate that TAGFM achieves a favorable balance between adaptation performance and resource efficiency at the evaluated scales.

W2. The spectral objective lacks a validated link to predictive utility. Equations (7)–(8) treat the magnitude of the model’s response to a frequency-localized perturbation as its preference for that frequency. A large response need not imply that the frequency helps classification. Moreover, structure optimization measures spectral energy using the original features X, whereas final prediction uses adapted features \tilde X. The paper does not show that alignment persists after feature adaptation.

To examine the connection between spectral responses and predictive utility, we separately masked the four highest- and lowest-response frequency bands and compared the resulting classification performance, as reported in Table 6 in Appendix D.5. Masking the highest-response bands yielded lower accuracy than masking the lowest-response bands across all tested combinations, providing empirical evidence linking high-response bands to predictive utility. Furthermore, using the original features XX for spectral alignment and the adapted features X~\tilde X for prediction has a coherent rationale. The band-masking experiment indicates that high-response bands estimated from the original features are associated with predictive utility, providing a task-relevant basis for graph structure optimization. Feature adaptation further enhances discriminability using the same nodes and their class information. Thus, the alignment target derived from XX has a task-relevant basis for supporting final classification. Both components serve the shared objective of effectively encoding class-discriminative information.

AuthorsOpenReview ↗

Response to Reviewer Ret3 (Part2)

W3. Theorem 2.1 provides a conditional upper bound on query risk. Its central link between spectral alignment and prediction risk is assumed in Assumption 2, while Assumption 3 assumes that the black-box spectral estimate is close to an unknown ideal preference. These assumptions are not justified or empirically assessed. Moreover, minimizing the spectral-alignment loss reduces only one term in the bound; the support risk and other terms may also change. The paper provides no comparable bound for competing adaptation methods, so it cannot conclude that TAGFM has a tighter upper bound. The MCR loss does not appear as an improving term in the theorem, leaving its claimed theoretical contribution to generalization unsupported. The authors should qualify the claim, examine the assumptions, and provide a matched comparison if they wish to claim a tighter bound.

We clarify that we do not claim that TAGFM has a universally lower bound than competing methods. Instead, our theorem decomposes the query-risk bound into the feature-adapter complexity, support risk, and structural transfer terms, showing that the proposed adaptation losses can reduce their corresponding terms under the stated assumptions. To address the assumptions, we have added their rationale and theoretical roles in the appendix. For Assumption 2, we explain that graph structure, node features, and labels may have an underlying task-dependent relationship, which induces a connection between frequency components and label information. For Assumption 3, we clarify the consistency between the implemented spectral-alignment loss and the spectral quantities used in the theorem. We further provide a mismatched-frequency supervision experiment, in which the probe-guided preference is compared with deliberately mismatched spectral supervision, to support the plausibility of Assumptions 2 and 3. Regarding the regularization losses, they do not invalidate the theorem because the resulting embeddings and adapter parameters remain within the admissible hypothesis class. In addition, our parameter-sensitivity experiments show that varying the regularization weights has only a limited effect on the final results.

W4. Figures 4–5 remove entire adapters but do not test MCR against an otherwise identical feature adapter, or model-guided spectral alignment against random or simpler structural objectives. Comparisons with input-side graph prompting and transformation methods also need sharper treatment. GPF and GTrans are cited but not directly evaluated; IGAP’s treatment of graph signal and structure gaps from a spectral perspective warrants explicit positioning.

To isolate the contribution of MCR, we conducted additional controlled experiments that removed only MCR while keeping the feature adapter and all other settings unchanged, as reported in Appendix D.7. Adding MCR improves average accuracy across the seven datasets by 5.10, 5.49, 6.40, 6.84, and 9.48 percentage points for BRIDGE, MDGFM, SAMGPT, GRAVER, and GCOPE, respectively. These results demonstrate that MCR provides additional gains with the same adapter, supporting the effectiveness of using target-node statistics to regularize feature adaptation.

We have also added direct comparisons with GPF and GTrans under the 1-shot setting across five frozen GFM backbones, as reported in Table 5. TAGFM achieves higher average accuracy than both methods on every backbone, with gains of 9.53–13.83 percentage points over GPF and 8.15–14.01 percentage points over GTrans, averaged across the seven target datasets. These results support the effectiveness of TAGFM beyond input-side prompting and graph transformation alone. We will discuss IGAP in the related work section, clarifying TAGFM’s use of the current frozen backbone’s spectral responses to guide alignment.

AuthorsOpenReview ↗

Response to Reviewer Ret3 (Part3)

W5. The paper does not discuss DAPrompt (AAAI 2026), which jointly adapts graph structure and semantics for few-shot learning, or Graph Sparse Prompting (TPAMI 2026), which provides input-side feature prompting for pretrained GNNs. These works are relevant to TAGFM’s claims about joint feature–structure adaptation and portability across backbones. Although their training settings are not identical, the paper should explain the differences explicitly and compare against them under matched backbones and label budgets where feasible. Without this positioning, it is difficult to determine which gains arise from TAGFM’s proposed objectives rather than from input-side adaptation more generally.

DAPrompt adapts graphs by jointly adjusting their structure and semantic information, while Graph Sparse Prompting adapts pretrained GNNs through sparse input feature prompts. In comparison, TAGFM aligns the target graph with the encoding characteristics of the frozen model and provides a unified adaptation mechanism across different GFMs, without requiring modules tailored to their internal architectures. This reduces the customization required when transferring to different backbones. We will discuss these distinctions in the related work section.

Furthermore, we conducted additional 1-shot comparisons, with results reported in Table 1. TAGFM improves average accuracy over the native adaptation strategies of DAPrompt and GSFP by 6.30 and 6.47 percentage points, respectively, and consistently outperforms GFMate across all settings in these comparisons. These results further demonstrate its performance advantages over both native adaptation strategies and a general-purpose adaptation method.

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

This paper proposes TAGFM, a target-side adaptation framework for frozen graph foundation models. It uses a feature adapter with maximal coding rate reduction and a structure adapter that aligns graph spectral information with the frozen model's estimated frequency preferences. Experiments across seven graph datasets and five pretrained backbones report improvements in few-shot node and graph classification.

Critical strengths
  1. The paper addresses the practical challenge of adapting frozen graph foundation models without modifying their internal architectures.

  2. The evaluation covers five pretrained backbones and multiple graph datasets. The reported average improvements are substantial, and the ablations support the usefulness of both adapters.

Critical weaknesses
  1. Previous methods, such as MDGFM, have already explored joint feature and structure adaptation. The contribution beyond combining existing adaptation and regularization techniques is not sufficiently distinguished.

  2. A strong model response to a frequency band does not necessarily indicate that this band contains more useful information for downstream prediction.

  3. Although the backbone remains frozen, optimizing the feature adapter appears to require gradients through the model, which is different from strict black-box access.

  4. On Computers and Chameleon, TAGFM sometimes performs worse than the original adaptation or GFMate, raising concerns about negative transfer.

  5. The computational overhead and scalability of spectral probing and graph structure optimization are insufficiently evaluated, particularly for larger graphs.

Questions for authors
  1. Provide evidence supporting the spectral preference assumption.

  2. Clarify the black-box access and gradient requirements.

  3. Explain the negative transfer cases and report computational costs.

2 comments
AuthorsOpenReview ↗

Response to Reviewer iunU (Part1)

We thank the reviewer for the careful reading of our paper and constructive comments in detail.

W1. Previous methods, such as MDGFM, have already explored joint feature and structure adaptation. The contribution beyond combining existing adaptation and regularization techniques is not sufficiently distinguished.

MDGFM uses pretrained domain tokens for feature prompting and graph structure learning to jointly adapt the target graph’s features and connections. Its adaptation mechanism relies on components learned during its own pretraining, making it difficult to directly reuse with other GFMs that lack these components.

In contrast, TAGFM provides a unified input-side adaptation mechanism without relying on specific pretrained tokens or modifying internal backbone modules. It also uses the frozen model’s spectral responses to guide structure optimization and incorporates target-node statistics through MCR. Thus, a key advantage of TAGFM is its applicability across backbones. We validate its effectiveness on five frozen GFMs; on MDGFM, TAGFM improves average accuracy over the native adaptation strategy by 7.21 percentage points.

W2. A strong model response to a frequency band does not necessarily indicate that this band contains more useful information for downstream prediction.

TAGFM uses spectral responses to characterize the frozen backbone’s sensitivity to different frequency bands and guide structural adjustments that better match its encoding characteristics. For example, a low-pass backbone may suppress class-discriminative differences between connected nodes. Adjusting the graph structure changes the spectral distribution of feature signals, providing a way to reduce the attenuation of useful information during encoding.

To directly examine the connection to predictive utility, we separately masked the four highest- and lowest-response frequency bands and compared classification accuracy (Table 6 in Appendix D.6). Masking the highest-response bands consistently yielded lower accuracy than masking the lowest-response bands. These results provide empirical support for using spectral responses as a task-relevant signal for adaptation.

W3. Although the backbone remains frozen, optimizing the feature adapter appears to require gradients through the model, which is different from strict black-box access.

Thanks for the constructive comments. We acknowledge that our use of “black-box” was imprecise. In the revision, we redefine our method throughout the manuscript as “architecture-agnostic adaptation of frozen GFMs” and explicitly clarify that feature adaptation requires computing input gradients through the frozen backbone.

W4. On Computers and Chameleon, TAGFM sometimes performs worse than the original adaptation or GFMate, raising concerns about negative transfer.

We thank the reviewer for highlighting these performance differences. We first clarify that Com denotes Computers and Co denotes Cornell in Table 1. On Computers, TAGFM outperforms both native adaptation and GFMate across all five backbones; its relative disadvantages occur mainly on Cornell and for certain backbones on Chameleon.

These differences may be related to the characteristics of the target datasets. Cornell has relatively few nodes and an imbalanced class distribution. In the one-shot setting, the representativeness of a single support example can substantially affect class prototypes and soft assignments, potentially reducing the stability of feature adaptation. Furthermore, heterophilous connections in Cornell and Chameleon may encode useful class relationships. If spectral alignment improves the match between the overall feature spectrum and the model response without adequately preserving these task-relevant connectivity patterns, classification gains may be limited. Despite these variations, TAGFM achieves the highest mean accuracy among the three adaptation strategies in all 25 combinations across the first five target datasets, demonstrating its applicability across diverse target graphs and frozen backbones.

AuthorsOpenReview ↗

Response to Reviewer iunU (Part2)

W5.The computational overhead and scalability of spectral probing and graph structure optimization are insufficiently evaluated, particularly for larger graphs.

We have added a time complexity analysis (Appendix D.4) and comparisons of adaptation time and peak GPU memory (Appendix D.5) between TAGFM and the native adaptation methods. On Pubmed, the largest graph by node count in our evaluation (19,717 nodes), TAGFM incurs comparable or lower adaptation costs. For BRIDGE, TAGFM and native adaptation take 10.25 s and 10.22 s, respectively, with peak GPU memory of 301 MB and 274 MB. For SAMGPT, TAGFM reduces adaptation time from 11.87 s to 9.75 s and peak memory from 446 MB to 311 MB. For MDGFM, it reduces time and memory by approximately 75% and 71%, respectively. These results demonstrate that TAGFM maintains manageable resource costs while delivering adaptation gains on the largest graph by node count in our evaluation.

In our implementation, the target spectral response is estimated before each graph-generation run and remains fixed throughout the subsequent structure optimization, eliminating the need to query the GFM again at every edge-weight update. Structure optimization uses Chebyshev filtering with sparse matrix operations, avoiding a full eigendecomposition of the graph Laplacian. These designs reduce repeated computation during structure optimization.

Q1.Provide evidence supporting the spectral preference assumption.

Please see our response in R2.

Q2.Clarify the black-box access and gradient requirements.

Please see our response in R3.

Q3.Explain the negative transfer cases and report computational costs.

Please see our response in R4 (negative transfer cases) and R5 (computational costs).

Reviewer_h5vYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes TAGFM, a target-side adapter for pretrained graph foundation models with frozen backbone parameters. It adjusts node features using a residual feature adapter and coding-rate objective, and adjusts graph structure by probing the frozen model’s response to frequency-localized perturbations and aligning spectral distributions. A conditional generalization bound and experiments on seven target datasets with five pretrained GFM backbones support the evaluation. The deployment motivation is relevant, but some assumptions about the model-access interface need clarification.

Critical strengths

S1. Adapting target inputs while keeping a backbone frozen is practically relevant for reusable GFMs, and the joint consideration of features and structure provides a unified design target (Sec. 2; Fig. 2). S2. The feature branch uses both labeled support examples and unlabeled node statistics, while the structure branch uses backbone-response probing instead of a fixed homophily assumption. These are plausible complementary signals (Secs. 2.3–2.4). S3. The evaluation covers five pretrained GFM backbones, seven target graphs, 50 random splits in the one-shot experiment, and feature/structure ablations. Table 1 reports favorable average gains for TAGFM over the original adaptations, although individual cases vary (Sec. 3).

Critical weaknesses

W1. The “black-box” access assumption requires a precise definition. Spectral preference is estimated through model-output probes, but the feature branch minimizes a supervised loss on embeddings produced by the frozen GFM and is updated iteratively (Eq. (14); Algorithm 1). Parameter freezing alone does not ensure compatibility with a forward-only service. The paper should clarify whether input gradients are assumed available, or report a derivative-free alternative and its query budget. W2. Spectral response magnitude is treated as a preference signal, but a strong response to frequency-localized perturbations does not necessarily indicate that the same frequency content helps classification. The connection between lower spectral-alignment loss and downstream accuracy should be tested more directly, especially when graph edges are modified (Sec. 2.4). W3. The generalization result is conditional on spectral-stability and calibration assumptions. Assumption 2 already relates prediction risk to distance from an ideal spectral preference, so Theorem 2.1 alone does not establish that real frozen GFMs satisfy this premise. The authors should discuss when these assumptions are plausible and how informative the bound is in the one-shot regime (Theorem 2.1; Appendix C). W4. The practical cost and scope are insufficiently established. Repeated probing, spectral computation, optimal-transport alignment, and feature optimization may add significant overhead. Moreover, the graph-classification extension uses ego-graphs derived from the same source graphs, rather than diverse independent graph-level benchmarks. Matched end-to-end cost and broader graph-level evaluation would strengthen the deployment claim (Sec. 3; Appendix D).

Questions for authors

Q1. What exact access to the frozen GFM is required for feature optimization: forward outputs only, input gradients, or another gradient estimator? Q2. How does estimated spectral preference correlate with downstream accuracy, and what happens when edge modifications are restricted or prohibited? Q3. What are the end-to-end wall time, peak memory, and forward-query costs relative to the compared adaptations, including on larger graphs?

2 comments
AuthorsOpenReview ↗

Response to Reviewer h5vY (Part1)

We thank the reviewer for the careful reading of our paper and constructive comments in detail.

W1. The “black-box” access assumption requires a precise definition. Spectral preference is estimated through model-output probes, but the feature branch minimizes a supervised loss on embeddings produced by the frozen GFM and is updated iteratively (Eq. (14); Algorithm 1). Parameter freezing alone does not ensure compatibility with a forward-only service. The paper should clarify whether input gradients are assumed available, or report a derivative-free alternative and its query budget.

Thanks for the constructive comments. We acknowledge that our use of “black-box” was imprecise. In the revision, we redefine our method throughout the manuscript as “architecture-agnostic adaptation of frozen GFMs” and explicitly clarify that feature adaptation requires computing input gradients through the frozen backbone (Appendix A). .

W2. Spectral response magnitude is treated as a preference signal, but a strong response to frequency-localized perturbations does not necessarily indicate that the same frequency content helps classification. The connection between lower spectral-alignment loss and downstream accuracy should be tested more directly, especially when graph edges are modified (Sec. 2.4).

The spectral response reflects the frequency bands that the GFM backbone can effectively encode, while spectral alignment aims to reduce the attenuation of discriminative information during encoding by frozen models. For example, a low-pass GFM smooths feature differences between neighboring nodes, which may blur class boundaries when neighbors belong to different classes. By adapting the graph structure, spectral alignment reshapes the spectral distribution of feature signals so that classification-relevant information is concentrated in frequency bands that the GFM can encode effectively. Intuitively, the adapted structure strengthens connections among same-class nodes, thereby improving classification performance. To further support this claim, we masked the four highest- and lowest-response frequency bands and compared the resulting classification performance (i.e., Table 6 in Appendix D.6). Masking the highest-response bands consistently yielded lower accuracy than masking the lowest-response bands, indicating that high-response bands contribute more to classification.

W3. The generalization result is conditional on spectral-stability and calibration assumptions. Assumption 2 already relates prediction risk to distance from an ideal spectral preference, so Theorem 2.1 alone does not establish that real frozen GFMs satisfy this premise. The authors should discuss when these assumptions are plausible and how informative the bound is in the one-shot regime (Theorem 2.1; Appendix C).

We clarify that Theorem 2.1 provides a target-dependent transductive PAC bound under Assumptions 1-3. Assumption 2 is motivated by the role of graph propagation in preserving or attenuating task-relevant frequency components. We view the relationship among graph structure, node features, and labels as intrinsic to the underlying data-generating process, rather than determined by the number of labeled support nodes. Accordingly, this underlying relationship does not depend on the shot count. We have added this rationale (i.e., Appendix C.1) and a controlled experiment (i.e., Table 6 in Appendix D.6) comparing probe-derived spectral supervision with deliberately mismatched spectral supervision. The experiment supports the task relevance of the probe-derived signal, while the assumption remains an explicit condition of the theorem.

AuthorsOpenReview ↗

Response to Reviewer h5vY (Part2)

W4. The practical cost and scope are insufficiently established. Repeated probing, spectral computation, optimal-transport alignment, and feature optimization may add significant overhead. Moreover, the graph-classification extension uses ego-graphs derived from the same source graphs, rather than diverse independent graph-level benchmarks. Matched end-to-end cost and broader graph-level evaluation would strengthen the deployment claim (Sec. 3; Appendix D).

In the manuscript, the time complexity of our method and the adaptation time cost of all other adaptation methods are reported in Appendix D4. In the revision, we further reported the memory cost of all GFM adaptation methods in Appendix D5.

Specifically, the time complexity of our method scales linearly with the sample size, similar to that of other GFM adaptation methods (e.g., [1,2,3]). Moreover, our method achieves the best adaptation performance, with lower memory costs than other GFM adaptation methods on most datasets.

[1] Yu X, Gong Z, Zhou C, et al. Samgpt: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. In WWW 2025.

[2] Wang S, Wang B, Shen Z, et al. Multi-domain graph foundation models: Robust knowledge transfer via topology alignment.In ICML 2025.

[3] Yuan H, Sun Q, Shi J, et al. How much can transfer? BRIDGE: Bounded multi-domain graph foundation model with generalization guarantees. In ICML 2025.

Q1. What exact access to the frozen GFM is required for feature optimization: forward outputs only, input gradients, or another gradient estimator?

Feature optimization requires access to the model’s forward embedding outputs and gradients with respect to the input features. Specifically, we use standard backpropagation through the frozen GFM to update the feature adapter while keeping all backbone parameters fixed. In revison, we clarify this access requirement in Appendix A.

Q2. How does estimated spectral preference correlate with downstream accuracy, and what happens when edge modifications are restricted or prohibited?

Please see our response in R2.

Q3. What are the end-to-end wall time, peak memory, and forward-query costs relative to the compared adaptations, including on larger graphs?

We have added comparisons of adaptation time (Table 7 in Appendix D.4) and peak GPU memory (Table 8 in Appendix D.5) between TAGFM and the native adaptation strategies of SAMGPT, BRIDGE, and MDGFM. Across the 21 backbone–dataset combinations, TAGFM reduces adaptation time in 15 cases and peak GPU memory in 12 cases. The gains are particularly pronounced for MDGFM: memory usage decreases on all seven datasets, and adaptation is faster on six. On Pubmed, the dataset with the largest number of nodes, TAGFM reduces MDGFM’s adaptation time from 46.76 s to 11.70 s and peak GPU memory from 1057 MB to 309 MB, corresponding to reductions of approximately 75% and 71%, respectively. These results show that TAGFM’s additional adaptation components do not incur greater practical overhead.

Our graph classification experiments follow the evaluation protocol of prior work [1, 2, 3], using ego-graphs extracted from the original graphs as graph-level samples to assess whether the adaptation method extends from node-level to graph-level prediction while maintaining comparability with existing works.

[1] Yu X, Gong Z, Zhou C, et al. Samgpt: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. In WWW 2025.

[2] Yuan H, Sun Q, Shi J, et al. How much can transfer? BRIDGE: Bounded multi-domain graph foundation model with generalization guarantees. In ICML 2025.

[3] Sun X, Cheng H, Li J, et al. All in One: Multi-Task Prompting for Graph Neural Networks. In KDD 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.