- 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
- Adapting an open-source, frozen GFM to an unseen graph with few labels is practically important.
- 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.
- The feature and structure adapters can be evaluated separately, and the paper includes initial component ablations.
Critical weaknesses
-
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.
-
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.
-
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.
-
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.
-
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
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.
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.
Response to Reviewer Ret3 (Part1)
We thank the reviewer for the careful reading of our paper and constructive comments in detail.
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.
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 X for spectral alignment and the adapted features 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 X has a task-relevant basis for supporting final classification. Both components serve the shared objective of effectively encoding class-discriminative information.