- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper presents GapTune, a graph prompting method for adapting pretrained GNNs without updating their parameters. The main idea is to identify differences between source and target graph contexts and use them to generate local prompts. The method uses learnable queries to capture node- and message-level information, then applies relevance weights and signed gates to construct the prompts. It works in both source-available and source-free settings, using generated proxy graphs when source data is unavailable. The paper also provides theoretical analysis of some properties of the method. Experiments are conducted on 9 datasets across 4 tasks, comparing GapTune with 12 prompting baselines and including a few ablation studies.
Critical strengths
- In stead of treating prompt values as entirely free parameters, GapTune derives them from paired source and target contexts. This provides a concrete way to incorporate transfer-related information into local prompting and distinguishes the method from other conventional feature- or message-prompt learning methods.
- The evaluation covers multiple graph tasks, and examine prompt values, insertion locations, local composition, and proxy construction. These experiments help clarify the behavior of different components, although they do not fully establish the claimed advantages.
Critical weaknesses
- Section 4.1 states that same-dataset prompting methods report their "best observed test mean" across a 7 x 6 architecture-objective grid. This appears to use test performance for configuration selection, introducing selection bias into Tables 2 and 15. Since these tables support several central empirical claims, I do not think their rankings can be considered reliable without validation-based selection and a fresh test evaluation.
- The main methodological distinction is the use of context differences instead of freely learned prompt values. Although Table 3 reports improvements on 4 datasets, the appendix (Table 11) shows that none of the 4 gap-versus-free comparisons remains significant after Holm correction at the 0.05 level. Moreover, the evaluation lacks accuracy comparisons with compute-matched adapter methods and cross-dataset full fine-tuning. These limitations make it difficult to determine whether the additional complexity of GapTune is justified.
Questions for authors
- Why are the same-dataset results selected using the best test mean? Could the authors provide results where architecture, pretraining objective, and hyperparameters are selected exclusively using validation data?
- Given that the gap-vs-free improvements do not survive the reported multiple-comparison correction, what evidence supports the conclusion that gap-derived values consistently outperform free prompt values?
- How does GapTune compare with full fine-tuning and parameter- or compute-matched adapter methods under identical pretrained checkpoints and evaluation budgets?
9 comments
Follow-up to Reviewer NSSF
Thank you again for your careful assessment. We would appreciate your feedback on our response, particularly the distinction between validation-based checkpoint selection within each configuration and best-test-mean aggregation across configurations. We also explain the fixed-checkpoint ablations and their statistical uncertainty, and point to the full fine-tuning comparisons and measured costs already included in the submission.
Could you let us know which concerns remain after considering these clarifications and results? A brief indication of the main unresolved issue would help us provide a focused follow-up while the discussion is open.
Follow-up on the reporting protocol and evidence in the original submission
Dear Reviewer NSSF,
With approximately three days remaining in the discussion period, could you please review our response to your three questions and let us know which concerns remain?
Regarding the first concern about experimental settings, our strategy for reporting results across configurations follows the same practice used in closely related work, including ProG (NeurIPS 2024) and MTG (ICML 2026).
For the second and third questions, our response identifies relevant experiments already included in the original submission, before the rebuttal period.
We would appreciate your feedback on these clarifications and results. If you have further questions or would like us to clarify a specific comparison, please let us know soon so we have time to respond before the discussion closes.
Thank you for your time.
Response to Reviewer NSSF
Summary: Checkpoints are validation-selected, and summary tables report the best test performance following the reporting convention of ProG and MTG. The submitted controls show positive mean gains from gap values, with uncertainty and full fine-tuning tradeoffs reported explicitly.
Q1. Configuration summaries and validation selection.
Tables 2 and 15 summarize each method's best result across configurations. Within each fixed configuration, every run selects its adaptation checkpoint by validation before testing. The summary then takes the best five-run test mean among completed configurations in the common 7×6 grid, with that configuration's sample standard deviation. We have clarified these two steps in revised Section 4.1, with the protocol in Appendix D.2 and the full grid in Appendix D.5.
This summary rule follows previous papers, including ProG (NeurIPS 2024), Tables 2-3 and MTG (ICML 2026), Table 1, which report the best across pretraining strategies. View Space (ICML 2026), Tables 1/19 and MASS (ICML 2019), Table 2 likewise aggregate the best pretrained variants.
Q2. Evidence for context-derived values.
Table 3, Tables 10-11, and Appendix C.7 of our submission provide the evidence: controlled value ablations, paired-run analysis, and controlled-shift tests. The conclusion supported by Table 3 is a positive mean improvement on each of the four targets.
Table 3 uses the fixed ZINC/GCN/EdgePred checkpoint, the same local composition and node/message insertion, and equally parameterized gap and free values. It therefore compares the value constructions within one configuration, independently of the summary-table selection in Q1.
Effects are percentage points in AUC for DBLP and accuracy otherwise. Gap values improve on free values in 5/5 paired runs on Photo and 4/5 on Chameleon. DBLP/MNIST use conservative bounds from marginal deviations, not exact paired tests. The four conservative Holm values exceed 0.05 within the twelve-comparison family, so these results support consistent positive means without establishing family-wise statistical superiority. The revised Introduction makes this mean-performance interpretation explicit.
The paired-mean control tests whether the benefit comes from subtraction while retaining both source and target information and the same query parameterization. Table 3 gives gap-minus-paired-mean gains of 4.14/2.52/1.65/0.51 points on Photo/Chameleon/DBLP/MNIST. Appendix C.7 provides a separate ten-repetition test using within-dataset GCN/DGI checkpoints. At joint shift 0.30, gap-minus-free effects are +2.20 points on Photo with pointwise interval [0.93, 3.47], and +1.20 on Chameleon with [-0.39, 2.79]. Its zero- and strong-shift results show that the advantage depends on the transfer condition. These controls support a task-dependent mean advantage with quantified uncertainty.
Q3. Full fine-tuning and matched adapters.
Appendices D.5 and C.10 of our submission provide full fine-tuning comparisons and measured computational costs, respectively. The performance comparison fixes each architecture and pretraining objective, while the timing study fixes the execution workload. Counting the submitted same-dataset means within each configuration gives:
Each setting has 7×6×9 comparisons, with lower MAE better for QM7b.
Appendix C.10 uses identical GCN batches, FP32, 500 updates and 20 evaluations. Values are means over five timing blocks, and inference is per batch.
Gap prompting reduces trainable parameters by 66.95% versus full fine-tuning, with additional time and memory for local prompting. Q2 gives the matched-parameter free-value accuracy comparison. For the additional adapter comparison, could you specify the GNN adapter methods and matching budget, such as trainable parameters or adaptation time?