acceptodds
Under review as a conference paper at ICLR 2027

GapTune: Learning Graph Prompts from Source–Target Context Gaps

Abstract

Graph prompting adapts *graph neural networks* (GNNs) pretrained on source graphs to target applications with few trainable parameters. Recent methods introduce feature, topology, and message prompts, but how source–target context gaps should inform local prompt values remains underexplored. In a controlled prompt-transfer study with a frozen GNN, we find that prompt transferability is associated with source–target context shifts in node representations and messages. Motivated by this association, we propose GapTune, a graph prompting framework that learns local prompts from context gaps in source-available and source-free settings. Source observations are obtained from pretraining graphs when available or approximated by responses to compact model-inverted proxy graphs. Shared learnable queries pool the frozen GNN's unprompted initial node states and layer-wise messages into paired source and target contexts, whose differences supply prompt values. Local relevance weights and shared signed gates combine these values into task-supervised local prompts. Our analysis establishes gap-dependent prompt bounds, task-compatible symmetries, and conditional prediction stability under source-context approximation for a fixed adapted predictor. Experiments against 12 prompting baselines on 9 target datasets and 4 tasks show task-dependent transfer gains. Joint gap-valued prompts outperform equally parameterized free values on four ablation targets, with benefits depending on insertion locations and graph shifts.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph prompt tuning, graph neural networks, source–target context gaps, few-shot learning, source-free adaptation.

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_NSSFOpenReview ↗
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
AuthorsOpenReview ↗

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 7×67\times6 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.

TargetGap minus freePointwise 95% intervalConservative Holm value
Photo+3.23[0.56, 5.90]0.2157
Chameleon+1.56[-0.77, 3.89]0.4114
DBLP+1.89[0.26, 3.52] (bound)0.2157
MNIST+1.92[-2.99, 6.83] (bound)0.4116

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:

SettingMatched configurationsGapTune+ better than full fine-tuning
5-shot378290
100-shot378255

Each setting has 7×6×97\times6\times9 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.

MethodTrainable parametersPreparation (s)Adaptation (s)Inference (ms)Peak memory (MB)
Full fine-tuning46,98404.851.81224.7
Free joint prompts15,52806.514.17333.1
GapTune+15,5282.227.584.96367.2
GapTune15,52841.737.584.96367.2

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?

AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

Reviewer_3sPMOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GapTune, a graph prompting framework that constructs node- and message-level prompts from source–target context differences. It uses shared queries, local relevance weights, and signed gates to adapt frozen GNNs. Experiments across multiple graph tasks demonstrate promising improvements over existing prompting methods.

Critical strengths
  1. Novel approach. Constructing prompt values from source–target context gaps offers an interesting perspective on graph adaptation.
  2. Comprehensive evaluation. The paper covers diverse tasks, datasets, and prompting baselines, with extensive ablations.
  3. Thoughtful design. The method is intuitive and supported by theoretical analysis of its properties.
Critical weaknesses
  1. Insufficient mechanistic understanding. The observed correlation between context gaps and prompt transfer loss does not directly explain why gap-derived directions benefit downstream optimization.

  2. Limited generalizability of gap-constrained adaptation. Restricting prompts to context-gap directions may limit adaptation when task mismatch does not align with distribution shifts, particularly when context gaps are small.

  3. Insufficient isolation of gap-specific benefits. Existing ablations do not fully distinguish the benefits of context subtraction from differences in conditioning and prompt magnitudes.

Questions for authors

Please address all of the weaknesses mentioned above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 3sPM

Summary: GapTune learns which source-target differences to extract and how to use them for each task. Its trainable head and evolving gap bank support adaptation across tasks, while the submitted controls compare subtraction with averaging and test prompt direction at fixed magnitude.

Thank you for the positive assessment and these helpful questions.

1. Why can gap-derived directions help optimization?

Subtracting contexts can remove shared response patterns and expose changes useful to the target task. Signed gates learn whether to follow or reverse these directions. Proposition B.5 in Appendix B.11 formalizes local descent: fix observations, queries, relevance maps and head, and write p=Ttanh⁡ϑp=T\tanh\vartheta and g=∇pL(0)g=\nabla_p\mathcal L(0). Under its smoothness and feasible-step assumptions, L(−ηTT⊤g)≤L(0)−η2∥T⊤g∥22\mathcal L(-\eta TT^\top g)\leq\mathcal L(0)-\frac{\eta}{2}\|T^\top g\|_2^2, so nonzero T⊤gT^\top g permits loss reduction. To preserve space for your other questions, we give the full comparison with alternative values in eEEC Q2 and revised Appendix B.11, Proposition B.6.

2. Small gaps and task changes without a context change.

GapTune can adapt to task changes with small dataset-level shifts through its trainable head and task-dependent context comparison. Target supervision jointly trains the head, queries, local relevance and signed gates. Contexts are recomputed after each query update, so the gap bank can change with the task, while local relevance selects mixtures for individual nodes and messages. The head learns the target prediction rule even when little prompting is needed. Appendix C.3 supports this design: learned queries improve on frozen queries in all four settings, and local mixtures and signed gates outperform their global and nonnegative counterparts.

The submission tests nine datasets and four tasks, including the seven-backbone, six-objective grid in Appendix D.5. In fixed-checkpoint cross-dataset comparisons, GapTune and GapTune+ exceed the strongest prompting-baseline mean in 29/36 and 32/36 cases, respectively (Section 4.2). Same-dataset adaptation can also have nonzero gaps because its pretraining reference and local prediction subgraph contain different observations.

For identical observation collections, all gaps vanish and adaptation reduces exactly to training the head. For small gaps, the prompt contribution is bounded by ∥po∥2≤max⁡k∥dk∥2\|p_o\|_2\leq\max_k\|d_k\|_2 and uses directions in the current gap span. Appendix C.7 tests this dependence with within-dataset GCN/DGI checkpoints, ten matched repetitions and validation-selected checkpoints:

Joint shiftPhoto gap minus freeChameleon gap minus free
0.00-0.40-0.50
0.30+2.20+1.20
0.50+0.40-0.20

Effects are accuracy percentage points, and the shifts preserve labels. Thus, task adaptation remains possible with small or zero gaps through the head, while gap prompts add value when their directions are useful for the target task.

3. Subtraction versus conditioning and magnitude.

Table 3 and Appendix C.5 of our submission already study these two concerns through complementary controls. The first compares subtraction with averaging while retaining both source and target information. Table 3 fixes the ZINC/GCN/EdgePred checkpoint, local descriptors, signed gates and node/message insertion locations. Paired means (cs+ct)/2(c_s+c_t)/2 retain both information sources and the same query parameterization. GapTune+ gives these mean gains:

ContrastPhotoChameleonDBLPMNIST
Gap minus paired mean+4.14+2.52+1.65+0.51
Gap minus free+3.23+1.56+1.89+1.92

Effects are percentage points in AUC for DBLP and accuracy otherwise. Free values replace query parameters with the same parameter count. All variants share a permitted value-norm envelope, while realized prompt norms may differ. Tables 10-11 report uncertainty and Holm analysis for gap versus free.

The second control changes direction while preserving every local prompt's length after fitting. Appendix C.5 applies a common orthogonal map per observation type, po′=Rpop'_o=Rp_o and ∥po′∥2=∥po∥2\|p'_o\|_2=\|p_o\|_2, with all learned parameters fixed.

ValuesPhoto 0∘0^\circPhoto 90∘90^\circChameleon 0∘0^\circChameleon 90∘90^\circ
Gap77.8668.9736.4731.04
Free74.6367.8634.9130.36

Angles denote the transformation parameter. Changing direction reduces accuracy for both constructions, showing that magnitude alone does not explain the fitted predictors. Table 3 separately supports subtraction over averaging under matched conditioning.

AuthorsOpenReview ↗

Follow-up to Reviewer 3sPM

We wanted to follow up on your helpful questions about the optimization mechanism, adaptation with small context gaps, and the controls for context subtraction. Our response now explains the mechanism intuitively before presenting the formal result, gives a clearer account of how the learned queries and prediction head support adaptation, and identifies the submitted controls for conditioning and prompt magnitude.

We would appreciate your thoughts on whether these explanations address your three concerns. If there is a particular aspect you would like us to explain further, please let us know so we can respond during the discussion. Thank you again for your time and constructive feedback.

Reviewer_eEECOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the relationship between source–target context shifts in node representations and messages and graph prompt transferability. Motivated by the observed correlations, the authors propose GapTune, a graph prompting framework that learns local prompts from source–target context gaps in both source-available and source-free settings. Extensive experiments demonstrate task-dependent transfer gains and the effectiveness of the proposed method.

Critical strengths
  1. The motivation is clear and supported by empirical evidence showing a correlation between source–target context shifts and prompt transferability.

  2. The experiments are comprehensive, covering multiple datasets and downstream tasks. The ablation studies further support the effectiveness of the proposed method and the rationale behind its design.

  3. The code is provided.

Critical weaknesses
  1. The methodological novelty is somewhat limited. Modeling source–target distribution differences has been extensively studied in domain adaptation, including graph domain adaptation. Moreover, several components of the proposed framework, such as learnable queries, attention-based aggregation, and gating mechanisms, are based on established techniques. The authors should further clarify the contribution and differences

  2. The propositions establish certain mathematical properties of the proposed framework but do not provide a theoretical justification for why gap-based prompts should be more effective than alternative prompt formulations.

I am not familiar with this research area, so I would like to read the other reviewers’ comments and open to adjust my rating.

Questions for authors

NA

4 comments
AuthorsOpenReview ↗

Response to Reviewer eEEC

Summary: To our knowledge, GapTune is the first graph prompting method to explicitly model a pretrained GNN's node/message response context and generate local prompts from source-target context differences. The revised analysis explains when these differences provide better task-directed prompt directions than alternative values under the same norm budget.

Thank you for recognizing the motivation and evaluation. We address the contribution and the reason gap-derived prompts can help below.

1. Contribution relative to graph prompting and domain adaptation.

GapTune starts from a context-based view of a pretrained GNN: its pretraining data and task define a source reference for downstream adaptation. Our contribution has two connected levels.

First, to our knowledge, GapTune is the first graph prompting method to explicitly investigate and model this source context through node/message responses. The context describes the response patterns associated with the pretrained model and its source data. Figure 1 studies how changes in these patterns relate to prompt transferability. To make this context usable for prompting, we summarize initial node states and layer-wise messages from the frozen GNN. Pretraining graphs supply the source observations, while message responses also reflect the encoder learned by the pretraining task. Model-derived proxy graphs supply the reference when source graphs are unavailable.

Second, GapTune is the first to use these paired source-target context differences as local node/message prompt values. Shared queries pool source and target observations under the same rule, giving dk=cs,k−ct,kd_k=c_{s,k}-c_{t,k}. Each node or message receives a task-trained mixture po=∑kao,ktanh⁡(ϑk)dkp_o=\sum_k a_{o,k}\tanh(\vartheta_k)d_k. Thus, the source-target comparison supplies the values used for adaptation, and target supervision learns how each difference contributes locally.

This gives a concrete distinction from earlier constructions. GraphPrompt unifies task instances through subgraphs and learns readout prompts, while SUPT assigns learned prompt features using target neighborhoods. These target-subgraph constructions do not model the pretrained model's paired source-target response contexts. GPF+ learns feature-prompt vectors, MTG learns message prototypes, and IGAP addresses signal and structural gaps through spectral prompting. GapTune instead makes the pretrained model's response context explicit and uses its source-target differences as node/message prompt values. Its adaptation objective learns how to use these differences rather than minimizing a source-target distribution distance.

2. Why can gap-derived directions help?

Source and target responses can share a large common component while differing in a component relevant to the target task. Subtracting their paired contexts cancels the common component and exposes the contrast. Signed gates can reinforce, reverse or suppress this contrast according to target supervision. This helps when the retained contrast is better aligned with a task-improving direction than the alternative prompt values.

Proposition B.5 in the submission establishes local descent along useful gap directions. We have added a direct comparison, its proof and a worked example to Appendix B.11 as Proposition B.6, marked with revision highlighting. Fix the observations, queries, relevance maps and head, and let TG,TAT_G,T_A map shared gate coefficients to stacked prompts for gap and alternative value banks. Let PiP_i project orthogonally onto col⁡(Ti)\operatorname{col}(T_i), g=∇L(0)g=\nabla\mathcal L(0), and Fi(r)=min⁡p∈col⁡(Ti), ∥p∥2≤rL(p)F_i(r)=\min_{p\in\operatorname{col}(T_i),\,\|p\|_2\le r}\mathcal L(p). Assume a β\beta-Lipschitz loss gradient on this ball and choose r>0r>0 with r∥Ti†∥2→∞<1r\|T_i^\dagger\|_{2\to\infty}<1 for both banks (†\dagger: pseudoinverse), so the compared prompts are realizable by their shared tanh gates. The comparison gives FG(r)−FA(r)≤−r(∥PGg∥2−∥PAg∥2)+βr2F_G(r)-F_A(r)\le-r(\|P_Gg\|_2-\|P_Ag\|_2)+\beta r^2. It follows that gap values attain a lower best local loss when their gradient-projection advantage exceeds βr\beta r. The proof bounds each subspace's best loss by its linearized optimum L(0)−r∥Pig∥2\mathcal L(0)-r\|P_i g\|_2, with Taylor error at most βr2/2\beta r^2/2.

For one query and one prompted object, take orthonormal u,vu,v with cs=u+v/2c_s=u+v/2, ct=u−v/2c_t=u-v/2 and g=vg=v. The gap is vv and the paired mean is uu, with equal norms, but only the gap has nonzero overlap with the task gradient. This example shows how context subtraction can improve the available direction without increasing its magnitude. The result is a conditional local training-loss comparison with fixed banks, making the task-dependent benefit explicit.

AuthorsOpenReview ↗

Follow-up to Reviewer eEEC

Thank you again for your thoughtful review. We have clarified the two main contributions in our response: explicitly modeling the source context of a pretrained GNN, and using source-target context differences to generate local prompts. We have also added a conditional comparison with alternative prompt values, together with its proof and an example, in revised Appendix B.11.

When you have a chance, could you let us know whether these clarifications address your concerns about novelty and theoretical motivation? We would be happy to discuss any point that remains unclear.

AuthorsOpenReview ↗

Further follow-up to Reviewer eEEC

Dear Reviewer eEEC,

With the discussion closing soon, we would be grateful if you could take a look at our response. Our response clarifies the two main contributions: explicitly modeling the source context of a pretrained GNN and using source-target context differences to construct local node/message prompts. We also added a conditional comparison, proof and example in Appendix B.11 showing when gap-derived prompts can achieve lower local training loss than alternatives under the same norm budget.

We have provided point-by-point responses to the other reviews, using evidence from the original submission and clarifications in the rebuttal. Reviewer 3sPM has since confirmed that they have no further questions and will maintain their positive assessment.

We would very much appreciate your consideration of these clarifications and the discussion so far in your final assessment. If anything remains unclear, please let us know so that we can respond before the discussion closes. Thank you again for your time and thoughtful review.

AuthorsOpenReview ↗

Final follow-up on novelty and theoretical justification

Dear Reviewer eEEC,

With less than two hours remaining in the discussion, we would like to briefly summarize our responses to your two main concerns.

1. Novelty. GapTune explicitly models a pretrained GNN’s source context through node and message responses and uses paired source–target context differences as local prompt values. To our knowledge, it is the first graph prompting method to systematically model this source context for local prompt construction. We also introduce model-derived proxy graphs to construct a source reference when the original source graphs are unavailable, extending this approach to source-free adaptation.

2. Theoretical justification. We added Proposition B.6, its proof, and a worked example to Appendix B.11. For fixed prompt banks under the same prompt norm budget, this result identifies conditions under which gap-derived directions permit a lower best attainable local training loss than alternative prompt values. The worked example shows how context subtraction can remove shared components and retain task-relevant directions, explaining when gap-derived prompts can help.

We would appreciate a brief acknowledgment that you have considered our responses and your assessment of whether they address your concerns. If any concerns remain, please let us know so that we can respond before the discussion closes.

Thank you for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Thank you for your positive assessment

Thank you for confirming that you have no further questions and for maintaining your positive assessment. Your constructive feedback helped us clarify the optimization mechanism, adaptation with small context gaps, and the controls for context subtraction. It also motivated the conditional comparison added to Appendix B.11. We appreciate your time and thoughtful review.

AuthorsOpenReview ↗

Rebuttal summary

Dear Reviewers and AC,

Thank you for the discussion. GapTune contributes a new way to use pretraining context in graph prompting. It models a pretrained GNN’s node and message responses and uses paired source–target differences to construct local prompts. Proxy graphs make the same approach available when source graphs are inaccessible. This gives a concrete way to use the relationship between pretraining and downstream data to guide adaptation, evaluated across four tasks and seven backbones.

The main empirical result remains after addressing the configuration-selection concern. Revised Tables 2 and 15 use validation loss for both checkpoint and configuration selection. GapTune+ achieves the best mean performance among prompting methods in all 18 same-dataset comparisons. The reported mean advantage therefore remains when test scores are excluded from these choices.

The original ablations also support the proposed construction. The free-value variant tests replacing context-derived values with learned vectors at the same parameter count. The paired-mean control separately compares subtraction with averaging while retaining both contexts, the query parameterization, and prompt composition. Subtraction gives higher mean performance on all four targets in Table 3. These controls address different questions.

We added Proposition B.6, its proof, and an example in Appendix B.11 to explain when the construction helps. For fixed prompt banks under the same norm budget, it identifies conditions under which gap-derived directions permit a lower best attainable local training loss.

The latest concern about implementation gaps does not identify a specific discrepancy. We have requested a concrete example so that it can be checked.

Reviewer 3sPM confirmed that they have no further questions and maintained their positive assessment. Reviewer NSSF acknowledged that the rebuttal strengthens the paper and increased their score, while retaining concerns.

We believe the paper merits acceptance because it contributes both a useful prompt construction and evidence about when source–target context differences help. The revised evaluation strengthens that case, and we ask the AC to weigh these contributions against the specific remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.