acceptodds
Withdrawn from ICLR 2027

GRACE: Gaussian Prototypical Consensus with Perturbation-aware Adaptation for Test-time Prompt Tuning in Graph Foundation Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

This paper studies the problem of test-time prompt tuning for graph foundation models (GFMs), which aims to improve GFMs at test time by optimizing prompts. Existing approaches typically learn prompts from a small number of support data to facilitate GFM adaptation. Despite recent progress, the performance of these approaches is far from satisfactory due to overconfident pseudo-labels of target nodes and thus unstable prompt optimization. Towards this end, we propose a novel approach named aussian Pototypicl onsensus with Prturbation-aware Adaptation (GRACE) for test-time prompt tuning in GFMs. The core of our GRACE is to improve the stability of test-time prompt tuning from two complementary perspectives, i.e., probabilistic prototype measurement and selective perturbation-aware adaptation. In particular, our GRACE first estimates Gaussian prototypes from support nodes to characterize the representation spaces across different layers in GFMs. More importantly, we aggregate covariance-aware Mahalanobis distances between each target node and its corresponding prototypes via majority voting, thereby filtering out unreliable pseudo-labels. Furthermore, we select target nodes with strong prototypical distance stability under graph perturbations and estimate prompt update directions through state-consistent symmetric perturbations, coupling reliability probing and prompt adaptation within a unified perturbation-aware procedure. Experiments on benchmark datasets validate the effectiveness of against extensive baselines. Our code is available at https://anonymous.4open.science/r/GRACE-03D3.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
GNNs, GFM, TTA, Few-Shot Graph 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_sBV2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces GRACE, which aims to alleviate optimization instability caused by pseudo-label overconfidence in graph-based model testing prompt tuning. The proposed method consists of two core designs. One is to improve the reliability of pseudo-labels by fusing multiple layers of evidence using Gaussian class prototypes and covariance-aware distances of each layer. The second is to predict stable target nodes by graph perturbation screening, and update hints using symmetric parameter perturbation in a fixed supervision state. Experiments show that GRACE can improve the adaptation performance compared with other baselines.

Critical strengths
  1. This paper focuses on pseudo-label error accumulation and optimization instability, and the connection between the problem and the method design is relatively clear.

  2. GRACE according to the performance of support nodes under graph perturbation, a category-dependent selection criterion is established, which considers distance stability, prediction flip, and confidence. The design makes full use of the limited supervision information of the target domain, and has good interpretability.

  3. The experiments are validated on several datasets and the number of experiments is rich.

Critical weaknesses
  1. Appendix A.1 shows that the paper only keeps one supporting node per class, but splits the remaining nodes into validation and test sets 1:9, without specifying whether the validation labels are used to select hyperparameters, adaptation rounds, or model configurations. If these labels are used, the actual labeling budget is more than "one per class". if not, explain what the validation set is for and how the parameters are determined. My question is whether there will be a label leak problem?

  2. The paper states that the experiments follow the setup of GFMate, but it is not clear whether the two share the exact same pre-training checkpoints, support sets, and test sets.

  3. Only five runs are not enough to fully support the statement "consistently outperforms the baseline on all datasets". For example, Squirrel's improvement is only 0.45 percentage points, and Cora's improvement is 0.71 percentage points, which corresponds to a large inter-run standard deviation.

  4. Missing some classical baseline comparisons of GFM, such as Unigraph, AnyGraph?

Questions for authors

Most of my questions focused on experimentation, so read the weekness section and respond.

Reviewer_anUdOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents GRACE, a test-time prompt tuning framework for few-shot adaptation of graph foundation models on unseen target graphs. The method integrates layer-wise Gaussian prototype modeling, covariance-aware distance computation, probabilistic cross-layer fusion, perturbation-calibrated pseudo-label filtering, and state-consistent zeroth-order prompt optimization. Experiments on nine benchmarks report improvements over the evaluated baselines, particularly GFMate. Overall, the work appears to be primarily an engineering contribution, as the methodology largely combines existing techniques and the claimed new perspective is not sufficiently substantiated.

Critical strengths
  1. GRACE combines multiple techniques into a unified adaptation framework and achieves the strongest reported performance among the evaluated methods for few-shot test-time adaptation of GFMs.

  2. The methodological description is detailed, with explicit equations and algorithms specifying prototype updates, probabilistic fusion, reliability filtering, and prompt optimization.

  3. The diagnostic experiments provide useful evidence regarding admission purity, retention rate, and class geometry, helping explain the behavior of the proposed components.

Critical weaknesses
  1. The first claimed contribution, “New Perspective,” is not adequately substantiated. The two motivating challenges—insufficient class-conditional distribution modeling and unreliable pseudo-labels during test-time adaptation—are well established in the literature. The paper reframes them as two coupled requirements, but does not sufficiently demonstrate a novel conceptual connection between them or explain why this formulation constitutes a new perspective. The novelty claim therefore appears overstated.

  2. The second claimed contribution, “Novel Methodology,” is also insufficiently justified. Gaussian prototypes, Mahalanobis distance, weighted cross-layer fusion, perturbation-based reliability filtering, and zeroth-order optimization are established techniques. Although their integration is technically plausible and yields empirical gains, the paper does not sufficiently identify a new methodological principle beyond this combination. The contribution appears to lie primarily in coordinating existing techniques for GFM adaptation, rather than introducing a fundamentally new mechanism.

  3. The experimental comparisons do not adequately establish the advantages of GRACE over existing test-time adaptation methods. The evaluation relies heavily on GFMate, and broader comparisons with graph-specific test-time adaptation methods are needed. Without these comparisons, it remains unclear whether the reported gains reflect a general advantage over existing adaptation strategies.

  4. The scalability of full covariance modeling is insufficiently evaluated. GRACE maintains a covariance matrix for each class at each selected layer and uses Cholesky factorization and triangular solves for covariance-aware distance computation. These operations may introduce substantial computational and memory costs as feature dimensionality and the number of classes increase. A complexity analysis and empirical scaling study are needed to establish the practicality of this design.

  5. The description of cross-layer aggregation as “majority voting” is technically inaccurate. The actual mechanism performs learnable weighted probability fusion and then selects layers according to their agreement with the fused pseudo-label. This differs from literal majority voting, and the terminology should be made consistent with the implemented mechanism.

  6. The reported efficiency advantage over GFMate requires further clarification. Despite introducing covariance-aware scoring, graph perturbations, reliability filtering, and additional prompt optimization, GRACE is reported to incur lower overhead. The paper does not provide sufficient evidence to explain this result or establish that the comparison accounts for all operations under matched conditions. The authors should report component-wise runtime, clarify the treatment of preprocessing and cached representations, and compare against additional baselines using a consistent measurement protocol.

Questions for authors

How do the authors distinguish the claimed contributions of “New Perspective” and “Novel Methodology” from established ideas in distributional prototype modeling and test-time adaptation?

Can the authors provide comparisons with additional graph-specific test-time adaptation methods under matched evaluation protocols?

Does the reported runtime include all graph perturbations, covariance computations, reliability filtering, and prompt optimization? Are preprocessing and cached representations treated consistently across methods?

What are the computational and memory scaling trends with respect to feature dimensionality, number of classes, and number of selected encoder layers?

Reviewer_jeMPOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on a few-shot test-time prompt tuning method for graph foundation models. The method combines layer-wise Gaussian prototypes, cross-layer consensus, and perturbation-based reliability filtering to enhance pseudo-label reliability and uses zeroth-order optimization to update prompts. Experiments on nine graph datasets show that GRACE achieves some performance improvements over existing methods.

Critical strengths
  1. This method is relatively comprehensive, integrating Gaussian prototype modeling, cross-layer consensus, and perturbation-aware adaptation into a unified framework. The overall workflow is clear.

  2. The method outperforms its closest baseline, GFMate, on all nine datasets. It also achieves consistent improvements across different GNN backbones.

Critical weaknesses
  1. Few-shot test-time prompt tuning has been extensively studied in prior work. The challenges identified by the authors, including pseudo-label reliability and adaptation instability, have also been widely discussed in existing research. The proposed method mainly combines well-known techniques, without clearly demonstrating new methodological contributions.

  2. The related work mainly provides a general overview of graph prompting and test-time adaptation, lacking in-depth comparisons with GFMate, DAGPrompt, and existing methods for uncertainty-aware prototype learning and consistency-based adaptation. The paper does not clarify the specific novelty of each component compared with existing techniques. It does not explain what new methodological contributions arise from combining these components.

  3. Each class has only one labeled support node, and PaCF uses its perturbation statistics to determine class-specific reliability thresholds. Moreover, the same support node is used both to initialize the Gaussian prototype and to construct the reliability thresholds. The stability of this in-sample calibration and its representativeness for unlabeled nodes remain unclear, and the paper lacks sufficient analysis of this issue.

  4. The abstract states that cross-layer information is aggregated through majority voting. However, Eq. (5) actually uses learnable weighted probability fusion, while Eq. (6) further applies layer-wise agreement filtering. These operations are not equivalent, and the description of the core mechanism should be more precise and consistent.

Questions for authors
  1. What are the specific challenges in few-shot test-time prompt tuning for GFMs that distinguish this problem from existing test-time adaptation settings?

  2. The abstract mentions majority voting, while Eq. (5) uses learnable weighted probability fusion. Could the authors clarify this inconsistency?

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

Related papers

Discussion (0)

Sign in to comment.