acceptodds
Under review as a conference paper at ICLR 2027

TI-GFM: Test-Time Invariant Prototype Learning for Text-Attributed Graph Foundation Models

Abstract

Graph foundation models (GFMs) aim to transfer pretrained graph encoders across domains and tasks. For text-attributed graphs, semantic class prototypes provide a direct interface for zero-shot prediction, but cross-domain shift can misalign them with target-domain class distributions. A shifted prototype may drift away from its corresponding class or approach confusing classes, leading to unreliable prototype-based prediction. Here, we introduce a Prototype Shift () metric to quantify this misalignment by combining intra-class deviation and nearest inter-class separation in the target domain. Based on this diagnosis, we propose TI-GFM, a test-time invariant prototype learning framework for text-attributed GFMs. TI-GFM learns perturbation-robust representations through Text-Attributed Graph Pre-training (TGP) and refines semantic prototypes at test time using target nodes with prediction consistency across perturbed graph views. This design injects invariant target-domain evidence into semantic prototypes without retraining the backbone encoder. Experiments on seven cross-domain target datasets show that TI-GFM achieves the best average rank and the best accuracy on six of seven zero-shot node classification benchmarks, with consistent gains in few-shot node classification and zero-shot link prediction. Ablations, analysis, and theoretical margin analysis further indicate that reducing prototype shift is a main factor behind cross-domain transfer performance in text-attributed GFMs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Foundation Models, Text-Attributed Graph

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_4JDhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents a graph foundation model framework named TI-GFM and focuses on improving cross-domain transfer by addressing the mismatch between pretrained class prototypes and target-domain graph representations. It introduces Prototype Shift (PS) as a measure of prototype reliability and proposes two main components: Text-Attributed Graph Pre-training (TGP) for improving representation robustness through contrastive learning and graph–text alignment, and Test-Time Invariant Prototype (TIP), which constructs target-aware prototypes from predictions that remain consistent across perturbed graph views. The framework supports both zero-shot and few-shot adaptation while keeping the main graph encoder frozen at test time.

The method is evaluated on 7 datasets for node classification, with additional experiments on link prediction. It achieves the best zero-shot accuracy on 6/7 datasets, outperforming GraphCLIP on several benchmarks. Ablation studies examine the contributions of TGP and TIP, while prototype-shift analysis, few-shot experiments, and hyperparameter sensitivity studies provide further validation of the approach.

Critical strengths
  • Improved performance across multiple datasets (6/7). It achieves the best zero-shot node classification accuracy on 6/7 datasets, with strong improvements over GraphCLIP on several benchmarks. The results are encouraging, particularly given the challenges of transferring graph representations across different domains. The additional link prediction and few-shot experiments further demonstrate the potential usefulness of the approach.

  • The motivation is clear and appropriate, and justified with the experimental analysis. To address the mismatch between pretrained semantic prototypes and target-domain representations is a relevant problem in graph foundation models. The combination of robust graph–text pretraining and test-time prototype refinement is practical, as it avoids updating the main encoder. The ablation studies and prototype-shift analysis provide useful evidence that both components contribute to the improvements, although the precise source of the gains requires further clarification.

Critical weaknesses
  • The paper has limited novelty and lacks comparison with the recent SOTA approaches. The key idea of refining class prototypes using reliable pseudo-labelled target samples closely resembles existing test-time adaptation approaches, such as T3A (Iwasawa and Matsuo, NeurIPS 2021) and ASSESS (Zhao et al., ICML 2025). The discussion also omits more recent graph foundation model adaptation methods, such as GFMate (Jiang et al., ICML 2026). Although the reported improvements are promising, the paper does not sufficiently establish how its proposed mechanism advances these approaches, weakening its originality claim.

  • The provided evidence is not good enough to isolate the proposed adaptation mechanism. The improvements could be from stronger graph–text pretraining rather than the invariant prototype refinement. Particularly, consistency across perturbed views does not guarantee correct pseudo-labels, and comparisons with simpler confidence-based prototype selection are insufficient. Furthermore, the reported correlation between Prototype Shift and accuracy does not establish a causal relationship. These limitations make it difficult to determine whether the proposed mechanism is responsible for the gains.

Questions for authors
  • The TIP mechanism appears closely related to earlier prototype-based test-time adaptation methods, such as T3A (Iwasawa and Matsuo, NeurIPS 2021) and ASSESS (Zhao et al., ICML 2025). Could you clarify what is fundamentally new in this approach beyond selecting reliable pseudo-labelled target samples and updating class prototypes? These relevant studies also appear to be missing from the related work.

  • The paper does not discuss several relevant recent methods, such as GFMate (Jiang et al., ICML 2026), TSA (Wang et al., IJCAI 2025), and ZPT (Parameswaran et al., WSDM 2026). How does TI-GFM differ from these approaches? Can any of them be compared under the same zero-shot or few-shot evaluation settings?

  • The improvements are good, but how much of the gain comes from perturbation-consistent prototype refinement instead of stronger pretraining? Do the existing experiments establish that TIP performs better than simpler confidence-based pseudo-label selection when using the same backbone?

  • Consistent predictions across perturbed views are not necessarily correct. Could you provide evidence on the accuracy and class-wise coverage of the selected pseudo-labels? Also, how should the effectiveness of the invariance criterion be interpreted when the best performance on some datasets is obtained with only one view?

  • Since PS requires ground-truth labels and is mathematically related to a class-separation margin, could you clarify its contribution beyond serving as an evaluation metric? Does the reported correlation with accuracy provide sufficient evidence that reducing prototype shift is responsible for the improvements?

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

This paper studies prototype drift in text-attributed graph foundation models (TAG-GFMs) under cross-domain transfer. The authors define a heuristic metric called Prototype Shift (PSPS) and propose TI-GFM, which uses contrastive Text-Attributed Graph Pre-training (TGP) and , Test-Time Invariant Prototype (TIP) refinement based on prediction consistency across perturbed graph views.

Critical strengths

Tackles a genuine problem: cross-domain misalignment between text-prompt prototypes and target graph embeddings.

Extensive empirical evaluation with 18 baselines across 7 target datasets.

The authors provide supplementary implementation details and complexity analyses.

Critical weaknesses

Despite the title claiming "Prototype Learning", no learning or gradient updates occur at test time. TIP is simply Test-Time Augmentation (TTA) combined with consistency-filtered pseudo-labeling and centroid averaging. This is a standard heuristic and does not introduce new representation learning principles.

Eq. (4) simply checks the signs of the first-order partial derivatives of an arbitrarily defined fraction Φ+−Φ−Φ++Φ−+ϵ\frac{\Phi^+ - \Phi^-}{\Phi^+ + \Phi^- + \epsilon}. Calling this an "analytical proof linking PSPS to transfer performance" is an overclaim. It provides no formal generalization bounds, excess risk bounds, or domain-divergence guarantees.

In Eq. (14), constructing PP requires transductively accessing \textit{all} NtN_t unlabeled target nodes. Comparing this against instance-level, inductive baselines that do not share information across test nodes is unfair. The paper needs to evaluate an inductive setup where test nodes arrive sequentially or in small mini-batches.

In Appendix B, the authors report a Pearson correlation r=0.7225r = 0.7225 between ΔPS\Delta PS and downstream gain. This is computed across only N=7N = 7 target datasets (df=5df = 5). With such a small sample size, the estimate has a wide confidence interval and is easily driven by a single outlier (e.g., CiteSeer). Claiming this establishes a quantitative foundation is statistically weak.

Figure 6 shows the model is sensitive to the fusion weight λ\lambda, with sharp drops when λ≠0.3\lambda \neq 0.3. In a true cross-domain zero-shot scenario, target labels do not exist to tune this parameter. The authors must explain how λ\lambda can be determined without implicitly snooping on the target set.

On Instagram---the only non-academic, user-generated web graph in the benchmark---the model fails to beat baselines (Table 1) and shows unstable few-shot scaling (Figure 9e). This indicates the framework relies heavily on clean, synthetic text and breaks down under typical web noise.

Questions for authors

How does zero-shot performance hold up if target nodes are evaluated in an inductive/streaming regime (e.g., mini-batch size = 32) where global test centroids cannot be computed?

Can the authors provide a principled, unsupervised heuristic to determine λ\lambda per dataset without target ground truth?

What is the actual end-to-end inference wall-clock time including graph perturbation, KK forward passes, and prototype aggregation, compared to standard one-pass inference?

Reviewer_WR25OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TI-GFM, a test-time invariant prototype learning framework for text-attributed graph foundation models. First, this paper introduces a Prototype Shift metric to quantify the misalignment between semantic prototypes and target-domain class distributions. This method contains two components: TGP and TIP. The former uses perturbations to improve the robustness during pre-training. The latter utilizes the consistency of different perturbed views to optimize the semantic prototype. Experiments are conducted on seven target datasets. The results show TI-GFM improve performance on zero-shot and few-shot node classification and zero-shot link prediction.

Critical strengths

S1. This paper studies a meaningful problem: how to improve the reliability of semantic prototypes when transferring GFMs to a unseen domain prototypes.

S2. The method is clear and easy to understand. It is an interesting idea that use prediction consistency across different graph view to refine the prototypes.

S3. The experiments are extensive. The paper evaluates several downstream tasks and provides useful ablation studies.

Critical weaknesses

W1. The invariant prototype construction process is not fully clear. According to Eq 13 and 14, nodes whose predicted labels differ across different views are assigned zero weights in matrix R. However, these nodes might participate in prototype construction because of using softmax to R (these positioin still gain positive weight).

W2. The fairness of the experimental comparison is not fully clear. Table 1 directly uses the baseline results reported in GraphCLIP, however Appendix H.3 shows that TI-GFM utilizes LLM-generated descriptions. These may directly affect the semantic prototypes. I believe adding a new comparison with GraphCLIP under exactly same settings will make these conclusion more convincing.

Questions for authors

See W1~2

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.