acceptodds
Under review as a conference paper at ICLR 2027

Adapting Tabular Foundation Models for Graph Node-Level Tasks

Abstract

While foundation models have revolutionized fields such as natural language processing and computer vision, their potential in graph machine learning remains largely unexplored. One of the key challenges in designing graph foundation models (GFMs) is handling diverse node features that can vary across different graph datasets. While many works on GFMs have focused exclusively on text-attributed graphs, the problem of handling arbitrary features of other types in GFMs has not been fully addressed. However, this problem is not unique to the graph domain, as it also arises in the field of machine learning for tabular data. In this work, motivated by the recent success of tabular foundation models (TFMs) like TabPFNv2 and LimiX, we propose G2T-FM, a simple framework that allows tabular foundation models to be applied to graph node-level tasks. Specifically, G2T-FM augments the original node features with neighborhood feature aggregation, adds structural embeddings, and then applies a TFM to the constructed node representations. Even in the in-context learning setting, our model achieves strong results when combined with a strong TFM, outperforming both prior GFMs and well-tuned GNNs trained from scratch. Moreover, after finetuning, G2T-FM consistently surpasses well-tuned GNN baselines, often by a significant margin. In summary, our paper reveals the potential of a previously overlooked direction: utilizing tabular foundation models for graph machine learning tasks. Our source code is available at https://anonymous.4open.science/r/8456b37f41900b.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
prior-data fitted networks, graph machine learning, graph neural network, foundation models, tabular foundation models, LimiX, TabPFN

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_HjETOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents G2T‑FM, a lightweight framework for adapting tabular foundation models to graph node‑level tasks. Instead of modifying the internal tabular foundation model, it augments node representations with graph topology information via three components: neighborhood‑feature aggregation (NFA), hand‑crafted structural features, and learnable PEARL structural encodings, converting graph‑node prediction into a tabular learning problem. Built upon off‑the‑shelf TFMs such as TabPFNv2 and LimiX, it supports both in‑context‑learning (ICL) and finetuning (FT) modes. Evaluated on multiple graph benchmarks including GraphLand covering classification and regression, G2T‑FM outperforms existing graph foundation models under ICL, and its finetuned version frequently surpasses well‑tuned vanilla GNN baselines by considerable margins. Ablation studies verify the necessity of each graph‑aware augmentation component; even when GNNs and LightGBM are fed with identical augmented features, G2T‑FM retains performance advantages. This work opens a new direction for solving graph‑learning problems using tabular foundation models, though it is currently limited to node‑level tasks and constrained by inherent limits of the underlying tabular foundation models on sample size and class number.

Critical strengths
  1. Propose a novel cross-modal migration approach: Pioneer research directions for using table foundation models for graph learning, without training graph foundational models from scratch, reusing mature table large models such as TabPFNv2 and LimiX. The framework features plug-and-play functionality, and future iterative upgrades of table foundation models will benefit G2T-FM directly without large-scale changes.
  2. Modular decoupled design, no changes to the table model kernel. All graph topology information is placed in the pre-feature augmentation stage, completed by three modules: NFA neighborhood aggregation, manual structure features, and PEARL encoding. The table base model section retains the original weights and inference logic; Supports two usage paradigms: Context-learned ICL and fine-tuning FT, offering flexible usage methods.
  3. Complete and fair experimental comparisons, with rigorous baseline configuration. Selected datasets such as GraphLand, which closely reflect the real distribution of industries, while also covering classification and regression tasks; The comparison objects include both traditional fully tuned GNNs (GCN, GraphSAGE, GAT, GT) and multiple public graph foundation models; Traditional GNN baselines perform complete hyperparameter searches, adding residuals and layer normalization enhancement components to avoid inflated effects caused by weak baselines; At the same time, I compared it with my work tag and also tested my speed.
  4. Sufficient ablation and controlled experimental evidence of the source of gain. Item-by-item ablation of the three graph augmentation modules proves that each component contributes to performance; Controlled experiment: feeding the full set of augmented features used by G2T-FM to GNN and LightGBM, it was found that even with identical features, G2T-FM still had the advantage, proving that the benefits do not come solely from feature engineering, but from the synergy between the table base model and graph augmentation components; It also validates the effect gains brought by integrated reasoning.
  5. Clarify common misconceptions in evaluating existing graph foundation models. Experiments have found that many published foundation graph models perform worse than fully tuned traditional GNNs on datasets with a higher proportion of real and training samples; It points out that some literature confuses the terminology between in-context learning and zero-shot, providing a reference benchmark for subsequent evaluation of graph foundation models.
  6. Support for node return tasks. Most public graph foundation models focus only on node classification. This framework, relying on the native capabilities of table foundation models, naturally supports node regression tasks, fills gaps in graph foundation models in regression scenarios, and better meets the diverse needs of industrial scenarios.
Critical weaknesses
  1. Graph information is injected only through prerequisite feature engineering, lacking a native deep interaction mechanism for the graph. The framework itself does not modify the table base model; the graph topology information relies entirely on feature augmentation (NFA, structural features, PEARL) during the preprocessing stage, and there are no native graph learning modules such as message passing or graph attention within the model; The upper limit of modeling capability for long-distance, dependent, complex high-order graph relationships is limited by manual/preprocessed features. The paper also mentioned that the initial experiment with multi-hop NFA did not yield stable returns.
  2. Strong performance depends on the inherent constraints of the underlying table foundation model. Inheriting the limitations of TFM itself: for example, TabPFNv2 natively has restrictions on the number of classes and the size of training samples, which can restrict graph tasks with many classes and a large number of nodes; The framework itself cannot solve these problems and can only wait for upstream table base model iterations.
  3. Task coverage is narrow, only supporting node-level tasks. Currently, only node classification and node regression have been completed, with no adaptation for edge prediction or full-map classification/regression tasks; The paper also acknowledges that migrating to graph-level tasks is not easy, lacking corresponding solutions and experimental validation, and that there is still considerable room for generalizability.
  4. Feature dimension expansion, some scenarios require PCA dimensionality reduction, introducing information loss risks. After encoding the original node feature concatenation with neighborhood statistics, degrees, PageRank, Laplace feature vectors, and PEARL, the feature dimensions increase significantly; Datasets such as avazu-ctr and city-reviews must use PCA compression to compress feature dimensions to avoid OOM. PCA compression loses some fine-grained feature information and may harm downstream performance.
  5. High inference overhead, and integrated inference further amplifies costs. By default, 10 forward-propagation rounds are used for integration to improve effectiveness; From the perspective of appendix running times, single-round inference on some datasets consumes a very long time; Although faster than TAG at the same time, compared to native lightweight GNN and LightGBM, inference speed is less disadvantaged, and deployment on large scales is under greater pressure.
  6. Insufficient interpretability and controllability of the PEARL module. In the ICL mode, PEARL encoding is generated directly by randomly initializing GCN, without downstream graph data training, yet it can generate partial dataset benefits; The mechanism behind this phenomenon is not clearly explained; the superparameters (GCN layers, output dimensions, number of random initializations) lack systematic mesh search tuning.
  7. Cross-map zero-sample generalization capability not fully verified. The paper mainly evaluates under same image forwarding settings (the test set and training set belong to the same graph), lacking true cross-dataset zero-sample/few-shot experiments across different graphs; ICL is in-context learning within the same graph, not cross-graph migration, making it difficult to prove that the framework can directly transfer to completely unseen new graphs.
  8. Insufficient testing of heterogeneous and extreme graph structures. Although the GraphLand benchmark is used, systematic testing of extreme scenarios is lacking, such as extremely sparse graphs, highly heterogeneous plots, and large-scale power-law distribution maps; It is unclear whether the NFA + manual structure feature augmentation scheme can still work stably when graph topology significantly deviates from the existing test set distribution.
Questions for authors
  1. G2T-FM relies entirely on feature augmentation injection during the preprocessing stage, and there is no native graph messaging mechanism within the model. Since the initial experiments with multi-hop NFA have not brought stable returns, will they consider embedding lightweight, learnable messaging modules within table foundation models in the future, rather than just at the feature engineering level to further capture high-level, long-distance graph dependencies?

  2. All current experiments are based on forwarding context learning settings within the same graph, without true cross-graph zero-shot generalization across different graph datasets. How do you distinguish between "ICL capability within the same map" and cross-map migration capability? How do you plan to design experiments to verify the G2T-FM migration effect on completely unseen new maps?

  3. Framework performance largely inherits the limitations of the underlying table base model, such as upper limits on the number of classes and training sample size. When faced with graph tasks with an extremely large number of nodes and many categories, what feasible strategies can alleviate the performance bottleneck caused by sample and class count without replacing the upstream TFM?

  4. Some high-dimensional datasets must use PCA to reduce dimensionality of stitched augmented features to prevent memory overflow, which can cause information loss. Is there a comparison of how different dimensionality reduction schemes (such as feature selection or sparse projection) affect the final metrics? Is there a risk that PCA compression could damage key topological features, leading to significant performance degradation?

  5. In the ICL mode, random initialization without training with PEARL encoding can still improve some dataset performance, but its internal mechanism still lacks explanation. PEARL's hyperparameters such as GCN layer count, output dimension, and random sampling frequency are only fixed configurations and do not perform a full mesh search. Will there be further ablation to clarify the conditions for PEARL to take effect and provide more general guidance for selecting superparameters?

5 comments
AuthorsOpenReview ↗

Official Comment (Part 1)

Dear Reviewer HjET,

Thank you for your detailed feedback and for the time and effort in reviewing our work. We address your concerns and questions below.

W1, Q1. First, we note that there is a follow-up work, GraphPFN, that implements your idea of adding native graph message passing to a TFM. Second, we argue that the lack of pretrained graph layers is both an advantage and a disadvantage for G2T-FM. On the one hand, G2T-LimiX performs worse than GraphPFN, which is based on the LimiX backbone. On the other hand, G2T-FM's plug-and-play design allows it to easily use any existing or future TFM, and using a more recent TFM, Kumo-Tabular, allows G2T-FM to outperform GraphPFN. In contrast, GraphPFN lacks such a plug-and-play capability, and migrating GraphPFN to a new TFM backbone would require adapting MPNN layers to the new architecture and pretraining them from scratch. Thus, we believe that reliance on feature engineering is an important trait of G2T-FM, which allows it to remain a relevant baseline even as GFMs develop further. We refer to our general response for further details on more recent TFM backbones and the comparison with GraphPFN.

W2, Q3. First, we note that the TFM field is rapidly developing. For example, more recent models like TabPFN-3.5 claim to handle datasets with up to a million training samples or natively handle more than 10 classes.

Second, we note that at least some limitations can be alleviated by our method. Specifically, G2T-FM supports error-correcting output codes, a common technique to extend TFMs beyond their native class limit. Below, we provide a comparison of G2T-FM with ECOC against supervised GNNs. We see that it performs competitively, matching or outperforming GNNs on lastfm-asia and roman-empire, though we also observe a substantial drop on hm-categories.

lastfm-asiaroman-empirehm-categories
GCN82.77 ± 0.5468.47 ± 0.6063.53 ± 0.25
GraphSAGE80.54 ± 0.2473.85 ± 0.4656.03 ± 0.65
GAT82.78 ± 0.2369.58 ± 0.6467.60 ± 0.51
GT83.33 ± 0.3571.49 ± 0.8069.27 ± 0.36
G2T-LimiX (ICL)84.68 ± 0.1373.20 ± 0.4258.10 ± 0.27

Third, one can potentially alleviate a TFM's limits on the number of training samples or features by subsampling them, though we would expect a considerable drop in performance in this case. Since scaling to larger datasets is not specific to graphs, we believe exploring such approaches is outside the scope of this work.

W3. We clearly acknowledge this limitation in the Limitations section.

W4, Q4. We acknowledge that PCA may lead to a loss of information, though we note that the original features and NFA are compressed, while all other features (PEARL embeddings, LapPE, etc.) are not compressed because they have a fixed, relatively low dimension. Since PCA is a temporary workaround rather than a principled component of G2T-FM, providing a detailed analysis of different compression techniques is beyond the scope of our work, and we believe that more scalable TFMs would reduce the need for such workarounds.

W5. Scalability and efficiency remain important directions for future work. However, we note that G2T-FM significantly outperforms all prior GFMs considered in this study, justifying the runtime and memory overhead.

W6, Q5. First, we note that we deliberately do not optimize PEARL hyperparameters. Optimizing them for each dataset would substantially increase the runtime of our method and introduce a tuning stage that we try to avoid in the ICL setting. Optimizing them based on validation scores on a collection of datasets is possible, but would introduce a risk of overfitting to these datasets and would limit our ability to evaluate transfer to fully unseen graph datasets.

Second, we note that message passing already introduces certain inductive biases into a model. Thus, while this phenomenon is indeed interesting, it is not completely surprising. For example, even the original GCN paper (Kipf and Welling, 2017) shows in Appendix A.1 that embeddings from a randomly initialized GCN could carry some signal.

AuthorsOpenReview ↗

Official Comment (Part 2)

W7, Q2. First, we highlight that we clearly state in our paper that G2T-FM works in the ICL (or FT) setting, and all relevant experiments follow this setting. We argue that the ICL setting is established and widespread in both the graph domain and other domains, including the tabular domain. We do not claim that our model works in a zero-shot setting.

Second, we note that true zero-shot transfer requires access to meta-information, such as a description of the task and target space in natural language. Without such meta-information or any labels, a model would not know what label "0" means compared to label "1", making predictions impossible. While this setup is valid, it falls far beyond the scope of our work.

Some confusion may arise from previous GFMs incorrectly claiming to work in a zero-shot setting, thus suggesting that it is a standard protocol for GFM evaluation. However, as we discuss in Section 2.2, these methods actually have access to labels from the same graph and thus cannot be called zero-shot. In fact, these methods use labels of training nodes from the same graph as the test nodes, matching the evaluation protocol we use. Thus, we believe our evaluation protocol is appropriate and common for GFM evaluation.

Finally, we note that the underlying TFMs have never seen the graphs considered here during pretraining, and the G2T-FM components are training-free. Therefore, we believe that our main experiments already show transfer to completely unseen graphs.

W8. We acknowledge that G2T-FM, like any other model, may fail under extreme conditions. However, we believe that, at present, GraphLand is the graph benchmark closest to real-world applications, and thus we focus on this benchmark in our evaluations.

Best regards,

Authors

Dear Reviewer HjET,

Since the discussion period is ending soon, please let us know if our response addresses your concerns, and if you have any further comments.

Thank you for your time and effort in reviewing our paper.

Best regards,

Authors

Reviewer_Hai9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how tabular foundation models can be applied to graph node-level prediction tasks. The authors propose G2T-FM, which augments the original node features with neighborhood feature statistics, node degree, PageRank, Laplacian positional encodings, and PEARL structural encodings. The resulting graph-informed tabular representations are then used by TabPFN models or LimiX for node classification and regression. The paper evaluates both in-context learning and fine-tuning on 13 datasets. The authors report that, with stronger tabular backbones, G2T-FM achieves better average rankings than the traditional GNN baselines considered, while fine-tuning generally provides further improvements.

Critical strengths
  1. Simple and adaptable framework. G2T-FM offers a straightforward way to use existing tabular foundation models for graph node-level tasks. Its graph-to-table preprocessing can be paired with different tabular backbones without graph-specific pretraining.

  2. Useful empirical comparisons. The evaluation covers both classification and regression across 13 datasets, compares against tuned GNN baselines, and includes in-context learning, fine-tuning, and component ablations. These results make the approach a useful baseline for future work.

Critical weaknesses

1.The label budget in the main comparisons is unclear. The paper states that TabPFNv2 is suited to training sets of roughly 10,000 samples or fewer, yet uses 10% of the nodes for training. Based on Table 1, this exceeds 10,000 nodes for city-reviews and twitch-views. The authors should report the number of labeled samples actually supplied to each model and any sampling rule. If the methods used different label budgets, the main comparisons would need to be repeated.

2.The contribution relative to closely related work is insufficiently established. The appendix reports performance comparable to TAG under the same evaluation setup, while both approaches construct graph-derived features for a tabular foundation model. Although G2T-FM reports a runtime advantage, the paper should clarify which design choices yield a consistent, attributable methodological contribution.

Questions for authors
  1. Clarification of the label budget. Under the stated 10% training split, city-reviews and twitch-views each have more training nodes than the roughly 10,000-sample range the paper identifies as suitable for TabPFNv2. How many labeled nodes were actually supplied to each G2T-FM variant during in-context inference and fine-tuning? Was any manual or automatic subsampling applied, and did the GNN baselines use the same labeled nodes?

  2. Clarification of the contribution relative to TAG. Table 6 reports broadly comparable performance for G2T-FM and TAG. Which specific design choices distinguish G2T-FM from TAG, and what evidence shows that they provide a consistent benefit? Were the reported runtimes measured on matched hardware, and do they include graph feature construction, PEARL computation, and the ten-member inference ensemble?

  3. Evidence needed to isolate the graph-to-table contribution. How does the same tabular foundation model perform using only the original node features, under identical data splits and inference settings? Every component-removal experiment in Table 4 retains at least one form of graph-derived augmentation, so this comparison would clarify how much of the gain comes from the proposed graph-to-table step.

  4. Robustness to key hyperparameters. Do the main performance conclusions relative to GNNs and TAG hold when the Laplacian positional encoding dimension, the number of random inputs used by PEARL, or the PCA dimension is varied? How were these values selected? Table 5 provides a comparison without ensembling, but does not address sensitivity to these other settings.

5 comments
AuthorsOpenReview ↗

Official Comment (Part 1)

Dear Reviewer Hai9,

Thank you for your review and valuable suggestions. We address your concerns and questions below. We also refer to our general response for additional experiments and discussion.

W1, Q1. All methods received the same set of labeled nodes, without any subsampling for any method. 10,00010{,}000 samples is a recommended limit for TabPFNv2, but TabPFNv2 handles datasets beyond this count.

W2, Q2. First, we would like to highlight that, as already mentioned in our paper, G2T-FM became publicly available before TAG. Thus, we believe that these methods should be treated as concurrent. That said, let us describe the key differences between the methods beyond what is already mentioned in Appendix D. First, TAG employs different graph-to-table features: feature smoothing over multiple hops instead of our one-hop NFA, RandomWalkPE instead of degree and PPR, and GPSE instead of PEARL. Second, in addition to a TFM, TAG also includes closed-form LinearGNNs in its ensemble. Third, TAG's ensembling is heavier: its 10-member ensembling is applied on top of built-in TFM ensembling and employs node and feature subsampling, out-of-fold prediction, and more sophisticated ensemble aggregation. We hypothesize that the main reason for TAG's longer runtime is its ensembling, though it is also likely that TAG produces more features than G2T-FM. Regarding runtime comparisons, they were conducted on matched hardware, and for G2T-FM they cover the end-to-end pipeline, including feature construction, PEARL, and ensembling.

Q3. Thank you for suggesting this ablation. We provide the results below. In short, we observe that G2T-FM models consistently outperform the corresponding TFM backbones, with more prominent gains on certain datasets (e.g., twitch-views).

artnet-expcity-reviewstolokers-2artnet-viewsavazu-ctrcity-roads-Mhm-pricestwitch-viewsAR
LightGBM-NFA46.13 ± 0.0478.53 ± 0.0156.34 ± 0.0656.10 ± 0.0231.71 ± 0.0161.18 ± 0.0370.84 ± 0.0460.14 ± 0.014.62
GCN44.86 ± 0.3677.81 ± 0.1556.27 ± 0.3156.03 ± 0.2532.00 ± 0.1658.82 ± 0.2568.02 ± 0.4275.51 ± 0.055.75
GraphSAGE45.14 ± 0.3678.17 ± 0.1054.43 ± 0.3449.32 ± 0.9131.44 ± 0.1659.44 ± 0.2770.00 ± 0.7466.29 ± 0.326.25
GAT45.06 ± 0.5277.74 ± 0.2157.41 ± 0.8553.60 ± 0.2432.63 ± 0.1759.86 ± 0.2072.07 ± 1.2272.89 ± 0.274.38
GT46.41 ± 0.7177.34 ± 0.2156.98 ± 0.5553.37 ± 0.4631.11 ± 0.4959.55 ± 0.2869.44 ± 0.9472.13 ± 0.135.62
LimiX (ICL, L)46.08 ± 0.0175.80 ± 0.0146.29 ± 0.0835.87 ± 0.0126.20 ± 0.0162.66 ± 0.0868.52 ± 0.0210.23 ± 0.077.75
G2T-LimiX (ICL, L)48.44 ± 0.2377.29 ± 0.5461.13 ± 0.2160.95 ± 0.0932.41 ± 0.1264.53 ± 0.0975.41 ± 0.0471.08 ± 0.073.12
Kumo-Tabular (ICL, L)48.38 ± 0.1277.18 ± 0.0646.35 ± 0.2539.74 ± 0.1328.21 ± 0.0962.33 ± 1.0072.18 ± 0.0713.32 ± 0.016.25
G2T-Kumo-Tabular (ICL, L)51.84 ± 0.2780.70 ± 0.0463.02 ± 0.3863.84 ± 0.0735.17 ± 0.0766.41 ± 0.1877.97 ± 0.2972.58 ± 0.021.25
AuthorsOpenReview ↗

Official Comment (Part 2)

Q4. Following your suggestion, we conducted ablations of G2T-LimiX (ICL, L) and provide the results and our key observations below.

Regarding the LapPE dimension, we provide the results in the table below. We see that G2T-LimiX is robust to changes in the LapPE dimension and, in particular, all variants outperform GNNs on average.

artnet-expcity-reviewstolokers-2artnet-viewsavazu-ctrcity-roads-Mhm-pricestwitch-viewsAR
LightGBM-NFA46.13 ± 0.0478.53 ± 0.0156.34 ± 0.0656.10 ± 0.0231.71 ± 0.0161.18 ± 0.0370.84 ± 0.0460.14 ± 0.015.62
GCN44.86 ± 0.3677.81 ± 0.1556.27 ± 0.3156.03 ± 0.2532.00 ± 0.1658.82 ± 0.2568.02 ± 0.4275.51 ± 0.056.38
GraphSAGE45.14 ± 0.3678.17 ± 0.1054.43 ± 0.3449.32 ± 0.9131.44 ± 0.1659.44 ± 0.2770.00 ± 0.7466.29 ± 0.327.12
GAT45.06 ± 0.5277.74 ± 0.2157.41 ± 0.8553.60 ± 0.2432.63 ± 0.1759.86 ± 0.2072.07 ± 1.2272.89 ± 0.275.00
GT46.41 ± 0.7177.34 ± 0.2156.98 ± 0.5553.37 ± 0.4631.11 ± 0.4959.55 ± 0.2869.44 ± 0.9472.13 ± 0.136.50
LapPE dim=448.29 ± 0.2077.34 ± 0.6560.68 ± 0.2359.24 ± 0.0732.66 ± 0.1465.00 ± 0.1575.38 ± 0.0370.32 ± 0.064.00
LapPE dim=848.35 ± 0.2977.29 ± 0.7061.05 ± 0.1560.18 ± 0.0632.63 ± 0.1464.63 ± 0.1375.39 ± 0.0370.76 ± 0.053.88
LapPE dim=1648.42 ± 0.2877.84 ± 0.3660.97 ± 0.2760.98 ± 0.0732.12 ± 0.2164.64 ± 0.1375.40 ± 0.0471.29 ± 0.072.88
LapPE dim=3248.84 ± 0.3377.74 ± 0.6661.22 ± 0.1861.62 ± 0.0730.57 ± 0.6764.62 ± 0.0875.53 ± 0.0771.38 ± 0.073.25

Regarding PCA, we note that it serves as a workaround to evaluate models on datasets that otherwise do not fit into memory, rather than as a principal G2T-FM component. That said, we vary the PCA dimension on the two GraphLand datasets that require it and note that other GraphLand datasets are evaluated without PCA. We see that increasing the PCA dimension from 64 (used in the main experiments) to 128 improves performance, though at the cost of increased runtime and memory consumption.

city-reviewsavazu-ctr
LightGBM-NFA78.53 ± 0.0131.71 ± 0.01
GCN77.81 ± 0.1532.00 ± 0.16
GraphSAGE78.17 ± 0.1031.44 ± 0.16
GAT77.74 ± 0.2132.63 ± 0.17
GT77.34 ± 0.2131.11 ± 0.49
PCA dim=1677.50 ± 0.0931.66 ± 0.19
PCA dim=3278.64 ± 0.0830.58 ± 0.28
PCA dim=6477.29 ± 0.5432.41 ± 0.12
PCA dim=12878.44 ± 0.2332.77 ± 0.14

For PEARL, we would like to highlight the ambiguity of the term "number of random inputs". If you meant the number of random features that PEARL receives, then it is always fixed at 1, according to the original paper. If you meant the number of samples (MM in the original paper's notation) over which PEARL is averaged, then please let us know, and we would be happy to conduct this ablation.

Best regards,

Authors

Thank you for the detailed response and the additional experiments. Your clarification that all methods use the same labeled nodes, without subsampling, resolves my concern about unequal label budgets in the main comparison.

The new raw-feature comparisons show that graph-derived features benefit LimiX and Kumo-Tabular. Two further experiments would help resolve my remaining concerns:

Raw-feature controls for the main backbones. Under identical data splits, labeled nodes, and inference settings, compare TabPFNv2 and TabPFNv2.5 using raw node features against their G2T-FM counterparts.

A stage-by-stage efficiency comparison with TAG. Confirm that the reported runtimes cover the same stages for both methods, and break them down into feature construction, model execution, and ensembling. If the paper attributes the efficiency difference to particular design choices, please provide a controlled comparison with the TFM backbone and ensemble size held fixed. Otherwise, present that attribution as a hypothesis.

Reviewer_a3Q1OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper demonstrates that tabular foundation models (TFMs) can also perform graph-based tasks by augmenting node features with graph information. By applying the proposed method to TabPFN and LimiX, the authors achieve superior performance compared to existing graph neural network (GNN) baselines and graph foundation models. The study is significant in that it introduces a "plug-and-play" paradigm, which remains highly versatile as TFMs continue to evolve, and applies it to node-level tasks.

Critical strengths
  1. In the rapidly evolving field of tabular foundation models (TFMs), this work makes a significant contribution by uncovering the value of TFMs beyond standard tabular tasks. The authors propose a universal interface designed to make any TFM applicable to node-level tasks.
  2. By modeling features at the local, global, and structural levels of a graph, the approach demonstrates strong and meaningful performance across both homophilous and heterophilous benchmarks when integrated with robust TFM architectures.
Critical weaknesses
  1. PFN-based graph foundation models, such as GraphPFN [1] and NodePFN [2], already exist. Compared to these models, the advantage of G2T-FM lies in its "plug-and-play" nature, which eliminates the need for pretraining; however, the experimental exploration of this plug-and-play capability remains insufficient. Furthermore, while the fine-tuned model achieves the highest average performance in the main tables, there is no analysis comparing its memory and computational costs with those of the baselines. Most importantly, it appears that GraphPFN and NodePFN were excluded from the baselines on the grounds that they require pretraining, as discussed in Appendix D. However, since other baselines already included in the comparison (e.g., AnyGraph and OpenGraph) also require pretraining, I believe GraphPFN and NodePFN should be included as baselines to ensure a fair comparison.
  2. The current experimental results do not sufficiently demonstrate the advantages of G2T-FM in terms of feature construction. To substantiate the advantages of the features employed by G2T-FM, it would be necessary to include baselines that use alternative feature constructions beyond TAG. Furthermore, the robustness of the approach should be demonstrated through feature-level ablations across various TFM backbones. Additionally, since the effectiveness of each feature varies across datasets, it would be valuable to include an analysis of these differences from the perspective of graph structure.

[1] https://openreview.net/forum?id=340Ep8sfJ3

[2] https://openreview.net/forum?id=FmxRzlu0rT

Questions for authors
  1. The paper uses only TabPFN-v2/2.5 and LimiX as tabular backbones. Could you include results with more recent state-of-the-art tabular backbones, such as TabPFN-3, LimiX-2, TabFM, and TabICLv2?
  2. Could you provide the detailed experimental settings for fine-tuning (e.g., epochs, steps, batch size, hardware, and memory), as well as a comparison of computational costs between G2T-FM and the baselines?
  3. Could you add GraphPFN and NodePFN as baselines?
  4. The contribution of each feature appears to vary across benchmarks. Analyzing the characteristics of each benchmark alongside the impact of these features could yield more insightful findings. Could you provide such an analysis?
  5. The proposed methodology may have limitations, either in terms of information that the features fail to capture or in terms of the model's inherent capabilities. Presenting failure cases could offer valuable insight. Could you include such failure cases, if any, and explain whether they stem from insufficient information in the features or from limitations of the model's capabilities?
3 comments
AuthorsOpenReview ↗

Official Comment (Part 1)

Dear Reviewer a3Q1,

Thank you for your thorough review and valuable suggestions. We address your concerns and questions below.

W1, Q1, Q3. First, let us note that GraphPFN and NodePFN were excluded from evaluation because they were published after the initial version of G2T-FM, not because they require pretraining. TAG was published after G2T-FM too, but we made an exception for it since it is much closer to G2T-FM than GraphPFN and NodePFN are.

That said, we agree that our paper would benefit from including comparisons between GraphPFN and G2T-FM with more recent TFM backbones. We refer to our general response for additional results for G2T-FM with more recent TFM backbones and comparisons with GraphPFN and NodePFN. These results support our claim that G2T-FM remains a strong baseline even with further development of GFMs, thanks to its plug-and-play ability to employ any existing or future TFM.

W1, Q2. Regarding finetuning settings, as we describe in our paper, we follow the recipe from Rubachev et al. (2025). Specifically, we used full finetuning with AdamW and zero weight decay, selecting the learning rate from 10 logarithmically spaced values between 5⋅10−65 \cdot 10^{-6} and 5⋅10−45 \cdot 10^{-4}. Each epoch comprised 10 optimizer steps, followed by validation. Training stopped after 16 epochs without validation improvement, with no fixed epoch limit. Each step used min⁡{1024,⌊0.25⋅Ntrain⌋}\min \{ 1024, \lfloor 0.25 \cdot N_{\text{train}} \rfloor \} query nodes as the supervised batch, with the remaining training nodes serving as context. The checkpoint with the best validation score was used for evaluation.

Regarding time and memory efficiency comparisons, we provide them below. For all methods, we perform a single evaluation with a single random seed, using hyperparameters from our main experiments. G2T-LimiX is evaluated without ensembling.

The table below contains peak memory consumption measured in GiB. We see that the memory consumption of G2T-LimiX is noticeably higher than that of GCN.

tolokers-2city-reviewsartnet-exphm-pricesavazu-ctrcity-roads-Mtwitch-viewsartnet-views
GCN0.565.852.12.533.692.176.591.93
G2T-LimiX3.549.4835.6146.736.8315.2140.1720.02

The table below shows the time in seconds required to preprocess a dataset ("simple" stands for loading and transforming it into the required format, while "G2T-FM" also includes computation of LapPE, node degree, etc.), to train or finetune a model, and to run inference on a dataset. Preprocessing and inference with G2T-LimiX together take at most about twice as long as a single GCN training run on the evaluated datasets. Thus, in the ICL regime, G2T-FM does not require orders of magnitude more runtime than GNN training. We also note that achieving the best results with GNNs requires extensive hyperparameter search, while ICL does not require tuning any hyperparameters. Finetuning G2T-FM is, however, noticeably more time-consuming.

tolokers-2city-reviewsartnet-exphm-pricesavazu-ctrcity-roads-Mtwitch-viewsartnet-views
Preprocessing: simple0.452.151.4611.1614.930.357.171.08
Preprocessing: G2T-FM1.1630.475.1427.8854.9242.4914.563.27
Inference: G2T-LimiX0.8727.2716.3920.7715.143.5821.325.54
Train: GCN25.7785.6641.5877.932.922.71327.4619.53
Finetune: G2T-LimiX139.521688.98832.071123.41013.35225.91680.59457.12

W2a. Regarding your point

To substantiate the advantages of the features employed by G2T-FM, it would be necessary to include baselines that use alternative feature constructions beyond TAG

Could you please clarify what you meant here? If you meant other methods based on an idea similar to G2T-FM and TAG, then we are only aware of one more such method, TabPFN-GN (Choi et al., 2025), for which no code is available at the time of writing. Moreover, we highlight that both TabPFN-GN and TAG became publicly available after the initial version of G2T-FM.

If you meant exploring alternative components for G2T-FM (e.g., replacing LapPE with InstantEmbeddings), please let us know, and we will do our best to conduct experiments during the discussion period.

AuthorsOpenReview ↗

Official Comment (Part 2)

W2b. Regarding your point

Furthermore, the robustness of the approach should be demonstrated through feature-level ablations across various TFM backbones

We conducted the ablations from Table 4 on one more tabular foundation model, Kumo-Tabular. The results appear to be consistent with the ablation results for G2T-LimiX in the paper.

artnet-expcity-reviewstolokers-2artnet-viewsavazu-ctrcity-roads-Mhm-pricestwitch-viewsAR
LightGBM-NFA46.13 ± 0.0478.53 ± 0.0156.34 ± 0.0656.10 ± 0.0231.71 ± 0.0161.18 ± 0.0370.84 ± 0.0460.14 ± 0.017.25
GCN44.86 ± 0.3677.81 ± 0.1556.27 ± 0.3156.03 ± 0.2532.00 ± 0.1658.82 ± 0.2568.02 ± 0.4275.51 ± 0.057.75
GraphSAGE45.14 ± 0.3678.17 ± 0.1054.43 ± 0.3449.32 ± 0.9131.44 ± 0.1659.44 ± 0.2770.00 ± 0.7466.29 ± 0.328.62
GAT45.06 ± 0.5277.74 ± 0.2157.41 ± 0.8553.60 ± 0.2432.63 ± 0.1759.86 ± 0.2072.07 ± 1.2272.89 ± 0.276.62
GT46.41 ± 0.7177.34 ± 0.2156.98 ± 0.5553.37 ± 0.4631.11 ± 0.4959.55 ± 0.2869.44 ± 0.9472.13 ± 0.137.88
G2T-Kumo-Tabular (ICL, L)51.84 ± 0.2780.70 ± 0.0463.02 ± 0.3863.84 ± 0.0735.17 ± 0.0766.41 ± 0.1877.97 ± 0.2972.58 ± 0.022.00
w/o NFA (ICL, L)50.15 ± 0.2080.37 ± 0.0662.95 ± 0.4263.31 ± 0.0634.74 ± 0.1164.53 ± 0.1174.83 ± 0.1471.71 ± 0.024.12
w/o SF & PEARL (ICL, L)51.15 ± 0.2180.18 ± 0.0559.21 ± 0.2660.95 ± 0.1234.88 ± 0.0965.70 ± 0.1978.06 ± 0.2862.40 ± 0.014.38
w/o SF (ICL, L)51.00 ± 0.1980.31 ± 0.0559.02 ± 0.4861.68 ± 0.0635.06 ± 0.1165.70 ± 0.2178.00 ± 0.1669.31 ± 0.034.00
w/o PEARL (ICL, L)51.90 ± 0.2080.66 ± 0.0662.85 ± 0.1663.81 ± 0.0635.21 ± 0.1266.47 ± 0.2178.20 ± 0.2168.47 ± 0.032.25

W2, Q4. We agree that such an analysis can be valuable and insightful. However, we do not see any obvious and straightforward patterns in the results we have, beyond those already mentioned in the paper. We believe that a more nuanced and fine-grained analysis would require far more than the 13 datasets we have to draw reliable conclusions. We are not aware of sources where we can obtain enough high-quality, reliable graph datasets for such an analysis.

Q5. First, as we acknowledge in the Limitations section, G2T-FM inherits all limitations and failure modes of a TFM backbone. For example, TabPFNv2 is limited to 10 classes, roughly 10,00010{,}000 training samples, and 500500 features. LimiX is likely to have similar constraints. Thus, we cannot scale G2T-FM with these backbones to large-scale graphs or graphs with sparse high-dimensional features like bag-of-words.

Second, our graph feature engineering naturally limits the complexity of graph patterns the model can capture. For example, G2T-FM cannot aggregate features beyond one hop, which may limit its performance on datasets that require long-range interactions. That said, a comparison of G2T-FM with GNNs and GraphPFN suggests that even the existing features already provide a surprisingly strong signal on generic real-world graph datasets.

Best regards,

Authors

As my concerns have been resolved, I'm increasing the rating to 3. Thanks to the authors for the detailed additional experimental results.

Best regards,

Reviewer a3Q1

Comments on the paper

1 comment

Dear Reviewers,

We thank you for your valuable feedback. In this general response, we provide additional results, comparisons, and discussion that we consider important for positioning our work.

Specifically, we argue that a key advantage of G2T-FM is its plug-and-play capability to employ any existing or future tabular foundation model. Given the rapid pace of TFM development, this capability can allow G2T-FM to remain a relevant baseline even with further development of GFMs. We now have an illustrative example of this.

Soon after the initial G2T-FM release, other PFN-based GFMs were proposed, including GraphPFN and NodePFN. In particular, GraphPFN adds message-passing adapters to the tabular foundation model LimiX and pretrains them on a synthetic prior. As one can see from the table below, more native handling of graph information allows GraphPFN to outperform G2T-LimiX on average.

However, GraphPFN lacks the plug-and-play capability to easily employ any TFM. Migrating GraphPFN to a new TFM backbone would require adapting MPNN layers to a new backbone architecture and pretraining them from scratch, without guaranteed success. In contrast, using a new TFM in the G2T-FM framework is straightforward, as G2T-FM simply converts graph tasks to tabular tasks.

When comparing GraphPFN and G2T-FM with more recent TFM backbones, TabFM [1] and Kumo-Tabular [2], we see that G2T-FM outperforms GraphPFN on average. In particular, G2T-Kumo-Tabular outperforms GraphPFN on 7 out of 8 datasets, with gains of about 2–4 percentage points on several datasets (tolokers-2, avazu-ctr, city-roads-M).

These results support our claim that G2T-FM can remain a strong baseline even with further development of GFMs, thanks to its easy integration with any TFM and the rapid pace of TFM development. We will revise our paper to include these results.

artnet-expcity-reviewstolokers-2artnet-viewsavazu-ctrcity-roads-Mhm-pricestwitch-viewsAR (cls)AR (all)
LightGBM-NFA46.13 ± 0.0478.53 ± 0.0156.34 ± 0.0656.10 ± 0.0231.71 ± 0.0161.18 ± 0.0370.84 ± 0.0460.14 ± 0.015.676.00
GCN44.86 ± 0.3677.81 ± 0.1556.27 ± 0.3156.03 ± 0.2532.00 ± 0.1658.82 ± 0.2568.02 ± 0.4275.51 ± 0.057.676.62
GraphSAGE45.14 ± 0.3678.17 ± 0.1054.43 ± 0.3449.32 ± 0.9131.44 ± 0.1659.44 ± 0.2770.00 ± 0.7466.29 ± 0.327.007.50
GAT45.06 ± 0.5277.74 ± 0.2157.41 ± 0.8553.60 ± 0.2432.63 ± 0.1759.86 ± 0.2072.07 ± 1.2272.89 ± 0.276.675.38
GT46.41 ± 0.7177.34 ± 0.2156.98 ± 0.5553.37 ± 0.4631.11 ± 0.4959.55 ± 0.2869.44 ± 0.9472.13 ± 0.136.337.00
NodePFN (ICL)24.19 ± 0.4170.58 ± 0.0643.79 ± 0.19----------10.00--
GraphPFN (ICL, L)51.46 ± 0.1180.01 ± 0.0360.97 ± 0.1362.07 ± 0.0531.43 ± 0.0664.03 ± 0.1177.33 ± 0.0672.75 ± 0.043.003.50
G2T-LimiX (ICL, L)48.44 ± 0.2377.29 ± 0.5461.13 ± 0.2160.95 ± 0.0932.41 ± 0.1264.53 ± 0.0975.41 ± 0.0471.08 ± 0.075.334.62
G2T-TabFM (ICL, L)50.73 ± 0.1180.50 ± 0.0663.00 ± 0.2061.36 ± 0.1034.01 ± 0.0564.62 ± 0.1874.58 ± 0.0572.02 ± 0.062.333.00
G2T-Kumo-Tabular (ICL, L)51.84 ± 0.2780.70 ± 0.0463.02 ± 0.3863.84 ± 0.0735.17 ± 0.0766.41 ± 0.1877.97 ± 0.2972.58 ± 0.021.001.38

Best regards,

Authors


[1] Kong et al., "TabFM: A Zero-Shot Foundation Model for Tabular Data," 2026.

[2] https://huggingface.co/blog/nvidia/kumo-tabular

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.