acceptodds
Under review as a conference paper at ICLR 2027

Graph Neural Diffusion as the Prior for In-Context Node Classification

Abstract

Prior-fitted networks (PFNs) are pretrained on synthetic tasks sampled from a prior and make predictions on a new dataset in a single forward pass. Recently, PFN-based node classifiers have emerged as competitive models that classify the nodes of a new graph without training on it. For such models, the prior determines how the graph structure shapes the node features, and thus what the network can learn about graphs. However, the priors of existing PFN-based node classifiers either generate the features independently of the graph or only aggregate them over neighbours, and therefore produce a narrow range of the relations between features and graph structure found in real graphs. Graph neural diffusion views message passing as the integration of a diffusion equation on the graph, and subsequent work has proposed a variety of propagation schemes by adding terms such as reaction and advection. In this work, we propose GadrPFN, a PFN for node classification whose prior generates each task through graph neural diffusion. In each task, a node state on a random graph evolves under an advection–diffusion–reaction equation for a randomly drawn time, so that diffusion smooths the state along the edges and reaction sharpens it against them, and the node features are read from this state. GadrPFN is trained from scratch on about half a million synthetic tasks in two hours on a single GPU, whereas the strongest in-context node classifiers build on tabular foundation models pretrained on far more data. On 28 benchmarks, GadrPFN achieves the best mean accuracy and average rank on both homophilous and heterophilous graphs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
node classification, graph neural diffusion, prior-fitted networks, in-context learning, synthetic priors, graph foundation models, heterophily

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_cm7FOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GadrPFN, a prior-fitted network for in-context node classification. To address the limitation that existing graph PFN priors generate narrow feature–structure relationships, the authors introduce concepts and operations from graph neural diffusion into the synthetic data prior. Specifically, hidden node states and labels evolve over random graphs for randomly sampled flow times. In tandem, the authors design a corresponding Advection–Diffusion–Reaction (ADR) graph module in the network architecture. The paper validates the efficacy of the proposed prior and the superiority of the model across various types of graphs through rigorous controlled ablation studies, prior task-space coverage analyses, and empirical probing of the network's learned non-linearity.

Critical strengths
  1. Incorporating advection, diffusion, reaction, and evolution time into the task generation pipeline provides a well-structured prior design space for graph PFNs. The separate label time is especially valuable, as it decouples feature smoothness from label homophily, directly reflecting the diverse feature-structure relationships observed in real-world graphs. This design offers actionable building blocks for understanding what synthetic tasks facilitate cross-graph transfer.
  2. GadrPFN outperforms TAG and GraphPFN on the 28 benchmarks by 3.30 and 3.62 percentage points, respectively, achieving the best average rank across both homophilous and heterophilous graph subsets. Furthermore, Figure 10 demonstrates that the full prior consistently yields higher accuracy curves under a unified inference configuration, providing additional evidence for the effectiveness of the prior design. Together, these results strongly support the practical potential of the method.
Critical weaknesses
  1. Experimental Fairness and Hyperparameter Selection:
  • In Appendix B, the authors state that in the controlled experiment of Table 2, "The two existing priors are evaluated under NodePFN's released settings on both networks and the GadrPFN prior under the per-dataset settings of Table 1." If GadrPFN benefits from a broader, re-tuned per-dataset hyperparameter search (including tuning the dual-head MLP weight tailored to "Ours" network) while the competing priors on the "Ours" network are locked to NodePFN's static released settings, the reported +5.1% gain cannot be considered a strictly fair apple-to-apple comparison attributed purely to the prior.
  • Additionally, tuning five inference hyperparameters per dataset requires a validation set. While the standard splits in Table 1 provide validation nodes, in true few-shot or zero-shot deployment scenarios (e.g., the ProG protocol in Appendix F.1, where the text explicitly states "there is no validation set"), how were these dataset-specific inference hyperparameters selected? It remains unclear whether there is test-set leakage or an over-reliance on heuristic defaults.
  1. Missing Mathematical Definitions and Hyperparameter Specifications for the Prior: Although the Reproducibility Statement claims that "The sampler of §3.1 with its coefficient mixtures is in Appendix A", Appendix A.1 is surprisingly brief and mostly qualitative. The manuscript does not provide the exact probability distributions or mixture weights for the PDE coefficients (α,κ,β,γ), the precise mathematical formulation of the velocity field V derived from node potentials, the exact functional forms of the reaction terms r(Z), or the exact mixture ratios among the different random graph generators. These details are essential for readers to accurately reproduce the results, assess the prior's suitability for target datasets, and evaluate real-world deployment costs.
Questions for authors
  1. Could you report the performance of all three priors in Table 2 under an identical, default (untuned) inference configuration on both networks to verify that the gain from the prior still holds?
  2. How were the inference hyperparameters chosen in settings without a validation set, such as the ProG few-shot protocol?
  3. Could you provide the exact mathematical expressions, functional forms, and parameter distributions omitted from Appendix A.1 regarding the synthetic data sampler?
3 comments

We thank Reviewer cm7F for providing a clear explanation of the method and for three questions that we could answer with new experiments. The changes in the revised paper are highlighted in red.


Response to W1 and Q1. The three priors under one untuned configuration on both networks.

The gain from the GadrPFN prior remains on both networks when no setting is tuned for it, including the head-mixing weight. The reviewer is right that the comparison of 75.3 with 70.2 in the submitted Table 2 mixed two inference settings. We added the comparison under common settings to Appendix J.2, and the paragraph of Section 4.3 now reports it. As the identical untuned configuration we use NodePFN's released settings, which were not chosen for the GadrPFN prior and keep the head-mixing weight at its trained value of 0.5 on every graph. Each column below applies one inference setting identically to all three priors of Table 2.

PriorOur network at NodePFN's released settingsNodePFN's network at NodePFN's released settingsOur network at the settings of Table 1
NodePFN's prior70.268.270.9
GadrPFN w/ GraphPFN agg.72.770.574.4
GadrPFN prior72.970.475.3

On our network at NodePFN's released settings the GadrPFN prior is more accurate than NodePFN's prior. On NodePFN's network both new priors are more accurate than NodePFN's prior. In Appendix J.2, we also report the per-graph results.


Response to W1b and Q2. Settings without a validation set in the ProG protocol.

In settings without a validation set, such as the ProG protocol, the revised paper chooses no inference hyperparameter, so no test label can enter the choice of settings. To this end, GadrPFN and NodePFN use NodePFN's released settings and GraphPFN uses its default in the revised Appendix F.1 and Appendix 4.2. We have also added the ProG protocol on all 28 graphs to Appendix J.6. With these fixed settings, GadrPFN has the highest mean accuracy of the three PFNs at every budget. The table below shows the mean accuracy over the six graphs of Appendix F.1.

Labels per class1351020
GadrPFN38.051.756.963.467.9
NodePFN33.745.852.460.065.6
GraphPFN23.442.549.758.664.5

On all 28 graphs, the mean accuracy over 83 graph-budget pairs is 49.4 for GadrPFN, 45.9 for NodePFN and 45.7 for GraphPFN. The submitted rows had tuned these settings on the evaluation nodes, which gave 41.6 with one label per class.


Response to W2 and Q3. Mathematical definitions of the sampler.

In the revised paper, we provide the exact parameter distributions and functional forms of the sampler, so that readers can reproduce the sampler. In the new Appendix J.7, we list the distributions and mixture weights of αα, κκ, ββ and γγ, the functional forms of the reaction terms r(Z)r(Z), the construction of the velocity field VV and the mixture ratios of the graph generators. In the revised paper, Appendix A.1 also refers to Appendix J.7. For the velocity field VV, we have also clarified its description in Section 3.1. The prior draws a random velocity for each edge, antisymmetric in direction and redrawn at every flow layer, which generalises the fields derived from a node potential. Only the ADR layer of the network uses a learned potential.

Dear authors,

Thanks for the response. The authors have addressed my concerns, and I will keep my original rating.

Best of luck.

AuthorsOpenReview ↗

Thank you for your feedback

Dear Reviewer cm7F,

Thank you for confirming that our response has addressed your concerns, and for your time and support.

Best regards, The authors

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

The paper proposes GADRPFN, a prior-data fitted network for in-context node classification. Its main contribution is a synthetic task prior that couples random nonlinear causal models with graph advection, diffusion, reaction, and restart dynamics. Randomized coefficients, evolution times, and label transformations generate varied relationships between graph structure, features, and labels. A pretrained attention-based network with an ADR graph module then predicts query labels from labeled context nodes without gradient-based adaptation on the target graph. Evaluation covers 28 benchmarks, few-shot classification, text-attributed graphs, and larger graphs and contexts.

Critical strengths
  1. A useful contribution to graph-task prior design, with evidence beyond the matching architecture. Randomized graph dynamics provide a concrete way to vary feature propagation and its relationship to labels. Appendix G.3 supports the benefit of the complete prior under the same architecture, training recipe, budget, and inference settings. Table 3 also retains 74.6% accuracy when replacing the ADR module with a GCN, versus 75.3% for the full model. Together, these results support a contribution beyond introducing another graph prediction architecture.

  2. Broad and consistently favorable empirical results. Across 28 benchmarks, the method achieves 75.31% mean accuracy versus 72.01% for TAG and 71.69% for GraphPFN, outperforming them on 25 and 23 benchmarks, respectively. Additional evaluations address limited labels, text features, and contexts larger than those encountered during pretraining. The appendix also usefully discloses baseline reproduction differences and acknowledges the dependence of some text-graph comparisons on the feature encoder.

Critical weaknesses
  1. The decomposition of the gains in Section 4.3 is not established by Table 2. Appendix B states that the existing-prior rows use NodePFN's released inference settings, whereas the proposed prior uses its own dataset-specific settings. Thus, the reported gains do not change only the prior. Moreover, the GraphPFN-aggregation variant simultaneously removes initial averaging, ADR flow, and separate label evolution, while adding degree/PageRank inputs. It cannot cleanly establish that graph/label generation explains approximately half the gain and diffusion explains the remainder. Appendix G.3 partially resolves this concern by demonstrating a substantial benefit of the complete prior under shared settings, but it does not validate the component-level attribution.

  2. Important parts of the central method and evaluation procedure remain underspecified. Appendix A gives qualitative coefficient choices but not the numerical distributions and mixture probabilities needed to reconstruct the proposed prior, nor a complete specification of its integration and inertia choices. These details define the contribution itself; the reproducibility statement defers the sampler and pretraining code until acceptance. Separately, Appendix D lists an extensive inference-configuration search without specifying the selection data, criterion, or budget. Appendix G.1 measures inference cost under NodePFN's settings rather than the proposed model's Table 1 settings, so the reported accuracy and cost do not describe the same configuration. These gaps limit reproducibility and assessment of the practical tradeoff; they do not establish evaluation leakage.

Questions for authors
  1. Additional evidence: attribution to the prior and its components. Could you report the Table 2 comparisons under common inference settings, extending the matched comparison already present in Appendix G.3? For the intermediate aggregation prior, can the contribution of ADR evolution be separated from the changes to label evolution, initial smoothing, and structural inputs? Alternatively, would you narrow the claim to a comparison between complete generator designs?

  2. Method specification: reconstructing the proposed prior. Could you provide the exact coefficient distributions and mixture probabilities, the construction of the transport operator, the reaction/inertia updates, and the rule for choosing integration steps, together with the pretraining configuration? An anonymized sampler or a complete specification would address this concern. Please also state whether clipping, resampling, or other numerical safeguards are used.

  3. Clarification of existing experiments: model selection and deployment cost. Which labels and selection criterion were used to choose the per-dataset inference settings, and how were settings selected for the few-shot experiments? Were comparable selection budgets available to the baseline priors? Please also report the inference time and memory of the actual Table 1 configuration, or its accuracy under the configuration measured in Table 14.

4 comments

We appreciate Reviewer HQik for the positive feedback and insightful comment. The changes in the revised paper are highlighted in red.


Response to W1 and Q1a. Table 2 under common inference settings.

When only the prior changes, the GadrPFN prior remains more accurate than NodePFN's prior on both networks. The reviewer is right that Table 2 reported the GadrPFN prior (75.3 and 72.0) at the settings of Table 1 and the two other priors at NodePFN's released settings, as Appendix B states. Appendix G.3 compared the two priors under the settings of Table 1, and we have extended this comparison to every prior of Table 2. In the below table, each column applies one inference setting identically to every prior.

PriorOur network at NodePFN's released settingsOur network at the settings of Table 1NodePFN's network at NodePFN's released settings
NodePFN's prior70.270.968.2
GadrPFN prior72.975.370.4

On our network the GadrPFN prior is more accurate on all 28 graphs at the settings of Table 1 and on 25 of 28 at NodePFN's released settings, which were set for none of the priors. On NodePFN's network the GadrPFN prior is trained with NodePFN's recipe for 30 epochs, as Appendix B states. In the revised paper, we have added the comparison of all three priors of Table 2, with the per-graph tests, to Appendix J.2, and the paragraph of Section 4.3 reports the gains under these settings.


Response to W1 and Q1b. Separating ADR evolution from the other changes of the aggregation variant.

Yes, the contribution of ADR evolution, together with its label time, can be separated from the changes to initial smoothing and structural inputs. ADR evolution in the flow of ② and the averaging of ① each improves the accuracy, while the structural inputs and the smoothing of the features in ① do not. Starting from our sampler with features independent of the graph, we added each change alone, retrained each prior with one network, one recipe and 60 epochs, and evaluated every prior at the settings of Table 1.

PriorMean accuracy
Our sampler with features independent of the graph73.3
+ structural inputs only73.3
+ smoothing of the features in ① only73.3
+ averaging of the state in ① only74.5
+ flow of ② and label time of ③ only74.6
GadrPFN prior75.3

The GadrPFN prior is more accurate than every other prior in the table. The averaging of ① and GraphPFN's GCN aggregation are each one Euler step of diffusion (Section 3.3), so the aggregation variant of Table 2 does not remove diffusion. Therefore, in the revised paper, we report these separations instead of attribution through aggregate variations in Section 4.3 and provide the full details in Appendix J.3.


Response to W2 and Q2. Specification of the prior.

In the revised paper, we added a detailed specification of the prior, which is the second option suggested by the reviewer, in Appendix J.7. This provides the exact distributions and mixture probabilities of the graph, the state and the flow coefficients. It also gives the construction of the transport operator, the reaction and inertia updates, the rule for choosing the number of integration steps and the pretraining configuration. Appendix A.1 now refers to Appendix J.7, and we have clarified the description of the transport operator in Section 3.1.

As for whether clipping, resampling or other numerical safeguards are used, the prior uses clipping and resampling. It clips the state in the reaction and the features before standardisation, and it redraws the classes of a task when its context and query nodes do not share them. We have added these safeguards to the specification of the prior in Appendix J.7.

Response to W2 and Q3a. Results that do not depend on the selection of settings.

The ranking of the three PFNs in Table 1 does not depend on the selection of the per-dataset settings. The reviewer is right that the selection budgets differ, since GadrPFN uses its own per-dataset settings and the baselines use their released settings.

  1. First, we removed the per-dataset selection of GadrPFN. GadrPFN then uses NodePFN's released settings, which were chosen for NodePFN, with fixed defaults on the 5 graphs that NodePFN did not evaluate. NodePFN and GraphPFN keep their released or default settings, as in Table 1. In this case GadrPFN reaches 72.9 against 71.7 for GraphPFN and 68.2 for NodePFN, and it has the best average rank.
  2. Second, we evaluated the Table 1 settings on splits that no selection used. For each of the 25 graphs that have more than one split, we froze these settings and applied them to 5 additional splits, which differ from the splits of Table 1. On these splits GadrPFN reaches 74.2 against 71.9 for GraphPFN at its default, and it is more accurate on 22 of the 25 graphs, a significant difference. We have added these results to Appendix J.1, which Section 4.2 cites.

Response to Q3b. Settings of the few-shot experiments.

The revised few-shot experiments select no setting, and the few-shot conclusions of Section 4.2 still hold. Because ProG provides no validation set, GadrPFN and NodePFN now use NodePFN's released settings, and GraphPFN uses its default (Appendix F.1). Averaged over the 6 graphs of Appendix F.1, GadrPFN has the highest accuracy of the three PFNs at every budget from 1 to 20 labels per class. Its lead over GraphPFN increases as the number of labels decreases, to 38.0 against 23.4 with one label per class. The submitted rows had tuned these settings on the evaluation nodes, which gave 41.6 with one label per class. For comparison, each pretrained ProG baseline copied in Appendix F.1 is the best of six pretraining variants. We also ran the ProG protocol on all 28 graphs with fixed settings for every model, where GadrPFN has the best mean accuracy at every budget and the best average rank.

Labels per class135Mean over 83 pairsAvg. rank over 83 pairs
GadrPFN40.151.956.449.41.64
NodePFN36.847.753.145.92.20
GraphPFN30.550.755.445.72.16

We have added this table to Appendix J.6.


Response to Q3c. Selection budgets of the baseline priors.

We have compared every prior of Table 2 again with comparable selection budgets, and added the results in Appendix J.2. Each comparison gives every prior identical settings, either NodePFN's released settings, which were chosen for NodePFN, or the settings of Table 1, which were chosen for GadrPFN. The GadrPFN prior is more accurate than NodePFN's prior under both, as the response to W1 and Q1a shows.


Response to W2 and Q3d. Cost and accuracy of one configuration.

The actual Table 1 configuration takes 247 s in total over the 28 graphs with a peak GPU memory of 15.0 GB. In the revised Appendix G.1, each configuration has both its accuracy and its cost, so every reported accuracy and cost refer to one configuration.

28 graphs on one GPUNodePFN's released settings (submitted Table 14)Table 1 settings
Total inference time221 s247 s
Peak GPU memory17.8 GB15.0 GB
Mean accuracy72.975.3

The Table 1 configuration reaches 75.3 with less peak memory than the configuration of Table 14.

Dear authors,

Thanks for the detailed responses, and I think they have addressed most of my concerns, I will conisder to update my score accordingly.

Best of luck, HQik.

Dear Reviewer HQik,

Thank you very much for your positive feedback and for carefully reading our responses.

Your constructive review has led to the two most important changes in the revised paper. First, Appendix J.2 compares the priors of Table 2 under identical inference settings, and Section 4.3 reports these comparisons. Second, Appendix J.7 specifies the prior, with the numerical details that you requested.

Thank you again for your time and for helping us improve the paper.

Best regards,

The authors

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

This paper proposes GADRPFN, a prior-fitted network (PFN) for in-context node classification that uses graph neural diffusion to generate synthetic pretraining tasks. Unlike existing graph PFNs, whose priors either generate node features independently of graph structure or use neighborhood aggregation, GADRPFN models the evolution of node features through an advection-diffusion-reaction equation. The proposed prior generates diverse relationships between graph structure, node features, and labels, including both homophilous and heterophilous patterns.

Critical strengths
  • The paper identifies an important limitation of existing graph PFNs.
  • The evaluation includes 28 graphs covering both homophilous and heterophilous settings.
Critical weaknesses
  • The central contribution is the graph neural diffusion prior, which combines diffusion, advection, reaction, restart, inertia, and a separate label time. However, the ablation studies do not establish which of these components are necessary or primarily responsible for the improvements.
  • Although pretraining is efficient, practical inference requires dimensionality reduction, feature smoothing, and potentially large ensembles of forward passes.
Questions for authors

See weaknesses.

  • How sensitive is performance to the prior's coefficient distributions, graph generator proportions, and diffusion-time sampling?
2 comments

We thank Reviewer Mpum for the concrete questions, which we answer with new controlled experiments. The changes in the revised paper are highlighted in red.


Response to W1. Which components of the prior are necessary?

The new ablation in Appendix J.3 shows which components are needed and which contribute most to the improvement. Removing any one of the 6 components that the reviewer lists lowers the accuracy. We retrained the prior without each component with one network, one recipe and 60 epochs, and evaluated every prior at the settings of Table 1.

PriorMean accuracy
GadrPFN prior75.3
without diffusion74.6
without advection74.7
without reaction74.7
without restart74.9
without inertia74.4
without the separate label time74.5

The mean score decreases from 75.3 to between 74.4 and 74.9. Therefore, each component is needed for the accuracy of the complete prior.


Response to W2. Does inference need smoothing and large ensembles?

The settings of Table 1 use these inference options, but GadrPFN keeps its accuracy without them. To test this, we removed the feature smoothing and the ensemble from NodePFN's released settings, so that GadrPFN runs a single forward pass. We kept their dimensionality reduction, which is applied only to graphs that have more features than the network accepts. We compare the test with NodePFN's released settings, because the settings of Table 1 differ in additional options.

ModelNodePFN's released or default settingsOne forward pass without smoothing
GadrPFN72.972.6
NodePFN68.261.3
GraphPFN71.7 with 10 members68.6 with its smallest ensemble

GadrPFN has the least accuracy loss among the three PFNs. Its single forward pass is still more accurate on average than NodePFN and GraphPFN at their released or default settings. With the options of Table 1, inference over the 28 graphs takes 247 s in total and at most 15.0 GB of memory on one GPU. We have added these results to Appendix J.5 and the cost of the Table 1 settings to Appendix G.1.


Response to Q1. How sensitive is the performance to the prior's distributions?

The sensitivity differs between homophilous and heterophilous graphs, and this difference supports the design of our prior. We retrained the model under 9 changes of the coefficient distributions, the graph generator proportions and the diffusion-time sampling. Every model uses one network, one recipe and 60 epochs, and we evaluated it at the settings of Table 1.

Prior changeAllHom.Het.
Submitted prior (reference)75.378.071.1
Diffusion coefficient range scaled by 0.574.977.970.2
Diffusion coefficient range scaled by 274.777.770.1
Advection coefficient range scaled by 0.574.877.670.4
Advection coefficient range scaled by 274.577.569.8
Reaction strength range scaled by 0.574.577.469.9
Reaction strength range scaled by 274.777.570.2
Five graph generators in equal proportions74.777.870.0
Maximum time 0.5 instead of 174.677.669.9
Maximum time 274.877.670.5

On homophilous graphs, the accuracy remains close to that of the submitted prior, and 7 of the 9 changes show no significant difference. On heterophilous graphs, every change lowers the accuracy. Existing priors generate few tasks whose labels run against the edges (Section 4.4), and our prior gains most on heterophilous graphs (Section 4.3). Therefore, the distributions of our prior matter on heterophilous graphs, for which they were designed. Every changed prior also remains above every baseline of Table 1, whose best mean accuracy is 72.0. We have added these results to Appendix J.4.

AuthorsOpenReview ↗

Kind reminder about our response

Dear Reviewer Mpum,

As the discussion period ends soon, we would be grateful to know whether our response has addressed your concerns. We hope that the new results in Appendix J.3 to J.5 resolve them, and we thank you again for your time.

Best regards, The authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Discussion for Reviewers and AC

Dear Reviewers and AC,

We thank the three reviewers for their careful reviews and for acknowledging the following contributions.

  • The paper identifies an important limitation of existing graph PFNs and evaluates on 28 graphs that cover homophilous and heterophilous settings (Mpum).
  • A useful contribution to graph-task prior design beyond the architecture, supported by the complete prior under one architecture and recipe (Appendix G.3) and by the GCN variant of Table 3 (HQik).
  • A well-structured prior design space in which the separate label time decouples feature smoothness from label homophily (cm7F).
  • Broad and favorable results on 28 benchmarks with the best average rank on homophilous and heterophilous graphs, and further evaluations with few labels, text features and large contexts (HQik and cm7F).

The table below summarises the main concerns of the reviewers and our responses. In the revised paper, the changes are highlighted in red and the new results are in Appendix J.

ConcernByOur response
Inference settings chosen per datasetHQikWe showed that GadrPFN stays first among the three PFNs without its per-dataset selection and on further splits (Appendix J.1).
Table 2 mixes inference settingsHQik and cm7FWe compared the priors of Table 2 under common settings and the GadrPFN prior improves over NodePFN's prior on both networks (Appendix J.2).
Which components of the prior are neededMpum and HQikWe separated the components with priors that differ by one change at a time (Appendix J.3).
Specification of the priorHQik and cm7FWe added the specification of the prior with its numerical safeguards (Appendix J.7).
Few-shot settings without a validation setHQik and cm7FWe fixed the settings of every model and GadrPFN stays first among the three PFNs on 6 and on 28 graphs (Appendix F.1 and J.6).
Smoothing and ensembles at inference and their costMpum and HQikWe showed that GadrPFN keeps its accuracy with one forward pass and no smoothing and reported the cost of the Table 1 settings (Appendix J.5 and G.1).
Sensitivity to the priorMpumWe retrained the model under eleven changes of the prior and every changed prior stays above all baselines of Table 1 in mean accuracy (Appendix J.4).

We thank the reviewers and the AC again for their time and for helping us improve the paper.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.