Evaluating Progress in Graph Foundation Models: A Comprehensive Benchmark and New Insights

Xingtong Yu; Shenghua Ye; Ruijuan Liang; Chang Zhou; Hong Cheng; Xinming Zhang; Yuan Fang

cs.CLFeb 28, 2026

Evaluating Progress in Graph Foundation Models: A Comprehensive Benchmark and New Insights

Authors: Xingtong Yu, Shenghua Ye, Ruijuan Liang, Chang Zhou, Hong Cheng, Xinming Zhang, Yuan Fang

Organizations: The Chinese University of Hong Kong* China · Singapore Management University Singapore · University of Science and Technology of China China

Abstract

Graph foundation models (GFM) aim to acquire transferable knowledge by pre-training on diverse graphs, which can be adapted to various downstream tasks. However, domain shift in graphs is inherently two-dimensional: graphs differ not only in what they describe (topic domains) but also in how they are represented (format domains). Most existing GFM benchmarks vary only topic domains, thereby obscuring how knowledge transfers across both dimensions. We present a new benchmark that jointly evaluates topic and format gaps across the full GFM pipeline, including multi-domain self-supervised pre-training and few-shot downstream adaptation, and provides a timely evaluation of recent GFMs in the rapidly evolving landscape. Our protocol enables controlled assessment in four settings: (i) pre-training on diverse topics and formats, while adapting to unseen downstream datasets; (ii) same pre-training as in (i), while adapting to seen datasets; (iii) pre-training on a single topic domain, while adapting to other topics; (iv) pre-training on a base format, while adapting to other formats. This two-axis evaluation disentangles semantic generalization from robustness to representational shifts. We conduct extensive evaluations of eight state-of-the-art GFMs on 33 datasets spanning seven topic domains and six format domains, surfacing new empirical observations and practical insights for future research. Codes/data are available at https://github.com/smufang/GFMBenchmark.

Figures & tables

Methods	Labeled source	Alignment						PLM/LLM	Adaptation
Methods	Labeled source	Structure	Feature	Codebook	Graph mod.	Feature mod.	Textual	PLM/LLM	Adaptation
Benchmarked graph foundation models
GCOPE ( Zhao et al., 2024a )	$\times$	$\times$	$\checkmark$	$\times$	$\checkmark$	$\times$	$\times$	$\times$	F, P
MDGPT ( Yu et al., 2024b )	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\times$	$\times$	P
MDGFM ( Wang et al., 2025 )	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\times$	$\times$	$\times$	P
SAMGPT ( Yu et al., 2025a )	$\times$	$\checkmark$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\times$	$\times$	P

Table 1. Comparison of representative GFMs.

Topic

Datasets

# Graphs

Avg. # nodes

Avg. # edges

Format

Used for pre-train

Task

Heterogeneous

Homophily ratio

Relational

Dynamic

Textual

Citation

Cora ( Liu et al., 2024a )

2,708

10,556

\times

0.81

\times

\times

\checkmark

\checkmark

\checkmark

\times

\times

Pubmed ( Liu et al., 2024a )

19,717

44,338

\times

0.80

\times

\times

\checkmark

\times

\checkmark

\times

\times

ogbn-arxiv ( Hu et al., 2020b )

169,343

1,166,243

\times

0.66

\times

\times

\checkmark

\checkmark

\times

\times

\times

ogbn-mag ( Hu et al., 2020b )

1,939,743

21,111,007

\checkmark

0.30

\times

\times

\times

\times

\checkmark

\times

\times

ACM ( Tang et al., 2024b )

10,942

547,872

\checkmark

0.88

\times

\times

\times

\checkmark

\checkmark

\times

\times

Table 2. Summary of the datasets by topics and formats.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCN	48.62 $\pm$ 8.40	5.83 $\pm$ 0.70	48.02 $\pm$ 39.11	26.50 $\pm$ 11.82	26.71 $\pm$ 4.33	21.12 $\pm$ 5.67	42.16 $\pm$ 39.45	32.88 $\pm$ 15.35	51.54 $\pm$ 13.11
GAT	45.50 $\pm$ 9.29	4.86 $\pm$ 0.60	45.32 $\pm$ 24.72	23.90 $\pm$ 11.13	26.24 $\pm$ 4.27	22.11 $\pm$ 6.28	53.86 $\pm$ 38.15	33.46 $\pm$ 16.12	47.93 $\pm$ 14.10
GCOPE	41.85 $\pm$ 8.34	0.27 $\pm$ 0.09	43.23 $\pm$ 25.98	24.16 $\pm$ 6.83	21.66 $\pm$ 3.41	OOT	45.42 $\pm$ 22.22	OOT	45.01 $\pm$ 12.73
MDGPT	40.50 $\pm$ 7.33	5.94 $\pm$ 0.63	51.02 $\pm$ 25.40	31.72 $\pm$ 8.93	28.36 $\pm$ 5.04	15.99 $\pm$ 3.48	58.95 $\pm$ 23.03	37.52 $\pm$ 12.02	50.72 $\pm$ 10.78
MDGFM	39.50 $\pm$ 7.20	2.93 $\pm$ 1.05	44.25 $\pm$ 28.21	19.80 $\pm$ 7.00	25.95 $\pm$ 3.75	8.11 $\pm$ 2.17	57.07 $\pm$ 19.12	36.08 $\pm$ 24.19	50.28 $\pm$ 8.80
SAMGPT	49.23 $\pm$ 8.07	5.32 $\pm$ 0.68	56.50 $\pm$ 27.26	27.09 $\pm$ 7.73	28.65 $\pm$ 5.12	22.68 $\pm$ 4.03	58.28 $\pm$ 15.86	39.42 $\pm$ 13.36	52.25 $\pm$ 11.97

Table 3. Accuracy of one-shot node classification on unseen datasets. Marco-F1 reported in Appendix H .

Methods	Edge classification						Graph classification
	DGraph		Wiki		WN18RR		PCBA		BZR
	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCN	10.13 $\pm$ 3.16	6.50 $\pm$ 0.99	9.35 $\pm$ 3.01	3.21 $\pm$ 0.19	15.45 $\pm$ 11.60	10.85 $\pm$ 4.39	55.09 $\pm$ 31.92	32.54 $\pm$ 15.72	52.73 $\pm$ 16.49	44.72 $\pm$ 10.30
GAT	9.55 $\pm$ 3.56	6.16 $\pm$ 1.05	2.65 $\pm$ 0.62	1.71 $\pm$ 0.15	10.60 $\pm$ 14.10	7.91 $\pm$ 3.93	56.65 $\pm$ 28.70	33.97 $\pm$ 13.34	52.36 $\pm$ 17.69	43.63 $\pm$ 10.93
GCOPE	OOT	OOT	OOT	OOT	16.59 $\pm$ 7.25	11.42 $\pm$ 3.29	48.99 $\pm$ 38.52	28.19 $\pm$ 19.08	54.31 $\pm$ 19.35	45.89 $\pm$ 12.33
MDGPT	9.92 $\pm$ 2.38	6.85 $\pm$ 0.95	10.72 $\pm$ 3.67	2.79 $\pm$ 0.21	13.16 $\pm$ 4.71	10.08 $\pm$ 2.01	55.54 $\pm$ 21.39	34.57 $\pm$ 9.90	53.88 $\pm$ 13.74	47.26 $\pm$ 9.46

Table 4. Evaluation of one-shot edge classification and graph classification on unseen datasets.

Methods

Cora

ACM

Wisconsin

Photo

Elliptic

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

GCN

41.76

\pm

10.38

37.85

\pm

9.98

38.31

\pm

8.08

24.12

\pm

10.36

50.90

\pm

33.56

32.75

\pm

16.73

34.20

\pm

7.94

22.49

\pm

4.55

53.87

\pm

12.35

52.05

\pm

11.90

41.74

\pm

18.55

32.74

\pm

11.50

GAT

43.48

\pm

9.64

39.92

\pm

8.96

35.82

\pm

5.75

22.37

\pm

8.94

48.68

\pm

30.94

32.30

\pm

15.74

34.18

\pm

8.57

23.06

\pm

5.94

46.38

\pm

11.11

45.49

\pm

9.65

38.73

\pm

20.26

29.95

\pm

11.87

DGI

43.82

\pm

9.33

40.78

\pm

8.06

36.01

\pm

7.24

28.11

\pm

9.48

46.92

\pm

20.14

33.27

\pm

9.98

27.29

\pm

8.32

21.34

\pm

5.51

47.26

\pm

7.47

45.23

\pm

6.57

44.56

\pm

15.54

34.25

\pm

10.64

GraphPrompt

55.51

\pm

10.51

52.53

\pm

9.76

42.18

\pm

9.78

35.12

\pm

12.51

48.32

\pm

14.02

35.11

\pm

6.97

32.73

\pm

4.80

26.41

\pm

3.88

57.06

\pm

9.09

55.81

\pm

7.99

50.55

\pm

17.03

39.50

\pm

13.23

HeCo

80.85

\pm

12.79

80.25

\pm

13.63

Table 5. Evaluation of one-shot node classification on seen datasets.

Methods	Edge classification		Graph classification
	FB15K-237		HIV		COX2		PROTEINS
	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCN	14.63 $\pm$ 1.4	13.35 $\pm$ 0.8	45.79 $\pm$ 27.1	31.22 $\pm$ 13.8	48.09 $\pm$ 14.2	42.15 $\pm$ 9.5	58.67 $\pm$ 9.2	52.33 $\pm$ 8.6
GAT	11.05 $\pm$ 1.2	10.90 $\pm$ 0.8	46.00 $\pm$ 27.4	31.50 $\pm$ 14.4	50.71 $\pm$ 14.9	43.67 $\pm$ 9.1	57.79 $\pm$ 9.4	51.44 $\pm$ 8.7
DGI	4.84 $\pm$ 1.4	1.61 $\pm$ 0.3	48.47 $\pm$ 13.2	34.93 $\pm$ 6.4	46.99 $\pm$ 12.8	41.99 $\pm$ 8.3	54.28 $\pm$ 8.7	51.56 $\pm$ 8.1
GraphPrompt	5.52 $\pm$ 1.6	1.42 $\pm$ 0.3	50.50 $\pm$ 9.3	36.15 $\pm$ 4.2	49.30 $\pm$ 10.9	44.18 $\pm$ 7.1	56.83 $\pm$ 7.0	54.32 $\pm$ 6.3

Table 6. Evaluation of one-shot edge classification and graph classification on seen datasets.

Methods	Citation		Social & Web			E-commerce	Finance		Proteins
Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	37.93 $\pm$ 5.54	0.26 $\pm$ 0.03	52.79 $\pm$ 27.79	23.16 $\pm$ 7.74	21.82 $\pm$ 3.43	OOT	33.94 $\pm$ 20.92	OOT	47.70 $\pm$ 13.36
MDGPT	47.93 $\pm$ 10.33	4.64 $\pm$ 0.58	51.80 $\pm$ 29.13	27.84 $\pm$ 10.71	27.26 $\pm$ 4.97	18.98 $\pm$ 5.42	59.88 $\pm$ 23.85	36.79 $\pm$ 12.12	51.07 $\pm$ 10.98
MDGFM	35.62 $\pm$ 6.59	1.21 $\pm$ 1.24	50.67 $\pm$ 28.53	18.19 $\pm$ 5.80	23.24 $\pm$ 3.59	3.07 $\pm$ 0.90	49.22 $\pm$ 21.44	41.96 $\pm$ 28.01	49.78 $\pm$ 8.57
SAMGPT	48.11 $\pm$ 6.83	6.57 $\pm$ 0.68	59.19 $\pm$ 30.30	25.77 $\pm$ 7.33	28.07 $\pm$ 4.76	24.32 $\pm$ 4.42	48.87 $\pm$ 17.22	38.37 $\pm$ 14.55	51.66 $\pm$ 10.73
G2P2	43.30 $\pm$ 8.57	-	44.27 $\pm$ 7.80	28.85 $\pm$ 9.66	-	-	59.72 $\pm$ 16.22	38.25 $\pm$ 10.39	-

Table 7. Accuracy of one-shot node classification on topic domain adaptation. Marco-F1 reported in Appendix J .

Methods	Finance (Edge Classification)		Common sense (Edge Classification)				Molecule (Graph Classification)
	DGraph		Wiki		WN18RR		PCBA		BZR
	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCOPE	OOT	OOT	OOT	OOT	16.05 $\pm$ 6.72	11.76 $\pm$ 3.33	61.25 $\pm$ 24.71	36.57 $\pm$ 10.62	57.37 $\pm$ 19.40	47.43 $\pm$ 11.66
MDGPT	9.10 $\pm$ 2.59	6.42 $\pm$ 0.94	7.90 $\pm$ 3.14	1.57 $\pm$ 0.23	14.48 $\pm$ 4.91	10.84 $\pm$ 2.25	62.51 $\pm$ 24.39	37.11 $\pm$ 11.11	57.01 $\pm$ 14.30	49.79 $\pm$ 9.77
MDGFM	9.35 $\pm$ 3.31	5.79 $\pm$ 1.16	4.34 $\pm$ 2.91	0.50 $\pm$ 0.09	14.14 $\pm$ 5.30	9.60 $\pm$ 2.52	58.16 $\pm$ 42.55	31.39 $\pm$ 21.02	52.12 $\pm$ 24.02	35.20 $\pm$ 9.12
SAMGPT	10.82 $\pm$ 3.14	6.97 $\pm$ 1.04	11.59 $\pm$ 3.03	3.52 $\pm$ 0.16	13.64 $\pm$ 5.11	10.03 $\pm$ 1.91	59.51 $\pm$ 24.76	35.91 $\pm$ 10.89	59.29 $\pm$ 15.52	50.81 $\pm$ 10.13

Table 8. Evaluation of one-shot edge classification and graph classification on topic domain adaptation.

Figure 1. Comparisons with Setting I, with higher values favoring Setting I more. (a) Single-topic (Setting III) vs. multi-topic pre-training (Setting I). (b) Base format (Setting IV) vs. multi-format pre-training (Setting I).

Methods

Heterogeneous

Heterophilic

Dynamic

Textual

ogbn-mag

Chameleon

Actor

Wikipedia

Pubmed

Products

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

GCOPE

0.32

\pm

0.04

0.11

\pm

0.01

22.46

\pm

2.93

19.99

\pm

3.38

20.66

\pm

8.01

16.05

\pm

4.58

58.69

\pm

29.29

36.89

\pm

14.91

39.06

\pm

7.55

35.09

\pm

8.07

OOT

MDGPT

6.01

\pm

0.81

4.42

\pm

0.37

27.90

\pm

5.13

25.06

\pm

5.03

27.23

\pm

8.64

22.22

\pm

5.09

54.17

\pm

22.13

35.72

\pm

10.71

45.80

\pm

10.14

45.10

\pm

10.47

21.23

\pm

4.96

11.50

\pm

1.88

MDGFM

0.92

\pm

0.54

0.56

\pm

0.41

21.50

\pm

3.12

17.54

\pm

4.00

20.20

\pm

7.73

13.83

\pm

3.12

50.66

\pm

29.90

32.51

\pm

14.61

34.51

\pm

5.84

28.24

\pm

4.55

3.92

\pm

2.17

1.78

\pm

0.47

SAMGPT

7.35

\pm

0.73

5.49

\pm

0.27

28.14

\pm

4.77

24.48

\pm

4.66

24.58

\pm

6.49

19.21

\pm

4.17

55.95

\pm

31.62

34.51

\pm

15.82

50.00

\pm

7.93

48.68

\pm

8.50

26.10

\pm

5.55

13.38

\pm

1.52

Table 9. Evaluation of one-shot node classification on format domain adaptation.

Methods	Relational (Edge Classification)				Textual (Graph Classif.)
	Wiki		WN18RR		PCBA
	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCOPE	OOT	OOT	17.76 $\pm$ 6.44	12.87 $\pm$ 3.03	50.93 $\pm$ 34.41	30.23 $\pm$ 16.26
MDGPT	10.07 $\pm$ 3.05	2.75 $\pm$ 0.18	14.71 $\pm$ 6.07	10.55 $\pm$ 2.46	55.25 $\pm$ 24.41	34.04 $\pm$ 11.43
MDGFM	3.85 $\pm$ 2.53	0.37 $\pm$ 0.09	13.77 $\pm$ 5.17	9.93 $\pm$ 2.07	51.68 $\pm$ 48.73	26.37 $\pm$ 24.37
SAMGPT	12.33 $\pm$ 3.59	3.69 $\pm$ 0.18	13.53 $\pm$ 4.88	9.97 $\pm$ 2.11	56.06 $\pm$ 22.56	34.63 $\pm$ 10.44

Table 10. Evaluation of one-shot edge classification and graph classification on format domain adaptation.

Appendix figures & tables26 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2. Graph foundation models pipeline.

Figure 3. Four evaluation settings in our benchmark.

Dataset	raw_texts	relation_texts
Cora, ogbn-arxiv	Title: "Further facts…" Abstract: "Previous results…" / Title: "Spring embedders…" Abstract: "Force-directed…"	cites
ACM	author, term, subject, paper	cites, mentions, writes, belongs_to, …
DBLP	venue, author, term, paper	published_in, mentions, writes, …
Reddit	user, post	interacts, rev_interacts
Texas, Cornell, Wisconsin	Content: Automated Theorem Proving Group… / Content: Steve Seitz’s View Interpolations…	links_to
IMDB	movie, director, actor, keyword	directed_by, stars, described_by, …

Appendix

Table 11. Examples of standardized raw_texts and relation_texts encapsulated in our unified data objects.

Method	GNN Architecture					Pretraining			Downstream Adaptation
Method	Backbone	Unified Input	Hidden Dim	Output Dim	Layers	LR	Epochs	Patience	LR	Epochs	Patience
GCOPE	FAGCN	100	128	128	2	$1\text{e-}4$	100	100	$1\text{e-}4$	100	100
MDGPT	GCN	50	256	256	3	$1\text{e-}3$	200	20	$1\text{e-}2$	400	10
MDGFM	GCN	50	256	256	3	$1\text{e-}3$	200	20	$1\text{e-}2$	400	20
SAMGPT	GCN	50	256	256	3	$1\text{e-}3$	200	20	$1\text{e-}2$	400	10
G2P2	GCN	128	128	128	2	$2\text{e-}5$	10	3	$1\text{e-}2$	50	50

Appendix

Table 12. General hyperparameter configurations.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCN	43.39 $\pm$ 11.32	4.49 $\pm$ 0.27	28.72 $\pm$ 19.65	20.24 $\pm$ 7.38	21.59 $\pm$ 4.32	12.17 $\pm$ 1.70	26.94 $\pm$ 21.92	23.42 $\pm$ 7.91	37.96 $\pm$ 9.95
GAT	39.05 $\pm$ 10.94	3.39 $\pm$ 0.25	31.13 $\pm$ 12.89	18.38 $\pm$ 7.29	21.96 $\pm$ 4.18	12.13 $\pm$ 1.85	33.67 $\pm$ 20.34	23.95 $\pm$ 8.20	31.99 $\pm$ 6.60
GCOPE	39.94 $\pm$ 8.76	0.06 $\pm$ 0.01	29.82 $\pm$ 13.47	19.02 $\pm$ 3.97	19.51 $\pm$ 3.61	OOT	32.62 $\pm$ 11.35	OOT	32.43 $\pm$ 7.98
MDGPT	38.80 $\pm$ 7.39	4.41 $\pm$ 0.32	33.70 $\pm$ 12.85	24.58 $\pm$ 5.67	25.98 $\pm$ 5.12	9.99 $\pm$ 1.54	42.96 $\pm$ 14.41	27.06 $\pm$ 6.17	47.88 $\pm$ 9.99
MDGFM	35.37 $\pm$ 6.65	2.06 $\pm$ 0.75	29.67 $\pm$ 14.59	14.80 $\pm$ 3.49	23.04 $\pm$ 3.75	5.09 $\pm$ 1.07	41.65 $\pm$ 11.32	21.77 $\pm$ 12.54	44.52 $\pm$ 6.79
SAMGPT	47.82 $\pm$ 8.75	3.77 $\pm$ 0.23	35.88 $\pm$ 12.86	21.59 $\pm$ 5.80	25.05 $\pm$ 5.33	12.53 $\pm$ 1.58	43.41 $\pm$ 10.06	27.45 $\pm$ 6.83	47.83 $\pm$ 10.66

Appendix

Table 13. Macro-F1 of 1-shot node classification on unseen domains.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCN	64.30 $\pm$ 6.69	11.32 $\pm$ 0.55	46.14 $\pm$ 28.17	40.51 $\pm$ 8.57	35.71 $\pm$ 3.48	34.95 $\pm$ 8.44	48.03 $\pm$ 39.90	40.84 $\pm$ 9.77	51.75 $\pm$ 12.54
GAT	65.71 $\pm$ 6.16	10.05 $\pm$ 0.66	48.14 $\pm$ 19.91	39.72 $\pm$ 9.00	35.33 $\pm$ 3.28	43.66 $\pm$ 5.16	59.05 $\pm$ 33.99	41.20 $\pm$ 9.92	49.96 $\pm$ 13.96
GCOPE	59.13 $\pm$ 5.40	0.25 $\pm$ 0.06	51.77 $\pm$ 18.38	29.71 $\pm$ 8.06	25.41 $\pm$ 2.19	OOT	35.86 $\pm$ 13.81	OOT	42.44 $\pm$ 11.57
MDGPT	53.58 $\pm$ 6.27	11.93 $\pm$ 0.64	54.48 $\pm$ 18.49	41.33 $\pm$ 5.26	36.33 $\pm$ 3.26	25.61 $\pm$ 5.47	58.90 $\pm$ 17.82	38.17 $\pm$ 7.42	54.00 $\pm$ 9.61
MDGFM	51.71 $\pm$ 7.34	OOM	49.81 $\pm$ 19.18	23.56 $\pm$ 5.66	31.70 $\pm$ 3.50	8.97 $\pm$ 3.85	54.14 $\pm$ 12.99	39.44 $\pm$ 13.73	50.98 $\pm$ 5.72
SAMGPT	62.21 $\pm$ 4.90	11.41 $\pm$ 0.81	58.08 $\pm$ 14.24	38.94 $\pm$ 5.48	37.86 $\pm$ 3.59	31.24 $\pm$ 6.93	60.00 $\pm$ 10.20	40.57 $\pm$ 10.19	54.51 $\pm$ 11.51

Appendix

Table 14. Accuracy of 5-shot node classification on unseen domains.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCN	63.59 $\pm$ 7.28	10.51 $\pm$ 0.29	31.07 $\pm$ 13.99	32.90 $\pm$ 5.64	34.91 $\pm$ 3.53	17.84 $\pm$ 3.39	31.14 $\pm$ 23.50	29.72 $\pm$ 4.80	41.13 $\pm$ 11.52
GAT	65.14 $\pm$ 6.32	9.08 $\pm$ 0.29	33.42 $\pm$ 9.54	30.96 $\pm$ 6.43	34.68 $\pm$ 3.40	24.08 $\pm$ 1.39	40.97 $\pm$ 22.11	30.11 $\pm$ 4.85	33.54 $\pm$ 6.99
GCOPE	59.15 $\pm$ 5.35	0.06 $\pm$ 0.01	35.54 $\pm$ 8.85	24.29 $\pm$ 4.88	23.83 $\pm$ 1.96	OOT	28.58 $\pm$ 8.11	OOT	30.93 $\pm$ 7.43
MDGPT	52.84 $\pm$ 6.44	9.39 $\pm$ 0.29	36.50 $\pm$ 8.66	33.60 $\pm$ 3.56	35.19 $\pm$ 3.19	13.15 $\pm$ 1.92	43.33 $\pm$ 10.49	28.41 $\pm$ 3.62	52.04 $\pm$ 9.65
MDGFM	50.99 $\pm$ 7.71	OOM	34.26 $\pm$ 9.04	18.46 $\pm$ 3.87	30.51 $\pm$ 4.16	4.77 $\pm$ 1.51	41.47 $\pm$ 7.71	28.49 $\pm$ 6.78	47.66 $\pm$ 4.13
SAMGPT	62.16 $\pm$ 4.87	8.91 $\pm$ 0.64	38.56 $\pm$ 5.89	31.67 $\pm$ 3.65	37.22 $\pm$ 3.77	15.60 $\pm$ 1.94	44.89 $\pm$ 5.95	29.47 $\pm$ 4.99	51.52 $\pm$ 10.31

Appendix

Table 15. Macro-F1 of 5-shot node classification on unseen domains.

Methods	Edge classification						Graph classification
	DGraph		Wiki		WN18RR		PCBA		BZR
	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCN	10.31 $\pm$ 1.77	7.68 $\pm$ 0.78	21.49 $\pm$ 3.38	7.40 $\pm$ 0.19	20.18 $\pm$ 13.60	20.14 $\pm$ 4.81	55.24 $\pm$ 30.77	32.74 $\pm$ 15.69	60.22 $\pm$ 13.99	49.94 $\pm$ 9.70
GAT	10.47 $\pm$ 1.75	7.77 $\pm$ 0.80	10.96 $\pm$ 1.28	4.93 $\pm$ 0.18	12.54 $\pm$ 10.60	16.66 $\pm$ 3.19	57.48 $\pm$ 24.48	34.82 $\pm$ 12.64	56.33 $\pm$ 13.48	49.00 $\pm$ 9.42
GCOPE	OOT	OOT	OOT	OOT	18.25 $\pm$ 4.67	14.35 $\pm$ 2.53	47.70 $\pm$ 32.64	28.97 $\pm$ 16.13	41.41 $\pm$ 16.23	38.46 $\pm$ 13.22
MDGPT	10.85 $\pm$ 1.90	8.02 $\pm$ 0.97	14.00 $\pm$ 2.58	5.34 $\pm$ 0.23	18.43 $\pm$ 3.62	16.03 $\pm$ 2.66	58.31 $\pm$ 16.94	36.25 $\pm$ 7.37	53.97 $\pm$ 11.76	50.35 $\pm$ 9.68

Appendix

Table 16. Evaluation of 5-shot edge classification and graph classification on unseen domains.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	4982.84	145726.60	3609.70	2399.45	2040.97	OOT	10597.74	OOT	30164.69
MDGPT	996.96	21295.45	578.18	559.66	621.26	41059.80	1677.46	6052.41	2246.63
MDGFM	4276.28	68365.23	520.57	571.15	977.57	206484.25	5429.48	4385.13	9203.33
SAMGPT	3100.42	37458.71	1302.22	1395.42	1413.17	101031.49	4254.17	7890.00	2561.89
G2P2	268.5	-	284.01	284.77	-	-	3150.19	1928.43	-
GraphCLIP	1143.31	-	549.77	603.42	-	-	1578.20	10320.75	-

Appendix

Table 17. Time of downstream tuning for 1-shot node classification on unseen domains.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	0.9966	29.1453	0.7219	0.4799	0.4082	OOT	2.1195	OOT	6.0329
MDGPT	0.0499	2.3057	0.0289	0.0280	0.0311	5.4761	1.1796	0.3600	2.2443
MDGFM	0.2952	4.5010	0.1541	0.0990	0.0744	12.7894	1.3986	0.6626	2.5256
SAMGPT	0.1554	2.7046	0.0721	0.0707	0.0711	6.5106	1.2225	0.4005	2.4340
G2P2	0.1074	-	0.1136	0.1139	-	-	1.2601	0.7714	-
GraphCLIP	0.2287	-	0.1100	0.1207	-	-	0.3156	2.0642	-

Appendix

Table 18. Time/epoch of downstream tuning for 1-shot node classification on unseen domains.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	4967.40	145716.29	3583.13	2419.86	2036.05	OOT	10599.70	OOT	30412.11
MDGPT	91.60	30592.08	573.91	559.37	620.14	29485.19	1983.93	6030.87	2309.75
MDGFM	4103.18	OOM	549.41	582.99	998.64	195360.13	4520.15	5207.33	11524.04
SAMGPT	3107.38	97086.74	1462.31	1375.35	1276.83	105312.52	4113.18	7786.37	2616.61
G2P2	279.28	-	284.19	278.05	-	-	3365.97	1920.82	-
GraphCLIP	1234.62	-	597.31	675.31	-	-	1624.89	10517.09	-

Appendix

Table 19. Time of downstream tuning for 5-shot node classification on unseen domains.

Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	0.9935	29.1433	0.7166	0.4840	0.4072	OOT	2.1199	OOT	6.0824
MDGPT	0.0503	4.0784	0.0287	0.0280	0.0311	6.6558	1.2596	0.3580	2.3765
MDGFM	0.2977	OOM	0.1592	0.0988	0.0819	16.6476	1.4073	0.5761	3.0493
SAMGPT	0.1554	5.5339	0.0731	0.0688	0.0638	9.0304	1.3162	0.3964	2.6281
G2P2	0.1117	-	0.1137	0.1112	-	-	1.3464	0.7683	-
GraphCLIP	0.2469	-	0.1195	0.1351	-	-	0.3250	2.1034	-

Appendix

Table 20. Time/epoch of downstream tuning for 5-shot node classification on unseen domains.

Methods	Edge classification						Graph classification
	DGraph		Wiki		WN18RR		PCBA		BZR
	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.
GCOPE	OOT	OOT	OOT	OOT	5366.42	1.0733	7849.48	1.5689	1999.33	0.3999
MDGPT	6691.48	0.6109	27480.67	2.1883	1482.20	0.0741	1171.83	0.0777	373.24	0.0254
MDGFM	16287.70	1.1188	79650.24	4.0031	12818.37	0.7522	9000.25	1.8664	194.41	0.0515
SAMGPT	14194.53	0.8092	52318.30	2.6658	5529.34	0.2765	2046.01	0.1056	413.24	0.0242

Appendix

Table 21. Time and time/epoch of downstream tuning for 1-shot edge classification and graph classification on unseen domains.

Methods	Edge classification						Graph classification
	DGraph		Wiki		WN18RR		PCBA		BZR
	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.
GCOPE	OOT	OOT	OOT	OOT	5377.28	1.0755	7869.82	1.5740	1990.91	0.3982
MDGPT	3239.32	0.5356	33128.83	3.1503	1467.97	0.0744	1159.48	0.0925	421.82	0.0232
MDGFM	14083.09	1.1941	78096.08	3.9079	12869.09	0.7497	9296.78	1.0705	233.65	0.0366
SAMGPT	17314.93	0.9869	53692.22	2.8210	5531.02	0.2766	1943.27	0.1037	297.63	0.0152

Appendix

Table 22. Time and time/epoch of downstream tuning for 5-shot edge classification and graph classification on unseen domains.

Methods

Cora

ACM

Wisconsin

Photo

Elliptic

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

GCN

69.86

\pm

4.93

68.39

\pm

4.21

52.49

\pm

12.80

44.46

\pm

16.02

51.02

\pm

20.66

35.75

\pm

10.64

42.85

\pm

4.74

31.75

\pm

3.62

76.09

\pm

7.03

74.45

\pm

6.73

51.90

\pm

12.45

42.21

\pm

7.92

GAT

68.57

\pm

4.38

67.55

\pm

4.03

39.28

\pm

7.90

26.88

\pm

11.74

49.74

\pm

15.65

35.81

\pm

7.48

40.35

\pm

4.80

30.58

\pm

4.14

76.11

\pm

4.54

75.25

\pm

3.57

50.43

\pm

14.12

40.17

\pm

8.47

Simple-HGN

66.31

\pm

20.71

62.29

\pm

26.23

FAGCN

59.42

\pm

9.09

47.42

\pm

6.96

DGI

58.70

\pm

6.05

55.59

\pm

5.33

42.48

\pm

10.45

38.31

\pm

11.38

50.85

\pm

9.54

36.41

\pm

4.31

31.93

\pm

7.41

25.35

\pm

5.89

56.69

\pm

7.18

55.48

\pm

6.59

44.95

\pm

14.57

35.26

\pm

10.62

Appendix

Table 23. Evaluation of 5-shot node classification on seen domains.

Methods	Edge classification		Graph classification
	FB15K-237		HIV		COX2		PROTEINS
	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCN	26.85 $\pm$ 1.55	25.90 $\pm$ 0.97	46.39 $\pm$ 28.97	30.91 $\pm$ 15.60	58.73 $\pm$ 11.32	50.45 $\pm$ 7.36	57.79 $\pm$ 8.91	55.05 $\pm$ 8.51
GAT	24.62 $\pm$ 1.33	24.57 $\pm$ 0.81	48.22 $\pm$ 20.26	34.07 $\pm$ 10.62	57.52 $\pm$ 10.63	50.90 $\pm$ 7.38	58.40 $\pm$ 8.52	56.20 $\pm$ 8.28
DGI	8.54 $\pm$ 1.23	9.40 $\pm$ 0.58	57.22 $\pm$ 17.97	38.35 $\pm$ 7.93	51.93 $\pm$ 10.63	45.77 $\pm$ 6.65	56.77 $\pm$ 8.38	52.55 $\pm$ 7.95
GraphPrompt	12.47 $\pm$ 1.16	10.42 $\pm$ 0.55	51.07 $\pm$ 9.62	36.41 $\pm$ 4.35	53.11 $\pm$ 7.82	47.14 $\pm$ 4.53	54.35 $\pm$ 8.31	52.96 $\pm$ 7.86

Appendix

Table 24. Evaluation of 5-shot edge classification and graph classification on seen datasets.

Methods	Cora		ACM		Reddit		Wisconsin		Photo		Elliptic
Methods	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.
GCN	31.31	0.0140	122.84	0.0636	127.02	0.1119	129.34	0.0255	828.39	0.0698	131.64	0.1707
GAT	50.53	0.0298	184.55	0.0955	159.57	0.1886	104.03	0.0406	410.57	0.0788	174.17	0.2399
Simple-HGN	-	-	3764.31	0.2510	-	-	-	-	-	-	-	-
TGN	-	-	-	-	1660.39	0.1660	-	-	-	-	-	-
FAGFN	-	-	-	-	-	-	217.31	0.0232	-	-	-	-

Appendix

Table 25. Time and time/epoch of downstream tuning for 1-shot node classification on seen datasets.

Methods	Cora		ACM		Reddit		Wisconsin		Photo		Elliptic
Methods	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.
GCN	257.61	0.0297	246.73	0.0681	318.01	0.1295	560.74	0.0310	196.12	0.0269	171.76	0.1286
GAT	296.70	0.0464	228.42	0.1224	220.35	0.1805	195.15	0.0213	500.67	0.0357	245.24	0.1822
Simple-HGN	-	-	3765.52	0.2510	-	-	-	-	-	-	-	-
TGN	-	-	-	-	1664.02	0.1664	-	-	-	-	-	-
FAGFN	-	-	-	-	-	-	282.46	0.0303	-	-	-	-

Appendix

Table 26. Time and time/epoch of downstream tuning for 5-shot node classification on seen datasets.

Methods	Edge classification		Graph classification
	FB15K-237		HIV		COX2		PROTEINS
	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.
GCN	1220.51	0.0639	166.80	0.0414	47.53	0.0250	44.65	0.0346
GAT	923.32	0.0782	145.08	0.0715	42.43	0.0280	39.23	0.0346
DGI	4468.84	3.4212	3769.68	2.8540	43.64	0.0248	119.66	0.2176
GraphPrompt	2239.06	0.2528	2495.85	2.5379	35.17	0.0640	181.10	0.3293

Appendix

Table 27. Time and time/epoch of downstream tuning for 1-shot edge classification and graph classification on seen datasets.

Methods	Edge classification		Graph classification
	FB15K-237		HIV		COX2		PROTEINS
	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.	Time	Avg. T.
GCN	1002.11	0.0796	332.07	0.0303	375.00	0.0277	127.64	0.0163
GAT	779.50	0.0876	290.26	0.0269	95.51	0.0110	119.32	0.0121
DGI	4640.11	3.6991	3807.46	2.9227	56.63	0.0302	131.69	0.2394
GraphPrompt	2599.97	0.2667	2651.73	2.8213	44.45	0.0808	181.88	0.3307

Appendix

Table 28. Time and time/epoch of downstream tuning for 5-shot edge classification and graph classification on seen datasets.

Methods	Citation		Social & Web			E-commerce	Finance		Proteins
Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	36.09 $\pm$ 6.14	0.14 $\pm$ 0.01	34.55 $\pm$ 14.27	17.67 $\pm$ 4.50	19.46 $\pm$ 3.48	OOT	26.43 $\pm$ 12.11	OOT	34.16 $\pm$ 7.92
MDGPT	45.38 $\pm$ 11.09	3.41 $\pm$ 0.39	33.40 $\pm$ 14.27	21.76 $\pm$ 5.68	24.47 $\pm$ 4.99	9.81 $\pm$ 1.54	43.76 $\pm$ 16.18	26.27 $\pm$ 6.42	44.41 $\pm$ 9.33
MDGFM	29.01 $\pm$ 5.05	0.83 $\pm$ 0.85	32.82 $\pm$ 14.06	13.97 $\pm$ 3.09	21.24 $\pm$ 4.12	1.63 $\pm$ 0.34	36.51 $\pm$ 12.34	21.65 $\pm$ 12.89	43.64 $\pm$ 6.17
SAMGPT	46.20 $\pm$ 7.43	4.98 $\pm$ 0.27	36.27 $\pm$ 14.67	19.74 $\pm$ 4.39	25.21 $\pm$ 4.37	13.33 $\pm$ 1.35	37.19 $\pm$ 9.70	27.00 $\pm$ 7.55	49.89 $\pm$ 10.20
G2P2	42.16 $\pm$ 8.28	-	21.80 $\pm$ 2.66	22.98 $\pm$ 6.09	-	-	44.30 $\pm$ 10.97	22.99 $\pm$ 3.99	-

Appendix

Table 29. Marco-F1 of 1-shot node classification on topic domain adaptation.

Methods	Citation		Social & Web			E-commerce	Finance		Proteins
Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	55.76 $\pm$ 5.66	0.26 $\pm$ 0.02	53.74 $\pm$ 20.62	30.36 $\pm$ 8.68	25.42 $\pm$ 2.32	OOT	29.14 $\pm$ 16.34	OOT	41.77 $\pm$ 9.42
MDGPT	62.51 $\pm$ 6.95	9.55 $\pm$ 0.90	58.41 $\pm$ 17.83	37.83 $\pm$ 7.13	33.79 $\pm$ 2.88	21.28 $\pm$ 9.81	61.10 $\pm$ 12.17	39.44 $\pm$ 9.47	53.75 $\pm$ 10.80
MDGFM	42.34 $\pm$ 7.76	OOM	49.89 $\pm$ 19.61	19.70 $\pm$ 4.81	27.50 $\pm$ 4.17	6.47 $\pm$ 2.25	49.76 $\pm$ 17.06	43.37 $\pm$ 17.01	50.50 $\pm$ 6.30
SAMGPT	61.36 $\pm$ 5.91	11.23 $\pm$ 0.66	54.74 $\pm$ 14.32	38.92 $\pm$ 5.53	37.75 $\pm$ 3.44	31.11 $\pm$ 6.32	49.73 $\pm$ 9.53	40.75 $\pm$ 8.97	55.17 $\pm$ 9.88
G2P2	61.21 $\pm$ 6.44	-	47.93 $\pm$ 8.18	38.05 $\pm$ 5.40	-	-	64.46 $\pm$ 11.21	44.04 $\pm$ 8.22	-

Appendix

Table 30. Accuracy of 5-shot node classification on topic domain adaptation.

Methods	Citation		Social & Web			E-commerce	Finance		Proteins
Methods	Pubmed	ogbn-mag	Wikipedia	Actor	Chameleon	Products	T-Finance	DGraph	ogbn-proteins
GCOPE	55.38 $\pm$ 5.62	0.14 $\pm$ 0.01	36.23 $\pm$ 10.58	24.36 $\pm$ 5.49	24.39 $\pm$ 2.11	OOT	24.05 $\pm$ 9.55	OOT	31.69 $\pm$ 7.16
MDGPT	62.55 $\pm$ 6.89	7.20 $\pm$ 0.51	38.30 $\pm$ 8.14	30.31 $\pm$ 4.24	31.96 $\pm$ 3.40	8.18 $\pm$ 2.32	45.30 $\pm$ 7.33	29.07 $\pm$ 4.80	47.79 $\pm$ 9.92
MDGFM	39.55 $\pm$ 7.68	OOM	34.33 $\pm$ 9.29	15.39 $\pm$ 2.36	26.50 $\pm$ 4.72	3.45 $\pm$ 0.80	38.26 $\pm$ 10.09	28.54 $\pm$ 6.77	47.36 $\pm$ 4.75
SAMGPT	61.06 $\pm$ 5.94	9.24 $\pm$ 0.32	37.08 $\pm$ 6.25	30.88 $\pm$ 3.47	37.01 $\pm$ 3.46	15.53 $\pm$ 1.90	39.05 $\pm$ 5.66	29.91 $\pm$ 4.46	53.46 $\pm$ 9.39
G2P2	60.53 $\pm$ 6.56	-	22.97 $\pm$ 2.58	32.22 $\pm$ 3.73	-	-	47.66 $\pm$ 6.83	25.55 $\pm$ 2.96	-

Appendix

Table 31. Marco-F1 of 5-shot node classification on topic domain adaptation.

Methods	Finance		Common sense				Molecule
	DGraph		Wiki		WN18RR		PCBA		BZR
	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCOPE	OOT	OOT	OOT	OOT	18.15 $\pm$ 4.25	13.59 $\pm$ 2.37	51.25 $\pm$ 15.99	33.35 $\pm$ 7.09	55.69 $\pm$ 12.07	50.56 $\pm$ 8.51
MDGPT	10.13 $\pm$ 2.59	10.13 $\pm$ 2.59	8.79 $\pm$ 1.63	2.24 $\pm$ 0.26	19.27 $\pm$ 3.32	15.54 $\pm$ 2.12	61.39 $\pm$ 15.70	37.73 $\pm$ 6.11	57.21 $\pm$ 9.53	53.11 $\pm$ 7.66
MDGFM	9.42 $\pm$ 2.35	6.47 $\pm$ 0.81	4.78 $\pm$ 2.52	1.07 $\pm$ 0.15	15.15 $\pm$ 3.33	10.90 $\pm$ 1.64	54.32 $\pm$ 30.48	32.64 $\pm$ 14.65	39.62 $\pm$ 19.77	31.45 $\pm$ 8.42
SAMGPT	10.41 $\pm$ 1.48	7.71 $\pm$ 0.57	15.11 $\pm$ 2.72	5.90 $\pm$ 0.24	18.06 $\pm$ 3.49	14.70 $\pm$ 1.84	57.40 $\pm$ 11.87	36.38 $\pm$ 4.76	66.34 $\pm$ 9.90	58.03 $\pm$ 6.84

Appendix

Table 32. Evaluation of 5-shot edge classification and graph classification on topic domain adaptation.

Methods

Heterogeneous

Heterophilic

Dynamic

Textual

ogbn-mag

Chameleon

Actor

Wikipedia

Pubmed

Products

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

Acc

MacroF

GCOPE

0.34

\pm

0.05

0.11

\pm

0.01

26.77

\pm

1.82

25.24

\pm

1.90

27.24

\pm

7.28

23.32

\pm

4.79

57.89

\pm

16.82

38.58

\pm

7.89

56.56

\pm

5.41

55.48

\pm

5.32

OOT

MDGPT

11.33

\pm

0.64

8.63

\pm

0.37

34.58

\pm

3.58

32.84

\pm

4.24

38.14

\pm

7.42

30.52

\pm

4.47

49.72

\pm

15.67

34.65

\pm

7.57

59.52

\pm

5.76

59.83

\pm

6.03

26.69

\pm

7.85

12.29

\pm

2.04

MDGFM

OOM

25.50

\pm

3.38

23.78

\pm

3.82

19.75

\pm

5.24

14.87

\pm

2.27

49.71

\pm

23.54

33.45

\pm

11.56

37.95

\pm

5.56

34.91

\pm

5.16

5.39

\pm

3.62

2.42

\pm

0.71

SAMGPT

12.33

\pm

0.71

9.86

\pm

0.33

37.27

\pm

3.48

36.35

\pm

3.50

38.52

\pm

5.37

30.69

\pm

3.36

52.67

\pm

12.71

36.30

\pm

5.89

63.88

\pm

5.30

63.78

\pm

5.29

32.07

\pm

8.02

15.53

\pm

1.98

Appendix

Table 33. Evaluation of 5-shot node classification on format domain adaptation.

Methods	Relational				Textual
	Wiki		WN18RR		PCBA
	Acc	MacroF	Acc	MacroF	Acc	MacroF
GCOPE	OOT	OOT	18.91 $\pm$ 4.13	14.92 $\pm$ 2.17	56.03 $\pm$ 26.67	33.97 $\pm$ 12.64
MDGPT	12.24 $\pm$ 2.74	4.14 $\pm$ 0.65	19.02 $\pm$ 3.24	15.73 $\pm$ 2.10	56.56 $\pm$ 18.54	35.36 $\pm$ 8.28
MDGFM	8.42 $\pm$ 3.77	1.49 $\pm$ 0.30	19.09 $\pm$ 5.56	13.34 $\pm$ 2.32	53.99 $\pm$ 44.41	28.92 $\pm$ 22.20
SAMGPT	15.15 $\pm$ 2.81	6.10 $\pm$ 0.22	17.06 $\pm$ 3.39	14.08 $\pm$ 1.74	56.29 $\pm$ 14.33	35.66 $\pm$ 6.18

Appendix

Table 34. Evaluation of 5-shot edge classification and graph classification on format domain adaptation.

Explore similar work

May 14, 2026

GFMate: Empowering Graph Foundation Models with Test-time Prompt Tuning

Graph prompt tuning has shown great potential in graph learning by introducing trainable prompts to enhance the model performance in conventional single-domain scenarios. Recent research has extended graph prompts to improve Graph Foundation Models (GFMs) by few-shot tuning auxiliary prompts. Despite their progress, most existing methods embed source-domain information into prompts, which serve either as input to GFMs or encoded during model pre-training. Such prompt entanglement with specific source domains and GFM pre-training strategy restricts their generalisability to other domains and different GFMs. Furthermore, existing GFM prompts merely rely on few-shot tuning for adaptation, neglecting the rich information in unlabelled target domain test data. Motivated by these insights, this paper aims to empower GFMs with pre-training-agnostic test-time graph prompt tuning, named GFMate. GFMate introduces centroid and layer prompts applied after pre-training on target domains, avoiding entanglement with specific source domains and model pre-training. In addition, a test-time complementary learning objective is devised to exploit both labelled and unlabelled target domain data for effective test-time prompt tuning. Extensive experiments on 12 benchmark datasets demonstrate the superior performance and efficiency of GFMate, achieving improvements of up to 30.63%. Code is available at https://github.com/YanJiangJerry/GFMate.

Yan Jiang, Ruihong Qiu, Zi Huang

School of Electrical Engineering and Computer Science, The University of Queensland, Brisbane, Queensland, Australia

Domain Adaptation Prompt Tuning Graph Foundation Models Test-Time Prompt Tuning

Ask PDF

May 28, 2026

When Do Graph Foundation Models Transfer? A Data-Centric Theory

Graph foundation models (GFMs) aim to reuse a single backbone across diverse graph domains, yet their transfer is often uneven and can exhibit negative transfer. While most prior work improves transfer through architectural or adaptation choices, we ask a data-centric question: which properties of two graph domains determine how much a fixed representation model changes its outputs? Using a graphon-based continuous limit for dense graphs, we show that for both set-based and message-passing tokenizations, any Lipschitz backbone admits an explicit decomposition of cross-domain output shift into (i) graph-specific finite-sample approximation terms and (ii) an intrinsic, relabeling-invariant domain discrepancy capturing structural mismatch. A key ingredient is positional-encoding (PE) stability: we establish stability guarantees for spectral PEs and highlight contrasting behaviors of eigenvector- versus subspace-based PEs. Experiments on synthetic and real graphs validate the theory and translate the decomposition into guidance for data curation in GFM transfer.

Jiajun Zhu, Ying Chen, Peihao Wang +4

University of Texas at Austin · Arizona State University · Georgia Institute of Technology

Graph Representation Learning Cross-Domain Transfer Learning Graph Foundation Models

Ask PDF

Jul 13, 2026

Surprisingly Simple and Effective Multi-Domain Graph Foundation Model through Graph-to-Table Alignment

Graph Foundation Models (GFMs) have emerged as a promising paradigm for learning transferable representations across diverse graph domains. Recent advancements in GFMs have been largely dominated by two paradigms: Graph Neural Network and Large Language Model (LLM) based methods. However, these methods often face a fundamental dilemma between training with limited data and a heavy reliance on textual attributes. Tabular foundation models (TFMs) offer a potential alternative, as node features and representations can be naturally organized in a tabular form. However, how to enable TFMs to effectively capture structural information of graphs remains largely unexplored. The key challenge is to learn a graph-to-table alignment mechanism that enables graph structural understanding for TFMs. To address this, we propose GTAlign, a surprisingly simple yet effective Graph-to-Table Alignment framework for text-free Graph Foundation Model. Specifically, we first pretrain a graph encoder that maps diverse graphs into a unified latent space to capture domain-agnostic graph representations. To further bridge the gap between graph topology and the tabular representation space, we propose community-guided continual pre-training, where pseudo-labels derived from graph community are used to construct few-shot prediction episodes. Lastly, we adapt the graph encoder for an unseen target domain and perform in-context inference. Extensive experiments on five benchmark datasets demonstrate that GTAlign significantly outperforms state-of-the-art baselines on both node and graph classification, offering a simple, effective, and text-free GFM model. Code will be released upon acceptance.

Chunyu Hu, Tianyin Liao, Ge Lan +4

Nankai University · Tsinghua University · Beihang University

Tabular Foundation Models Graph Representation Learning Graph Foundation Models

Ask PDF