The same node attributes can imply different labels in different structural contexts. Yet fast graphless models typically treat topology as additional evidence supplied to a globally shared predictor. This assumes that structure augments the available information, rather than determining how node attributes should be interpreted. We introduce SCOPE (Structure-COnditioned PrEdictor), in which a small generator maps each node's cached structural fingerprint to coefficients over a shared low-rank basis, producing a context-specific update to a global weight matrix. The base predictor, basis, and generator remain shared across all nodes, while the resulting transformation adapts how attributes are processed in each relational environment. Once fingerprints are cached, inference requires only a forward pass and no online message passing. This design combines context-specific specialization with global parameter sharing, compact low-rank adaptation, and message-passing-free inference. On two large OGB node-classification graphs, conditioning the weights with raw positional signals yields clear gains over supplying the same signals through input concatenation. Further results under a structure-transductive held-out-node protocol and in teacher-free graph classification show that the mechanism extends beyond a particular distillation setting. These findings establish graph structure not only as predictive evidence, but as context that determines the predictor used to interpret that evidence.
Figures & tables
Figure 1. Existing methods inject structure via inputs, supervision, or initialization, or generate weights from an unconditioned latent; SCOPE instead maps cached structural fingerprints to context-dependent effective weights without online neighborhood aggregation.
Figure 2. 2D PCA of the per-node generator output s(zv) on ogbn-arxiv at four training epochs (shared axes), colored by degree percentile (top) and class label (bottom). Organization by degree is consistent with sensitivity to structural context, while class organization indicates task-aligned variation.
Variant
Acc.
Δ vs. GLNN
( M1 ) structure-free
GLNN (plain MLP + KD)
58.98 ± 0.53
0.00
( M2 ) structure as input
+ concatenate zv
63.46 ± 0.30
+4.48
( M3 ) structure as adaptation context
without conditioning (controls)
Table 1. Mechanism ladder on ogbn-arxiv under ho-full evaluation (%). The axis swept is how structural information enters the model, ( M1 )–( M3 ). M3 includes variants of SCOPE .
ogbn-arxiv
ogbn-products
Signal (dim)
Used as
tran
ho-full
Δ
tran
ho-full
Δ
Teacher (SAGE)
message passing
71.08 ± 0.42
70.52 ± 0.36
78.15 ± 0.39
78.18 ± 0.26
none (GLNN)
–
64.81 ± 0.28
58.98 ± 0.53
0.0
64.60 ± 0.34
64.16 ± 0.19
0.0
node-wise spectral structural summaries
HKS (8)
generation
66.11 ± 0.36
59.51 ± 0.63
+0.5
63.70 ± 0.30
63.15 ± 0.26
−1.0
RWSE (8)
generation
66.04 ± 0.25
59.75 ± 0.44
+0.8
63.68 ± 0.46
63.15 ± 0.29
−1.0
Table 2. Changing the structural signal and its role ( SCOPE rank 32 on arxiv , 64 on products , %). Parenthesized values are signal dimensions; Δ is the ho-full gain over GLNN.
Figure 3. Rank behaviour and node-specificity on ogbn-arxiv . (a) Transductive accuracy (%) versus adaptation rank. (b) ECDF of per-node Δ KL to the teacher ( SCOPE r32 ); positive favours a node’s own fingerprint. Left: accuracy versus conditioning rank for Diag and \method. Right: empirical CDF of per-node KL differences under own, swapped, and disabled fingerprints.
Config
KLown
struct-gain
frac. helped
node-spec.
frac. own better
r16
0.388
0.647
0.854
1.896
0.891
r32
0.358
0.691
0.870
2.041
0.900
r64
0.315
0.726
0.888
2.074
0.915
r32 thick
0.329
0.764
0.893
2.366
0.922
Table 3. Functional node-specificity on ogbn-arxiv . Struct-gain is KLoff−KLown and node-spec. is KLswap−KLown .
Figure 4. Measured weight-space diagnostics on ogbn-arxiv . (a) Per-node generator outputs with chart-restricted Fisher directions. (b) A node’s replayed optimization path. (c) Convergence of the structure-reachable subspace.
Transductive
Held-out nodes (ho-full)
Method
Cora
Citeseer
Pubmed
A-computer
A-photo
ogbn-arxiv
ogbn-products
Cora
Citeseer
Pubmed
A-computer
A-photo
ogbn-arxiv
ogbn-products
Graph-based task references
Teacher (SAGE)
80.49 ± 1.90
70.23 ± 2.15
75.42 ± 1.85
82.11 ± 1.88
90.71 ± 1.26
71.08 ± 0.42
78.15 ± 0.39
81.24 ± 1.91
70.00 ± 2.91
75.37 ± 2.74
82.91 ± 1.56
91.04 ± 1.21
70.52 ± 0.36
78.18 ± 0.26
SGC
80.81 ± 1.37
70.91 ± 2.36
74.53 ± 1.91
83.74 ± 1.58
91.13 ± 1.12
69.31 ± 0.31
68.26 ± 0.15
80.33 ± 2.13
70.83 ± 3.02
72.94 ± 2.63
83.58 ± 1.84
91.31 ± 0.81
69.07 ± 0.69
68.45 ± 0.16
Learned positional-input reference (CE, no teacher)
DW+MLP
75.50 ± 1.25
62.35 ± 2.80
67.98 ± 2.75
76.08 ± 1.86
79.87 ± 1.51
69.45 ± 0.25
78.06 ± 0.23
71.38 ± 2.78
63.92 ± 3.62
65.34 ± 2.96
80.30 ± 1.43
81.13 ± 1.72
69.81 ± 0.35
78.88 ± 0.43
Table 4. Node classification accuracy. Graph-based and positional rows are task-level references rather than matched mechanism controls.
ogbn-arxiv
ogbn-products
Method
Structural route
tran
ho-obs
tran
ho-obs
TINED ( Zhou et al., 2025 )
teacher-weight transfer
64.44 ± 0.72
59.79 ± 0.46
69.48 ± 0.25
68.68 ± 0.27
SCOPE
conditioned weights
68.87 ± 0.40
63.73 ± 0.45
70.52 ± 0.30
69.44 ± 0.26
NOSMOG ( Tian et al., 2023 )
DeepWalk input
71.65 ± 0.29
70.09 ± 0.55
78.45 ± 0.38
77.02 ± 0.19
SCOPE + DeepWalk
DeepWalk-conditioned weights
71.56 ± 0.30
69.98 ± 0.47
78.93 ± 0.44
78.63 ± 0.34
TINED+ ( Zhou et al., 2025 )
teacher transfer + DeepWalk input
71.52 ± 0.34
70.42 ± 0.35
78.59 ± 0.28
79.31 ± 0.29
Table 5. System-level context on the two OGB graphs. External numbers are as published; SCOPE uses our protocol under ho-obs (§ 4.1 ). Bold marks the higher entry of a pair.
Figure 5. Utility and cost on the two OGB graphs. (a) Increment from SCOPE over the matching plain MLP under CE and KD. (b) Prediction latency versus graph density; dotted lines mark the two datasets. Left: accuracy increment from \methodwith and without distillation on two OGB graphs. Right: inference latency versus graph density for message passing, a shared MLP, and \method.
ROC-AUC
10-fold accuracy
Arm
molhiv
PROTEINS
DD
IMDB-B
REDDIT-B
NCI1
no structural signal
plain MLP
68.3 ±
0.7
60.6 ±
2.2
70.3 ±
3.8
70.2 ±
3.6
62.8 ±
1.6
57.1 ±
1.4
structure as input
fingerprint-only zG
70.4 ±
0.6
74.2 ±
3.8
76.2 ±
2.9
70.1 ±
1.6
83.3 ±
2.2
70.0 ±
2.4
input-concat zv
76.7 ±
0.9
68.4 ±
4.0
72.6 ±
3.3
71.8 ±
4.0
78.8 ±
2.0
68.7 ±
2.1
Table 6. Graph classification accuracy. ogbg-molhiv reports ROC-AUC over seeds; the five TU columns report nested-cross-validation accuracy over folds. Generation and broadcast variants use rank 64 .
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
nodes
edges
feat. dim
classes
teacher
student
lr
wd
drop.
r / k
Cora
2,485
5,069
1,433
7
2 × 128
2 × 128
0.010
5e-3
0.6
8 / 18
Citeseer
2,110
3,668
3,703
6
2 × 128
2 × 128
0.010
1e-3
0.1
8 / 18
Pubmed
19,717
44,324
500
3
2 × 128
2 × 128
0.005
1e-3
0.4
8 / 18
A-computer
13,381
245,778
767
10
2 × 128
2 × 128
0.003
5e-3
0.1
8 / 18
A-photo
7,487
119,043
745
8
2 × 128
2 × 128
0.002
1e-3
0.3
8 / 18
ogbn-arxiv
169,343
1,166,243
128
40
3 × 256 BN
3 × 256 BN
0.010
0
0.2
32 / 7
Appendix
Table 7. Node-classification datasets, architectures, and training hyperparameters. lr/wd/dropout are the student’s; r is SCOPE ’s rank, k the fingerprint dimension.
Figure 6. SCOPE overview: a cached structural fingerprint zv is mapped to coefficients s(zv) that select a context-specific update to the shared base weight. The graph-level variant broadcasts one zG per graph.
Dataset
graphs
avg. nodes
hidden
layers
batch
epochs
ogbg-molhiv
41,127
25.5
300
3
256
100
PROTEINS
1,113
39.1
128
2
128
100
DD
1,178
284.3
128
2
64
100
IMDB-B
1,000
19.8
128
2
128
100
REDDIT-B
2,000
429.6
128
2
64
100
NCI1
4,110
29.9
128
2
128
100
Appendix
Table 8. Graph-classification datasets and student hyperparameters (all use rank r=64 ). TU splits use 10-fold CV, ogbg-molhiv the official scaffold split.
Method
Cora
Citeseer
Pubmed
A-computer
A-photo
Transductive
TINED ( Zhou et al., 2025 )
82.63 ± 1.57
74.43 ± 1.53
77.09 ± 2.14
85.18 ± 1.12
93.97 ± 0.53
SCOPE
80.60 ± 2.16
72.11 ± 1.46
76.25 ± 1.88
82.42 ± 1.89
92.06 ± 0.93
NOSMOG ( Tian et al., 2023 )
83.04 ± 1.26
73.78 ± 1.54
77.34 ± 2.36
84.04 ± 1.01
93.36 ± 0.69
SCOPE + DeepWalk
80.57 ± 2.45
72.18 ± 1.32
76.25 ± 1.96
82.60 ± 2.02
91.23 ± 0.94
TINED+ ( Zhou et al., 2025 )
83.70 ± 1.02
75.39 ± 1.59
77.75 ± 3.14
84.82 ± 1.58
94.05 ± 0.39
Appendix
Table 9. System-level context on the five CPF benchmarks, transductive (top) and ho-obs (bottom), paired as in Table 5 . External numbers as reported by TINED ( Zhou et al., 2025 ) ; SCOPE rows are per-dataset tuned.
Variant
tran
ho-full
(a) conditioned: rank sweep on ogbn-products
SCOPE r32
70.10
69.64
SCOPE r64
70.63
70.06
(b) unconditioned: Low-rank control vs. GLNN, both graphs
Low-rank r8 (arxiv)
53.60
53.45
Low-rank r32 , matched (arxiv)
60.08
58.05
Appendix
Table 10. Rank behavior on the OGBN graphs (%). (a) SCOPE rank sweep on ogbn-products . (b) The unconditioned Low-rank control never improves on GLNN, and collapses at r8 .
Constraint added (cumulative)
tran
ho-full
none ( SCOPE , all learned)
69.00
64.50
W0← teacher
68.85
64.01
+ W0 frozen
65.21
63.35
+ U,V← SVD( W0 )
65.54
63.45
+ U,V frozen (fully constrained)
61.87
60.99
control: SVD basis, orthogonalized
64.72
63.71
Appendix
Table 11. Design-constraint spine on ogbn-arxiv ( r32 , %).
Config
κ=8
16
32
64
128
r16
1.41
1.10
1.48
1.15
1.09
r32
0.33
0.52
0.51
0.45
0.46
r64
0.23
0.24
0.21
0.21
0.21
r32 thick
0.67
0.70
0.60
0.52
0.56
Appendix
Table 12. Subspace-overlap test on ogbn-arxiv (transductive). Overlap between the structure-reachable subspace and the FIM top- κ eigenspace, normalized so 1.0= random. The r32 thick row is exploratory.
Figure 7. PCA trajectories of vec(Ut) and vec(Vt) over training on ogbn-arxiv ( 21 snapshots, every 10 epochs; each factor on its own top-2 basis, color light to dark).
Figure 8. Illustrative, not evidentiary: nodes at a fixed, graph-only Laplacian-eigenvector layout (identical in all panels), colored by the basis-invariant m(v) (top) and struct-gain (v) (bottom), under LapPE (left) and degree/PageRank (right).
Figure 9. ROC-AUC vs. rank for the per-node SCOPE student on ogbg-molhiv , against the GIN teacher and plain-MLP references.
val
test
(a) main comparison
GIN teacher, plain (ref)
80.56
73.57
plain MLP (CE)
72.18
68.64
plain MLP (KD)
72.07
66.89
per-graph SCOPE (CE, r64 )
79.67
71.58
per-graph SCOPE (KD)
79.45
70.90
Appendix
Table 13. Per-graph spectral SCOPE on ogbg-molhiv (ROC-AUC); student configuration as in Table 8 . The r64 (CE) and edge-aware GIN rows are those reported in Table 6 ; (b) is a separate sweep.
Achieving strong performance with graph neural networks (GNNs) typically requires training and hyperparameter tuning for each dataset, incurring repeated costs and effort. Graph in-context learning (ICL) avoids this by using a single pretrained model to predict unknown node labels directly from labeled context nodes. Existing approaches, however, rely on dense attention across nodes, making inference increasingly expensive as graphs grow. In this work, we present Ephris, a new graph in-context learner built on sparse message passing, scaling linearly with the number of node-feature entries and graph edges. Ephris is pretrained entirely on synthetic graphs generated from structural causal models with diverse graph structures and relational dynamics, exposing the model to varied dependencies among topology, features, and labels. We evaluate Ephris on 51 node-classification datasets against 15 extensively tuned GNNs and existing graph ICL methods under both high- and low-label train/validation/test splits. Across both settings, Ephris ranks first on all four aggregate measures: Elo, improvability, average rank, and accuracy. Its inference cost remains comparable to training a single GNN once, while being over 10 times faster than previous graph ICL models. Together, these results advance the performance-runtime Pareto frontier, demonstrating that strong graph ICL does not require dense attention. Code and model weights are available at https://github.com/nums-ai/ephris.
Graph foundation models aim to transfer across graphs, feature spaces, relational schemas, and prediction tasks, yet existing approaches typically generalize only within particular graph modalities or tasks. We propose Wander, a graph foundation model designed to operate across these settings within a single pretrained checkpoint. Following the prior-predictive perspective, we formulate graph learning as completion of a partially observed graph. We realize this task-general view through a common interface based on random walks, allowing the same model to operate across homogeneous and multi-relational graphs with varying features, labels, and relational schemas. Wander can increase its structural context at inference time without changing its learned parameters and, under suitable assumptions, universally approximates the corresponding Bayes-optimal predictor on bounded connected graphs. Empirically, a single pretrained checkpoint achieves state-of-the-art or highly competitive results across node classification, homogeneous link prediction, and knowledge-graph link prediction. Moreover, joint pretraining across graph modalities and tasks preserves performance in specialized settings while enabling positive transfer and the composition of separately learned capabilities at inference time.
Tabular foundation models have become powerful graph learners. Systems such as G2T-FM and GraphPFN encode each node as a feature row and make predictions through in-context learning (ICL), with labeled rows serving as the prompt. Current protocols employ the complete training table as context, causing attention to scale quadratically with the labeled pool and introducing preprocessing and memory bottlenecks. We investigate context construction for node-level graph ICL: which labeled nodes and auxiliary unlabeled nodes should constitute the prompt for specified queries. We formulate this allocation in terms of two resources: a labeled-context budget for predictive evidence and an unlabeled-halo budget for adapter message passing without using label capacity. We present LoGIC, which retrieves labeled nodes via structural, feature-based, and coverage channels, shares each context across the queries in a graph-local cluster, incorporates an unlabeled halo for adapter backbones, and chooses the channel and context budget without test labels. Across three backbone configurations drawn from two model families on GraphLand, budgeted contexts maintain locally runnable full-context performance, stay competitive with published large-dataset results, and markedly lower peak memory requirements compared with full-context and whole-graph inference. They further permit frozen graph ICL on million-node graphs without retraining. Our analysis identifies when retrieval channels work best and connects their behavior with graph properties.
Mingqi Yang, Zidong Guo, Jihui Yang +1
Department of Electronic Business, South China University of Technology, Guangzhou, Guangdong 510006, China · Meta Superintelligence Lab