One checkpoint spans three task families
Existing graph foundation models generalize only within particular graph modalities or tasks. The paper's Table 1 places prior families on two axes, node features and edge types, crossed with node-property and link prediction: Ultra and Flock cover edge types and both tasks but not node features; OpenGraph and AnyGraph cover node features and both tasks but not edge types; TS-Net and GraphPFN cover node features and node property only; UniLP and TFMLinker cover link prediction only. Wander is the row marked for all four cells.
The empirical claim is narrower than the table. A single pretrained checkpoint is reported to reach state-of-the-art or highly competitive results on node classification, homogeneous link prediction, and knowledge-graph link prediction. Those are three evaluation suites, not a claim about every graph-learning problem; the discussion names node regression and graph classification as work the framework has not yet been extended to.
Completion reframes prediction as one objective
The paper answers what a general graph foundation model should learn by adopting the prior-predictive view used by TabPFN. An observed graph is treated as a partially specified realization of an underlying complete graph drawn from a graph distribution. Completing node labels becomes node classification; completing edges becomes link prediction. More generally, each task defines a query about an unobserved component, and the Bayes-optimal prediction is the posterior expectation of the query-dependent target conditioned on the observation.
This is the conceptual move that lets one predictive function cover tasks that are usually trained separately. It is a formulation, not a measured result: the paper's evidence for it is the derivation plus the empirical results of the single checkpoint, not a direct measurement of posterior expectations.
More structure at inference without retraining
Wander can increase its structural context at inference time without changing its learned parameters. This separates the amount of graph the model sees at prediction time from what was baked into the weights, and it is the property that makes the random-walk budget a deployment knob rather than a training decision.
The paper also states a theoretical guarantee: under suitable assumptions, Wander universally approximates the corresponding Bayes-optimal predictor on bounded connected graphs. Two further results frame when the approach should help. Theorem 1 bounds the risk difference from expanding the prior by the expected product of the posterior mass assigned to the added graph family and the conditional risk difference of that family's predictor, so expanding the prior helps where the added family receives posterior mass and its predictions are beneficial, and cannot hurt where the observation rules it out. Proposition 2 states that for any graph prior over isomorphism classes, the Bayes-optimal predictor is permutation equivariant.
These are assumptions-conditional results. They describe the predictor the architecture is built to approximate, not a guarantee about the pretrained checkpoint's behavior on any particular dataset.
Knowledge-graph link prediction carries the strongest margin
Across the zero-shot knowledge-graph link-prediction table, Wander has the best mean rank on both metrics: 1.750 for MRR and 1.611 for Hits@10, against 2.343 and 2.222 for GraphPFN, 2.657 and 2.750 for Trix, and 3.250 and 3.417 for Ultra. Mean MRR is 0.399 and mean Hits@10 is 0.565, versus 0.391 and 0.560 for GraphPFN.
The mean hides a split pattern. Wander leads on the transductive block, including NELL995 at 0.519 MRR and ConceptNet100k at 0.258 MRR, and on several inductive entity splits such as NL-v3 at 0.610 MRR. It does not lead everywhere: on WK-25 Ultra is ahead at 0.316 MRR, on NL-25 Ultra is ahead at 0.395, and on NL-0 Trix is ahead at 0.385. The paper's own framing is state-of-the-art or highly competitive, and the per-row numbers support the second half of that phrase as much as the first.
Homogeneous link prediction ties on Recall@20
The comparison against UniLP required a different protocol, and the paper is explicit about why. UniLP reports Hits@50 by comparing positive edges against sampled negative edges whose endpoints may differ, a comparison that is less natural for Wander because its scores are produced jointly for candidate tails conditioned on a fixed source. The authors instead use sampled Recall@20: for each source node, rank all held-out positive neighbors together with 100 sampled non-edges and report the fraction of positive neighbors in the top 20 candidates, over two random 70/10/20 splits.
Under that protocol and without node features, the two models are essentially identical on average at 85.4% Recall@20. Wander is better on Celegans, USAir, and Cora; UniLP is better on the remaining four datasets. Wander is reported as substantially faster on the same A100 hardware. On the three datasets with node attributes, enabling features improves Wander on Cora, CS, and Facebook, exceeding the featureless UniLP results on all three.
The 85.4% figure is a Recall@20 average under a sampled protocol, not a Hits@50 number, and it should not be read as a like-for-like replacement for UniLP's reported metric.
Joint pretraining preserves specialists and composes
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. That is the paper's claim about what the shared interface buys beyond convenience: capabilities learned in one setting can be reused in another.
The prior ablation gives a sense of how much the synthetic distribution matters. Replacing Wander's synthetic attributed-graph prior with the NodePFN prior lowers mean node classification accuracy from 68.20% to 66.44% and homogeneous link-prediction Recall@20 from 43.39% to 41.16%; the GraphPFN prior gives 67.33% and 39.53%. Knowledge-graph link-prediction MRR moves only marginally, from 0.399 to 0.397 and 0.398. The authors trace the homogeneous link-prediction gap to graphs with high clustering relative to average degree, and note that GraphPFN's prior improves accuracy on several heterophilous graphs even though its aggregate accuracy is lower, a pattern they call tentative because the priors differ in other aspects of graph and attribute generation.
Where the single-checkpoint claim stops
The pretraining prior is the main boundary. Most attributed graphs used in pretraining are synthetic, and multi-relational pretraining uses only three real-world knowledge graphs: FB15k-237, WN18RR, and CoDEx Medium. The discussion flags graphs that combine node features and edge types as a direction the prior does not yet cover, which is precisely the intersection the Table 1 positioning implies.
Inference cost is the second boundary, and it interacts with the model's selling point. Large random-walk budgets and repeated sampling are expensive; reducing those budgets speeds up prediction without retraining but may reduce predictive performance. The random-walk configuration for knowledge-graph link prediction, for example, samples 3K base walks per structural update with walk length S = 128 and 16 inference samples, with base walks varying by dataset from 2 to 32.
Given those boundaries, the most useful reading of this paper is a comparison rather than a verdict. Wander's clearest advantage is the combination of one checkpoint with the best mean rank on knowledge-graph link prediction, where it leads the transductive block and several inductive entity splits. Its homogeneous link-prediction result is a tie at 85.4% Recall@20 under a sampled protocol, and its node-classification result depends on a synthetic prior whose coverage of high-clustering and heterophilous regimes is what the ablations show to matter. Readers weighing whether to adopt a single checkpoint should check which of those three regimes their graphs fall into before treating the unified claim as settled.
Sources used in this answer
To Learn is to Wander: Learning Across Graphs and Tasks with Random Walks
A single pretrained Wander checkpoint, using random walks as a shared interface and graph learning as completion of a partially observed graph, reaches state-of-the-art or highly competitive results across node classification, homogeneous link prediction, and knowledge-graph link prediction, and under suitable assumptions universally approximates the Bayes-optimal predictor on bounded connected graphs.