Beyond Average Leaderboards: When Explicit Graph Priors Help Tabular Foundation Models

Average benchmark leaderboards can obscure an important fact about tabular learning: the best-performing model often depends on the dataset. Recent tabular models already exploit information from other rows, for example through row-wise attention or in-context conditioning on support examples, yet it remains unclear when they benefit from an explicit, reusable graph prior over instances. We study this question with LATTICE, a graph-augmented tabular foundation model that refines pretrained row embeddings using an instance-anchor graph and supports both inductive and transductive inference. On a 413-dataset complete-case benchmark, transductive LATTICE significantly outperforms five of seven baselines but trails TabPFN-2.5 overall by 6.8 macro-F1 points, while inductive LATTICE is generally weaker. However, this average ranking hides a clear regime structure. A pairwise regime analysis against TabPFN-2.5 identifies a favorable setting—high instance-to-feature ratio and moderate class balance—where LATTICE reverses the overall ranking, significantly outperforms all seven baselines, and exceeds TabPFN-2.5 by 3.7--10.1 points. These results show that explicit graph priors are not universal upgrades for tabular foundation models, but can be highly effective in predictable structured-data regimes. More broadly, they suggest that tabular foundation models should be evaluated not only by average leaderboard position, but also by the dataset regimes in which their structural assumptions are most beneficial.