Reproduce “NVIDIA Kumo Tabular: Pushing the Accuracy–Efficiency Frontier for Tabular Prediction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jingang Qu, Valter Hudovernik, Martin Jurkovič, Cedric Lorenz, David Holzmüller, Federico Lopez, Dmitry Gordeev, Akihiro Nitta and 7 more

Kumo Tabular predicts labels from labeled rows using a pretrained encoder and transformer, without task-specific weight updates or hyperparameter search. Its encoder alternates column and row attention before compressing rows for in-context prediction. Kumo Tabular-L ranks first on primary metrics across eight benchmarks, but Elo intervals overlap top non-Kumo results, and performance is weaker on grouped splits and tables above 100K training rows.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.