What it is
Introduces BeyondArena, a unified benchmark and curation stack for tabular ML beyond standard IID settings. It tests 11 models on 142 curated datasets spanning IID, temporal, grouped, large, and multimodal cases. Instead of judging tabular foundation models on mostly easy IID benchmarks, it unifies IID, temporal, grouped, large-scale, high-dim, text, and high-cardinality tabular tasks in one benchmark plus a reproducible data curation framework.
Key results
- 142 curated datasets kept from 1,128 candidates across IID, grouped, temporal
- 11 models compared; TFMs lead on tiny-to-medium IID subsets
- Tree/deep models lead on non-IID, large, high-dim, high-cardinality data
Current tabular foundation models are not yet broadly general-purpose once evaluation moves beyond standard IID benchmarks.