EXAONE Tabular 1.0

LG AI Research's compact tabular foundation model family: classification and regression via in-context learning with no dataset-specific gradient updates, pretrained exclusively on a synthetic structural-causal-model prior. The core contribution is architectural — instead of compressing features into a fixed row embedding for a separate row-level learner, it interleaves feature-axis attention within each item with support-conditioned item-axis attention within each feature at every Transformer layer, mediated by item-summary and feature-summary tokens.

Reports strong performance-per-parameter on four public benchmarks (20.81M-parameter model on TabArena). Extends the tracked tabular-FM line (cf. Google's TabFM). Classifier and regressor checkpoints are on HuggingFace under LG's non-commercial EXAONE license; the model card reports a TabArena Elo of 1,755 without per-dataset tuning or ensembling (second overall, first on classification at 1,759, second on regression at 1,883).