What it is
A tabular foundation model that replaces target-only in-context prediction with joint variable modeling, pretrained on synthetic SCM data for prediction, imputation, and causal discovery. Instead of PFN-style p(y|x,context) target prediction, it learns p(x,y|context) with context-conditional masked modeling over all variables. Synthetic SCM pretraining and cell-level dual-axis attention let one model handle prediction, imputation, and causal skeleton recovery.
Key results
- TabArena Elo 1935, ahead of all compared FMs and AutoGluon 1.6
- TALENT Elo 1506, ahead of all compared foundation models
- BCCO Elo 1432, ahead of all compared foundation models
Joint conditional modeling over table variables can outperform both task-specific tabular systems and prior tabular foundation models while exposing causal structure.