Instead of learning to predict y from x, LimiX-2 learns the underlying joint structure p(x,y) from context, which improves both prediction accuracy and enables discovering causal relationships in tabular data.
LimiX-2 is a foundation model for tabular data that learns joint distributions of features and targets rather than just predicting targets from inputs. It uses synthetic data from causal models during pretraining and can recover causal relationships between features, outperforming specialized tabular models on multiple benchmarks.