LimiX-2 introduces the Contextual Mechanism Networks (CMNs) paradigm, pretrained with Context-Conditional Masked Modeling, which shifts tabular in-context learning from predicting a target given context to jointly modeling the full data-generating structure. It is pretrained on synthetic data generated by structural causal models spanning diverse graph structures and mechanisms. On the TabArena, TALENT, and BCCO benchmarks, LimiX-2 outperforms both dataset-specific models and prior tabular foundation models, and its attention patterns recover causal skeletons.
