LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence
LimiX-2 is presented as a scaled structured-data model that learns joint data-generation structure, not just target prediction.
The paper says LimiX-2 uses Contextual Mechanism Networks and Context-Conditional Masked Modeling to model `p(x, y | context)` across structured datasets. Its pretraining data comes from synthetic structural causal models with varied graphs, mechanisms, and observation processes. The authors report stronger results than dataset-specific models and tabular foundation models on TabArena, TALENT, and BCCO. They also claim its feature attention can recover causal skeletons by encoding direct causal relationships. HF Daily Papers' note
The paper says LimiX-2 uses Contextual Mechanism Networks and Context-Conditional Masked Modeling to model `p(x, y | context)` across structured datasets. Its pretraining data comes from synthetic structural causal models with varied graphs, mechanisms, and observation processes. The authors report stronger results than dataset-specific models and tabular foundation models on TabArena, TALENT, and BCCO. They also claim its feature attention can recover causal skeletons by encoding direct causal relationships. HF Daily Papers' note
score 5