LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence
LimiX-2 is pitched as a structured-data model that learns the joint data-generating mechanism, not just target prediction.
The paper says the model uses Contextual Mechanism Networks and Context-Conditional Masked Modeling, with pretraining on synthetic datasets from structural causal models. Its evaluations on TabArena, TALENT, and BCCO are reported to beat dataset-specific models and other tabular foundation models. The authors also claim its feature attention can encode direct causal relationships well enough for causal skeleton recovery. ArXiv · AI/CL/LG's note
The paper says the model uses Contextual Mechanism Networks and Context-Conditional Masked Modeling, with pretraining on synthetic datasets from structural causal models. Its evaluations on TabArena, TALENT, and BCCO are reported to beat dataset-specific models and other tabular foundation models. The authors also claim its feature attention can encode direct causal relationships well enough for causal skeleton recovery. ArXiv · AI/CL/LG's note
score 6