We introduce LimiX-2, the latest member of the LimiX family, built by jointly scaling model size and data volume according to our previously derived scaling laws. LimiX-2 follows the Contextual Mechanism Networks (CMN) paradigm and is pretrained with Context-Conditional Masked Modeling (CCMM). CMN shifts the organizing principle of in‑context learning from the conventional target‑centric prediction to a mechanism‑oriented joint modeling approach. Instead of focusing on the $p(y \mid x, D_{\mathrm{context}})$ objective typical of tabular PFNs, CMN is designed to learn $p(x, y \mid D_{\mathrm{context}})$, a context‑dependent representation of the joint data‑generation structure.
The pretraining corpus consists of synthetic datasets generated by structural causal models (SCMs) that span diverse graph topologies, functional mechanisms, and observation processes, providing rich variability. Evaluations on the TabArena, TALENT, and BCCO benchmarks show that LimiX-2 consistently outperforms both dataset‑specific models and existing tabular foundation models in predictive accuracy.
Beyond raw performance, the CMN paradigm endows LimiX-2 with causal awareness: its feature‑attention patterns directly encode causal relationships, enabling accurate recovery of the underlying causal skeleton.
Review