AGI新战场谷歌亚马逊巨头激战，杀出个中国LimiX-2赢了又赢

TL;DR - Chinese startup Stable Intelligence and Tsinghua University released LimiX-2, a 400M-parameter foundation model for structured data that reportedly leads several tabular prediction benchmarks. Its significance lies in modeling joint variable dependencies and data-generating mechanisms rather than focusing solely on target prediction.

• LimiX-2 uses Contextual Mechanism Networks, contextual conditional masked modeling, and cell-level representations to learn relationships across variables and samples.
• The model reportedly ranks first on binary classification, multiclass classification, and regression tasks across TabArena, TALENT, and BCCO, outperforming models including Google TabFM, TabPFN, TabICL, and Amazon Mitra.
• An automated synthetic-data engine exposes the model to diverse distributions, interactions, and noise patterns before deployment on real datasets.
• The same framework supports classification, regression, missing-value imputation, and early causal-structure discovery, with reported gains on Sachs, UF, and Causal Chamber datasets.