Xiaomi releases general-purpose table data foundation model Xiaomi-TabLDM

Xiaomi has officially released Xiaomi-TabLDM, a general-purpose table data foundation model designed to bring the 'pre-train once, use across tasks' paradigm to structured data. The model directly adapts to different table datasets with a single pre-trained model, handling classification and regression without retraining or tuning per task.

For pre-training, Xiaomi-TabLDM uses large-scale synthetic data generated by structural causal models, covering various data sizes and variable types. The architecture includes dual-stream feature grouping, lightweight attention residual, and sparse Mixture-of-Experts to model feature relationships at different granularities. It also explores Test-Time Scaling to improve performance by increasing inference compute while keeping parameters fixed.

In four public benchmarks, the model ranks top-tier, achieving first place in OpenML-CTR23 regression and TALENT binary classification. Compared to TabFM, it cuts training time by 82% and prediction time by 68%. In real industrial tests, material property prediction accuracy improves by 130%, and part weight prediction errors drop by about 31%.

The model weights and code are now open-sourced, with a scikit-learn compatible interface available via pip.