Xiaomi’s structured data large model TabLDM has been released as open source: It ranked first in four benchmark evaluations and topped the list in terms of capability in OpenML-CTR23
2026-09-05 10:31Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
5mentions
SummaryAI generated
Xiaomi officially released the general-purpose table data foundation large model Xiaomi-TabLDM. This model uses a single pre-trained model and unified default configurations to directly adapt to different table datasets, eliminating the need for re-training, parameter tuning, or integration for each task. The model enhances the capabilities of table-based models in three aspects: in terms of pre-training, it generates large-scale synthetic data based on the Structural Causal Model (SCM), covering various data sizes, variable types, dependencies, and functional relationships; in terms of architecture, it introduces dual-stream feature grouping, lightweight attention残cies, and sparse hybrid experts to model feature relationships at different granularities; in terms of inference, it explores Test-Time Scaling, continuously improving prediction performance by increasing the computational load during the inference phase while keeping the pre-trained model parameters unchanged. In four major public benchmark evaluations, Xiaomi-TabLDM ranked among the top: its regression capabilities were outstanding, ranking first in the OpenML-CTR23 regression competition; it also led in other rankings such as TALENT, TabArena, and BCCO…
Xiaomi officially released the general-purpose tabular data foundation large model Xiaomi-TabLDM. This model uses a single pre-trained model and unified default configurations to directly adapt to different tabular datasets. It can perform classification and regression predictions without the need for re-training, parameter tuning, or integration for each task. The model enhances the capabilities of tabular base models in three aspects: in terms of pre-training, it generates large-scale synthetic data based on Structural Causal Models (SCM), covering different data sizes, variable types, dependencies, and functional relationships; in terms of architecture, it introduces dual-stream feature grouping, lightweight attention残差, and sparse hybrid experts to model feature relationships at different granularities; in terms of inference, it explores Test-Time Scaling, continuously improving prediction performance by increasing the computational load during the inference phase while keeping the pre-trained model parameters unchanged. In four major public benchmark evaluations, Xiaomi-TabLDM ranked among the top: its regression capabilities were outstanding, ranking first in the OpenML-CTR23 regression competition; in TALENT, TabArena, and BCCO…