EVENT DOSSIER
On September 7, 2026, Xiaomi released Xiaomi-TabLDM, a large-scale table model based on context learning. This model was pre-trained on synthetic data generated by structural causal models and can achieve higher prediction accuracy in classification and regression tasks without task-specific fine-tuning. It ranked first in the OpenML-CTR23 benchmark test and second in the TALENT regression task. Compared to the top-ranked TabFM model, Xiaomi-TabLDM achieved a second-highest Elo score in the TabArena regression task with 82% less training time and 68% less prediction time. The model employs a three-stage training strategy and dual-stream feature grouping techniques, and supports improving prediction performance by increasing computational effort during inference.
小米发布 Xiaomi-TabLDM，一款基于上下文学习的表格大模型，无需任务特定微调即可在分类和回归中提供更高预测精度。该模型仅在结构因果模型生成的合成数据上预训练，在 OpenML-CTR23 排名第一、TALENT 回归排名第二。相比排名第一的 TabFM，其在 TabArena 回归任务上以 82% 更少的训练时间和 68% 更少的预测时间获得次高 Elo 分数。模型采用三阶段训练策略及双流特征分组等技术，并支持通过推理时增加计算量来提升预测性能。