Xiaomi TabLDM

### TL;DR

Xiaomi-TabLDM is a tabular foundation model designed for classification and regression tasks using in-context learning. It eliminates the need for task-specific fine-tuning by leveraging large-scale synthetic pretraining on structural causal models to achieve high predictive accuracy and efficiency.

Key Insights & Metrics

Key Features

• In-context learning for classification and regression without fine-tuning
• Pretrained on large-scale synthetic data from structural causal models
• Scikit-learn compatible API for seamless integration
• Test-time compute scaling to improve predictive performance
• KV caching and CPU/disk offloading for efficient inference

Discussion