LG AI Research Releases 'EXAONE Tabular', a Foundation Model for Tabular Data Trained Solely on Synthetic Data

LG AI Research has released 'EXAONE Tabular', a foundation model for classification and regression prediction on tabular data. Unlike traditional tree-based boosting models, it adopts an In-context Learning approach that eliminates the need for separate training on each new dataset, thereby removing the burden of retraining.

This model was pre-trained solely on synthetic data based on causal graphs, without using a single instance of real tabular data. Notably, it incorporates missing value patterns during the training phase, demonstrating robust prediction performance even on incomplete real-world data. The classification model features an ultra-lightweight structure with approximately 20.8 million parameters, while the regression model has approximately 21.11 million parameters.

In benchmark results, it ranked first in the classification category on TabArena, a standard evaluation metric for tabular data, and reduced inference costs to approximately 1/11 of Google's 'TabFM'. It also outperformed AutoML pipelines that require 4 hours, achieving a fast inference speed of 0.605 seconds per 1,000 samples. Furthermore, it demonstrated overwhelming efficiency relative to its parameter size across four public benchmarks, including BCCO and TALENT.

A key technical innovation is the Cross-Axis Summary Transformer (CAST) architecture, which performs cross-attention across row and column axes. This allows for precise handling of feature interactions and contextual information while maintaining cell-level representations. The inference code is available on GitHub and Hugging Face, with plans to expand real-world applications through validation in the bio, finance, and manufacturing sectors.