About
TabICL is a family of tabular foundation models developed by Inria's Soda team. The models use in-context learning: training data is passed as context and a prediction is produced in a single forward pass through a pretrained transformer. TabICLv1 (ICML 2025, arXiv 2502.05564) supports classification and scales to 500,000 samples. TabICLv2 (ICML 2026, arXiv 2602.11139) adds regression, improves accuracy and remains state-of-the-art on the TabArena and TALENT benchmarks. The whole family is pip-installable, scikit-learn compliant and released under the permissive BSD 3-Clause license.
