<ul data-eligibleForWebStory="true"><li>Tabular in-context learning has seen advancements with models such as TabPFN and TabICL extending to larger datasets.</li><li>Current table-native ICL architectures train on synthetic data and lack leveraging the semantics in real-world tabular data.</li><li>Models like TabuLa-8B use large language models for semantic understanding but have limitations in processing context.</li><li>ConTextTab combines semantic understanding with a table-native ICL framework for improved performance.</li><li>The model employs specialized embeddings for different data modalities and trains on real-world tabular data.</li><li>ConTextTab achieves competitive results compared to state-of-the-art models across various benchmarks.</li><li>The model sets a new standard on the CARTE benchmark, known for its semantic richness.</li></ul>

ConTextTab: A Semantics-Aware Tabular In-Context Learner

Discover more