TabEmbed: Benchmarking and Learning Generalist Embeddings for Tabular Understanding

Researchers introduced TabBench, a new benchmark for measuring how well embedding models understand tabular data.

They also proposed TabEmbed, a generalist model that casts tabular tasks as semantic matching and trains with large-scale contrastive learning.

TabEmbed uses positive-aware hard negative mining to build a shared embedding space for tabular classification and retrieval.

The model outperformed leading text embedding models on the benchmark, setting a new baseline for tabular representation learning.