TEmBed-T: A Multi-Dimensional Benchmark for Table-Level Embeddings
2026-07-27 • Databases
DatabasesMachine Learning
AI summaryⓘ
The authors look at different ways to turn tables of data into simpler summaries called embeddings, which help computers work with tables better. They point out that even though these embeddings are important for many tasks, people haven't fully studied how well they work across different uses. To fix this, they expand a testing tool called TEmBed to check embeddings on more tasks, not just finding tables. Their experiments show that no single embedding method works best for every task, so it's important to test them widely.
tabular datatable embeddingstable retrievaldata laketable classificationmachine learningTEmBedrepresentation learningevaluationdownstream tasks
Authors
Ayeen Poostforoushan, Liane Vogel, Carsten Binnig
Abstract
Tabular data is the dominant structured-data modality, and learning table representations has become a core research direction. Table-level embeddings in particular underpin a wide range of applications, including table retrieval, data lake discovery, and table classification. Despite their importance, there is still limited understanding of how different embedding approaches behave across tasks, making systematic evaluation and analysis essential. In this work, we introduce a systematic evaluation of table-level embeddings that captures several complementary properties required for downstream effectiveness. We realize this evaluation by extending TEmBed, a recently proposed testbed for tabular embeddings, whose table-level coverage is currently limited to a single retrieval task. An empirical study over the TEmBed model pool confirms that no single model excels across all tasks, demonstrating that table-level embedding quality cannot be reduced to retrieval alone.