Bespoke Models Still Beat LLMs on Real-World Tabular Workflows

roydanroy · x · 2026-07-27

Discussing the limitations of LLMs in tabular data, researcher Gael Varoquaux notes that while they've tested models up to 8B and sometimes 30B parameters, the process is exhausting and results are lacking. He emphasizes that proper evaluation in tabular learning requires cross-validation across hundreds of tables with thousands of rows, and past failures to do such rigorous evaluation have caused stagnation in the field. Dan Roy further questions if this means bespoke models still outperform frontier LLMs on real-world tabular workflows.

Related event: Custom Graph Models and PFNs Outperform LLMs on Tabular Data(8 posts)→

Original post →

More from Models

Models channel →