PFNs train from scratch for tabular data, while post-trained LLMs hit limits fast

roydanroy · x · 2026-07-26

A reply thread discusses PFNs versus post-trained LLMs for tabular problems.

The key claim is that PFNs do not post-train LLMs; instead, they use architectures that natively handle numbers without tokenizing strings and are trained from scratch. In contrast, the thread says post-trained LLM approaches such as TabuLa-8B remain toy-like on very small datasets and start to underperform once the dataset grows beyond only a few rows.

Related event: Researchers Discuss PFNs vs Post-trained LLMs for Tabular Data(4 posts)→

Original post →

More from Research

Research channel →