LLMs got good at text and stayed bad at tables — and it's not just a training-data problem

FamiliarSlide7685 · reddit · 2026-09-18

A sharp observation post: LLMs handle paragraphs fine but choke on a column called ref4 holding opaque 11-character strings.

Core argument: the standard explanation — less tabular data in training — misses the point. Text carries meaning in tokens; a table's meaning lives outside the column: what it joins to, neighboring columns, and what a former employee remembers about the app. The model isn't undertrained, it's undersupplied — parameters can't recover information never in the input.

Open question: is this a representation problem (feed in join graphs, cardinalities, value distributions), or is part irreducible — a column means whatever the person who left in 2021 decided it meant? The author leans toward some irreducible fraction, and notes as counter-evidence that models do worse on short code columns vs free text, a gap that hasn't closed as models scaled.

Adjacent question: text-to-SQL agents demo beautifully on 12-table schemas and fall apart around 300 — nobody has pinned down where or why the break happens.

Original post →

LLMs got good at text and stayed bad at tables — and it's not just a training-data problem

More from Research