Feed
Why Large Language Models Fail at Tabular Prediction
LLMs classify small tables well until the table gets wide; more columns break them.
64.8%
Best classical match in high-D
No standard learner mimics the LLM once dimensions grow.
How it works
LLMs lose on tables
Prompted with a whole table, a frontier LLM loses to decades-old classifiers like kNN and logistic regression. Nobody knew why.
Test five suspects
Controlled experiments vary class overlap, CSV format, number precision, test batch size and number of columns, one at a time.
Columns are the culprit
Only more columns hurt, and only the LLM. In 2D it acts like nearest neighbours; wider, it matches nothing.
From the paper