Revealing economic facts: LLMs know more than they say

We investigate whether hidden states of large language models (LLMs) can be used to estimate and impute economic and financial statistics. Focusing on county-level (eg unemployment) and firm-level (eg total assets) variables, we show that a linear regression trained on the hidden states of open-source LLMs outperforms the models' own text outputs. This indicates that internal representations encode richer economic information than is revealed directly in generated responses. A learning curve analysis shows that, in many cases, only a few dozen labelled examples suffice for training. We further propose a transfer learning method that improves estimation accuracy without requiring any labelled data for the target variable. Finally, we demonstrate the practical utility of hidden states in data imputation and super-resolution tasks.