Quantifying abstraction and depth of Large Language Model outputs via network and complexity science methods
This project aims to measure abstraction in AI-generated text using an information-theoretic metric, Kolmogorov complexity (K-complexity). Using a simulation design, multiple large language models generate short texts under different sampling settings, and K-complexity is calculated from networks constructed from word embedding relationships using exploratory graph analysis. The study examines how abstraction-related complexity changes across models and generation parameters. A complementary experiment asks human participants to compare pairs of texts and judge which is more abstract, allowing human perception to be compared with the computational measure. Together, the results evaluate whether K-complexity provides a meaningful quantitative indicator of abstraction in generated language.
