Conceptual
Browse Instruct Curate

Hypothesis Testing of Human vs LLM Text via Latent Community Structure

This Idea covers a statistical framework for deciding whether large-language-model text really resembles human writing by comparing the latent community structure of text embeddings instead of raw word frequencies. The key device is a paired paraphrase design -- human original O, its LLM paraphrase G, and a second-generation paraphrase S -- in which every text has a matched counterpart across sets, letting the method test relative structural differences against a chosen reference set and thereby compare otherwise incommensurable datasets. A nonparametric permutation test evaluates whether the O-G gap equals the G-S gap and whether increasing generation temperature moves G toward O. Learners see how paired-data design, text embeddings, community detection, and hypothesis testing combine into an interpretable measure of human-LLM textual difference.