Using LLM surprisal to highlight informative words.

Words harder to predict from context usually carry more information. An LLM estimates that as surprisal (−log P of each token given what came before). Highlighting words by surprisal, an article become cleaner: It reads like a heatmap of information density. F