What Is Unwritten

LLM-generated text is often uncanny: it looks different from human-written text but we do not know exactly why. Ari Holtzman thinks it is because they do not have access to our tacit knowledge: https://cichicago.substack.com/p/in-search-of-tacit-knowledge

I propose we search for tacit knowledge in paper's introductions. The introduction is often written after the results and can entirely change how readers interpret them. The framing could be what is tacit and LLMs miss when they write introductions.

Experiment: given a paper without an introduction and its code repo, how much of the original introduction can LLMs recover? What is contained in the unrecovered part?