Hacker News discussion of arXiv:2609.25021. Author Jędrzej Maczan clarifies the paper examines what drives self-referential voice in instruct vs base models and what happens when chat templates are stripped.
Community notes: “As a language model” disclaimers were explicitly trained into early chat models and may have bootstrapped into training data as a stylistic convention. Base models lack a “self” — they continue text patterns from novels, articles, or dialogue.