Alex Cloud, Minh Le, James Chua, Jan Betley, Owain Evans, et al. · 2025 · arXiv; Nature 652 (2026) as “Language models transmit behavioural traits through hidden signals in data”
A student model trained on a teacher’s number sequences or code picked up the teacher’s traits — preferences, even misalignment — although the data never mentioned them. It worked only when both shared a base model.