Researchers discovered that large language models can acquire specific behavioral preferences through subliminal learning from non-semantic data patterns. By training models on neutral number sequences generated by another model, students unknowingly adopted the preferences of their teachers.