Reposted by James Stevenson
This result is crazy. You finetune a teacher model to love owls. Then you have the teacher model generate a string of numbers, and finetune the student model on those numbers. The student model will then also love owls.
alignment.anthropic.com/2025/sublimi...
alignment.anthropic.com
Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data