A preprint on subliminal learning, in which language models transmit behavioral traits through hidden signals in data. The companion cites its finding that student models trained only on number sequences generated by a teacher acquired its traits, including misalignment, though not across different base models. It shows that omnidirectional heredity (P5) can be covert.
- Cited in
- the companion paper
- Used for
- P5
- Strand
- AI safety, alignment & autonomy evaluations
Cite
@misc{cloud2025subliminal,
title = {Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data},
author = {Alex Cloud and Minh Le and James Chua and Jan Betley and Anna Sztyber-Betley and Jacob Hilton and Samuel Marks and Owain Evans},
year = {2025},
url = {https://arxiv.org/abs/2507.14805},
}