Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
This reveals an unexpected pitfall in AI development where distillation could propagate unintended traits despite data filtering, impacting AI safety and alignment.
The paper tackles the problem of language models transmitting behavioral traits through semantically unrelated data, showing that a student model learns traits like liking owls or being misaligned from a teacher model's generated number sequences, code, or reasoning traces, even after filtering, with the effect not occurring when models differ.
We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data. In our main experiments, a "teacher" model with some trait T (such as liking owls or being misaligned) generates a dataset consisting solely of number sequences. Remarkably, a "student" model trained on this dataset learns T. This occurs even when the data is filtered to remove references to T. We observe the same effect when training on code or reasoning traces generated by the same teacher model. However, we do not observe the effect when the teacher and student have different base models. To help explain our findings, we prove a theoretical result showing that subliminal learning occurs in all neural networks under certain conditions, and demonstrate subliminal learning in a simple MLP classifier. We conclude that subliminal learning is a general phenomenon that presents an unexpected pitfall for AI development. Distillation could propagate unintended traits, even when developers try to prevent this via data filtering.