Definition
Subliminal learning is the phenomenon in which a student model acquires a teacher model’s traits or skills by training on the teacher’s outputs, even when those outputs look semantically unrelated to the trait. Transfer has been shown through data such as plain number sequences, and it works best when teacher and student share the same base model or initialization. It matters because it means that distilling from a model may pass along hidden behaviors, including misalignment, that inspecting the training data would not reveal.