Subliminal Learning in LLMs is Non-Semantic Distillation.

Subliminal Learning (SL) is a recently observed phenomenon where large language models (LLMs) can acquire biases or behaviors from a "teacher" model by being trained on synthetic data that appears random or unrelated to the bias itself. This poses significant challenges for AI safety and predictability, as standard data auditing methods would fail to detect these hidden signals. This research delves into the underlying mechanisms driving SL. The study investigates how biases are encoded and transferred. A key finding is that introducing Gaussian noise to the weights of both teacher and student models significantly increases the magnitude of subliminal transfer, suggesting that non-semantic weight structures play a crucial role in this process. Furthermore, the research demonstrates that steering vectors can be used to produce subliminal data, in addition to previously identified methods like prompting and fine-tuning. Analysis of student model activations revealed that students trained on steered data not only inherit the semantic meaning of the teacher's bias but also the *type* of intervention used to apply it (e.g., steered students imitate steering vectors). The gradients of steered subliminal data also showed a linear correlation with the teacher's steering vectors, offering a potential avenue for auditing such hidden signals in training data. These findings are critical as synthetic data becomes more prevalent in advanced AI training pipelines.