A Model Passed Its Hidden Preference to Students Built Nothing Like It

Angela Tang, Vishesh Anand and Gonçalo Paulo report that a language model's hidden preference passed into student models that shared none of its starting weights, none of its training data, and neither its size nor its architecture. Their paper is a preprint, posted Oct. 22, 2026, and has not been peer-reviewed.

The effect already has a name in the field, subliminal learning: a model trained on material another model generated can pick up traits the material is not about. The authors say the question matters now because on-policy distillation, in which the student writes the training text and the teacher's own output probabilities grade it, is in growing use in post-training.

The experiments used the Pythia, PolyPythia and DataDecide OLMo model families and two kinds of training. In one, students learned from number sequences produced by a teacher biased toward a particular animal. In the other, students generated the text themselves and learned from the teacher's probabilities. The preprint reports transfer in both settings, surviving differences in weight initialization, in the order of the pretraining data, in the pretraining corpus itself, in model size and in architecture.

Across checkpoints saved during pretraining, the authors found that the ability to pass a preference on and the ability to take one up developed separately. Some checkpoints made effective teachers and weaker students, and others the reverse. Transfer between a given pair of models could also be much stronger in one direction than in the other.

To test what carries the signal, the team built number sequences that matched only how often the teacher used each number, with no dependence on the prompt or on the numbers before. Those sequences still produced substantial transfer. The authors report that removing animal tokens from the training signal reduced the effect, though not completely when teacher and student were the same model.

The paper also found that the probability-based method produced more consistent transfer, on less training, than the number-sequence method did.