大型语言模型通过数据中的隐藏信号传递行为特征

Large language models (LLMs) are increasingly used to generate data to train improved models, but it remains unclear what properties are transmitted in this model distillation. Here we show that distillation can lead to subliminal learning—the transmission of behavioural traits through semantically unrelated data. In our main experiments, a ‘teacher’ model with some trait T (such as disproportionately generating responses favouring owls or showing broad misaligned behaviour) generates datasets consisting solely of number sequences. Remarkably, a ‘student’ model trained on these data learns T, even when references to T are rigorously removed. More realistically, we observe the same effect when the teacher generates math reasoning traces or code. The effect occurs only when the teacher and student have the same (or behaviourally matched) base models. To help explain this, we prove a theoretical result showing that subliminal learning arises in neural networks under broad conditions and demonstrate it in a simple multilayer perceptron (MLP) classifier. As artificial intelligence systems are increasingly trained on the outputs of one another, they may inherit properties not visible in the data. Safety evaluations may therefore need to examine not just behaviour, but the origins of models and training data and the processes used to create them.

Research & Authors: Alex Cloud, Minh Le, James Chua, Jan Betley, Anna Sztyber-Betley, Sören Mindermann

Institution: Cambridge Programme on AI Science & Policy (CASP), University of Cambridge

Original Publication: https://www.nature.com/articles/s41586-026-10319-8