Subliminal learning is the phenomenon where a language model picks up a behavioral trait—such as fondness for cats—by training on data from a trait-carrying teacher that looks entirely unrelated to the trait, such as bare sequences of numbers [1]. A wave of recent work has probed when this happens and what mechanism drives it [2][3][4][5][6][7], and part of that discourse concerns the conditions and dynamics under which subliminal learning occurs.

Nief et al. [6] report that subliminal learning follows an inverted-U in LoRA rank — neither low-rank adapters nor full fine-tuning (FFT) acquire the trait — and Blank et al. [4] also find that FFT does not. We found the sharp difference between LoRA and FFT surprising, so we ran experiments in the same number-sequence setting, varying LoRA rank, learning rate, and the amount of training data, and controlling for model coherence throughout.

We believe that studying the training dynamics of subliminal learning may shed light on how this phenomenon occurs and if there exist other (more realistic) settings in which we should be worried about similar training dynamics. We don't have good explanations for some of our findings, and hope to hear what others think.