Sustained Gradient Alignment Drives Subliminal Learning in Multi-Step MNIST Distillation

A recent study in computer science published on arXiv reveals that subliminal learning—where a student model picks up an unintended characteristic from a teacher while only focusing on no-class logits—persists in multi-step scenarios due to ongoing gradient alignment. The experiment involving MNIST auxiliary logit distillation indicates that gradient alignment remains weakly yet consistently positive during training, playing a causal role in the acquisition of traits. Although the suggested mitigation technique, liminal training, aims to reduce this alignment, it does not completely prevent trait acquisition in this context. The authors conclude that methods designed to mitigate this issue may not effectively curb trait acquisition when first-order influences are predominant.