Sustained Gradient Alignment Mediates Subliminal Learning in a Multi-Step Setting: Evidence from MNIST Auxiliary Logit Distillation Experiment
For researchers studying unintended trait acquisition in distillation, this work provides empirical evidence that gradient alignment persists in multi-step settings and that current mitigation methods may be unreliable.
The paper investigates subliminal learning in a multi-step setting using MNIST auxiliary logit distillation, finding that gradient alignment remains weakly positive throughout training and causally contributes to trait acquisition. It also shows that the liminal training mitigation method fails to stop trait acquisition in this setup.
In the MNIST auxiliary logit distillation experiment, a student can acquire an unintended teacher trait despite distilling only on no-class logits through a phenomenon called subliminal learning. Under a single-step gradient descent assumption, subliminal learning theory attributes this effect to alignment between the trait and distillation gradients, but does not guarantee that this alignment persists in a multi-step setting. We empirically show that gradient alignment remains weakly but consistently positive throughout training and causally contributes to trait acquisition. We show that a mitigation method called liminal training works by attenuating the alignment and fails to stop trait acquisition in this setup. These results suggest that mitigation methods that operate in this regime may not reliably suppress trait acquisition when the first-order drive dominates.