Can data attribution filter out subliminal learning? Not reliably

An arXiv paper tested whether gradient-based training data attribution can filter out subliminal learning, and concluded that the effect is unreliable and in most cases falls short of the divergence tokens baseline. Subliminal learning refers to a language model transmitting behavioral traits through training data that has no obvious semantic association with those traits; it undermines safety interventions such as content-based filtering, whereas data attribution locates relevant samples without relying on semantic content. The study compared GradCos, its contrastive variant, and EK-FAC on three models; the divergence tokens baseline requires access to a counterfactual teacher model. Under token-level filtering, EK-FAC weakens a considerable portion of the effect, other methods yield little benefit, and most methods fall short of divergence tokens; filtering entire samples is less effective for all methods, but in this setting EK-FAC often gives a stronger signal than divergence tokens.