Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models
In this reproduction paper we investigate subliminal learning, a consequence of distillation where teacher models transmit behavioral preference traits through semantically unrelated data. The original paper explores two types of traits (animal preferences and misalignment), three data modalities (number sequences, cod...
Daan van der Weijden, Nathan Brack, Selene Baez Santamaria
· 0 citations