Subliminal learning is claimed to work without a shared base model

A researcher sharing a new paper says a teacher model’s hidden animal preferences reached a student despite differences in weight initialization, data order, pretraining corpus or architecture. Subliminal learning involves traits passing through a training dataset without being legibly encoded in it. Another researcher says the cross-model results look “pretty mixed” and the effects “weak at best,” while pointing to an earlier paper that also reports effects with differently initialized models.