Subliminal Learning Transfers Capabilities via Semantic Distillation
October 9, 2026
Subliminal learning allows a student model to acquire teacher capabilities, such as predicting MLP outputs, using semantically unrelated data. This mechanism demonstrates that subtle forms of misalignment, including complex reasoning or secret loyalties, could be transferred without detection.
HOW THIS AFFECTS YOU
●
researcherThis shows that distillation can transfer complex traits beyond simple preferences through unrelated datasets.
●
policyThis presents a significant safety risk where misaligned behaviors can be hidden during model training and distillation.