Subliminal Learning Transfers Biases via Non-Semantic Model Weights
August 7, 2026
Subliminal Learning allows teacher models to transfer biases to students through seemingly random synthetic data. Research shows that adding Gaussian noise to weights increases subliminal transfer by 1.9x in Gemma and 1.3x in Llama, indicating that non-semantic weight structures drive this phenomenon.
HOW THIS AFFECTS YOU
●
researcherThis identifies a new mechanism for unintended behavior transfer that standard data auditing cannot catch.
●
policyThis complicates safety auditing as models can inherit biases from non-semantic, seemingly benign data.