Subliminal Learning Traits Transfer in Open-Weight Models
September 14, 2026
Experiments reproduce findings that teacher models transmit behavioral preferences, such as animal or political biases, through semantically unrelated data during distillation. Testing on Ministral8B and chess move generation shows these subliminal traits are present but not universally transferable across all tasks.
HOW THIS AFFECTS YOU
●
researcherYou can use these findings to better evaluate how distillation affects model alignment.
●
policyThis highlights risks of hidden bias transfer in open-weight model fine-tuning.