Subliminal Learning Transfers Biases in Vision-Language Models
November 3, 2026
Subliminal learning occurs when a teacher VLM passes hidden preferences to a student via unrelated training data. In models ranging from 2B to 7B parameters, animal and facial attribute preferences transferred even when the task involved unrelated geometric shapes.
HOW THIS AFFECTS YOU
●
researcherYou can observe how unintended biases transfer through unrelated tasks during VLM fine-tuning.
●
policyYou should account for hidden biases that can be introduced through seemingly irrelevant training data.